Unveiling spatiotemporal patterns of wildfire risk: a transformer-based earth system analysis

Wildfires profoundly influence ecosystems, human societies, and economic activities as a global phenomenon. The continuing climate change increases the frequency and intensity of wildfire, resulting in urgent needs in searching for better fire management strategies. This paper aims to pave a pathway towards meeting this challenge through accurately predicting spatiotemporal pattern of global wildfire risk, using an improved Transformer model that integrates information theory and full-attention mechanisms. Experimental results demonstrate that the proposed SimAM modulated Full Attention Network shows superior performances in terms of Accuracy, Recall, and Area Under the Precision-Recall Curve. Furthermore, new discoveries based on the model find out that the wildfire risk in the northern forest region of Australia is influenced by the seasonality of the climate in North America and the Pacific and the dry winter climate in the Canadian region, illuminating the intricate relationship between the global climate and regional wildfire risk. These findings provide new tools and knowledges for understanding the mechanisms in global wildfire risk.