AttentionFire_v1.0: interpretable machine learning fire model for burned-area predictions over tropics

African and South American (ASA) wildfires account for more than 70 % ofglobal burned areas and have strong connection to local climate forsub-seasonal to seasonal wildfire dynamics. However, representation of thewildfire-climate relationship remains challenging due to spatiotemporallyheterogenous responses of wildfires to climate variability and humaninfluences. Here, we developed an interpretable machine learning (ML) firemodel (AttentionFire_v1.0) to resolve the complex controls ofclimate and human activities on burned areas and to better predict burnedareas over ASA regions. Our ML fire model substantially improvedpredictability of burned areas for both spatial and temporal dynamicscompared with five commonly used machine learning models. More importantly,the model revealed strong time-lagged control from climate wetness on theburned areas. The model also predicted that, under a high-emission future climate scenario, the recently observed declines in burned area will reversein South America in the near future due to climate changes. Our studyprovides a reliable and interpretable fire model and highlights the importanceof lagged wildfire-climate relationships in historical and futurepredictions.