Stage-wise detection of late blight in field-grown potato: a comparative evaluation of deep learning object detectors

Potato late blight (Phytophthora infestans) progresses rapidly, yet most image-based approaches treat it as a single category; this study evaluated direct disease-stage identification by object detection under field conditions. Field RGB images (1,561) of potato cv. Kufri Jyoti, collected at IIT Kharagpur from a single experimental plot during one naturally occurring late-blight episode, were annotated at leaf level (22,216 instances) into Healthy, Initial Late Blight (ILB) and Advanced Late Blight (ALB) and split 70:20:10 for training, validation and testing. YOLOv8, YOLOv12 and YOLO26 shared the same data partitions, input resolution and training budget, each with its native optimizer and loss formulation and were benchmarked against six reference detectors. On the validation subset, YOLO26 performed best (mAP@0.50, 0.961; precision, 0.953; recall, 0.944; 91 frames s−1), versus 0.903 for YOLOv5l, the best-performing of the six reference detectors. As the YOLO models differed in capacity and optimizer and used a fixed random seed, this ranking is indicative. Residual errors were concentrated at the ILB–ALB boundary, where reciprocal confusion fell from 13% (YOLOv8) to 6% (YOLO26). Within this single-field, single-cultivar setting, direct stage detection provided leaf-level disease-stage information; its reliability under other field conditions remains to be established.