Performance Analysis of Deep Learning-Based Fusion of Sentinel-2 and MODIS Thermal Anomaly Data for Small-Scale Forest Fire Detection

Forest fires constitute one of the most destructive natural disasters affecting ecosystem integrity, biodiversity, and human settlements worldwide. The early and accurate detection of small-scale fires below 1 hectare is critically important for preventing large-scale catastrophes. Existing threshold-based methods rely on low spatial resolution thermal data and prove inadequate for detecting small fire hotspots. This study examines the performance of a deep learning-based fusion framework that integrates Sentinel-2 multispectral imagery with MODIS thermal anomaly data for real-time detection of small-scale forest fires. The proposed approach employs U-Net and Attention-based convolutional neural network (CNN) architectures for multi-source data fusion, with a comparative analysis against traditional thresholding methods (MODIS Active Fire, Brightness Temperature Thresholding). Experimental findings demonstrate that the deep learning-based fusion approach achieves statistically significant superiority over traditional methods in terms of F1 score, precision, and recall metrics. Notably, the proposed model attained a recall rate of 89.3% for fire hotspots in the 0.5–1.0 hectare range, whereas traditional thresholding methods remained at 52.7%. This study highlights the potential of integrating remote sensing and artificial intelligence disciplines in disaster management and provides a roadmap for future research.