The Congo Basin in Central Africa remains one of the few regions globally where the Intergovernmental Panel on Climate Change has neither assessed changes in hot extremes and extreme precipitation since the 1950s nor attributed such changes to anthropogenic influences, primarily due to the sparsity of in situ observational data. Although extensive daily weather records exist, spanning from the 1900s to the early 2000s and covering numerous stations across the basin, the majority of these remain archived on paper, limiting their accessibility for climate analysis. Here we present and analyse over 1 million temperature and precipitation observations from 37 weather stations in the Democratic Republic of the Congo, which have until now been unavailable to the research community. To produce this dataset, we first digitized (imaged) 9885 paper sheets stored in local archives during two dedicated field campaigns. We subsequently apply an improved version of the MeteoSaver software (version 1.1) to transcribe the imaged sheets, yielding daily time series of maximum, minimum, and average temperatures, precipitation, as well as dry bulb and wet bulb temperatures measured at three times per day (06:00, 15:00, and 18:00 local time). After quality control, analysis of multi-decadal temperature data across the basin reveals a consistent and accelerating warming signal since the 1960s, characterized by a warming shift in the distribution of daily maximum, minimum, and average temperatures with each successive decade. Median trends across 21 of the 37 stations with sufficient data availability are 0.22, 0.10, and 0.15 °C per decade for daily maximum, minimum, and average temperatures, respectively, corresponding to approximately 0.7, 0.3, and 0.5 °C of warming during the 1961–1990 period. We further find an increasing frequency of hot extremes and a decreasing frequency of cold extremes with each successive decade, with the most recent decade exhibiting nearly twice as many hot days per year and fewer cold days compared to the earliest decade. Analysis of precipitation at three stations with sufficient data indicates an increased frequency of heavy precipitation events at two stations and no substantial change at the third; however, limited spatial coverage due to data availability restricts broader conclusions for the DRC. Overall, this analysis of rescued weather data from Central Africa highlights the urgent need to close the knowledge gap on climate trends in the Congo Basin, one of the world's most data-sparse yet climatically significant regions.
Fonds Wetenschappelijk Onderzoek11M8825N11M8823NHORIZON EUROPE Framework Programme101076909101081369Introduction
Recent studies have unequivocally identified anthropogenic climate forcing as the main driver for the observed increase in global mean air surface temperature over recent decades . Similarly, the increase in frequency and intensity of hot extremes, such as heatwaves, in most regions of the world since the 1950s has also been attributed to human-induced climate change . However, unlike most regions, Central Africa remains one of the few areas where the Intergovernmental Panel on Climate Change (IPCC) has neither assessed changes in hot extremes and extreme precipitation since the 1950s nor attributed such changes to human influence, primarily due to the sparsity of in situ observational data .
The persistent data sparsity across Central Africa has been a significant impediment to studies assessing the regional effects of human-induced climate change on ecosystems, societies, and the water cycle , despite the region hosting the Congo Basin, which contains the world's second-largest river and tropical rainforest . For example, while observed changes in the hydrological cycle and water resources have been robustly attributed to anthropogenic climate change in many regions e.g.,, the Congo Basin, remains a major observational blind spot.
Moreover, the Congo Basin is home to some of the world's most vulnerable populations and is experiencing rapid demographic growth in both urban and rural areas . These populations are increasingly exposed to climate-related hazards, such as landslides and gully erosion, which are further exacerbated by human-induced land transformation associated with population growth .
Furthermore, satellite data suggest that the Congo Basin has experienced a drying trend in recent decades, characterized by reductions in precipitation, canopy water content, and forest photosynthetic capacity . Although Land Surface Models (LSMs) and Global Hydrological Models (GHMs) can be used to study hydroclimatic trends within the region, the quality of these models remains uncertain due to the limited availability of ground-based observations for validation .
Substantial records of observed weather within the Congo Basin, stored in both local and international archives, present an opportunity to overcome these data limitations and thus facilitate the analysis of hydroclimatic trends in the basin . However, as in many parts of the world, millions of in-situ weather records from various stations remain preserved within archives as hard copies or in image form e.g.,. To date, daily temperature and precipitation records originally collected by the Institut National pour l'Etude Agronomique du Congo Belge (INEAC), with some stations dating back as early as 1908, though the majority begin in the 1930s, have been preserved in the State Archives of Belgium. These records, which extend up to 1960, were digitized under the Congo Basin Eco-climatological Data Recovery and Valorization (COBECORE) project, which involved photographing all available climatological and ecological records for 575 climate stations spread throughout the Congo Basin . Additionally, post-1960 daily temperature and precipitation records for 37 weather stations in the Democratic Republic of the Congo (DRC) are hosted in the archives of the lnstitut National pour l'Etude et la Recherche Agronomiques (INERA), formerly INEAC, in Yangambi. These records span from 1960 to the 1990s for most stations, with a few extending into the early 2000s.
This paper describes the data rescue efforts within the INERA archives in Yangambi, for post-1960 handwritten daily temperature and precipitation records taken at 37 weather stations in the DRC. This process involves: (i) compiling a data inventory, (ii) digitizing (that is, imaging) the handwritten weather records from paper to digital format, and (iii) transcribing the digitized image into machine-readable format using MeteoSaver, an open-source software we developed based on machine learning (ML) algorithms . MeteoSaver takes the digitized images of tabular observation sheets as input and transcribes the data, in this case, handwritten daily weather data, into a standardized, machine-readable format known as the Station Exchange Format (SEF), as prescribed by the Copernicus Data Rescue Service . By automating transcription, MeteoSaver saves numerous man-years of manual data entry and demonstrates its potential for wider application in similar historical data rescue efforts.
In addition to this newly automatically transcribed temperature dataset, we incorporate three recently available, manually transcribed datasets for the region. First, presents daily temperature and precipitation data for one station, Yangambi-Km.5, spanning 1960–2020. As one of the main findings, reports a long-term increase of approximately 1 °C in the annual average mean daily temperature over seven decades, and an increase in frequency of hot extremes, specifically warm days and nights, recorded at that station. Second, since 2000, INERA has also transcribed daily temperature and precipitation data for 6 of the 37 weather stations. Most of these records cover the period from the 1990s to the mid-2010s, with only two stations having data extending back to the 1960s, and one continuing up to the early 2020s. Third, we incorporate pre-1960 data from the COBECORE project for three stations that are also part of the 37-station post-1960 INERA dataset. We therefore include the , INERA and COBECORE datasets in our combined temperature and precipitation dataset, enabling the construction of long, continuous time series.
The resulting time series forms the largest in situ temperature and precipitation dataset available for the Congo Basin, and represents a major step toward advancing hydroclimatic and climate change research in this data-sparse region.
Using this consolidated temperature and precipitation dataset, we analyze temperature trends across the Congo Basin from the 1930s to the early 2000s. We examine long-term trends in daily maximum, minimum, and average temperatures; investigate shifts in their distributions across decades; and assess changes in the frequency of hot and cold extremes. We further investigate precipitation extremes through changes in daily precipitation return periods. Lastly, we compare the observed temperature trends with those derived from the state-of-the-art ERA5-Land reanalysis product developed by the European Centre for Medium-Range Weather Forecasts (ECMWF).
Data and MethodsData
The INERA – Section de Climatologie archives in Yangambi (hereafter referred to as INERA archives) host daily meteorological observations from 37 weather stations throughout the DRC, including records of temperature, evaporation, and precipitation, along with various other weather data variables, with most spanning from the 1960s to the mid 1990s, and only a few through the early 2000s. Historically, these archives have served as the central repository for all weather data recorded at INERA stations throughout the country (Fig. ). These records consist of paper bundles containing handwritten observations on tabular sheets, with each bundle representing weather data from a specific station for a given year, and each sheet representing one month's data from that station (Fig. ).
Each sheet contains daily weather observations, their pentad totals and averages (5 and/or 6 d), as well as monthly totals and averages. Specifically, for temperature records, each sheet provides daily maximum, minimum, and average values, along with the diurnal temperature range. There is also dry bulb and wet bulb temperatures measured at three times per day (06:00, 15:00, and 18:00 local time). Additionally, each sheet provides metadata, including the station name, year, month, observer's names, authorization signature, and occasionally extra comments at the bottom of the sheet (Fig. ). No information on the location of the weather stations is given on these sheets, however, the data inventory we performed at the archives provided us with information on the station code, latitude, longitude, and altitude of each of the weather stations (see Fig. ).
While the paper size, quality, colour, and maintenance condition vary across the bundles and stations, the format of the sheets and tables is generally consistent across these post-1960 records. However, it is important to note that in a small fraction of cases, the tables on the sheets were hand-drawn, possibly when the weather station ran out of standard paper, which presents an additional challenge for non-manual transcription (see Fig. ). In this dataset, 225 sheets were hand-drawn and were therefore excluded from automated transcription in this study.
The pre-1960 records digitized (imaged) during the COBECORE project follow a different table structure than the post-1960 sheets. Although they similarly provide daily maximum, minimum, and average temperatures for a given month and station, they do not include pentad totals or averages. Instead, these earlier sheets report only end-of-month totals and averages. They also lack the dry bulb and wet bulb temperatures recorded at three times per day as in the later records, and the table formats vary from decade to decade, and in a few cases between stations. Examples of these variations can be seen in Figs. , , and , corresponding to the 1930s, 1940s, and 1950s, respectively.
In addition to the weather data sheets, additional archive documents available through the COBECORE records , namely Instructions aux observateurs – INEAC and Guide pratique à l'usage des observateurs – Réseau Météorologique du Congo Belge, provide information on some of the instruments and observation procedures prescribed for use at the stations, including thermometer types, measurement techniques, and rain-gauge information. These documents demonstrate that standardised instrumentation and observation procedures were prescribed to ensure rigorous and consistent measurements across the INEAC/INERA observational network. Examples of these documents are shown in Figs. –.
Two photographs showing the interior of the historical weather data archives at the INERA – Section de Climatologie site in Yangambi, Democratic Republic of the Congo. These shelves contain thousands of paper-based meteorological records spanning several decades, and form the foundation of our climate data rescue efforts. Photographs taken on 9 March 2023.
An example One Month observed weather data sheet for Station Luki-Plateau (5°37′ S, 13°06′ E) in DRC available within the INERA – Section de Climatologie archives in Yangambi. The sheet contains the following information: (i) the station name (“Station”), year (“Année”) and month (“Mois”) in the top right corner, (ii) the data owner (“Institut National pour l'Étude Agronomique du Congo”) in the top left corner, (iii) a large table (center) showing the pentad number (“No. de la pentade”), date (“Date”), a measure for total solar radiation (“Bellani”), extreme temperatures in degree Celsius (“Températures extrêmes”) consisting of maximum, minimum and average temperatures, as well as the diurnal temperature range (“Ampl.”), Evaporation in cubic centimeters (“Evaporation”), Rainfall (“Pluies”) in millimeters per day, and Dry- and Wet-bulb Temperatures and Humidity (“Température et Humidité”) recorded at three times of the day, namely 06:00, 15:00, and 18:00 local time, and (iv) observer's names (“Les Observateurs”) in the bottom left, authorization signature (“Visa du Chef hiérarchique”) in the bottom right, and extra comments written in the bottom center of the sheet. Additionally, the table also shows 5 d (and/or 6 d) totals (“Tot.”) and means (“Moy.”), as well as monthly totals and mean values. Note that the values in these sheets are all handwritten, and the handwriting styles and paper maintenance conditions vary throughout the dataset. In this case, the image represents a well-preserved example relative to the range of paper conditions present in the archive.
Data Rescue Missions
We undertook two data rescue missions: (i) a two-week campaign to construct a comprehensive data inventory (7–17 March 2023), and (ii) a two-month campaign for the digitization of the archived historical weather data sheets (August–September 2023). Throughout both the data inventory and digitization phases, we followed the guidelines for climate data rescue established by the World Meteorological Organization . In the following subsections, we describe our inventory and digitization campaigns in the INERA archives, DRC.
Data Inventory
The inventory process within the INERA archives involved two main steps: (i) collecting metadata, including information on all stations such as station names, codes, and location (latitude, longitude, and altitude), and (ii) cataloging the available data sheets and variables present for each station by year, along with the paper formats and maintenance condition.
The first step is important as it ensures that the locations of the recorded weather data are correctly linked to the accurate locations (latitude, longitude, and altitude) for each weather station (Figs. and ). This is crucial for any research, analysis and conclusions that would be drawn from the final transcribed data. The next step involved determining the total number of available weather data sheets in the archives for all 37 stations. A total of 9885 data sheets (in a similar format to Fig. ) were identified in the archives and checked for the presence of daily temperature and precipitation records (see Fig. ). Additionally, the placement and order of the sheets within the bundles were verified to ensure they were correctly organized by month and year, and properly assigned to the corresponding station. These two steps are essential to prevent errors during the subsequent digitization and transcription processes, such as the loss of metadata or the incorrect assignment of data to the wrong station or date. The results from the inventory process were collected in a spreadsheet (“INERA_Stations_Data.xlsx”) which is published alongside the database (see Code and data availability).
Map showing the location of the 37 weather stations in the DRC whose data is hosted by INERA – Section de Climatologie archives in Yangambi (Background map courtesy of ESRI National Geographic).
Inventory of available daily precipitation (P) and temperature (T) data for weather stations across the DRC, based on records from the INERA – Section de Climatologie archives in Yangambi and the State Archives of Belgium. Blue bars indicate months where both daily P and T data are available in hard copy at the INERA – Section de Climatologie archives in Yangambi; hatched blue bars indicate such availability at the State Archives of Belgium. Green markers denote periods where only one of the two variables (P or T) is available in hard copy. Orange bars represent months for which data have previously been transcribed, either by INERA, COBECORE, or as compiled in .
Digitization
To preserve and enable analysis of historical weather data stored at the INERA archives, we digitized all available data sheets using a high-resolution digital camera (Canon EOS 2000D with an EF-S 18–55 mm IS II lens) mounted on a tripod, combined with a photo box (with a black background) equipped with uniform lighting (Fig. ). This digitization step was essential not only for the subsequent transcription steps, but also to prevent further data loss due to pests or fires, such as the recent February 2023 fires reported in one of the wings of the INERA Yangambi archives . Given the lack of electricity at the archive site, the set-up was powered by two rechargeable batteries for the camera.
