The consideration of gross land changes, meaning all area gains and losses within a pixel or administrative unit (e.g. country),
plays an essential role in the estimation of total land changes. Gross land changes affect the magnitude of total land changes, which
feeds back to the attribution of biogeochemical and biophysical processes related to climate change in Earth system models. Global
empirical studies on gross land changes are currently lacking. Whilst the relevance of gross changes for global change has been
indicated in the literature, it is not accounted for in future land change scenarios. In this study, we extract gross and net land
change dynamics from large-scale and high-resolution (30–100
Land change dynamics (e.g. changes in land cover or land use) play a major role in the Earth system. They have far-reaching consequences by altering many biophysical and biogeochemical ecosystem processes (e.g. albedo, greenhouse gas fluxes, transpiration, water balance, and surface roughness), which directly or indirectly drive the climate on the continental to global scale (Ciais et al., 2013; Gaillard et al., 2010; Houghton et al., 2012; Shevliakova et al., 2009; Teuling et al., 2017; Zaehle and Dalmonech, 2011). Earth system models (ESMs) are used to explore the impacts of land changes on future climate, biogeochemical cycling, and vegetation dynamics. Information on the extent and amount of land changes is usually provided by land use change models (LUCMs) or the land use modules of integrated assessment models (IAMs). Land change dynamics can be provided by LUCMs and IAMS either by a “net change approach”, i.e. area gains minus area losses per grid cell, or by a “gross change approach”, i.e. area gains plus area losses per grid cell. Not accounting for gross land changes has been shown to substantially underestimate the amount of land changes and related climate effects (Arneth et al., 2017; Bayer et al., 2017; Fuchs et al., 2015, 2016; Peng et al., 2016; Prestele et al., 2017). Thus, gross changes need to be considered in future model development.
The implementation of gross land changes, however, faces various difficulties. First, LUCMs and IAMs mostly have limited abilities to
account for gross land changes at the scale of modelling. Most land use models only account for land changes in one direction. For
instance, if the model has to allocate increasing area for a specific land cover type, it is often not able to model area losses of the
same class at the same time. Second, land use models typically simulate land changes at a spatial resolution of 5 arcmin
(ca. 10
As part of the Coupled Model Intercomparison Project Phase 6 (CMIP6) many ESMs are potentially able to account for gross land changes (Arneth et al., 2017). However, empirically based gross land change data that can directly be implemented in assessment models are currently lacking on a global scale. This lack of data availability hampers a comprehensive integration of gross land change information in LUCMs and, since LUCMs often feed into ESMs and IAMs, also in ESMs and IAMs (Bayer et al., 2017; Prestele et al., 2017). Moreover, in recent years, the focus of assessing the impact of gross land changes on the climate was mainly based on the historical period (Bayer et al., 2017; Fuchs et al., 2015, 2016; Hurtt et al., 2006; Wilkenskjeld et al., 2014). The role of gross land changes in future land use projections remained unclear, mostly because of the unknown magnitude of present-day gross land changes, but also the lack of understanding of how gross land change dynamics would develop with time (Arneth et al., 2017; Hurtt et al., 2011; Stocker et al., 2014). This inhibits a precise appraisal of future mitigation and adaptation potentials (Arneth et al., 2017). Currently, the Land Use Harmonization data (LUH; Hurtt et al., 2011, and its updated CMIP6 version LUH2; G. C. Hurtt et al., personal communication, 2018) are the only global datasets accounting for gross land changes. However, in these datasets gross land changes are assumed to only occur in shifting cultivation areas of the tropics (Bayer et al., 2017). A global quantification of other bidirectional changes, like cropland expansion and abandonment or afforestation and deforestation within the grid cell sizes of ESMs or IAMS, is missing completely (Prestele et al., 2017).
Empirical data, such as from remote sensing or land cover statistics, that contain information on area gains and area losses can be
used to inform LUCMs and IAMs about land changes below their native resolution (further on referred to as land changes on a “sub-pixel”
scale). Such empirical data has recently become available at very high spatial resolutions (30–100
The objective of this paper is to improve the current representation of gross land changes in LUCMs and IAMs by conducting an empirical
analysis of gross land use changes and proposing an approach that implements empirically derived gross land changes in a global
land use model. We account for both the gross land changes at the model scale (5 arcmin spatial resolution) and the gross land
changes at the sub-pixel scale. Specifically, we (1) characterize global-scale relationships between gross and net change by analysing
empirical data, (2) apply these findings to a future land use change simulation, and (3) demonstrate how the consideration of gross
land changes, in contrast to net land changes, can lead to substantially different results with respect to land use composition, future
land change dynamics, and consequences for global change studies, e.g. on the global carbon cycle. Moreover, we translate the total
gross and net land change into metrics that ESMs are able to use at a common resolution of 0.5
Spatial coverage of high-resolution land change datasets based on remote sensing that were used in this study to derive
gross land change dynamics within 0.5
Overview of high-resolution land change datasets used in this study.
