Design of an iterative vision transformer and LSTM-attention framework for multiple temporal wildfire detection and severity mapping using Sentinel-2 and MODIS data samples

Detection of wildfires and severity mapping in forest ecosystems are critical for ecological preservation and risk mitigation. Traditional methods face challenges such as low spatial resolution, delayed reporting and limited integration of multi-source satellite data. This study proposes a hybrid wildfire detection framework that integrates multi-temporal Sentinel-2 composites and MODIS thermal and burned‑area products, accessed via Google Earth Engine (GEE). Vision Transformers (ViTs) are employed to extract spatial and spectral features, while Long Short-Term Memory (LSTM) networks with an attention mechanism capture temporal dynamics and interval-specific changes. The model generates fire probability maps and employs K-Means clustering to classify fire-prone zones, while NDVI loss and Root Mean Square Error (RMSE) are used to assess vegetation degradation and severity. Experimental results show strong alignment with MODIS burned-area masks, achieving 92.8% accuracy, 90.4% precision, 89.1% recall and an F1 Score of 89.7%. Severity overlays and fire masks further provide actionable insights for risk mitigation. The proposed framework demonstrates significant improvements in spatial granularity, temporal sensitivity and operational applicability, making it a scalable solution for near real-time wildfire detection and management.