High-Resolution Satellite Data Improve Insights Into Landscape Fires and Their Drivers in Southeastern Africa

Abstract

High-resolution time series of burned area derived from Sentinel-2 can advance understanding of the determinants and dynamics of fire by incorporating small fires previously excluded from regional analyses. Here, we assessed the drivers of fire frequency, size, and seasonality across Southeastern Africa comparing fine (Sentinel-2 MSI) and moderate (MODIS) resolution data. Twenty-six predictors of ignition patterns, fuel load, flammability, and fire spread were incorporated into machine learning models to evaluate their predictive capacity, relative importance, and directional relationships with fire regime attributes. We found large differences between fine- and moderate-resolution estimates of fire frequency, size, and to a lesser extent seasonality. Models using Sentinel-2 showed better predictive performance than those using MODIS with R2 values of 0.24 and 0.13, respectively, for fire frequency when validated in regions outside the training areas. However, the shapes of the relationship curves between fire regime attributes and predictors were generally consistent between sensors. High fire frequency was positively associated with fuel load and environmental seasonality, whereas low fire frequency was associated with interannual stability in land cover, livestock density, and human population. Fire sizes were generally small at both the high and low extremes of the precipitation and vegetation productivity gradient, as well as in highly transformed areas. The fraction of fire outside of the fire season was higher in low seasonality environments and under strong human influence. We demonstrate the general applicability of existing theory of fire dynamics derived via moderate-resolution fire data to fine-resolution data, while providing more nuanced insights into fire drivers.