The resulting images were saved in JPG format, as it preserves high image quality while remaining an open and widely supported format. The images were organized into station-specific folders, and each image file was named using a standardized convention encoding key metadata: station code, year, month and page format. Thus, each file name follows the format STN_YYYYMM_PF, where STN is the station code, YYYY is the year, MM is the month, and PF is the page format (SF for standard format and HD for hand-drawn format). This naming convention ensures easy identification and retrieval of the 9885 digital data copies.
Digitization campaign setup at the INERA – Section de Climatologie archives in Yangambi, including a digital camera mounted on a tripod and a photobox equipped with lighting fixtures to ensure high-quality digital copies of the historical weather records. Photograph taken on 15 August 2023.
Transcription
To automatically transcribe all the digitized images of historical weather data sheets, we developed and applied a new version of MeteoSaver, version 1.1 . In the following subsections, we first briefly introduce the original software, and subsequently describe three main improvements implemented in the new release (v1.1).
Original MeteoSaver Description
MeteoSaver is an open-source machine learning (ML)-based software specifically developed to transcribe handwritten tabular data into machine-readable format ready for analysis . MeteoSaver takes images of tabular weather data as input and processes each image through six modules: (i) configuration, (ii) image pre-processing, (iii) table and cell detection, (iv) transcription, (v) quality assessment and quality control (QA/QC) and (vi) data formatting and upload .
In the configuration module, the software requires user-specific settings, including, but not limited to, the choice of computational environment (personal computer or High Performance Computing (HPC) infrastructure), directory paths for input and output files, and a description of the general table structure (e.g., number of rows and columns). These user settings enable the software to run smoothly across the subsequent modules. The image-preprocessing module enhances the quality of input images using adaptive threshold binarization thereby improving the performance of the subsequent table and text recognition steps. Following this, the table and cell detection module applies ML-based algorithms to detect only the tables from the images, correct for skew or tilted tables/images, and identify individual table cells. These detected cells are then subsequently passed to the transcription module, which uses a pre-trained Tesseract optical character recognition (OCR) model to recognise handwritten values. All transcribed values are then evaluated in the Quality Assessment and Quality Control (QA/QC) module, where each value is individually checked. These checks include maximum and minimum threshold tests, logical consistency checks, inter-variable relationship checks, checksum validation, and outlier identification. Finally, QA/QC “confirmed” transcribed data are formatted per station into both the Station Exchange Format (SEF), as prescribed by the Copernicus Data Rescue Service, and the Comma-Separated Values (CSV) format, making the data ready for upload to open-access repositories. For a detailed description of the software, refer to .
In the following subsections, we describe three main improvements implemented in the new release (v1.1).
Improved Table Detection
As an improvement to the Table and Cell Detection module in MeteoSaver v1.0, version 1.1 introduces more precise cell detection by using the pre-existing horizontal and vertical line detection based on structuring elements generated with the OpenCV (Open Source Computer Vision) library .
Within the MeteoSaver version 1.0 framework, once the table boundaries were identified using OpenCV, individual table cells were detected through further image processing. This involved dilating the inverted binary image of the table (with horizontal or vertical lines removed), thereby turning the handwritten values into white blobs on a black background. Digits in close proximity, as expected for multi-digit values, were merged into a single text blob, which was then treated as an individual cell . While this approach generally performs well in identifying individual cells, it occasionally misclassified text written near boarders, and close to text in adjacent cells, as a single blob.
To address this limitation, we introduce in version 1.1 an additional step in the module that re-erases the pre-identified vertical and horizontal lines after text blob identification. This serves as a secondary check to avoid merging of text blobs across different cells.
Additionally, we have optimized the overall table and cell detection process in our framework by using arrays to store the detected text components and noise features (e.g., dots), which reduces this module's processing time by more than half.
Improved Transcription
Building upon the transcription module in MeteoSaver v1.0, we further train the Tesseract optical character recognition (OCR) model on a broader range of handwriting styles to increase its accuracy in text recognition. This is achieved by feeding the tesstrain model with ground truth data in the form of images and their corresponding transcriptions as text files .
With every MeteoSaver v1.1 run, we generate an extensive training database from the INERA handwritten monthly data sheets, comprising more than 5 million clipped individual cell images in total and their corresponding confirmed transcriptions (post QA/QC checks), where available. By iteratively adding newly confirmed transcribed cell images, and their respective values, from preceding runs to this database, and retraining the model using tesstrain, the Tesseract OCR model continuously learns new handwriting styles across the dataset. This improvement makes use of Tesseract's character pattern recognition capabilities (Tesseract 3) and its Long Short-Term Memory (LSTM) neural network for line-based recognition (Tesseract 4) . For example, after five consecutive runs of MeteoSaver over the entire INERA database, the total number of QA/QC confirmed transcribed temperature and precipitation values increased from 321 456 to 1 019 153, corresponding to a 215 % increase in confirmed values, highlighting the iterative learning capability of the software.
QA/QC extension to additional weather variables
demonstrates the use of the QA/QC module in MeteoSaver v1.0 to perform quality checks on transcribed daily maximum, minimum, and average temperatures (Tx, Tn, and Tavg). In version 1.1, we extended the QA/QC module to other recorded variables on the INERA sheets, namely: daily total precipitation (P, mmd-1), and dry bulb temperature (T, °C), wet bulb temperature (T′a, °C), actual vapour pressure (e, hPa), relative humidity (U, %) and vapour pressure deficit (Δe, hPa) recorded at three times of the day (06:00, 15:00, and 18:00 local time) (Fig. ). The following logic checks are now included in the QA/QC module.
Similar to the checks performed on the transcribed values for Tx, Tn, and Tavg in v1.0 (Sect. ), the latest version applies user-defined (i.e., region-specific) maximum and minimum temperature thresholds to instantaneous T and T′a as a first check . The multi-day (pentad) totals and averages are then calculated for transcribed T and T′a values within these thresholds, as well as for P, and compared with their corresponding transcribed pentad totals and averages. If the transcribed pentad total and/or average is equal to (or, for T and T′a, within the set uncertainty margin of 0.2 °C) the calculated pentad total and/or average, all the respective transcribed daily values for T, T′a, and P, leading up to that pentad total and/or average, are marked as confirmed. An uncertainty margin of 0.2 °C is applied to temperature checks to account for potential human rounding errors in the original pentad averages or Tavg values. Such rounding may result in small discrepancies (typically <0.2 °C) between the transcribed and QA/QC-calculated values.
In addition, logic checks are carried out on the transcribed values for T, T′a, e, U, and Δe for each day as follows:
T must be greater than or equal to T′a. Values that do not meet this condition are flagged as inconsistent.
We calculate U (using Eq. ), and compare it with the transcribed U. If the transcribed U is equal to (or within the set uncertainty margin of 0.2 %) the calculated U, we mark the transcribed U, e and Δe values for that day and time as confirmed.U=ees⋅100where, e is the actual vapour pressure and es is the saturation vapour pressure. Given that es values are not recorded in the DRC sheets, we derive them using e and Δe as follows:Δe=es-e
Rearranging Eq. (), es can be expressed in terms of Δe and e as:es=Δe+e
Thus, by substituting terms in Eq. (), we calculate U in terms of Δe and e as:U=eΔe+e⋅100
Following this, if the U, e and Δe values at a particular time have passed the internal consistency check in step (ii), we subsequently calculate the T using a modified Magnus-Tetens formula (Eqs. –). If the calculated T is equal to (or within the set uncertainty margin of 0.2 °C) the transcribed value, we mark the transcribed value as confirmed; otherwise it is replaced with the calculated value. In the case of replacement, this asymmetric correction is applied because, at this stage, U, e, and Δe have been shown to be mutually consistent, and T, in that case, is the only variable inconsistent with the reconstructed relationship. We therefore consider a single transcription error in T more likely than compensating errors in U, e, and/or Δe that nevertheless preserve their mutual consistency. However, this step does not assume that e and Δe are free of uncertainty. Small discrepancies arising from rounding or derived-variable precision of T are accounted for through the 0.2 °C margin. The correction should therefore be interpreted as a rule-based QA/QC consistency correction, rather than as a quantitative estimate of the propagated uncertainty in the calculated T.es=6.112exp17.67TT+243.5
By changing the subject to T and substituting es,T=243.5⋅ln(Δe+e6.112)17.67-ln(Δe+e6.112)
If both the transcribed values of T and T′a at a particular time and day are already marked as confirmed from their pentad totals and/or averages, we calculate the e using the psychometric equation from the Smithsonian meteorological tables (Eq. ) and subsequently the U and Δe values for that day and time are calculated (using Eqs. –).
However, if only one of the two transcribed values (T and T′a) at a particular time and day is marked as confirmed by either the pentad total or average, the confirmed value is used to check the relationship between the transcribed e, U and the other of the two (applying the Eqs. –). If the calculated value of either T or T′a matches the transcribed value (or within the uncertainty margin of 0.2 °C), the respective transcribed value is marked as confirmed. This relationship check indirectly checks the transcribed U and Δe values and corrects the incorrectly transcribed one if the other is marked as confirmed.7es(T′a)=6.112exp17.67T′aT′a+243.58e=es(T′a)-γ(T-T′a)where:γ=0.000661+0.00115T′ap0and p0 is the standard atmospheric pressure at mean sea level (1013.25 hPa)
Application of MeteoSaver to the DRC data
We apply MeteoSaver v1.1 to 9885 images of historical weather data sheets from 36 stations across the DRC (one of the 37 stations was previously transcribed by ), to automatically transcribe daily records of Tx, Tn, Tavg, P, T, and T′a. As a prerequisite for executing MeteoSaver, we input the user-defined settings in its configuration module, for example: (i) we set the computational environment to hpc, as we run the software on HPC infrastructure due to the large volume of images. Processing all images requires approximately 4.5 d on the VSC (Flemish Supercomputer Center) infrastructure using 18 parallel CPU cores. (ii) We specify all required input and output directories. (iii) We set region-specific maximum and minimum temperature thresholds to 40 and 5 °C, respectively . (iv) We define the uncertainty margins for the QA/QC checks as 0.2 °C for temperature values and 0.2 percentage points for relative humidity values. In the MeteoSaver configuration module, these margins are not estimates of the full observational uncertainty, but acceptance criteria used to account for rounding in the handwritten records and in derived quantities during internal consistency checks. Temperature values in the sheets are reported to one decimal place; therefore, a margin of 0.2 °C allows small discrepancies arising from rounding when comparing transcribed values with recalculated values, pentad averages or consistency relationships such as that between Tx, Tn, and Tavg. Similarly, the 0.2 percentage-point margin for relative humidity accounts for rounding differences in the pentad averages and derived relative-humidity consistency checks. These margins were therefore selected based on reporting precision and rounding considerations, rather than through a separate sensitivity analysis. We detail the full list of our configuration settings in Table , and in the configuration.ini file available in the Zenodo repository .
Although most of the tables in our dataset share a similar structure, applying the software to such a large and diverse dataset, characterized by multiple handwriting styles, different paper sizes, varying paper quality and maintenance conditions, demonstrates its robustness and flexibility.
The final output consists of QA/QC confirmed daily temperature and precipitation time series formatted in the Station Exchange Format (SEF), ready for subsequent analysis. This resulting dataset has been uploaded to the Copernicus Global Land and Marine Observations Database, where it will be made available to users via the Copernicus Climate Change Service Data Store, as well as on Zenodo .
Statistical AnalysisValidation
Following the transcription and quality control of the daily temperature and precipitation data from 37 weather stations across the DRC using MeteoSaver, we validate the automatically transcribed data against manually transcribed records by INERA for the five stations where such data were available and overlapped with our automatically transcribed time series, namely: Kiyaka-Plateau, Luki-Plateau, Gandajika, Mulungu-Mulohe, and Gimbi-Plateau. While these five stations may not constitute a fully representative sample of all climatic regions in the DRC, handwriting styles, and temporal periods in the broader archive, they provide a useful validation subset because they are geographically distributed across the DRC (Fig. ), originate from different stations and observers, and span different periods from the 1960s to the 2010s, including both short and multi-decadal records. They therefore capture differences in handwriting styles, table completion practices, and paper maintenance conditions. The validation should therefore be interpreted as an assessment of MeteoSaver performance across the available manually transcribed subset, rather than as a complete evaluation of all possible archive conditions.
We assess the accuracy of the automatic transcription by calculating the percentage of correctly transcribed daily temperature values (Tx, Tn, and Tavg) relative to their respective manually transcribed values. This comparison assumes that the manual transcriptions are accurate and therefore does not account for potential errors in the manually transcribed data itself. Similar to , we apply an uncertainty margin of 0.2 °C during this comparison. In cases where a value is incorrectly transcribed, we also examine the manually transcribed values for the preceding and following days to identify potential 1 d shifts caused by row or cell misalignment during table and cell detection. We, therefore, present both the percentage of correctly transcribed values (hereafter referred to as the match rate) and the percentage of values correctly transcribed but shifted by ±1 d (hereafter referred to as the 1 d shift percentage).
Finally, we compute the Mean Absolute Error (MAE) of the automatically transcribed values, using the corresponding manually transcribed values on the same day as the reference.
Trend Analysis
We undertake trend analysis of Tx, Tn, Tavg, and P values. The analysis primarily covers the period from 1960 to the 1990s, with 11 stations with data pre-1960 and 6 stations extending into the 2000s.
It is important to note that some stations are excluded from this analysis due to substantial data gaps, incomplete years, or insufficient quality in the transcribed temperature records. Specifically, stations with fewer than 25 years of data and/or with significant missing periods (>5 years) that would bias trend estimation are omitted. 21 of the 37 stations are considered for trend analysis in the DRC (Table ).
