Adaptively spatial feature fusion network: an improved UAV detection method for wheat scab

Scab is one of the most important diseases in wheat. Rapid and accurate detection of wheat scab under farmland conditions is essential for timely and effectively managing the disease. This study proposes a method for automatically detecting wheat scab by using remote sensing from unmanned aerial vehicles (UAVs). In the method, contrast enhancement was carried out on acquired RGB images of wheat to highlight the diseased spots, and then an adaptively spatial feature fusion network (ASFFNet) was constructed to detect wheat scab in the images. ASFFNet used the feature enhancement module to combine the global and local features of RGB images of wheat to improve the expression ability of these features. In addition, the feature fusion module in ASFFNet adaptively fused the enhanced features at multiple scales to solve the inconsistency of features at different scales during fusion caused by too small disease areas, which improved the detection precision. The results show that the proposed method has a higher AP (average precision) than the existing object detection algorithms, single shot MultiBox detector (SSD), RetinaNet, YOLOv3 (you only look once version 3) and YOLOv4 (you only look once version 4). The proposed method can be a practical way to handle the scab detection task using UAV images. It also can provide technical references for farmland-level wheat phenotype monitoring.