Comparison of regional and continental remote sensing products (left) with global Globeland30 (right) for overlapping areas and
roughly the same time spans, showing
In total, we used 13 independent empirical datasets based on remote sensing to assess land changes on the sub-pixel scale. The spatial
coverage of all datasets used in this study is depicted in Fig. 1. The individual features, accuracies, and available years are shown in
Table 1. Since our objective was to describe future land change dynamics, we first focused on datasets that cover the most recent years
(from 2000 onwards). Some datasets contain data for years before 2000. They were preprocessed and analysed but not used in this
study. Secondly, we selected datasets that had a minimum spatial resolution of 100
Land system map for the baseline year 2000 used by the CLUMondo model.
To assess gross and net land changes for a future scenario, we used simulation output from the land system model CLUMondo (van Asselen
and Verburg, 2012, 2013). This model uses a land system classification instead of the more traditional land
cover classification (van Asselen and Verburg, 2012). Land systems are described by a set of fractional land cover classes consisting
of built-up, cropland, grassland, forest, and other land co-occurring in spatial simulation units of
The model allocates the land systems to fulfil the demand for goods described by a scenario. The model is able to simulate gross land changes inherently by expanding a land system at one place while contracting it at another place. Within the model algorithm, each location is assigned the land system with the highest competitive power at that place. For some land systems, conversion restrictions are applied, e.g. to prevent urban development from being converted back to agricultural use and to account for conversion costs. Allowing such co-occurring area gains and losses in the same land system within a world region the model accounts for gross change dynamics at the scale of modelling. However, gross changes at the sub-pixel scale are not taken into account.
In this study, we used a reference scenario for the period 2000–2040 to demonstrate the feasibility of our approach to include gross land change dynamics in a land use model. This reference scenario is driven by the demand for crop production, ruminant livestock production, and the provision of built-up areas (Eitelberg et al., 2016). The scenario is based on the United Nations Food and Agriculture Organization (FAO) report World Agriculture Towards 2030–2050, the 2012 revision (Alexandratos and Bruinsma, 2012), and characterizes the development of crop and livestock systems from 2010–2050. Regional-level future demands for crop production and livestock are provided by the integrated assessment model IMAGE (Stehfest et al., 2014). Further details on the scenario can be found in Eitelberg et al. (2016).
To assess the gross or net land change dynamics at a spatial resolution relevant for ESMs and IAMs, we analysed all land changes (at
the scale of modelling and sub-pixel scale) at the common spatial resolution of ESMs and IAMs (0.5
Detailed overview of the approach. The approach is divided into three major steps: preprocessing (top), processing (middle), and the post-processing of the results (bottom). The left panel explains the individual steps for the analysis at the scale of modelling using a land system model and a reference scenario for the period 2000–2040. The right panel shows the individual steps for the analysis at the sub-pixel scale using empirical data.
In order to assess gross and net land changes at the level of land cover types commonly used in Earth system models, the land systems
had to be translated back into their land cover components (e.g. grassland, cropland, forest, etc.). For the sake of efficiency, we
focused on areas where land systems have changed (Fig. 4, upper left box). Within each 0.5
We first reprojected all original empirical datasets into an equal area projection (WGS84 Eckert IV). Subsequently, we aggregated all class legends for each product into five IPCC land categories (Intergovernmental Panel on Climate Change (IPCC), 2003): settlement, cropland (including orchards and agro-forestry), forest, grassland (pastures and natural grassland), and other land. In the Supplement S1 Sect. 1, we give an overview of how each legend was aggregated. The Globeland30 dataset had a class for “cultivated land” that in addition to cropland also contains managed pastures that could not be separated properly. In this study, we considered “cultivated land” as cropland.
We calculated one change dataset for every time step of each product with the original spatial resolution. For the CORINE product, we used the available change layers. For Indian land cover 1995 to 2005 and for some regions in the Globeland30, we recognized a shift by 1 pixel between the individual years when calculating land changes. This caused problems in generating the change layers. In the Supplement S1 Sect. 2, we explain in detail how we solved these problems.