To assess long-term changes in temperature, we apply the widely used Theil–Sen estimator to determine multi-decadal trends in Tx, Tn, and Tavg for each station across its respective period of record. The Theil–Sen estimator was applied directly to the available time series, with missing values ignored in the trend estimation. Here, we express trends in degrees Celsius (°C) per decade. Additionally, we apply spatial interpolation of station-based trends using a Radial Basis Function (RBF) approach, a commonly used method for spatial interpolation of in situ meteorology data , to visualize their spatial patterns across the DRC. The interpolation is performed using stations with a minimum record length of 25 years within the most overlapping observation window and is constrained to the DRC boundary.
In parallel, we apply the kernel density estimation (KDE) method to construct probability density functions of Tx, Tn, and Tavg, grouped by decade per station. This approach enables us to (i) examine shifts in the overall temperature distributions over time, assessed through the changes in the mode and the 95th percentile of each variable's distribution per decade, and (ii) evaluate changes in the temperature extremes per station, based on the frequency of days falling outside the middle 90 % of the distribution (i.e., below the 5th percentile and above the 95th percentile). Similar to , we define these extremes for each variable using KDE-derived percentiles calculated across the entire station record and then count how many days per decade fall below (cold days/nights) or above (hot days/ warm nights) these thresholds. This approach enables a robust assessment of both the central shift in temperature and the changing frequency of temperature extremes over time at each station. For the aggregated temperature distributions across stations, we pool all QA/QC-confirmed daily observations from the stations into a single regional dataset and restrict the analysis to the most overlapping observation window, 1961–1990 (Fig. ). Decade-specific KDEs are then calculated from the pooled daily values of Tx, Tn, and Tavg. The aggregated KDEs should therefore be interpreted as distributions of available QA/QC-confirmed station-day observations across the DRC during 1961–1990, rather than as equal-weighted averages of station-level distributions.
Furthermore, we investigate the changes in daily precipitation return periods at stations with records spanning over 30 years. For each station, we split the daily precipitation time series into an earlier and a later half by dividing the record at the midpoint of the valid daily observations, so that both halves contain a similar number of data points. For each half, we estimate empirical return-period curves using the Weibull plotting position, focusing on the upper tail of the distribution, to highlight changes in heavy precipitation events. We plot these curves on the same figure to visualise how the intensity of heavy precipitation events has changed between the earlier and later periods.
Lastly, we compare the observed temperature trends at each station with those derived from the ERA5-Land dataset. ERA5-Land provides hourly 2 m air temperature data over land at a spatial resolution of 0.1° × 0.1° (approx. 9 km), from 1950 to present . We derive daily Tx, Tn, and Tavg from the ERA5-Land hourly dataset. For each station, we extract ERA5-Land values from the grid cell whose center lies closest to the station coordinates, using the nearest-neighbor method. To ensure direct comparability, we compute the trends in ERA5-Land only over time instances where both the station observations and the corresponding ERA5-Land timeseries contain data.
ResultsDatabase characteristics
In total, 1 011 297 daily and sub-daily temperature records, as well as daily precipitation observations, across the 37 weather stations achieved the highest internal quality flag, meaning that they passed all applicable QA/QC checks described in Sect. . These records are hereafter referred to as QA/QC confirmed values. Together with manually transcribed daily temperature and precipitation records from previous efforts, namely (95 395 records) and the COBECORE project (246 815 records), they form the foundation of the first large-scale, digital historical climate dataset for the Democratic Republic of the Congo. The dataset spans the period from 1908 to 2023. The longest station record is from Eala (1908–1991, with eight missing years), while the shortest record is from Lubarika (1961–1963). Most station records cover the 1960s through the 1990s (Table , and Figs. and ). The percentage of QA/QC confirmed values out of the total transcribed values for the daily temperature records ranges from 29.1 %–97.5 % across the 37 stations, with a mean of 61 %.
This percentage is lower for daily precipitation records ranging from 10.3-89.4 % with a mean of 30 % across the stations. This lower confirmation rate does not necessarily imply that all unconfirmed precipitation values are erroneous but rather reflects the more limited set of automated QA/QC checks available for precipitation, in comparison to temperature. The QA/QC module includes only pentad total consistency checks for precipitation, whereas daily temperature values are subjected to additional consistency checks, including the physical relationship between Tx, Tn, and Tavg, their consistency with the diurnal temperature range (Ampl.), and maximum and minimum threshold checks (see Fig. ) . These additional checks increase the likelihood of detecting and correcting transcription errors, thereby leading to a higher percentage of confirmed temperature values. Consequently, the precipitation analyses in this study are restricted to station records for which more than 30 years of data are available and at least 80 % of the transcribed daily precipitation values were confirmed through the QA/QC procedure, and the results should therefore be interpreted as station-specific examples rather than as a spatially representative assessment of precipitation extremes across the DRC.
Summary of station data sheets transcribed using MeteoSaver v1.1, including also already manually transcribed data.
where: Tx: daily maximum temperature, Tn: daily minimum temperature, Tavg: daily average temperature, and P: precipitation. Altitude is given in meters above sea level (m a.s.l.). Confirmed values are calculated with an uncertainty margin of 0.2 °C.
Validation
We compare the automatically transcribed daily temperature values (Tx, Tn, and Tavg) from five stations with their corresponding manually transcribed values produced by INERA. The validation results indicated that across these stations, the automatically transcribed daily temperature records had different accuracies across the stations ranging from 81.2 % to 92.3 %, with a median match rate of 87.4 % and a median 1 d shift percentage of 14.7 % across all the five stations (Fig. ).
Finally, we compute the mean absolute error (MAE) of the automatically transcribed values relative to the manually transcribed values on the same day. The MAE across the transcribed sheets of these five stations ranges from 0.18–0.51 °C, with a median of 0.35 °C (Fig. ).
Time series plots of the daily maximum (red), average (orange), and minimum (blue) temperatures for stations: (a) Kiyaka-Plateau (5°16′ S, 18°57′ E), (b) Luki-Plateau (5°37′ S, 13°06′ E), (c) Gandajika (6°45′ S, 23°57′ E), (d) Mulungu-Mulohe (2°18′ S, 28°47′ E), and (e) Gimbi-Plateau (5°31′ S, 13°22′ E). Automatically transcribed values are shown as solid markers, while manually transcribed values by INERA are illustrated as lighter time series bands, with a 0.2 °C uncertainty margin applied during QA/QC checks (as done in ). The match rate of the automatic transcription relative to the manual transcription, the 1 d shift percentage (±1 d), and the mean absolute error (MAE) are indicated in the lower right corner of each panel.
Trend AnalysisWarming Trends and Shifting Temperature Distributions
Of the 21 stations considered in the trend analysis (see Sect. ), 16 exhibit an increasing trend in daily maximum temperature (Tx) over their respective observation periods (Table , Figs. and –), with rates ranging from 0.14 to 0.51 °C per decade (equivalent to 0.4–1.5 °C over 30 years, with a median of 0.8 °C warming over this period). Four stations show no change in Tx, while one record shows a decreasing trend of -0.20 °C per decade (-0.6 °C over 30 years). Across all 21 stations, the median Tx trend is 0.22 °C per decade which corresponds to approximately 0.7 °C of warming over a 30-year period.
For daily minimum temperature (Tn), 13 stations show an increasing trend of 0.03–0.33 °C per decade (0.1–1 °C over 30 years, with a median of 0.6 °C), six show no change, and two record decreases of -0.21 and -0.44 °C per decade (-0.6 and -1.3 °C over 30 years, respectively). The median Tn trend across all the stations is 0.10 °C per decade, equivalent to 0.3 °C increase over 30 years (Table , Figs. and –).
Average daily temperature (Tavg) increases at 18 stations, with trends of 0.05–0.41 °C per decade (0.2–1.2 °C over 30 years, with a median of 0.5 °C). Two stations show no change in Tavg, and one shows a decline of -0.14 °C per decade (-0.4 °C over 30 years). Across these stations, the median Tavg trend is 0.15 °C per decade, corresponding to approximately 0.5 °C of warming across a 30-year period.
Overall, at 20 of the 21 assessed stations, at least one of Tx, Tn, or Tavg exhibits a positive trend, i.e., at each of these stations, at least one of the three variables shows an increase (Table , Figs. and –). Spatial interpolation of these temperature trends at the stations during the most overlapping observation window (1961–1990) illustrates a widespread warming signal throughout the network (Fig. ).
Daily temperature trends at stations: (a) Luki-Plateau (5°37′ S, 13°06′ E), (b) Mulungu-Mulohe (2°18′ S, 28°47′ E), (c) Yangambi-Km.5 (0°46′ N, 24°29′ E), and (d) Eala (0°03′ N, 18°18′ E) in the DRC. Colored dots represent daily values of maximum (Tx, red), average (Tavg, orange), and minimum (Tn, blue) temperatures. Solid lines indicate Theil–Sen trend estimates for each variable, with the corresponding slope (°C per decade) shown at the end of each line. This data includes both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. Note that these stations have different start and end years.
Daily temperature trends at 21 stations selected for the climate analysis in the DRC. These trends are estimated using the Theil–Sen estimator . The data used to estimate these trends includes both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA, and through the COBECORE project. Start and end years indicate the period of available observations for each station. Missing years represent the number of years without data within this period.
No.Station nameStart yearEnd yearMissing yearsTrend (°C per decade) TxTnTavg1Yangambi-Km.51950202040.190.100.142Yaekama1970201110.320.130.223Gimbi-Plateau1960201220.220.110.174Luki-Plateau1960201120.210.030.125M'Vuazi-Poste (Plateau)1960199540.250.000.156Binga195919928-0.200.220.007Eala1908199180.000.190.098Bambesa1960199230.14-0.44-0.149Boketa1960199120.310.000.1710Kutubongo1960198410.000.000.0011Kiyaka-Plateau1967201220.000.190.1212Gandajika1960199010.200.000.0513Kisanga-Plateau1960199020.280.290.3914Kaniama1960199130.510.000.1915Kipopo1960199330.38-0.210.0616Mulungu-Mulohe1960200110.230.300.2717Mulungu-Nyamunyunye1960198620.000.280.1618Mulungu-Tshibinda1960199110.480.330.4119Mulungu-Bukulumisa1960198420.340.000.1120Nioka-Drusi1960199050.430.060.2221Rumangabo1960198630.190.260.26
Where: Tx: daily maximum temperature, Tn: daily minimum temperature, and Tavg: daily average temperature.
Spatial interpolation of temperature trends in daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures across the DRC, based on observations from 19 climate stations during the period 1961–1990. Trends are expressed in °C per decade and were derived using Theil–Sen slope estimates for each station. The interpolated fields were generated using the Radial Basis Function (RBF) method with a linear kernel for scattered data interpolation, and visualized using a continuous color scale. Station locations are indicated by black dots outlined with white circles. Lakes are visualised in dark blue. Note that this map serves as a visualization based on the limited number of available stations (n=19) and may not fully represent the spatial variability across the region.
We then examine the decadal temperature distributions both at individual stations and for the aggregated dataset across stations. For the latter, all available QA/QC-confirmed station-day observations are pooled across stations, and the analysis focuses on the most overlapping observation window (1961–1990).
In recent decades, the distributions of Tx, Tn and Tavg have shifted toward higher values, reflecting warmer modal temperatures and higher 95th percentile values. This pattern is observed at both individual stations (Figs. a–d and –) and in the aggregated dataset for the DRC (Fig. ). For example, at Station Mulungu-Mulohe, the mode of Tx and Tn during the last two decades (1981–1990 and 1991–2000) is approximately 1 °C higher than in the first decade (1961–1970) (Fig. b). Similarly, the 95th percentile values of Tx and Tn at this station are about 1 °C higher in the last decade (1991–2000) compared to the first decade (1961–1970) (Fig. b).
In the aggregated dataset, the distributions of Tx, Tn, and Tavg values over the overlapping window (1961–1990) shift progressively to warmer values with each successive decade (Fig. ). The bimodal distribution of temperatures observed across all stations in the first decade (1961–1970) becomes less pronounced, evolving toward a more unimodal distribution in Tx and Tavg. This corresponds to an increase in the density of higher temperatures during the second decade (1971–1980) and an even greater increase in the last decade (1981–1990). An increase in the 95th percentile values for Tx and Tavg of just under 1 °C is also observed in the last decade compared to the first (Fig. ).
Decadal shifts in temperature distributions at stations: (a) Luki-Plateau (5°37′ S, 13°06′ E), (b) Mulungu-Mulohe (2°18′ S, 28°47′ E), (c) Yangambi-Km.5 (0°46′ N, 24°29′ E), and (d) Eala (0°03′ N, 18°18′ E) in the DRC. Kernel density estimates of daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures are shown for each decade: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). Vertical dashed lines indicate the 95th percentile values for each decade, with panels (a), (b) and (c) illustrating the progressive rightward shift towards higher temperatures and more extreme values over time.
Decadal shifts in temperature distributions based on an aggregated dataset of 37 stations in the DRC for the period 1961 to 1990. The aggregated distributions were obtained by pooling all available QA/QC-confirmed station-day observations across stations during 1961–1990; no additional station-level weighting was applied. Kernel density estimates of daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures are shown for each decade: 1961–1970 (yellow), 1971–1980 (red), and 1981–1990 (maroon). Vertical dashed lines indicate the 95th percentile values for each decade, illustrating the progressive rightward shift in temperature extremes over time.
More Hot Days and Warm Nights, Fewer Cold Events
The average number of hot days per year, defined here as days with Tx exceeding the 95th percentile threshold, has more than tripled over the period 1961–2000 at several stations, including Luki-Plateau, Mulungu-Mulohe, and Yangambi-Km.5 (Fig. ). Across all stations, during the most overlapping observation window (1961–1990), the number of hot days has on average increased from about 3 d per year with Tx above 33.3 °C (the 95th percentile threshold) in the first decade to about 5 d yr−1 in the last decade (Fig. ). Similarly, the average number of days with Tavg exceeding 26.9 °C (the 95th percentile threshold) increased from approximately 3 to 6 d yr−1, corresponding to a doubling over this period.
Regarding the 97.5th percentile, the most recent decade exhibits on average one additional hot day and warm night per year compared to the earliest decade. A similar increase of approximately one hot day is also observed at the 99th percentile for both Tx and Tavg (Fig. ).
