Spatiotemporal attention ConvLSTM networks for predicting and physically interpreting wildfire spread

Wildfire is one of the billion-dollar disasters that manifest complex climate and vegetation dynamics. It has been one of the most headache problems in the Western United States and many other countries in recent years. Predicting its spreading over time after initial discovery is the key information for the subsequent response and containment. This chapter discusses modeling challenges and presents deep learning approaches to better capture spatiotemporal interrelationships between a fire's space-time evolution and underlying biophysical factors. Specifically, we showcase two spatiotemporal attention-based sequence-sequence modeling frameworks using Convolutional Long Short-Term Memory (ConvLSTM) networks to model fire progression dynamics in space-time. The effectiveness of these approaches is examined with a very high spatiotemporal resolution wildfire spread dataset generated by a semiempirical percolation model that mimics the biophysical process of wildfire spread. Additionally, we examine the model interpretability using Integrated Gradients and attention maps to identify spatiotemporal biophysical feature importance for wildfire spread. Results show that attention-based models are superior in predicting fire-front movement and support fire-ecologically meaningful physical interpretations. This chapter provides important insights on the effectiveness of attention mechanisms for capturing complex spatiotemporal dynamics of wildfire spread, which can help fire management efforts.