Regionally optimized fire parameterizations using feed-forward neural networks

Abstract. The fire weather index (FWI) is a widely used metric for fire danger based on meteorological observations. However, due to its empirical formulation based on a specific regional relationship between the meteorological observations and fire intensity, the ability of the FWI to accurately represent global satellite-derived fire intensity observations is limited. In this study, we propose a fire parameterization method using feed-forward neural networks (FFNNs) for individual grids. These FFNNs for each grid point utilize four daily meteorological variables (2-meter relative humidity (RH2m), precipitation, 2-meter temperature, and wind speed) as inputs. The outputs of the FFNNs are satellite-derived fire radiative power (FRP) values. Applying the proposed FFNNs for fire parameterization during the 2001–2020 period revealed a marked enhancement in cross-validated skill compared to parameterization solely based on the FWI. This improvement was particularly notable across East Asia, Russia, the eastern US, southern South America, and central Africa. The sensitivity experiments demonstrated that the RH2m is the most critical variable in estimating the FRP and its regional differences via the FFNNs. Conversely, the FWI-based estimations were primarily influenced by precipitation. The FFNNs accurately captured the observed nonlinear correlations between FRP and RH2m, as well as precipitation. In contrast, FWI-based estimations exhibit an excessively negative relationship between FRP and precipitation.