A Remote Sensing Spectral Index Guided Bitemporal Residual Attention Network for Wildfire Burn Severity Mapping

Abstract

Wildfires cause substantial damage and present considerable risks to both natural ecosystem and human societies. A precise and prompt evaluation of wildfire-induced damage is crucial for effective postfire management and restoration. Considerable advancements have been made in monitoring and mapping fire-affected areas through feature engineering and machine learning techniques. However, existing methods often exhibit several limitations, such as complicated and time-intensive procedures on manual labeling, and a primary focus on binary classification, which only distinguishes between burned and nonburned areas. In response, this study develops a wildfire burn severity assessment model, BiRAUnet-NBR, which can not only accurately identify fire-affected areas, but also assess the burn severity levels (low, moderate, and high) within those areas. Built upon the standard U-Net architecture, the proposed BiRAUnet-NBR first incorporates bitemporal Sentinel 2 Level-2A remote sensing imagery, captured before and after a wildfire, which enables the model to better distinguish burned areas from the background and identify the severity level of the resulting burns. In addition, it further enhances the standard U-Net architecture by fusing additional spectral layers, such as the normalized burn ratio (NBR) derived from post- and prefire images, therefore, informing the detection of burn areas. Moreover, BiRAUnet-NBR also integrates attention mechanism, enabling the model to pay more attention to meaningful features and burn areas, and residual blocks in the decoder module, which not only significantly improves segmentation results but also enhances training stability and prevents the issue of vanishing gradients. The experimental results demonstrate the superiority of the proposed model in both multiclass and binary mapping of wildfire burn areas, achieving an overall accuracy over 95%. Furthermore, it outperforms baseline algorithms, including support vector machine, random forest, eXtreme gradient boosting, and fully convolutional network, with an average improvement of 18% in F1-score and 15% in mean intersection over union.

Keywords

• Deep learning (DL)
• image segmentation
• machine learning (ML)
• semiautomatic labeling
• U-Net
• wildfire mapping