In contrast, the average number of cold days per year across all stations, defined as days with Tx below the 5th percentile, decreased from around 7 to 5 d yr−1 between the first and last decades considered (Fig. ). At more extreme thresholds (2.5th and 1st percentiles), cold days declined by approximately 1 d yr−1 over the same period. Likewise, the average number of cold nights (Tn<5th percentile) declined from about 5 to 4 d yr−1 over the same period (Fig. ).
These differences in decadal average number of hot days, warm nights and cold days are much more difficult to interpret on a station by station basis (as shown in Figs. and –), as the extreme percentiles tend to be noisier, partially due to variations in time series completeness, internal climate variability, station data inhomogeneities (e.g., due to station location, instrument changes or observation practices), and possible local-scale effects (e.g., land-use and land-cover changes). However, the station-based analysis still shows a general tendency towards rising frequency of hot days/nights and declining cold events across decades, illustrating the shift in temperature distribution towards higher values. Moreover, when aggregated across all the stations in the DRC dataset (as shown in Fig. ), we observe a clear trend of more hot days and fewer cold days across the DRC from 1960 to 1990.
Frequency of hot and cold temperature extremes at stations: (a) Luki-Plateau (5°37′ S, 13°06′ E), (b) Mulungu-Mulohe (2°18′ S, 28°47′ E), (c) Yangambi-Km.5 (0°46′ N, 24°29′ E), and (d) Eala (0°03′ N, 18°18′ E) in the DRC. The plots show the average number of days per decade falling below the 5th percentile (cold events) and above the 95th percentile (hot events) for daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures, similar to . Values are based on kernel density estimates calculated across the entire period. The number of days per year within each decade are shown: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). The middle 90 % of the distribution (5th–95th percentile range) is omitted for clarity. The 95th percentile temperature for each variable is noted in the top-right corner of each plot.
Frequency of hot and cold temperature extremes aggregated across all the 37 stations in the DRC from 1961 to 1990. The plots show the average number of days per decade falling below the 5th percentile (cold events) and above the 95th percentile (hot events) for daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures, similar to . Values are based on kernel density estimates calculated across the entire period. The average number of days per year within each decade are shown: 1961–1970 (yellow), 1971–1980 (red), and 1981–1990 (maroon). The middle 90 % of the distribution (5th–95th percentile range) is omitted for clarity. The 95th percentile temperature for each variable is noted in the top-right corner of each plot. The rising frequency of hot days/nights and declining cold events across decades illustrates the intensifying shift in temperature extremes.
Changes in Precipitation Return Periods
For precipitation we analyse three stations with more than 30 years of daily precipitation records, namely: Yangambi-Km.5, Mulungu-Mulohe, and Gimbi-Plateau. The former two stations show a clear increase in the frequency of heavy precipitation in the later periods compared to the earlier periods. For instance, at Mulungu-Mulohe, a daily precipitation amount of 63 mm d−1 has a return period of about 2 years in the earlier period (1960–1985), but about 1 year during in the later period (1986–2011) (Fig. b), indicating that such an event has become twice as frequent. Similarly, at Yangambi-Km.5, a daily precipitation amount of 83 mm d−1 is twice as frequent in the later period (1990–2020) than in the earlier period (1960–1989) (Fig. a). These increases are much more pronounced at Yangambi-Km.5 and Mulungu-Mulohe in comparison to Gimbi-Plateau where there are no substantial changes in the return periods between the two time periods (Fig. c). In general, precipitation events greater than 50 mm d−1 are observed to occur more frequently in the later periods than in the earlier periods at Yangambi-Km.5 and Mulungu-Mulohe.
Changes in daily precipitation (P) return periods recorded at three stations: (a) Yangambi-Km.5 (0°46′ N, 24°29′ E), (b) Mulungu-Mulohe (2°18′ S, 28°47′ E) and (c) Gimbi-Plateau (5°31′ S, 13°22′ E). The blue lines represent the return periods of daily precipitation (in mm d−1) during the first half of the available time series at each station, while the red lines represent those during the second half of the period. Note that while two of the stations have the same starting year (1960), one starts from 1977. All three stations differ in their end years.
Comparison To Reanalysis
Comparison of the trends in transcribed observations of Tx, Tn, and Tavg with those in the ERA5-Land dataset reveals no systematic differences across the DRC (see Figs. –). Although ERA5-Land generally captures the seasonal patterns well compared to the observations (as shown, for example in Fig. a), the trends in Tx, Tn, and Tavg are sometimes either underestimated or overestimated relative to the observed values, depending on the station and variable (see Fig. –). These differences are typically of the same order of magnitude as the trends themselves, generally 10−1 °C per decade, with some discrepancies on the order of 10−2 °C per decade. While most stations show the same sign of difference in trend as their nearby stations, a few exhibit opposite signs compared to their closest station, for example Kipopo and Kisanga-Plateau (see Fig. ), further underscoring the non-systematic nature of differences between the observations and ERA5-Land, which may partly reflect non-climatic inhomogeneities in the station records.
We also observe that during the early 1950s, the ERA5-Land data displays a pronounced “jump” of approximately 4.5–5 °C in temperature timeseries for Tx, Tn, and Tavg that is not captured by stations with observations during that period, such as Yangambi and Eala (see Figs. c–d and a–b). This suggests that the ERA5 reanalysis may be weakly constrained in this region during the early reanalysis period, highlighting the need for more historical observational data for this region to improve this product.
Comparison of trends in daily maximum, minimum and average temperatures at four example stations: (a) Luki-Plateau (5°37′ S, 13°06′ E), (b) Mulungu-Mulohe (2°18′ S, 28°47′ E), (c) Yangambi-Km.5 (0°46′ N, 24°29′ E), and Eala (0°03′ N, 18°18′ E) in the DRC. Blue dots represent daily values of observed maximum, minimum and average temperatures, while solid yellow lines show the ERA5-Land timeseries. The dotted dark blue and orange lines indicate Theil–Sen trend estimates for each variable for observations (Obs) and ERA5-Land, respectively, with the corresponding slope (°C per decade) shown in the top left legend. The observations include both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. The trends are estimated using values where both the ERA5-Land data and transcribed observations are available. Note that these stations have different start and end years.
Map showing the differences between ERA5-Land and observed (Obs) trends in (a) daily maximum (Tx), (b) daily minimum (Tn), and (c) average (Tavg) temperatures across the DRC, based on observations from 21 climate stations and their corresponding ERA5-Land time series obtained using the nearest neighbor. The green shading represents the Natural Earth shaded background used for geographic context (http://www.naturalearthdata.com/, last access: 25 September 2026). Lakes are visualised in dark blue and occur primarily along the eastern border of DRC, while the Atlantic Ocean is shown in light blue to the west of the domain. Note that the stations have different start and end years.
Graphs comparing the observed trends in (a) daily maximum (Tx), (b) daily minimum (Tn), and (c) daily average (Tavg) temperatures from 21 weather stations across the DRC with their corresponding ERA5-Land time series. Note that these stations have different start and end years.
DiscussionTranscription with MeteoSaver and its Limitations
In this study, we apply MeteoSaver v1.1 to transcribe 9885 handwritten tabular historical weather data sheets from 36 INERA stations across the DRC. We obtain over one million QA/QC confirmed values for daily and sub-daily temperature and precipitation. Validation against previously manually transcribed data from five stations yields match rates ranging from 81.2 % to 92.3 %, with a median match rate of 87.4 %, a median 1 d shift percentage of 14.7 %, and a median MAE of 0.35 °C.
These results indicate that MeteoSaver performs robustly across diverse handwriting styles, variations in ink intensity of recorded values and table borders, paper maintenance conditions, and paper sizes, across this large dataset. An MAE of 0.35 °C is of the same order of magnitude as typical observational uncertainties reported for historical thermometer measurements of 0.2 °C (1σ) for stations within the latitude range of 20° S–20° N . Moreover, the mismatches associated with 1 d shifts (row shifts), rather than incorrect digit recognition, suggest row identification errors as one of sources of disagreement between the manually and automated transcription. Additionally, the validation checks we undertake against the manually transcribed data assume that those human transcriptions were performed accurately and therefore do not account for potential human keying errors, which have been shown to occur in manual transcription projects . Therefore, part of these differences observed in the validation here might be a result of the manually transcribed reference dataset. Overall, we consider the transcribed dataset sufficiently reliable for decadal-scale temperature trend analysis.
Nevertheless, we acknowledge that the following improvements need to be made to MeteoSaver to enhance its robustness and maximise the volume of transcribed data: (i) The transcription module can be improved through continued expansion of the OCR training dataset. This is because the iterative ML training strategy used in this study resulted in a 215 % increase in QA/QC confirmed values after five consecutive runs over the INERA dataset. This substantial improvement illustrates the importance of feeding a large dataset of diverse handwriting images to the TesseractOCR tesstrain model. This approach would also improve the model for scalability, especially in the context of additional data rescue projects in data-sparse regions, through the use of multi-station datasets; (ii) Improvements to the row identification algorithm within the Table and Cell detection module would reduce the ±1 d shift of correctly transcribed values and further improve alignment between detected values and calendar dates.
These software improvements would also increase its potential for further data rescue. For instance, while the output dataset presented here contains over one million QA/QC confirmed values for daily and sub-daily temperature and precipitation, approximately 300 000 additional temperature records and 200 000 precipitation records were transcribed but did not pass the current QA/QC checks. Engaging volunteers to review and verify such records could enhance data recovery while simultaneously generating additional training data to further refine the transcription module.
Similarly, the 225 hand-drawn sheets (as in Fig. ), which we excluded from automatic transcription due to their structural complexity, could benefit from the improved software, as well as citizen science efforts, to expand both the training dataset and the observations dataset.
In addition, the extensive pre-1960 climate records from 575 stations digitised (imaged) during the COBECORE project, as reported in , represent a major opportunity for further data rescue. Automatic transcription of these records could substantially expand the observational dataset, including additional meteorological variables (e.g., wind speed and direction, evaporation, and humidity). Although not all of these stations contain continuous long-term historical weather records, their transcription could potentially add millions of additional historical in-situ records for the Congo Basin.
Finally, this pipeline could also be extended to other national and international archives hosting historical climate data for the Congo Basin, such as the African Centre of Meteorological Applications for Development (ACMAD) database and national meteorological and hydrological services in the region.
Warming Trends and Shifting Temperature Distributions
Twenty of the 21 stations considered in this study show an increasing trend in at least one of the Tx, Tn, or Tavg values. The median warming trends are 0.22, 0.10, and 0.15 °C per decade for Tx, Tn, and Tavg, respectively, which correspond to approximately 0.7, 0.3, and 0.5 °C of warming over 30 years. When scaled to multi-decadal periods, these station-based trends are of similar magnitude to the reported warming of 0.75–1.05 °C since 1960 for Central Africa , indicating consistency between in-situ observations and prior regional-scale assessments that did not have access to these data.
These station-based trends are also consistent with the broader expectation of warmer temperature distributions under a warming climate. Similarly, reported warming trends in Central African temperature-extreme indices, including the warmest day, warmest night, coldest day, and coldest night, of 0.25, 0.21, 0.13, and 0.23 °C per decade, respectively, over the period 1955–2006, based on station observations from countries in the region, including additional stations in the DRC that are distinct from those analysed in this study.
These warming trends are also reflected in the aggregated dataset across all stations. Each successive decade within the period 1961–1990 displays a shift toward higher temperature values in distributions of Tx, Tn, and Tavg, showing higher modal temperatures, and higher extremes, here represented by the 95th percentile values (Fig. ). This pattern is also consistent at individual stations with longer time series extending beyond the common overlapping period i.e., stations with data before 1960 and/or continuing into the 2000s (see Fig. a–d).
Together, these results show a clear and consistent warming signal across the DRC during 1961–1990, with evidence of continued warming at the few stations with records extending into the early 2000s. Given that the DRC constitutes a substantial portion of the Congo Basin (approx. 62 % of the basin; ), these findings are likely broadly representative of regional conditions, although they may not fully represent spatial variability across the entire Congo Basin.
However, it is important to note that, while QA/QC checks were applied to identify and correct transcription errors and physically inconsistent values, homogenisation was not performed to detect and adjust for potential non-climatic inhomogeneities in the station records. Such inhomogeneities, well-rehearsed throughout the literature, may arise from changes in station location, instruments, calibration procedures, measurement precision, observation times or practices, and could introduce artificial (non-climatic) shifts in the temperature time series . While digitized records documenting the types of instruments used and some measurement procedures are available in the COBECORE records (Figs. –; ), the available metadata are not sufficiently complete to reconstruct station-specific histories of such changes throughout the full observational period.
Possible physical contributors may also have influenced differences among the station records. These include differences in elevation , changes in the immediate station environment, proximity to forests or water bodies, and local and remote land-use or land-cover change . For instance, report a loss of approximately 9.7 % of tropical forest in the Tshopo province, where three stations in our dataset are located (Yangambi-Km.5, Yaekama, and Gazi; the former two analysed), over the period 1900–2023 due to deforestation for agriculture and settlement. Within Yangambi alone, approximately 24 ha of primary forest was lost between 2002 and 2025, accounting for only about 2 % of the total tree-cover loss in the area recorded over the same period . Between 2001 and 2025, permanent agriculture was associated with approximately 19 % of the tree-cover loss in Yangambi . Such changes can lead to localised warming effects .
More generally, differences in trend magnitude and sign among stations may reflect a combination of large-scale climate variability, local environmental conditions, and station-specific non-climatic effects. However, because this study did not perform a formal attribution analysis, we could not quantify the relative contributions of such possible drivers to the observed station-level trends. Therefore, part of the distribution changes may reflect these local physical and/or non-climatic influences from one or more stations, although the consistency with previous regional assessments suggests that the broad warming pattern is robust.
Moreover, most station records stop in the mid 1990s because of interruptions in observations during the First Congo War (1996–1997) and subsequent Second Congo War (1998–2003) that led to nationwide institutional breakdown . This lack of sufficient data post 1990 is limiting given that recent studies report an increased global warming rate since the 1990s , suggesting that the full magnitude of recent warming in the region may not be captured within this available station data.
