Detection of yam leaf diseases and pest damage under complex field conditions using LIMN-YOLO

Detecting yam leaf diseases and pest damage in field imagery remains difficult because small, densely distributed lesions have weak boundaries and are frequently affected by occlusion and uneven imaging conditions. We investigated whether non-uniform degradation augmentation, response-guided cross-scale fusion, and scale-sensitive bounding-box regression could improve detection without substantially increasing model complexity. A four-category dataset containing 1,434 images and 9,221 annotated instances was divided into training (1,149), validation (142), and held-out test (143) partitions. We developed LIMN-YOLO from YOLO11n and evaluated its components through controlled ablations, fixed-coefficient validation sensitivity analysis, size-stratified detection metrics, and three-seed comparisons. On the test set, LIMN-YOLO achieved mean average precision (mAP) at an intersection over union (IoU) of 0.5 and across IoU 0.5–0.95 of 0.8792 and 0.7182, respectively, versus 0.8100 and 0.6250 for YOLO11n. For targets with width and height at most 64 pixels at the evaluation input scale, average precision at IoU 0.50 (AP50) increased from 0.5652 to 0.7383. Across three training seeds, LIMN-YOLO maintained higher mAP@0.5:0.95 (0.7182 ± 0.0076 versus 0.6253 ± 0.0063; mean ± SD), whereas the additional benefit of Latin hypercube sampling (LHS) over random shadow–blur augmentation was modest. These findings demonstrate the potential of lightweight, image-based detection for yam leaf disease and pest monitoring under heterogeneous field acquisition conditions.