We used a mask of each land system for the base year 2000 to clip our various change products by land system in order to retrieve land
changes per land system. In parallel, we created a 0.5
The tabulation of change areas per land system within each 0.5
From the various change products and time steps, we calculated spatially weighted averages for the change parameters to account for the different spatial coverages of each product. In the Supplement S1 Sect. 5, we provide an overview for each dataset and its fractional contribution per land system to the final weighted average. Further, we annualized every time step of the individual products to make datasets with different time spans comparable in their change dynamics (Fig. 4, upper right box). Time steps 1 year are the regular time intervals of many land use models, including the one used here.
Subsequently, we applied our derived change parameters (net change fraction, gross
In the post-processing phase, we combined our results from the land changes derived at the scale of modelling (Sect. 2.3.2) and land
changes derived on the sub-pixel scale (Sect. 2.3.3). We aggregated both datasets at 0.5
Empirical gross
In Table 2, we list all gross
Gross
If we separate the different land systems based on their land cover composition, it can be seen that on average homogeneous land
systems have lower gross
Overall area gains (blue) and losses (red) per land cover class shown as change
rate per year in percent for the period 2000–2040. Left figure panels
To demonstrate our approach for deriving gross change in future scenarios, we used a reference scenario for the period 2000–2040 based on the United Nations Food and Agriculture Organization (FAO) report World Agriculture Towards 2030–2050 (Alexandratos and Bruinsma, 2012; see Sect. 2.2. for details). In Fig. 6, we show the different area gains and losses based on this scenario for each land cover component over the entire modelling period (2000–2040). The left panels (Fig. 6a, c, e, and g) refer to gains and losses derived by land system changes (scale of modelling). The right panels (Fig. 6b, d, f, and h) refer to the combination of changes at both the scale of modelling and the sub-pixel scale.
Based solely on the simulated changes at the model scale (Fig. 6, left panels), the main areas of land use change were found on the east coast of the US, in Brazil and Argentina, the Sahel zone in Africa, large parts of Europe, and some regions of the Middle East, India, China, and South-east Asia. Except for eastern Europe, some parts of India, China, and Mexico, these changes led to widespread cropland area gains. On the east coast of the US, eastern Europe, India, Argentina, and South-east Asia, this came at the expense of forest. In Brazil, the Sahel zone, the Middle East, and northern China, cropland gains occurred at the cost of grassland and other land losses. However, large parts of the world remained unaltered in the modelled scenario (see the yellowish colours for each land cover class in the left panels of Fig. 6).
The combination of changes at the scale of modelling and sub-pixel scale changed the overall picture of area gain and loss (Fig. 6, right panels). Spatial patterns of changes appeared more diversified and subtle, depending on the occurrence and the empirical parameterization of land change dynamics of each individual land system type. For instance, forest changes (Fig. 6b) appeared more widespread. Large parts of Africa and South America outside of the tropical rainforest, the boreal region, China, and Australia were now subject to strong forest dynamics including reforestation
Likewise, areas of large forest losses on the east coast of the US and South-east Asia were amplified at the same time as some reforestation is happening in these regions. Similar results were found for all other land cover classes. Additionally, the magnitude of changed area per pixel increased considerably by adding sub-pixel processes. When combining changes at the scale of modelling and sub-pixel scale, the land system model and scenario implementation accounted for 20 % of all gross and net land changes, while the other 80 % of changes originated sub-pixel changes. For forest and grassland, this led to larger area gains, while for cropland and other land this led to a higher magnitude of area losses. Moreover, the overall trends of gains and losses for some land cover classes in some regions (for instance, grassland in the US) even reversed by adding sub-pixel processes (Fig. 6d) compared to the approach without including these (Fig. 6c).
Global patterns of combined land change rates (at the scale of modelling and sub-pixel scale) per year (in percent) for
a reference scenario for the period 2000–2040. Panel
Major land cover classes that cause gross land changes. Land changes comprise land changes at the scale of modelling and sub-pixel changes depicted as RGB composite, with forest (red), grassland (green), and cropland (blue). Note: pink, turquoise, and yellow colours refer to changes between two of these three classes (pinkish indicates cropland and forest, turquoise indicates cropland and grassland, and yellowish indicates forest and grassland). Brighter colours refer to higher gross land changes, and darker colours refer to lower gross land changes.
