PPCS-YOLO: a lightweight multiscale feature fusion network for forest fire detection in satellite imagery
Abstract
Forest fire detection in satellite imagery remains challenging because of substantial target-scale variations, weak visual features of small fire regions, complex background interference, and the need for lightweight deployment. We propose PPCS-YOLO, a lightweight forest fire detection model based on YOLOv11n. The proposed CSS-CA-PConv module combines channel splitting and shuffling, Coordinate Attention, and Partial Convolution to improve feature representation while reducing redundant computation. A Pixel-attention-guided Fusion Module (PagFM) adaptively integrates shallow spatial details and deep semantic features. C2PSA-Mona enhances global contextual and local multiscale representations. A lightweight Detect-Efficient head and Scylla Intersection over Union (SIoU) loss are adopted to improve prediction efficiency and bounding-box regression. Experimental results show that PPCS-YOLO achieves an mAP 50 of 0.859 and an mAP 50:95 of 0.618. Comparative and ablation experiments demonstrate favorable detection performance across target scales and complex background conditions while maintaining relatively low model complexity. The complementary feature extraction, adaptive multiscale fusion, and global-local enhancement mechanisms help preserve weak fire-related features and improve target localization. PPCS-YOLO provides a promising lightweight approach for forest fire detection in large-scale satellite imagery.