Climate change intensifies wildfires by worsening droughts and heatwaves, causing catastrophic impacts on ecosystems, economies, and human health. Forecasting wildfire danger is challenging due to its stochastic nature, yet critical for effective management. This study employs advanced Deep Learning (DL) models in a data-driven approach to forecast wildfire danger probabilities at a 1 km \(^2\) spatial resolution over a 1–5 day forecast horizon. Our study focuses on the seven regions in Australia: Western Australia (WA), New South Wales (NSW), Northern Territory (NT), Queensland (QL), South Australia (SA), Tasmania (TA), and Victoria (VI). We evaluate and compare three temporal DL models: Temporal Convolutional Network (TCN), Convolutional Long Short-Term Memory (ConvLSTM), and Gated Recurrent Unit (GRU), to identify the most effective method under varying wildfire conditions. To enhance transparency and model trustworthiness, we integrate Explainable Artificial Intelligence (xAI) techniques to identify key drivers of wildfire danger predictions. The results indicate that TCN and ConvLSTM models outperform the GRU model, with TCN achieving the highest average F1 score of 0.7982 in 2019 and ConvLSTM achieving the highest average F1 score of 0.661 in 2020. While all models showed strong predictive capabilities in 2019, their performance declined in 2020, particularly in Precision and F1 scores, due to more severe and unpredictable conditions. ConvLSTM demonstrated greater resilience and stability across different forecast horizons compared to TCN and GRU, suggesting its potential for more consistent wildfire danger prediction under varying environmental conditions. The study identifies solar radiation and relative humidity as the most influential factors. These findings contribute significantly to the refinement of forecasting models and hold important implications for enhancing wildfire response strategies.
