White mold and Sclerotinia sclerotiorum detection: An ensemble deep learning approach

The white mold disease, caused by the soilborne pathogen Sclerotinia sclerotiorum , is an epidemic and widespread disease affecting common beans and over 400 host plants. Its early recognition at different growth stages is crucial for disease prevention and effective management, as well as for mitigating yield losses and ensuring crop sustainability. This work introduces a novel, high-quality image dataset annotated by plant pathologists named the Sclerotinia and White Mold Dataset (SWM), focusing on the pathogen's survival and reproductive structures, respectively, mature sclerotia and apothecia, and the symptomatic stage when white mold is visible. The dataset was explored by several traditional and state-of-the-art deep learning models, including SSD, Faster R-CNN, YOLO Family, DETR, and TransUNet, focusing on object recognition in challenging field conditions. These evaluated models were employed as baselines for ensemble strategies based on affirmative, consensus, and unanimous voting schemes. Among the individual object detection models, DETR achieved the highest F1-score (0.931). The consensus voting strategy further improved performance, reaching an F1-score of 0.953. The results demonstrated that the ensemble approach effectively addresses challenges posed by small object detection and class imbalance at the field level, offering a promising tool for precision agriculture and early intervention strategies targeting Sclerotinia sclerotiorum .