Semi-supervised segmentation of forest fires from UAV remote sensing images via panoramic feature fusion and pixel contrastive learning

Yuchen Ma¹Haifeng Lin^2*

• ¹Nanjing University of Posts and Telecommunications, Nanjing, China
• ²College of Information Science and Technology, Nanjing Forestry University, Nanjing, China

Unmanned aerial vehicle (UAV) remote sensing has become an important tool for forest fire monitoring due to its high spatial resolution and flexible data acquisition. However, existing forest fire segmentation methods still face significant limitations in terms of accuracy and robustness. These limitations are mainly caused by the complex imaging environment, unclear fire boundaries, diverse fire shapes, and limited availability of high-quality annotated data. To address these, we propose a semi-supervised segmentation method for UAV remote sensing forest fire, named PPCNet, which combines panoramic feature fusion and pixel-level contrastive learning. Specifically, a panoramic feature fusion (PFF) module is designed to dynamically integrate multi-scale and multi-level features, which enhances the joint expression of global structure and local details and alleviates the problem of insufficient boundary representation under complex backgrounds. To further improve the segmentation of fire edges and texture details, a dual-frequency feature enhancement (DFFE) module is constructed, which effectively combines high-frequency and low-frequency information to strengthen boundary and detail features. In addition, a pixel contrastive loss (PCL) is introduced based on pseudo-labels and feature contrast constraints, which reduces the misclassification and feature degradation in the unsupervised branch and significantly improves segmentation accuracy and stability under limited labeled data. Experimental results on four typical forest fire datasets demonstrate that the proposed PPCNet (Panoramic Feature Fusion and Pixel Contrastive Learning Network) achieves better segmentation performance than existing mainstream methods. This method provides an effective technical solution for intelligent forest fire segmentation based on UAV remote sensing and shows good application potential.