Capturing and interpreting wildfire spread dynamics: attention-based spatiotemporal models using ConvLSTM networks