Exploring the impacts of dual-polarized vegetation indices and U-shaped deep learning architectures on SAR-based burned area mapping

Abstract

Frequent and severe wildfires driven by climate change intensify the need for accurate burned area (BA) mapping, which can be effectively addressed using synthetic aperture radar (SAR) due to its cloud- penetration capability and sensitivity to vegetation and moisture changes. However, BA mapping based on SAR-only approaches relies on U-Net with ResNet50 backbone or fully convolutional neural network, while the potential of advanced architectural components remains underexplored. Moreover, prior research primarily emphasizes log-ratio features, with limited focus on standalone capacity of dual polarized vegetation indices (VIs). This study addresses these gaps by evaluating the performance of five U-Net variants (U-Net, Attention U-Net, Residual Attention U-Net, U-Net++, and U-Net 3+) using four input schemes: log-ratio, log-ratio without cross-ratio, VIs, and a combined feature set of all. Three combinations of loss function - binary cross entropy (BCE), dice, and focal - were also applied to the best model of all schemes. Experimental results showed that U-Net++ with log-ratio inputs under BCE loss function achieves the highest performance, yielding an F1 score of 0.8218 and an Intersection of Union (IoU) of 0.6795. Further analysis revealed that VIs alone can effectively delineate burned areas (F1: 0.8244; IoU: 0.7013) with focal loss and combining them with log-ratio features delivered the best performance (F1: 0.8364; IoU: 0.7188), when dice and focal loss functions were applied. Overall, this study provided a quantitative evaluation of how dual-polarized VIs and deep learning architectures affect SAR-based BA mapping performance and suggested promising directions for future enhancement through advanced feature extraction techniques.