In Fig. 7, we added the absolute area gains and losses of all land cover classes together, comprising changes at the scale of modelling and sub-pixel scale. We depict the total net changes (Fig. 7a), total gross changes (Fig. 7b), and their difference (Fig. 7c) expressed as change rate per year and pixel in percent. Major change areas (net and gross) occurred in the eastern US, Mexico, Colombia, Argentina, the Sahel zone, the Atlas region in northern Africa, eastern and southern Europe, Turkey, central Asia, northern India, and China. The implementation of gross changes into the future scenario led to higher change rates. While net land changes had a global average of 0.92 % area change per year in this scenario, the consideration of gross land changes yielded 1.36 % per year. The difference in net and gross land changes occurred mostly in large farming regions because of the scenario conditions under which new agricultural areas were established (Fig. 7c). Hot spots with larger differences between net and gross land changes appeared in Mexico, Spain, eastern Europe, parts of the Sahel zone, central Asia, India, and China mostly due to the high rates of land cover change in these regions.
In Fig. 8, we illustrate the relative contribution of land cover classes to the gross change rates at the modelling scale and sub-pixel scale. The individual contributions of land cover classes varied quite strongly over the whole globe. Forest changes (reddish colours) contributed most to the changes in the boreal region. Grassland changes (greenish colours) occurred most dominantly in the western US, the Andes region, major parts of sub-Saharan Africa, central Asia, and Australia. Cropland and forest conversions (pinkish colours) can be seen on the east coast of the US, Europe, India, and South-east Asia. While at the east coast of the US and in South-east Asia the main change processes comprise cropland expansions at the cost of forests, the picture in eastern Europe and India is the opposite (see also Fig. 6). High gross land changes in cropland expansion at the expense of grasslands (turquoise colours) occurred mostly in heterogeneous agricultural areas like Mexico, the Sahel zone, the Mediterranean region, and northern China. These regions are known for their smallholder, mosaic land systems, which have in general a high gross change rate due to their regional land management practice and shifting cultivation. Additionally, due to the scenario forcing, many new cropland areas were established on former grassland areas. In the wider Amazon region in South America, in regions around the Congo, and in southern China, contributions to gross land changes came from all three land cover classes (darker brownish colours).
In this paper, we presented a first estimate of global gross land change
parameters to account for gross land changes in global assessments. Our work
was largely based on empirical high-resolution data derived from remote
sensing (30–100
In the near future, many more datasets suitable for implementation in our approach from remote sensing can be expected, for example
the new land cover change product on a yearly basis, recently released by the Land Cover Climate Change Initiative (LC-CCI; 1992–2015, 300
In this study, we considered only data from the last 2 decades. However, additional land change data exist dating back to before the
year 2000. For Africa, a few RCMRD datasets provided information for 1990 (RCMRD, 2016). The NLCD and CORINE data also provided
land change data for the 1990s, although with lower accuracies (Vogelmann et al., 2001). The Indian land cover dataset has land change
data available back to 1980 (Meiyappan et al., 2016). More regional land change datasets certainly exist. Landsat satellite archives
provide data back to the 1970s (USGS, 2017). For some countries, for example the Netherlands, data back to 1900 exist, which in principle
could allow us to retrieve gross land changes (Kramer and Dorland, 2009). These historic data would allow us to generate time-period-dependent
gross
Applying existing and upcoming datasets will help to further extend our database and strengthen the reliability of the land change parameters. The use of multiple datasets for every world region would allow for more robust and region-specific estimates. For this study, we had to average our gross change parameters globally for each land system due to the limited amount of data for some regions. Averaging all empirical datasets for these land systems globally may lead to incorrect or inaccurate regional characterizations. Especially for grassland systems that occur over a very wide range of biomes (from tundra to the Sahel zone), such averaging is not correct. Therefore, we choose to use for these different grassland systems different averages for the Northern Hemisphere and the subtropical grassland systems (see the Supplement S1, Sect. 4). Nonetheless, an overestimation of this particular region, the tundra, may remain. This should be taken into account in applications using our data.
