UAV remote sensing for rice false smut detection via improved YOLOv12n

Rice false smut (RFS) has become widespread in major rice-growing regions. It not only reduce rice production, but more importantly, it produces toxins on panicles—posing a significant threat to food safety. Rapidly and accurately detecting RFS plays a key role in ensuring rice production. This study proposes a lightweight detection model named HSFF-YOLO, which is based on an improved YOLOv12n and utilizes unmanned aerial vehicle (UAV) remote sensing images. To address the challenge of small target detection during model development, the original loss function and Neck structure were replaced with the Shape-NWD loss function and the High-level Selective and Feature Fusion (HSFF) multi-scale feature fusion structure, respectively. This replacement effectively enhances the model’s detection accuracy through semantic screening and multi-scale feature fusion. To further highlight the lesion information of RFS, the Convolutional Block Attention Module-Parallel (CBAM-P) was incorporated to emphasize semantic and spatial information. Additionally, the lightweight DySample (Upsampling by Dynamic Sampling) upsampling module was introduced to reduce the model’s computational resource overhead. The results showed that the HSFF-YOLO model achieved an average accuracy of 80.7%, with parameters of 1.89 M, a Floating Point Operations (FLOPs) of 5.6 G, and a model size of 4.09 MB. Compared with YOLOv12n, it has decreased 25% and 21% of parameters and model size, respectively, while still presented relatively better mAP0.5%, precision and recall, with approximate 3%, 3.4% and 2.4% improvement. This study also conducted evaluations of the severity levels of RFS in the two regions. The results showed that RFS in the Longyan region was significantly more severe. This study indicated that UAV imaging coupled with HSFF-YOLO has excellent potential for RFS detection and completed the entire process from the rapid detection of RFS to disease severity assessment.