Fig. 23

The visualisations depict the ground truth fire mask, the prior prediction, and the current prediction for a wildfire scenario, accompanied by the relevant generated textual analysis. Both models comprehend the spatiotemporal dynamics of the fire; LLaMA3-8B demonstrates superior semantic accuracy and contextual depth in delineating the spread pattern, intensity, and directional progression. Mistral provides a clear overview that includes significant statistics such as spatial fragmentation and percentage change.