More Hot Days and Fewer Cold Events
Analysis of the aggregated dataset across all stations during the common overlapping period (1961–1990) reveals a clear increase in the frequency of hot days, represented here as the average number of days with Tx and/or Tavg values above the 95th percentile thresholds, with each successive decade (Fig. a, c). On average, the number of hot days per year nearly doubled between the earliest and most recent decades. These changes in hot extremes in the region have not been assessed in the IPCC Sixth Assessment Report (AR6), largely due to the sparsity of in situ observational data in Central Africa . Our results, therefore, contribute new observational evidence for changes in hot extremes in the region and underscore the importance of data rescue efforts in the region to enable robust detection and attribution studies.
In contrast, the frequency of cold days, here defined as days with Tx values below the 5th percentile, declined over the same period (Fig. a). Taken together, the increase in frequency of hot days and the decrease in frequency of cold days provide evidence of a warming signal across the DRC during 1961–1990, consistent with the distributional changes discussed in Sect. . A similar pattern is also observed in more recent decades (2000s and 2010s) at stations with data extending beyond 1990 (Fig. a–c).
Consistent with our findings, report increases in the frequency of warm extremes accompanied by decreases in cold extremes over the period 1955–2006 across Central Africa, based on observations from stations in Cameroon, the Central African Republic, São Tomé and Príncipe, the Republic of the Congo, Gabon, and eight additional stations in the DRC that are distinct from those analysed in this study. This indicates that these changes are not limited to the DRC but reflect broader regional trends across Central Africa.
Changes in Precipitation Return Periods
Unlike the transcribed daily temperature values, which were subjected to multiple QA/QC checks to confirm and correct transcription errors, the transcribed daily precipitation values were subjected to only one QA/QC check, namely the pentad total consistency check (Sect. ). Consequently, the number of QA/QC-confirmed daily precipitation values were substantially lower than for temperature. For this reason, we analyse changes in precipitation return periods for only three stations with more than 80 % QA/QC-confirmed values (Yangambi-Km.5, Mulungu-Mulohe, and Gimbi-Plateau), with the former two indicating an increase in the frequency of heavy precipitation events.
However, given that only three out of 37 stations in our dataset have more than 30 years of daily precipitation data, it is difficult to draw conclusions about the broader changes in frequency of heavy precipitation across the DRC. For this reason, our discussion mainly focuses on the changes in temperature across the region, for which we have a larger spatial dataset.
Comparison To Reanalysis
In this study, we compare temperature trends derived from station observations with those from ERA5-Land to assess the consistency between in-situ data and the reanalysis. As this comparison involves station (point-based) observations and gridded ERA5-Land data (matched using the nearest-neighbor method), differences in absolute temperature values are expected. Therefore, we focus on comparing temporal trends over time periods where both datasets contain data.
Our results show that ERA5-Land generally captures the seasonal variability well relative to the station records and exhibits no systematic differences in trends across the DRC. However, ERA5-Land trends sometimes underestimate or overestimate the trends, depending on the station and variable. In some cases, even nearby stations show opposite signs in their differences in trend (Figs. – and G1–G5). This may be partly due to non-climatic inhomogeneities in the station data (see Sect. ), which were not explicitly accounted for in this analysis.
A notable discrepancy is observed in the early period of the record, where ERA5-Land exhibits a pronounced “jump” of approximately 4.5–5 °C in temperature variations for Tx, Tn, and Tavg during the 1950s that is not evident in the station observations (Fig. c and d). This likely results from the limited number of assimilated observations available to ERA5-Land during that period compared to more recent decades .
Overall, these results indicate that ERA5-Land captures large-scale variability and long-term warming trends in the region, particularly from the 1960s onward. However, caution should be taken when interpreting trends in the early-period (1950s), where the reanalysis may be weakly constrained. These findings underscore the importance of continued data-rescue efforts to recover and digitize archived historical observational records, which are essential for evaluating and constraining reanalysis products in data-sparse regions, especially in the early periods.
Potential of Our Data Rescue Approach and Opportunities
Throughout our study, we demonstrate our climate data rescue workflow, starting with data inventory of archived weather sheets, followed by the digitization (imaging) of these sheets, transcription using the open-source machine-learning software MeteoSaver, automated QA/QC and data formatting, and finally the upload of the rescued datasets to open-access repositories.
By documenting each step, we aim to provide a replicable workflow that can be applied to similar data rescue projects across Africa and other data-sparse regions where historical weather observations are still stored in local or international archives. These techniques can also be used by the broader hydro-climate research community to recover additional historical records such as river discharge, lake level, and coastal water level measurements (e.g., ).
The rescued dataset, in addition to daily maximum, minimum, and average temperatures and daily precipitation, also includes dry- and wet-bulb temperatures recorded three times per day (06:00, 15:00, and 18:00 local time). These sub-daily temperature measurements enable the derivation of additional variables, including actual vapour pressure, relative humidity, and vapour pressure deficit (see Fig. and Sect. ). This dataset therefore enables additional analyses beyond those presented here, including but not limited to extreme event attribution analyses, climate model evaluation, and constraining of regional climate models for the Congo Basin.
Conclusions
In this study, we describe our climate data rescue mission in the Democratic Republic of Congo where we: (i) digitised (imaged) 9885 historical weather sheets presently stored in the INERA archives in Yangambi, (ii) automated transcription of the data using an open-source machine-learning based software MeteoSaver version 1.1, resulting in over one million daily and sub-daily temperature and precipitation records from 37 stations across the DRC, (iii) uploaded the data to open-access repositories, and finally (iv) analysed the temperature trends at 21 stations and changes in precipitation return periods at three stations, as well as comparing the temperature trends in the station observations with ERA5 reanalysis.
This newly transcribed dataset represents one of the first large-scale digital historical climate datasets for the DRC and is made available to the wider research community. Our results show clear and consistent warming trends in daily maximum, minimum, and average temperatures across the DRC, with median increases of approximately 0.7, 0.3, and 0.5 °C over a 30-year period, respectively. The warming shift in distributions of daily maximum, minimum and average temperatures with each successive decade during the period 1961–1990 further highlights the warming pattern in the region, with higher modal temperatures and more extreme values. This decadal shift in distributions is also evident in stations with data extending into the early 2000s and 2010s.
Consistent with these trends, we observe an increasing frequency of hot extremes across the region with each successive decade. During the period 1961–1990, the average number of hot days per year nearly doubled between the earliest and most recent decades. Simultaneously, fewer cold days were observed across all the aggregated station database over the same period. This pattern persists in more recent decades at stations with data extending beyond 1990. This further strengthens the evidence of an unequivocal warming trend in the region.
Comparison between station observations and ERA5-Land reanalysis shows that, while ERA5-Land captures large-scale variability and long-term warming trends, some discrepancies with the station observations remain, particularly during the early reanalysis period (1950s), when observational constraints were limited.
In conclusion, our results demonstrate consistent warming trends across the Congo Basin during the period 1961–1990 at all the considered stations, with continued warming during the 2000s at those with extended data, accompanied by increasing frequency of hot extremes. These findings provide new observational evidence of climate change in this data-sparse region and underline the critical need for more climate data rescue efforts to improve assessments of observed changes in climate, as well as to better constrain reanalysis products.
Data Inventory and Digitization Campaign at INERA – Section de Climatologie archives in Yangambi, DRCINERA Station Metadata
Station metadata document photographed at INERA – Section de Climatologie archives in Yangambi, DRC, during the data inventory and digitization campaign. The sheet contains the following information about all the INERA weather stations: (i) station name, (ii) classification order (ORDRE), (iii) station code (CODE), (iv) region (SECTEUR), (v) Latitude, (vi) Longitude (East) (LONGITUDE EST), (vii) altitude (ALTIT. ±), and (viii) other station specific information, classification and remarks (t, Aa, An and PARTICULARITÉS).
Example of a Hand-Drawn Post-1960 Observation Sheet
Example of a hand-drawn monthly weather observation sheet for Station Yaekema (0°46′ N, 24°14′ E) in the Democratic Republic of the Congo (DRC), as found in the archives of INERA, Yangambi.
Location of the INERA stations
Locations of weather stations in the DRC whose records are archived at the INERA – Section de Climatologie in Yangambi. Red circles show the full station network, while orange circles highlight the subset of stations for which manually transcribed reference data were available and used for validation in Sect. (Background map courtesy of ESRI National Geographic | Powered by Esri).
Pre-1960 weather data sheets at the State Archives of Belgium digitized by the COBECORE projectExample of a Pre-1960 Sheet Format: 1930s
Representative example of a hand-drawn monthly weather observation sheet from the 1930s, showing the typical format used during that period. This sheet corresponds to Station Eala (0°03′ N, 18°18′ E) in the Democratic Republic of the Congo (DRC), and is preserved in the State Archives of Belgium. Digitized by the COBECORE project .
Example of a Pre-1960 Sheet Format: 1940s
Representative example of a hand-drawn monthly weather observation sheet from the 1940s, showing the typical format used during that period. This sheet corresponds to Station Eala (0°03′ N, 18°18′ E) in the Democratic Republic of the Congo (DRC), and is preserved in the State Archives of Belgium. Digitized by the COBECORE project .
Example of a Pre-1960 Sheet Format: 1950s
Representative example of a hand-drawn monthly weather observation sheet from the 1950s, showing the typical format used during that period. This sheet corresponds to Station M'Vuazi (5°27′ S, 14°54′ E) in the Democratic Republic of the Congo (DRC), and is preserved in the State Archives of Belgium. Digitized by the COBECORE project .
Example of archived ancillary documents, digitized under the COBECORE project, describing types of instruments and observation procedures prescribed for use at the stations
Example of COBECORE archive documents describing thermometers used to measure daily maximum and minimum temperatures and observation procedures used at INEAC/INERA stations. These documents provide useful contextual metadata and demonstrate that standardised instrumentation and observation procedures were prescribed across the INEAC/INERA observational network. However, they do not constitute complete station-specific histories of instrument changes, calibration procedures, relocations, or observing practices over the full observational period (Source: Guide pratique à l'usage des observateurs – Réseau Météorologique du Congo Belge; available in the COBECORE records; ).
Example of COBECORE archive documents describing rain gauges used to measure daily precipitation and observation procedures used at INEAC/INERA stations. These documents provide useful contextual metadata and demonstrate that standardised instrumentation and observation procedures were prescribed across the INEAC/INERA observational network. However, they do not constitute complete station-specific histories of instrument changes, calibration procedures, relocations, or observing practices over the full observational period (Source: Guide pratique à l'usage des observateurs – Réseau Météorologique du Congo Belge; available in the COBECORE records; ).
Configuration User Settings in MeteoSaver v1.1
MeteoSaver configuration settings, descriptions and default values used in this project.
Setting/ValueDescriptionDefault ValueGeneral run_modeDefines the environment where the software will be run. Options: local for personal computers or hpc for High Performance Computing.hpcnum_processorsSpecifies the number of CPUs to use when running on an HPC infrastructure (in hpc mode).18Directories full_datadirDirectory containing folders of historical weather data sheet images, organized by station number.data/00_post1960...pre_QA_QC_transcribed_hydroclimate_data_dirDirectory for storing pre-QA/QC transcribed hydroclimate data.results/01_pre...post_QA_QC_transcribed_hydroclimate_data_dirDirectory for storing post-QA/QC transcribed hydroclimate data.results/02_post...validation_dirDirectory for validation results.results/03_valid...final_refined_daily_hydroclimate_data_dirDirectory for final refined hydroclimate data after all quality checks.results/04_final...transient_transcription_output_dirDirectory to store transient transcription outputs during processing.results/05_trans...manually_transcribed_data_dirDirectory for manually transcribed data, used for validation purposes.results/06_manual...result_maps_dirDirectory for results overlain on maps and combined plots.results/07_maps...metadata_file_pathDirectory for all the stations metadata.data/01_meta...formatted_already_digitized_data_dirDirectory for previously transcribed data to add to our new transcribed database.data/02_form...region_shapefile_pathShapefile of study regiondata/03_shap...era5_dirDirectory for ERA5-Land data.data/04_era5...Table and Cell Detection clip_up, clip_down, clip_left, clip_rightClipping values (in pixels) to remove headers and row labels from detected tables.430, 280, 200, 120max_table_width, max_table_heightMaximum expected table width and height (in pixels) to ensure correct table detection.3900, 3600min_cell_width_threshold, max_cell_width_threshold, min_cell_height_threshold, max_cell_height_thresholdMinimum and maximum width and height (in pixels) allowed for detected table cells.60, 140, 30, 50no_of_rows, no_of_columns, no_of_rows_including_headersExpected number of rows and columns in detected tables.43, 24, 46space_height_threshold, space_width_threshold, max_cell_height_per_boxMinimum height and width space (in pixels) between bounding boxes to detect missing cells, with height for newly added (missing) bounding boxes/cells.50, 120, 40Transcription ocr_modelDefines which OCR/HTR model to use (e.g., Tesseract-OCR, EasyOCR, PaddleOCR).Tesseract-OCRtesseract_pathPath to the Tesseract-OCR executable file, needed for Tesseract-OCR model..../tesseract.exesystem_tessdata_dirDirectory path to trained OCR/HTR language models for transcription (e.g., Tesseract tessdata)..../tessdataQA/QC daily_temperature_columns; daily_temperature_columns_and_diurnal_temperature_ rangeColumns to focus on for quality checks, here columns with Daily Maximum Temperature, Minimum Temperature,Average Temperature and the Diurnal Temperature Range.D, E, F; D, E, F, Gdaily_precipitation_columnColumn with daily precipitation.Kdry_and_wet_bulb_temperature_columnsColumns with dry bulb and wet bulb temperature readings recorded at specific times of the day.L, M, Q, R, V, Wmax_temperature_threshold, min_temperature_thresholdTemperature thresholds (in °C) to flag invalid transcribed temperature readings.40, 5decimal_placesNumber of decimal places in the data sheets.1uncertainty_marginDefines the uncertainty margin to be used in temperature calculations and quality checks.0.2header_rowsNumber of header rows in the sheet.3multi_day_totals, multi_day_averagesFlags to specify whether multi-day totals or averages (e.g., 5 d totals) are present in the sheets.True, Truemax_days_for_multi_day_totalMaximum number of days contained in multi-day totals (e.g., 5 or 6 d).6multi_day_totals_rows; final_totals_rowsRows where multi-day totals or final totals are located (if applicable).9, 16, 23, 30, 37, 45; 45excluded_rows; excluded_columnsRows and columns to exclude during QA/QC checks involving multi-day totals (e.g., headers, date).1, 2, 3, 9, 16, 23, 30, 37, 45; 1, 2, 3, 15, 20, 25, 26, 27additional_excluded_rowsExtra rows to exclude during QA/QC checks involving multi-day averages (if applicable).10, 17, 24, 31, 38, 46
Note: Paths with “…” indicate truncated directory names that are relative to the repository root; see the Zenodo repository for full names.
