A novel salient object detection network for burned area segmentation in high-resolution remote sensing images

Abstract

Burned area segmentation (BAS) in remote sensing images (RSIs) is critical for forest fire monitoring, as it helps locate and extract damaged areas, providing a scientific basis for post-disaster recovery. However, existing BAS methods underperform on high-resolution RSIs due to diluted location information and blurred edges during sampling. To this, we propose PANet, a novel salient object detection (SOD) network designed for BAS in high-resolution RSIs. PANet introduces two key modules: Path Aggregation Decoder (PAD) and Progressive Multi-level Aggregation Predictor (PMAP). PAD integrates multi-level features for richer semantics, using detail feature flow to enhance edge quality and refined location feature flow to improve spatial accuracy. PMAP progressively fuses features from PAD to predict saliency maps, leveraging higher-level features to complement lower-level ones. We also constructed a new dataset for high-resolution BAS. Experiments on two BAS datasets show that PANet outperforms state-of-the-art methods. Code is available at: https://github.com/Voruarn/PANet.