Our empirical analysis has confirmed that gross land changes occur globally in every world region. Applied to our future reference
scenario, net land changes led globally to an average of 0.92 % area change per year. Based on gross land changes the average
change rate was 1.35 % per year, which is an increase of roughly 50 % compared to the net change approach. In earlier
approaches that covered Europe only, a similar magnitude of difference between gross and net land changes could be proven outside
shifting cultivation areas (Fuchs et al., 2015). Approximately 20 % of all gross and net land changes originated from the scenario
implementation. The other 80 % of changes can be explained by sub-pixel changes identified from empirical data. This points to the
significance of empirical data and sub-pixel processes. Over the entire modelling period, the gross
In the simulated gross changes and in the sub-pixel gross changes, the main areas of change were related to regions with heterogeneous land systems, such as in shifting cultivation areas of Central America, the Sahel zone, and India. Mediterranean land systems (e.g. agro-forestry) and smallholder farming systems like in China or eastern sub-Saharan Africa also showed major changes.
Regrouping of our empirical data to five IPCC land categories for continental
regions as an example of adaptation potential to new legends or focus areas. The table shows
region-specific gross
The empirical data we used were subject to uncertainties as well. Occasionally it happened that the net change fraction was very
small,
resulting in very high gross
The approach presented in this paper used a specific future simulation model (CLUMondo) as an illustration. Other models using
a different land cover class aggregation of the original classes may need to further aggregate our classes. For example, ESMs like the
coupled LPJ (Smith et al., 2001) or ORCHIDEE (Ciais et al., 2005; Krinner et al., 2005) the IAMs like IMAGE (Stehfest et al.,
2014) or MESSAGE-GLOBIOM (Havlík et al., 2014) are able to account for cropland, grasslands, and forests. Urban areas and other
land are considered as well, but neutral in terms of fluxes. Additionally, all these models are able to work on at least 0.5
We aggregated our data to five common Intergovernmental Panel on Climate Change (IPCC) categories: settlement, cropland, forest,
grassland, and other land (IPCC, 2003) in order to show the potential of our approach. For each continental region, we averaged the
individual land cover components across all available land systems and calculated the same land change dynamic parameter as explained
in the methods. In Table 3, we show the gross
Similar to Fig. 6, in Figs. 7 and 8 we see the highest gross
Regrouping of our empirical data to five IPCC land categories for continental regions as an example of adaptation potential to new legends or focus areas. The table shows averaged land transition matrices for these continental regions (conversion matrix on the left, change matrix on the right). Note: T0 refers to time step 0 and T1 refers to time step 1, indicating the direction of change in time. These values may serve as proxies for ESMs and IAMs to account for gross land changes. Note: bold numbers refer to the totals per class for time step 0 and 1.
Using our gross change data may have various implications for Earth system modelling, since the amount of changed area determines the
dynamics and quantity of carbon fluxes, and the land conversion types determine to which carbon stocks the land changes have to be
allocated (Bayer et al., 2017; Fuchs et al., 2016). The same applies to other biogeochemical and biophysical variables (e.g. methane,
In this study, we could show based on empirical data that gross land changes occur globally in every world region. This finding contradicts earlier studies, which assumed gross land changes to appear in shifting cultivation areas only. Applied to our future reference scenario, net land changes led globally to an average of 0.92 % area change per year, while for gross land changes the average change rate was 1.35 % per year. This is an increase of roughly 50 % compared to the net change approach. Empirical data contributed ca. 80 % of changes in the future scenario we used. This highlights the importance of accounting for sub-pixel processes in global assessments. In our scenario, gross land changes appeared in regional patterns that are most dominant in eastern Europe, Turkey, the Sahel zone, the US, and development countries in transition, like the BRICS states (Brazil, Russia, India, China, and South Africa). Large-scale and high-resolution remote sensing data were crucial for this kind of assessment. This highlights the increasing importance of land-related remote sensing data in global assessments. With our approach, it is possible to further decrease uncertainties in land change dynamics and related land atmosphere fluxes in ESMs. This again helps to improve accuracies for future mitigation and adaptation scenarios.
All results are based on CLUMondo simulations (see
Eitelberg et al., 2016). The CLU-Mondo source code and simulation results are
available from
The supplement related to this article is available online at:
The authors declare that they have no conflict of interest.
The work reported in this study was supported financially by the European Union Seventh Framework LUC4C project and ERC grant agreement no. 311819 – GLOLAND to VU University Amsterdam. We would like to thank Atul Jain, Prasanth Meiyappan, Wenbin Wu, Qiangyi Yu, Daniel Murdiyarso, and Ahmad Basyiruddin Usman for sharing their data, expertise, and support with us.The article processing charges for this open-access publication were covered by a Research Centre of the Helmholtz Association. Edited by: Somnath Baidya Roy Reviewed by: two anonymous referees