Adaptive multimodal transformer for wildfire spread prediction using feature-weighted attention

Abstract

Hazard modelling technology has evolved quickly in recent years to predict wildfires, which are an important class of natural disaster for accurate forecasting and proactive control. The combination of remote sensing and machine learning has been playing an important role in this area. Yet traditional models fail to accommodate the sophisticated spatiotemporal dynamics in fire-affected areas and tend to be agnostic toward individual predictors. This study utilizes a multimodal dataset comprising normalized difference vegetation index (NDVI), wind vectors, surface temperature, humidity, land cover, and elevation from sources such as moderate resolution imaging spectroradiometer (MODIS), Sentinel-2, ERA5, and shuttle radar topography mission (SRTM). These inputs were normalized within spatial grids, and temporally-aligned for day-ahead prediction. We introduce an adaptive multimodal transformer (AMT) with a feature weighting module (FWM) to adaptively emphasize informative features. The transformer architecture allows for long-range spatial learning, and the FWM strengthens contextual sensitivity. The novelty of this approach lies in its interpretable feature reweighting mechanism for dynamic environmental conditions. Model performance was evaluated using F1-score, intersection over union (IoU), mean absolute error (MAE), and mean average precision (MAP). Results show that the proposed model outperforms the MA-Net baseline, achieving a 5% improvement in F1-score and a 14.7% reduction in MAE, demonstrating superior accuracy, generalization, and interpretability.