A Ground-Penetrating Radar Method for Detecting Underground Plant Diseases Based on an Improved YOLOv13 Approach

To address the challenges of low accuracy, missed detections, false positives, and difficulties in detecting small targets during underground defect detection using Ground Penetrating Radar (GPR), this paper proposes an improved detection algorithm, CDD-YOLO, based on YOLOv13n. First, the original downsampling convolutions are replaced with the Context-Guided Block module to enhance the fusion of local and global context and improve the multi-feature representation capability of objects. Second, the DynamicWeightCBAM module introduces dynamic weights into the Convolutional Block Attention Module (CBAM) to enhance recognition and localization accuracy for small targets. Finally, the DySample module replaces traditional upsampling to improve the model's detail recovery and content-aware capabilities. Experimental results demonstrate that the CDD-YOLO model achieves improvements of 1.8%, 3.6%, 3.9%, and 5.7% over YOLOv13n in accuracy, recall, mAP50, and mAP50: 95 on the data-augmented GPR public dataset, while reducing parameter count by 8%, validating the effectiveness of the proposed improvements.