Daily Temperature Trends at Individual Stations
Daily temperature trends at stations: (a) Yaekama (0°46′ N, 24°14′ E), (b) Gimbi-Plateau (5°31′ S, 13°22′ E), (c) M'Vuazi-Poste (Plateau) (5°27′ S, 14°54′ E), and (d) Binga (2°18′ N, 20°30′ E) in the DRC. Colored dots represent daily values of maximum (Tx, red), average (Tavg, orange), and minimum (Tn, blue) temperatures. Solid lines indicate Theil–Sen trend estimates for each variable, with the corresponding slope (°C per decade) shown at the end of each line. This data includes both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. Note that these stations have different start and end years.
Daily temperature trends at stations: (a) Bambesa (3°27′ N, 25°43′ E), (b) Boketa (3°11′ N, 19°46′ E), (c) Kutubongo (4°01′ N, 19°09′ E), and (d) Kiyaka-Plateau (5°16′ S, 18°57′ E) in the DRC. Colored dots represent daily values of maximum (Tx, red), average (Tavg, orange), and minimum (Tn, blue) temperatures. Solid lines indicate Theil–Sen trend estimates for each variable, with the corresponding slope (°C per decade) shown at the end of each line. This data includes both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. Note that these stations have different start and end years.
Daily temperature trends at stations: (a) Gandajika (6°45′ S, 23°57′ E), (b) Kisanga-Plateau (11°44′ S, 27°25′ E), (c) Kaniama (7°25′ S, 24°09′ E), and (d) Kipopo (11°34′ S, 27°24′ E) in the DRC. Colored dots represent daily values of maximum (Tx, red), average (Tavg, orange), and minimum (Tn, blue) temperatures. Solid lines indicate Theil–Sen trend estimates for each variable, with the corresponding slope (°C per decade) shown at the end of each line. This data includes both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. Note that these stations have different start and end years.
Daily temperature trends at stations: (a) Mulungu-Nyamunyunye (2°18′ S, 28°48′ E), (b) Mulungu-Tshibinda (2°19′ S, 28°45′ E), (c) Mulungu-Bukulumisa (2°20′ S, 28°43′ E), and (d) Nioka-Drusi (2°09′ N, 30°39′ E) in the DRC. Colored dots represent daily values of maximum (Tx, red), average (Tavg, orange), and minimum (Tn, blue) temperatures. Solid lines indicate Theil–Sen trend estimates for each variable, with the corresponding slope (°C per decade) shown at the end of each line. This data includes both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. Note that these stations have different start and end years.
Daily temperature trends at station Rumangabo (1°21′ S, 29°22′ E) in the DRC. Colored dots represent daily values of maximum (Tx, red), average (Tavg, orange), and minimum (Tn, blue) temperatures. Solid lines indicate Theil–Sen trend estimates for each variable, with the corresponding slope (°C per decade) shown at the end of each line. This data includes both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. Note that these stations have different start and end years.
Decadal shifts in temperature distribution at Individual Stations
Decadal shifts in temperature distributions at stations: (a) Yaekama (0°46′ N, 24°14′ E), (b) Gimbi-Plateau (5°31′ S, 13°22′ E), (c) M'Vuazi-Poste (Plateau) (5°27′ S, 14°54′ E), and (d) Binga (2°18′ N, 20°30′ E) in the DRC. Kernel density estimates of daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures are shown for each decade: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001-2010 (light purple) and 2011–2020 (dark violet). Vertical dashed lines indicate the 95th percentile values for each decade, with panels (a), (b) and (c) illustrating the progressive rightward shift towards higher temperatures and more extreme values over time.
Decadal shifts in temperature distributions at stations: (a) Bambesa (3°27′ N, 25°43′ E), (b) Boketa (3°11′ N, 19°46′ E), (c) Kutubongo (4°01′ N, 19°09′ E), and (d) Kiyaka-Plateau (5°16′ S, 18°57′ E) in the DRC. Kernel density estimates of daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures are shown for each decade: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). Vertical dashed lines indicate the 95th percentile values for each decade, with panels (a), (b) and (c) illustrating the progressive rightward shift towards higher temperatures and more extreme values over time.
Decadal shifts in temperature distributions at stations: (a) Gandajika (6°45′ S, 23°57′ E), (b) Kisanga-Plateau (11°44′ S, 27°25′ E), (c) Kaniama (7°25′ S, 24°09′ E), and (d) Kipopo (11°34′ S, 27°24′ E) in the DRC. Kernel density estimates of daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures are shown for each decade: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011-2020 (dark violet). Vertical dashed lines indicate the 95th percentile values for each decade, with panels (a), (b) and (c) illustrating the progressive rightward shift towards higher temperatures and more extreme values over time.
Decadal shifts in temperature distributions at stations: (a) Mulungu-Nyamunyunye (2°18′ S, 28°48′ E), (b) Mulungu-Tshibinda (2°19′ S, 28°45′ E), (c) Mulungu-Bukulumisa (2°20′ S, 28°43′ E), and (d) Nioka-Drusi (2°09′ N, 30°39′ E) in the DRC. Kernel density estimates of daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures are shown for each decade: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). Vertical dashed lines indicate the 95th percentile values for each decade, with panels (a), (b) and (c) illustrating the progressive rightward shift towards higher temperatures and more extreme values over time.
Decadal shifts in temperature distributions at station Rumangabo (1°21′ S, 29°22′ E) in the DRC. Kernel density estimates of daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures are shown for each decade: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). Vertical dashed lines indicate the 95th percentile values for each decade, with panels (a), (b) and (c) illustrating the progressive rightward shift towards higher temperatures and more extreme values over time.
Frequency of hot and cold temperature extremes at individual stations
Frequency of hot and cold temperature extremes at stations: (a) Yaekama (0°46′ N, 24°14′ E), (b) Gimbi-Plateau (5°31′ S, 13°22′ E), (c) M'Vuazi-Poste (Plateau) (5°27′ S, 14°54′ E), and (d) Binga (2°18′ N, 20°30′ E) in the DRC. The plots show the average number of days per decade falling below the 5th percentile (cold events) and above the 95th percentile (hot events) for daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures, similar to . Values are based on kernel density estimates calculated across the entire period. The number of days per year within each decade are shown: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). The middle 90 % of the distribution (5th–95th percentile range) is omitted for clarity. The 95th percentile temperature for each variable is noted in the top-right corner of each plot.
Frequency of hot and cold temperature extremes at stations: (a) Bambesa (3°27′ N, 25°43′ E), (b) Boketa (3°11′ N, 19°46′ E), (c) Kutubongo (4°01′ N, 19°09′ E), and (d) Kiyaka-Plateau (5°16′ S, 18°57′ E) in the DRC. The plots show the average number of days per decade falling below the 5th percentile (cold events) and above the 95th percentile (hot events) for daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures, similar to . Values are based on kernel density estimates calculated across the entire period. The number of days per year within each decade are shown: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). The middle 90 % of the distribution (5th–95th percentile range) is omitted for clarity. The 95th percentile temperature for each variable is noted in the top-right corner of each plot.
Frequency of hot and cold temperature extremes at stations: (a) Gandajika (6°45′ S, 23°57′ E), (b) Kisanga-Plateau (11°44′ S, 27°25′ E), (c) Kaniama (7°25′ S, 24°09′ E), and (d) Kipopo (11°34′ S, 27°24′ E) in the DRC. The plots show the average number of days per decade falling below the 5th percentile (cold events) and above the 95th percentile (hot events) for daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures, similar to . Values are based on kernel density estimates calculated across the entire period. The number of days per year within each decade are shown: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). The middle 90 % of the distribution (5th–95th percentile range) is omitted for clarity. The 95th percentile temperature for each variable is noted in the top-right corner of each plot.
Frequency of hot and cold temperature extremes at stations: (a) Mulungu-Nyamunyunye (2°18′ S, 28°48′ E), (b) Mulungu-Tshibinda (2°19′ S, 28°45′ E), (c) Mulungu-Bukulumisa (2°20′ S, 28°43′ E), and (d) Nioka-Drusi (2°09′ N, 30°39′ E) in the DRC. The plots show the average number of days per decade falling below the 5th percentile (cold events) and above the 95th percentile (hot events) for daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures, similar to . Values are based on kernel density estimates calculated across the entire period. The number of days per year within each decade are shown: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). The middle 90 % of the distribution (5th–95th percentile range) is omitted for clarity. The 95th percentile temperature for each variable is noted in the top-right corner of each plot.
Frequency of hot and cold temperature extremes at station Rumangabo (1°21′ S, 29°22′ E) in the DRC. The plots show the average number of days per decade falling below the 5th percentile (cold events) and above the 95th percentile (hot events) for daily maximum (Tx), minimum (Tn), and average (Tavg) temperatures, similar to . Values are based on kernel density estimates calculated across the entire period. The number of days per year within each decade are shown: 1911–1920 (blue), 1921–1930 (light green), 1931–1940 (dark olive green), 1941–1950 (olive green), 1951–1960 (gold), 1961–1970 (yellow), 1971–1980 (red), 1981–1990 (maroon), 1991–2000 (indigo), 2001–2010 (light purple) and 2011–2020 (dark violet). The middle 90 % of the distribution (5th–95th percentile range) is omitted for clarity. The 95th percentile temperature for each variable is noted in the top-right corner of each plot.
Comparison of trends in daily maximum, minimum and average temperature at Individual Stations with ERA5-Land
Comparison of trends in daily maximum, minimum and average temperatures at stations: (a) Yaekama (0°46′ N, 24°14′ E), (b) Gimbi-Plateau (5°31′ S, 13°22′ E), (c) M'Vuazi-Poste (Plateau) (5°27′ S, 14°54′ E), and (d) Binga (2°18′ N, 20°30′ E) in the DRC. Blue dots represent daily values of observed maximum, minimum and average temperatures, while solid yellow lines show the ERA5-Land timeseries. The dotted dark blue and orange lines indicate Theil–Sen trend estimates for each variable for observations (Obs) and ERA5-Land, respectively, with the corresponding slope (°C per decade) shown in the top left legend. The observations include both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. The trends are estimated using values where both the ERA5-Land data and transcribed observations are available. Note that these stations have different start and end years.
Comparison of trends in daily maximum, minimum and average temperatures at stations: (a) Bambesa (3°27′ N, 25°43′ E), (b) Boketa (3°11′ N, 19°46′ E), (c) Kutubongo (4°01′ N, 19°09′ E), and (d) Kiyaka-Plateau (5°16′ S, 18°57′ E) in the DRC. Blue dots represent daily values of observed maximum, minimum and average temperatures, while solid yellow lines show the ERA5-Land timeseries. The dotted dark blue and orange lines indicate Theil–Sen trend estimates for each variable for observations (Obs) and ERA5-Land, respectively, with the corresponding slope (°C per decade) shown in the top left legend. The observations include both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. The trends are estimated using values where both the ERA5-Land data and transcribed observations are available. Note that these stations have different start and end years.
Comparison of trends in daily maximum, minimum and average temperatures at stations: (a) Gandajika (6°45′ S, 23°57′ E), (b) Kisanga-Plateau (11°44′ S, 27°25′ E), (c) Kaniama (7°25′ S, 24°09′ E), and (d) Kipopo (11°34′ S, 27°24′ E) in the DRC. Blue dots represent daily values of observed maximum, minimum and average temperatures, while solid yellow lines show the ERA5-Land timeseries. The dotted dark blue and orange lines indicate Theil–Sen trend estimates for each variable for observations (Obs) and ERA5-Land, respectively, with the corresponding slope (°C per decade) shown in the top left legend. The observations include both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. The trends are estimated using values where both the ERA5-Land data and transcribed observations are available. Note that these stations have different start and end years.
Comparison of trends in daily maximum, minimum and average temperatures at stations: (a) Mulungu-Nyamunyunye (2°18′ S, 28°48′ E), (b) Mulungu-Tshibinda (2°19′ S, 28°45′ E), (c) Mulungu-Bukulumisa (2°20′ S, 28°43′ E), and (d) Nioka-Drusi (2°09′ N, 30°39′ E) in the DRC. Blue dots represent daily values of observed maximum, minimum and average temperatures, while solid yellow lines show the ERA5-Land timeseries. The dotted dark blue and orange lines indicate Theil–Sen trend estimates for each variable for observations (Obs) and ERA5-Land, respectively, with the corresponding slope (°C per decade) shown in the top left legend. The observations include both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. The trends are estimated using values where both the ERA5-Land data and transcribed observations are available. Note that these stations have different start and end years.
Comparison of trends in daily maximum, minimum and average temperatures at station Rumangabo (1°21′ S, 29°22′ E) in the DRC. Blue dots represent daily values of observed maximum, minimum and average temperatures, while solid yellow lines show the ERA5-Land timeseries. The dotted dark blue and orange lines indicate Theil–Sen trend estimates for each variable for observations (Obs) and ERA5-Land, respectively, with the corresponding slope (°C per decade) shown in the top left legend. The observations include both automatically transcribed records using MeteoSaver and manually transcribed records previously available at INERA. The trends are estimated using values where both the ERA5-Land data and transcribed observations are available. Note that these stations have different start and end years.
