FDTNet: Multi-Disease Detection for Preliminary Visual Screening in Soybean–Maize Strip Intercropping Systems

Soybean–maize strip intercropping is a representative sustainable cereal–legume production system, but dense canopy closure, interspecific leaf occlusion, and heterogeneous field backgrounds hinder timely and accurate disease scouting. Rapid identification and localization of multiple foliar diseases are critical for supporting field inspection and preliminary targeted management. To address the challenges of small lesions, variable lesion scales, and complex field backgrounds, this study proposes FDTNet, an improved object detection model based on YOLOv11n. We introduce the FasterNet backbone for optimized feature extraction, and construct a Feature Diffusion Pyramid Network (FDPN) combined with a Dimension-Aware Selective Integration (DASI) module as the joint fusion block. A Task-Aware Decoupled Detection Head (TADDH) is further designed to enhance disease classification and bounding box localization. The model was evaluated on a dataset of 2196 images containing 10,574 annotated instances across nine soybean and maize foliar diseases. On the test set with image-level random split, FDTNet achieved a precision of 94.4%, a recall of 92.6%, an mAP@0.5 of 95.2%, and an mAP@[0.5:0.95] of 69.5%, which is 5.2 percentage points higher in mAP@0.5 than the baseline YOLOv11n. The AP@0.5 values for the nine categories ranged from 87.1% to 99.6%, showing favorable performance for most disease types. This study demonstrates that FDTNet can serve as a preliminary visual screening tool for multi-disease identification and symptomatic region localization. The model does not provide outputs for disease severity, treatment thresholds, pesticide selection or application timing; its performance under plot-level splits, cross-region evaluation and different imaging devices still needs further verification.