ASCA-YOLO: Adaptive Sparse and Context-Aware YOLO Algorithm for Forest Wildfire Detection

Combining Unmanned Aerial Vehicle (UAV) remote sensing with computer vision has become an efficient approach to detect forest wildfires. Nevertheless, existing methods still face several challenges, including missed detection of small fire spots and slender smoke under limited computational resources, false alarms caused by complex forest backgrounds, and insufficient adaptability to the irregular and dynamic morphology of fire and smoke. To address these issues, an improved YOLO26-based model, termed ASCA-YOLO, is proposed. Specifically, FWAMSConv module is introduced to improve multi-scale indicator representation of small and sparse targets. In addition, FWSCSAttention mechanism is designed to reduce background interference by modeling contextual feature distributions. Moreover, FWASIoU loss is developed to improve bounding box regression for non-rigid targets. The experimental evaluation indicates that, relative to YOLO26, the proposed model decreases the parameter count and FLOPs by 19.2% and 21.3%, respectively. Meanwhile, recall reaches 0.809 and precision reaches 0.870, indicating improved detection performance under complex conditions. In addition, mAP50-95 is improved by 12.9%, reflecting more stable localization for irregular wildfire targets. Overall, ASCA-YOLO attains a better equilibrium between detection quality and computational cost than several mainstream object detection models, indicating its potential for real-time UAV-based wildfire monitoring.