ERA5-Land timeseries for daily maximum, minimum and average temperatures at stations: (a) M'Vuazi-Poste (Plateau) (5°27<sup>′</sup> S, 14°54<sup>′</sup> E), and (d) Binga (2°18<sup>′</sup> N, 20°30<sup>′</sup> E) in the DRC
ERA5-Land timeseries for daily maximum, minimum and average temperatures at stations: (a) Yangambi-Km.5 (0°46′ N, 24°29′ E), and (b) Eala (0°03′ N, 18°18′ E) in the DRC. The solid yellow lines show the ERA5-Land timeseries highlighting a pronounced jump in the 1950s for both stations that is not captured by stations with observations during that period (see Fig. c–d). Note that these stations have different end years dependent on the transcribed in situ observations in the INERA dataset (as in Fig. c–d).
Code and data availability
MeteoSaver v1.1 is available on the GitHub repository of the Department of Water and Climate at the Vrije Universiteit Brussel (VUB) (https://github.com/VUB-HYDR/MeteoSaver, last access: 6 August 2026) and on Zenodo (10.5281/zenodo.21818879; ). Additionally, all scripts used for the analysis of the transcribed data are available on the same GitHub repository (https://github.com/VUB-HYDR/2026_Muheki_et_al, last access: 25 September 2026) and on Zenodo (10.5281/zenodo.21808003, ). The transcribed dataset is available on Zenodo (10.5281/zenodo.18770063; ). Digital copies of the historical weather records taken at INERA stations across the DRC are available on Zenodo (10.5281/zenodo.21274281; ).
Author contributions
DM and WT designed the study. DM undertook the digitization of the post-1960 weather data sheets at the INERA archives. DKN, OKM, JM and TBL facilitated access to the INERA dataset (both hardcopy and manually-digitized data), while KHu, KJ, and HV provided access to the COBECORE dataset. DM, WT, KHu, and SL developed the scripts and conducted the development of MeteoSaver with assistance from BV, CV, EH, HV, JMB, KKTC, PB, and PT. EH, PT, KHe, SN, KHu, and KJ contributed expertise in data rescue throughout the project. All authors provided feedback during the data analysis and the writing of this paper.
Competing interests
The contact author has declared that none of the authors has any competing interests.
Disclaimer
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. The authors bear the ultimate responsibility for providing appropriate place names. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.
Acknowledgements
We would like to thank the Institut National pour l'Etude et la Recherche Agronomiques (INERA) situated in the Democratic Republic of the Congo (DRC) for granting us access to the extensive historical weather database available in the archives in Yangambi, DRC. We also thank the COBECORE project for providing access to pre-1960 DRC data, including manually transcribed records for the Eala, Bongabo, and Yangambi-Km.5 stations by Innocent Banzi Ngulu-kulu. DM is a research fellow at the Research Foundation – Flanders (11M8823N). WT received funding from the European Research Council (ERC) under the European Union's Horizon Framework research and innovation programme (grant agreement No. 101076909; “LACRIMA” project), the Interuniversity BOF projects (iBOF) (grant reference: iBOF/25/081; Cong'eau project), and the European Union's Horizon 2020 (grant agreement No. 101081369; 'SPARCCLE'). Compute and storage resources and services used in this work were provided by the VSC (Flemish Supercomputer Center), funded by the Research Foundation – Flanders (FWO) and the Flemish Government.
Financial support
This research has been supported by the Fonds Wetenschappelijk Onderzoek (grant nos. 11M8825N and 11M8823N), the HORIZON EUROPE European Research Council (grant no. 101076909), the Interuniversity BOF projects (grant reference: iBOF/25/081), and the European Union's Horizon 2020 (grant no. 101081369).
Review statement
This paper was edited by Kai Kornhuber and reviewed by Stephan Pietsch and two anonymous referees.
ReferencesACP: La toiture de l’atelier mécanique de l’INERA/Yangambi détruite par le feu, Agence Congolaise De Presse, https://acp.cd/communication/environnement/la-toiture-de-latelier-mecanique-de-linera-yangambi-detruite-par-le-feu/ (last access: 17 July 2024), 2023.Aguilar, E., Aziz Barry, A., Brunet, M., Ekang, L., Fernandes, A., Massoukina, M., Mbah, J., Mhanda, A., do Nascimento, D. J., Peterson, T. C., Thamba Umba, O., Tomou, M., and Zhang, X.: Changes in temperature and precipitation extremes in western central Africa, Guinea Conakry, and Zimbabwe, 1955–2006, J. Geophys. Res.-Atmos., 114, 10.1029/2008JD011010, 2009.Alsdorf, D., Beighley, E., Laraque, A., Lee, H., Tshimanga, R., O'Loughlin, F., Mahé, G., Dinga, B., Moukandi, G., and Spencer, R. G. M.: Opportunities for hydrologic research in the Congo Basin, Rev. Geophys., 54, 378–409, 10.1002/2016RG000517, 2016.Bell, B., Hersbach, H., Simmons, A., Berrisford, P., Dahlgren, P., Horányi, A., Muñoz-Sabater, J., Nicolas, J., Radu, R., Schepers, D., Soci, C., Villaume, S., Bidlot, J.-R., Haimberger, L., Woollen, J., Buontempo, C., and Thépaut, J.-N.: The ERA5 global reanalysis: Preliminary extension to 1950, Q. J. Roy. Meteor. Soc., 147, 4186–4227, 10.1002/qj.4174, 2021.Bindoff, N., Stott, P., AchutaRao, K., Allen, M., Gillett, N., Gutzler, D., Hansingo, K., Hegerl, G., Hu, Y., Jain, S., Mokhov, I., Overland, J., Perlwitz, J., Sebbari, R., and Zhang, X.: Detection and Attribution of Climate Change: from Global to Regional, in: Climate Change 2013: The Physical Science Basis. Contribution of Working Group I to the Fifth Assessment Report of the Intergovernmental Panel on Climate Change, edited by: Stocker, T., Qin, D., Plattner, G.-K., Tignor, M., Allen, S., Boschung, J., Nauels, A., Xia, Y., Bex, V., and Midgley, P., book section 10, 867–952, Cambridge University Press, Cambridge, United Kingdom and New York, NY, USA, ISBN 978-1-107-66182-0, 10.1017/CBO9781107415324.022, 2013.Bolton, D.: The Computation of Equivalent Potential Temperature, Mon. Weather Rev., 108, 1046–1053, 10.1175/1520-0493(1980)108<1046:TCOEPT>2.0.CO;2, 1980.Bradski, G. and Kaehler, A.: Learning OpenCV: Computer Vision with the OpenCV Library, Computer Programming/Robotics, O'Reilly Media, Incorporated, ISBN 9780596516130, https://books.google.be/books?id=J0GbAgAAQBAJ (last access: 9 February 2026), 2008.Brohan, P., Kennedy, J. J., Harris, I., Tett, S. F. B., and Jones, P. D.: Uncertainty estimates in regional and global observed temperature changes: A new data set from 1850, J. Geophys. Res.-Atmos., 111, 10.1029/2005JD006548, 2006.COBECORE: Congo basin eco-climatological data recovery and valorisation, https://cobecore.org/index.html (last access: 25 June 2025), 2018.Coppieters, G.: Inventaris van het archief van de Rijksplantages en de Regie der Plantages van de Kolonie (REPCO), het Nationaal Instituut voor de Landbouwkunde in Belgisch-Congo (NILCO/INEAC) en de Documentatiediens, State Archives of Belgium, https://agatha.arch.be/en/data/ead/BE-A0510_003812_005470#I381254706050 (last access: 5 August2026), 2013.CSC: Climate Service Centre Report No. 11, in: Climate Change Scenarios for the Congo Basin, edited by: Haensler, A., Jacob, D., Kabat, P., and Ludwig, F., Hamburg, Germany, https://www.climate-service-center.de/imperia/md/content/csc/csc-report11_optimized.pdf (last access: 7 April 2026), 2013.Depicker, A., Jacobs, L., Mboga, N., Smets, B., Van Rompaey, A., Lennert, M., Wolff, E., Kervyn, F., Michellier, C., Dewitte, O., and Govers, G.: Historical dynamics of landslide risk from population and forest-cover changes in the Kivu Rift, Nature Sustainability, 4, 10.1038/s41893-021-00757-9, 2021.de Smeth, K., Comer, J., and Murphy, C.: Hydrometric data rescue and extension of river flow records: Method development and application to catchments modified by arterial drainage, Geosci. Data J., 11, 176–196, 10.1002/gdj3.206, 2024.Dille, A., Dewitte, O., Handwerger, A. L., d’Oreye, N., Derauw, D., Ganza Bamulezi, G., Ilombe Mawe, G., Michellier, C., Moeyersons, J., Monsieurs, E., Mugaruka Bibentyo, T., Samsonov, S., Smets, B., Kervyn, M., and Kervyn, F.: Acceleration of a large deep-seated tropical landslide due to urbanization feedbacks, Nat. Geosci., 15, 10.1038/s41561-022-01073-3, 2022.Doherty, A., Pearce, M., Calow, R., Daoust, G., Higazi, A., Burgin, L., and Osborne, R.: Climate risk report for the Central Africa region, Met Office, https://www.metoffice.gov.uk/binaries/content/assets/metofficegovuk/pdf/services/government/central-africa-climate-risk-report-final.pdf (last access: 18 February 2026), 2022.Eyring, V., Gillett, N., Achuta Rao, K., Barimalala, R., Barreiro Parrillo, M., Bellouin, N., Cassou, C., Durack, P., Kosaka, Y., McGregor, S., Min, S., Morgenstern, O., and Sun, Y.: Human Influence on the Climate System, in: Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change, edited by: Masson-Delmotte, V., Zhai, P., Pirani, A., Connors, S. L., Péan, C., Berger, S., Caud, N., Chen, Y., Goldfarb, L., Gomis, M. I., Huang, M., Leitzell, K., Lonnoy, E., Matthews, J. B. R., Maycock, T. K., Waterfield, T., Yelekçi, O., Yu, R., and Zhou, B., book section 3, 423–551, Cambridge University Press, Cambridge, UK and New York, NY, USA, 10.1017/9781009157896.005, 2021.Folland, C. K., Rayner, N. A., Brown, S. J., Smith, T. M., Shen, S. S. P., Parker, D. E., Macadam, I., Jones, P. D., Jones, R. N., Nicholls, N., and Sexton, D. M. H.: Global temperature change and its uncertainties since 1861, Geophys. Res. Lett., 28, 2621–2624, 10.1029/2001GL012877, 2001.Global Nature Watch: Interactive charts and maps summarising statistics about forests in Yangambi, Tshopo, Democratic Republic of the Congo, https://gfw.global/SAko14 (last access: 5 August 2026), 2026.Grant, L., Vanderkelen, I., Gudmundsson, L., Tan, Z., Perroud, M., Stepanenko, V. M., Debolskiy, A. V., Droppers, B., Janssen, A. B. G., Woolway, R. I., Choulga, M., Balsamo, G., Kirillin, G., Schewe, J., Zhao, F., del Valle, I. V., Golub, M., Pierson, D., Marcé, R., Seneviratne, S. I., and Thiery, W.: Attribution of global lake systems change to anthropogenic forcing, Nat. Geosci., 14, 849–854, 10.1038/s41561-021-00833-x, 2021.Gudmundsson, L., Boulange, J., Do, H. X., Gosling, S. N., Grillakis, M. G., Koutroulis, A. G., Leonard, M., Liu, J., Schmied, H. M., Papadimitriou, L., Pokhrel, Y., Seneviratne, S. I., Satoh, Y., Thiery, W., Westra, S., Zhang, X., and Zhao, F.: Globally observed trends in mean and extreme river flow attributed to climate change, Science, 371, 1159–1162, 10.1126/science.aba3996, 2021.Hansen, M. C., Potapov, P. V., Moore, R., Hancher, M., Turubanova, S. A., Tyukavina, A., Thau, D., Stehman, S. V., Goetz, S. J., Loveland, T. R., Kommareddy, A., Egorov, A., Chini, L., Justice, C. O., and Townshend, J. R. G.: High-Resolution Global Maps of 21st-Century Forest Cover Change, Science, 342, 850–853, 10.1126/science.1244693, 2013.Hawkins, E., Burt, S., McCarthy, M., Murphy, C., Ross, C., Baldock, M., Brazier, J., Hersee, G., Huntley, J., Meats, R., O’Grady, J., Scrimgeour, I., and Silk, T.: Millions of historical monthly rainfall observations taken in the UK and Ireland rescued by citizen scientists, Geosci. Data J., 10, 246–261, 10.1002/gdj3.157, 2022.Hawkins, E., Brohan, P., Burgess, S. N., Burt, S., Compo, G. P., Gray, S. L., Haigh, I. D., Hersbach, H., Kuijjer, K., Martínez-Alvarado, O., McColl, C., Schurer, A. P., Slivinski, L., and Williams, J.: Rescuing historical weather observations improves quantification of severe windstorm risks, Nat. Hazards Earth Syst. Sci., 23, 1465–1482, 10.5194/nhess-23-1465-2023, 2023.Hock, R., Rasul, G., Adler, C., Cáceres, B., Gruber, S., Hirabayashi, Y., Jackson, M., Kääb, A., Kang, S., Kutuzov, S., Milner, A., Molau, U., Morin, S., Orlove, B., and Steltzer, H.: High Mountain Areas. In: IPCC Special Report on the Ocean and Cryosphere in a Changing Climate, Cambridge University Press, 10.1017/9781009157964.004, 2019.Hua, W., Zhou, L., Chen, H., Nicholson, S. E., Raghavendra, A., and Jiang, Y.: Possible causes of the Central Equatorial African long-term drought, Environ. Res. Lett., 