Quantitative Comparison with Mainstream Models.

Forest fires pose a severe threat to ecosystems, and accurate burn scar extraction is critical for post-disaster recovery and ecological management. This study proposes an attention mechanism enhanced deep learning model for semantic segmentation of burn scars in Karst regions, aiming to address challenges such as fragmented terrain and complex vegetation patterns. The model integrates ResNet50 as the backbone network to leverage its robust feature extraction capability and residual connections, mitigating gradient vanishing problem. To enhance multi-scale feature learning while avoiding grid artifacts, we optimize the Atrous Spatial Pyramid Pooling (ASPP) module by reducing dilation rates to (1, 3, 5). Furthermore, a novel Global Attention Module (GAM) is introduced after the decoder branches to dynamically recalibrate channel-spatial dependencies, enabling precise segmentation in heterogeneous backgrounds. Experiments demonstrate the model’s superiority with a mean Intersection over Union (mIoU) of 91.82% and mean accuracy (mAcc) of 95.73%, outperforming mainstream models (e.g., DeepLabV3 + , SegFormer, Mask2former) and traditional methods. The model demonstrates outstanding extraction accuracy and strong generalization capabilities; however, there remains room for optimization in terms of parameter quantity and inference speed. Future work will further explore lightweight design and real-time performance enhancement strategies. This study combines deep learning with GIS and remote sensing technology to construct a single region dataset for typical fire events in Huaxi District, Guiyang City, Guizhou Province in 2024. An efficient framework for extracting burn spots from karst landforms is proposed, which can provide real-time reference for the impact assessment, ecological restoration, and carbon flux estimation of this fire event in the region.