Self-Supervised Deep Learning Model for Plant Disease Classification Under Limited Labeled Agricultural Data

Abstract

Accurate plant disease classification is constrained by the limited availability of labeled agricultural images. This study presents a contrastive self-supervised framework that learns transferable leaf representations before supervised fine-tuning. The pretraining corpus contains 19,456 unique unlabeled PlantVillage images from apple, maize, grape, potato, and tomato; two independently augmented views of every image produce 38,912 contrastive inputs per epoch. A balanced labeled set of 6,240 images covering 14 crop species and 26 crop-condition classes was separated at source-image level into 4,992 training, 624 validation, and 624 test images. A ResNet-50 encoder and a 2048–512–128 projection head were optimized with the normalized temperature-scaled cross-entropy loss and then fine-tuned using a 26-way classifier. The proposed model correctly classified 602 of 624 internal test images, corresponding to 96.47% accuracy (95% Wilson confidence interval: 94.72%-97.66%), compared with 91.51% for the strongest identically trained baseline. On an independent 360-image PlantDoc field subset containing eight overlapping classes, it achieved 337 correct classifications (93.61%). The study therefore demonstrates improved visible-symptom disease classification under limited-label conditions; it does not claim pre-symptomatic disease forecasting.