11, 1–13, 10.1088/1748-9326/11/12/124002, 2016.IPCC: Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change, Cambridge University Press, Cambridge, UK and New York, NY, USA, 10.1017/9781009157896 (last access: 17 September 2024), 2021a.IPCC: Summary for Policymakers, in: Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change, edited by: Masson-Delmotte, V., Zhai, P., Pirani, A., Connors, S. L., Péan, C., Berger, S., Caud, N., Chen, Y., Goldfarb, L., Gomis, M. I., Huang, M., Leitzell, K., Lonnoy, E., Matthews, J. B. R., Maycock, T. K., Waterfield, T., Yelekçi, O., Yu, R., and Zhou, B., 1–31, Cambridge University Press, Cambridge, UK and New York, NY, USA, 10.1017/9781009157896.001, 2021b.Jacobsen, K., Van Hirtum, L., Amara, M., Beeckman, H., Meeus, S., Vandelook, F., Van den Bulcke, J., Stoffelen, P., Verbeeck, H., and Hufkens, K.: Climate data rescue from the Belgian colonial archives: helping to close the data-gap over Central Africa, in: Early Instrumental Meteorological Series, Conference abstracts, https://biblio.ugent.be/publication/8587593 (last access: 25 June 2025), 2018.Jaczewski, A., Marosz, M., and Miętus, M.: PL1GD-T: a high-resolution gridded daily air temperature dataset for Poland, Earth Syst. Sci. Data, 17, 3857–3871, 10.5194/essd-17-3857-2025, 2025.Kasongo Yakusu, E., Van Acker, J., Van de Vyver, H., Bourland, N., Mbifo Ndiapo, J., Besango Likwela, T., Lokonda Wa Kipifo, M., Mbuya Kankolongo, A., Van den Bulcke, J., Beeckman, H., Bauters, M., Boeckx, P., Verbeeck, H., Jacobsen, K., Demarée, G., Gellens-Meulenberghs, F., and Hubau, W.: Ground-based climate data show evidence of warming and intensification of the seasonal rainfall cycle during the 1960–2020 period in Yangambi, central Congo Basin, Clim. Change, 176, 28, 10.1007/s10584-023-03606-0, 2023.Latapy, A., Ferret, Y., Testut, L., Talke, S., Aarup, T., Pons, F., Jan, G., Bradshaw, E., and Pouvreau, N.: Data rescue process in the context of sea level reconstructions: An overview of the methodology, lessons learned, up-to-date best practices and recommendations, Geosci. Data J., 10, 396–425, 10.1002/gdj3.179, 2022.List, R.: Smithsonian meteorological tables, Smithsonian miscellaneous collections, Smithsonian Inst., https://books.google.be/books?id=wws50AEACAAJ (last access: 2 February 2026), 1966.Lorenz, R., Stalhandske, Z., and Fischer, E. M.: Detection of a Climate Change Signal in Extreme Heat, Heat Stress, and Cold in Europe From Observations, Geophys. Res. Lett., 46, 8363–8374, 10.1029/2019GL082062, 2019.Mamba, F.: Tshopo: une partie de la toiture de la station climatologique de l’Inera/Yangambi réduite en cendres, mediacongo.net, https://www.mediacongo.net/article-actualite-117493_tshopo_une_partie_de_la_toiture_de_la_station_climatologique_de_l_inera_yangambi_reduite_en_cendres.html (last access: 17 July 2024), 2023.Mangaza, L., Batsi, G., Peroches, A., Masson, C., Sonwa, D. J., Lhoest, S., Makana, J.-R., Hubau, W., Lejeune, P., and Fayolle, A.: Dynamics and determinants of forest cover changes in the inner Congo basin, Trees, Forests and People, 23, 101126, 10.1016/j.tfp.2025.101126, 2026.Marvel, K. and Bonfils, C.: Identifying external influences on global precipitation, P. Natl. Acad. Sci. USA, 110, 19 301–19 306, 10.1073/pnas.1314382110, 2013.Marvel, K., Cook, B. I., Bonfils, C. J. W., Durack, P. J., Smerdon, J. E., and Williams, A. P.: Twentieth-century hydroclimate changes consistent with human influence, Nature, 569, 59–65, 10.1038/s41586-019-1149-8, 2019.Mawe, G. I., Landu, E. L., Dujardin, E., Imwangana, F. M., Bielders, C., Hubert, A., Michellier, C., Nzolang, C., Poesen, J., Dewitte, O., and Vanmaercke, M.: Mapping urban gullies in the Democratic Republic of the Congo, Nature, 644, 10.1038/s41586-025-09371-7, 2025.Megevand, C.: Deforestation Trends in the Congo Basin: Reconciling Economic Growth and Forest Protection, World Bank, 10.1596/978-0-8213-9742-8, 2013.Morice, C. P., Kennedy, J. J., Rayner, N. A., and Jones, P. D.: Quantifying uncertainties in global and regional temperature change using an ensemble of observational estimates: The HadCRUT4 data set, J. Geophys. Res.-Atmos., 117, 10.1029/2011JD017187, 2012.Muheki, D., Hufkens, K., Jacobsen, K., Verbeeck, H., Boeckx, P., Kankonde Ntumba, D., Kapalay Moulasa, O., Vercruysse, B., Birkholz, J. M., Verbruggen, C., Hawkins, E., Lampe, S., Kasongo Yakusu, E., Imwangana, F. M., Mbifo, J., Besango Likwela, T., Meunier, F., Dewitte, O., Thorne, P., and Thiery, W.: VUB-HYDR/MeteoSaver: MeteoSaver v1.1 (Version v1.1), Zenodo [software], 10.5281/zenodo.21818879, 2026a.Muheki, D., Hufkens, K., Jacobsen, K., Verbeeck, H., Boeckx, P., Kankonde Ntumba, D., Kapalay Moulasa, O., Vercruysse, B., Birkholz, J. M., Verbruggen, C., Hawkins, E., Lampe, S., Kasongo Yakusu, E., Imwangana, F. M., Mbifo, J., Besango Likwela, T., Meunier, F., Dewitte, O., Thorne, P., and Thiery, W.: VUB-HYDR/2026_Muheki_et_al: Scripts_Revealing_Congo_Basin_Warming (Version v1), Zenodo [software], 10.5281/zenodo.21808003, 2026b.Muheki, D., Hufkens, K., Jacobsen, K., Verbeeck, H., Boeckx, P., Kankonde Ntumba, D., Kapalay Moulasa, O., Vercruysse, B., Birkholz, J. M., Verbruggen, C., Hawkins, E., Lampe, S., Kasongo Yakusu, E., Imwangana, F. M., Mbifo, J., Besango Likwela, T., Meunier, F., Dewitte, O., Thorne, P., and Thiery, W.: Daily and Sub-daily Temperature and Daily Precipitation Records from 37 INERA stations across the Democratic Republic of the Congo (Version 1.0), Zenodo [data set], 10.5281/zenodo.18770063, 2026c.Muheki, D., Hufkens, K., Jacobsen, K., Verbeeck, H., Boeckx, P., Kankonde Ntumba, D., Kapalay Moulasa, O., Vercruysse, B., Birkholz, J. M., Verbruggen, C., Hawkins, E., Lampe, S., Kasongo Yakusu, E., Imwangana, F. M., Mbifo, J., Besango Likwela, T., Meunier, F., Dewitte, O., Thorne, P., and Thiery, W.: Digital Copies of Historical Daily Weather Observations taken at INERA stations across the Democratic Republic of the Congo, Zenodo [data set], 10.5281/zenodo.21274281, 2026d.Muheki, D., Vercruysse, B., Chandrasekar, K. K. T., Verbruggen, C., Birkholz, J. M., Hufkens, K., Verbeeck, H., Boeckx, P., Lampe, S., Hawkins, E., Thorne, P., Ntumba, D. K., Moulasa, O. K., and Thiery, W.: MeteoSaver v1.0: a machine-learning based software for the transcription of historical weather data, Geosci. Model Dev., 19, 3213–3255, 10.5194/gmd-19-3213-2026, 2026e.Muñoz Sabater, J.: ERA5-Land hourly data from 1950 to present, Copernicus Climate Change Service (C3S) Climate Data Store (CDS) [data set], 10.24381/cds.e2161bac (last access: 25 July 2025), 2019.Noone, S., D'Arcy, C., Donegan, S., Durkan, W., Essel, B., Healion, K., Hersbach, H., Madden, S., Marshall, J., McConnell, L., Mensah, I., Scroxton, N., Thiesen, S., and Thorne, P.: Investigating the potential for students to contribute to climate data rescue: Introducing the Climate Data Rescue Africa project (CliDaR-Africa), Geosci. Data J., 11, 758–774, 10.1002/gdj3.248, 2024.Peterson, T. C., Easterling, D. R., Karl, T. R., Groisman, P., Nicholls, N., Plummer, N., Torok, S., Auer, I., Boehm, R., Gullett, D., Vincent, L., Heino, R., Tuomenvirta, H., Mestre, O., Szentimrey, T., Salinger, J., Førland, E. J., Hanssen-Bauer, I., Alexandersson, H., Jones, P., and Parker, D.: Homogeneity adjustments of in situ atmospheric climate data: a review, Int. J. Climatol., 18, 1493–1517, 10.1002/(SICI)1097-0088(19981115)18:13<1493::AID-JOC329>3.0.CO;2-T, 1998.Ryu, S., Song, J. J., and Lee, G.: Interpolation of Temperature in a Mountainous Region Using Heterogeneous Observation Networks, Atmosphere, 15, 10.3390/atmos15081018, 2024.Samset, B. H., Zhou, C., Fuglestvedt, J. S., Lund, M. T., Marotzke, J., and Zelinka, M. D.: Steady global surface warming from 1973 to 2022 but increased warming rate after 1990, Commun. Earth Environ., 4, 10.1038/s43247-023-01061-4, 2023.Seneviratne, S., Zhang, X., Adnan, M., Badi, W., Dereczynski, C., Di Luca, A., Ghosh, S., Iskandar, I., Kossin, J., Lewis, S., Otto, F., Pinto, I., Satoh, M., Vicente-Serrano, S., Wehner, M., and Zhou, B.: Weather and Climate Extreme Events in a Changing Climate, in: Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change, edited by: Masson-Delmotte, V., Zhai, P., Pirani, A., Connors, S. L., Péan, C., Berger, S., Caud, N., Chen, Y., Goldfarb, L., Gomis, M. I., Huang, M., Leitzell, K., Lonnoy, E., Matthews, J. B. R., Maycock, T. K., Waterfield, T., Yelekçi, O., Yu, R., and Zhou, B., book section 11, 1513–1765, Cambridge University Press, Cambridge, UK and New York, NY, USA, 10.1017/9781009157896.013, 2021.Sims, M. J., Stanimirova, R., Raichuk, A., Neumann, M., Richter, J., Follett, F., MacCarthy, J., Lister, K., Randle, C., Sloat, L., Esipova, E., Jupiter, J., Stanton, C., Morris, D., Melhart Slay, C., Purves, D., and Harris, N.: Global drivers of forest loss at 1 km resolution, Environ. Res. Lett., 20, 074027, 10.1088/1748-9326/add606, 2025.Sonwa, D. J., Nkem, J. N., Idinoba, M. E., Bele, M. Y., and Jum, C.: Building regional priorities in forests for development and adaptation to climate change in the Congo Basin, Mitigation and Adaptation Strategies for Global Change, 17, 10.1007/s11027-011-9335-5, 2012.Tesseract OCR: tesstrain, https://github.com/tesseract-ocr/tesstrain (last access: 7 October 2025), 2025a.Tesseract OCR: tesseract, https://github.com/tesseract-ocr/tesseract (last access: 7 October 2025), 2025b.The World Bank Group: Climate Risk Profile: Congo, Democratic Republic, World Bank Publications, https://climateknowledgeportal.worldbank.org/sites/default/files/2021-06/15883-WB_Congo%2C Democratic Republic Country Profile-WEB.pdf (last access: 11 April 2026), 2021.Turubanova, S., Potapov, P. V., Tyukavina, A., and Hansen, M. C.: Ongoing primary forest loss in Brazil, Democratic Republic of the Congo, and Indonesia, Environ. Res. Lett., 13, 074028, 10.1088/1748-9326/aacd1c, 2018.Węglarczyk, S.: Kernel density estimation and its application, ITM Web of Conferences: XLVIII Seminar of Applied Mathematics, 23, 8, 10.1051/itmconf/20182300037, 2018. Weiss, H.: War and Peace in the Democratic Republic of the Congo, Current African issues, Nordiska Afrikaninstitutet, ISBN 9789171064585, 2000.Wilcox, R. R.: Theil-Sen estimator, in: Fundamentals of Modern Statistical Methods: Substantially Improving Power and Accuracy, 194–197, Springer Nature, ISBN 978-1-4419-5524-1, 10.1007/978-1-4419-5525-8, 2001. Williams, C. N., Menne, M. J., and Thorne, P. W.: Benchmarking the performance of pairwise homogenization of surface temperatures in the United States, J. Geophys. Res.-Atmos., 117, 10.1029/2011JD016761, 2012.WMO: Guidelines on Best Practices for Climate Data Rescue, World Meteorological Organisation, https://library.wmo.int/idurl/4/55395 (last access: 30 September 2024), 2016.Wongchuig, S., Papa, F., Fleischmann, A. S., Sierra, J. P., Boucharel, J., Espinoza, J. C., Kitambo, B., Oliveira, R. J., Paris, A., Paiva, R., Casas, P., and Tshimanga, R.: Recent significant drying in Central Congo Basin linked to weakened Walker circulation and warmer Atlantic, npj Climate and Atmospheric Science, 8, 10.1038/s41612-025-01225-3, 2025.Yuh, Y. G., N’Goran, K. P., Kross, A., Heurich, M., Matthews, H. D., and Turner, S. E.: Monitoring forest cover and land use change in the Congo Basin under IPCC climate change scenarios, PLoS ONE, 19, 10.1371/journal.pone.0311816, 2024.Zeppetello, L. R. V., Parsons, L. A., Spector, J. T., Naylor, R. L., Battisti, D. S., Masuda, Y. J., and Wolff, N. H.: Large scale tropical deforestation drives extreme warming, Environ. Res. Lett., 10.1088/1748-9326/ab96d2, 2020.Zhang, X., Zwiers, F. W., Hegerl, G. C., Lambert, F. H., Gillett, N. P., Solomon, S., Stott, P. A., and Nozawa, T.: Detection of human influence on twentieth-century precipitation trends, Nature, 448, 461–465, 10.1038/nature06025, 2007.Zhou, L., Tian, Y., Myneni, R. B., Ciais, P., Saatchi, S., Liu, Y. Y., Piao, S., Chen, H., Vermote, E. F., Song, C., and Hwang, T.: Widespread decline of Congo rainforest greenness in the past decade, Nature, 508, 86–90, 10.1038/nature13265, 2014.