Disease detection in potatoes using deep learning with YOLOv12

Abstract

Potato is an essential food crop, but its production is threatened by diseases such as early and late blight, whose delayed detection leads to significant economic and environmental losses. Early and accurate disease detection in crops is crucial for ensuring food security and agricultural sustainability. In this work, we propose an innovative deep learning–based approach for the automatic identification of diseases in potato leaves (Solanum tuberosum) using the standard YOLOv12 architecture. The model was trained on accessible computational resources (2× NVIDIA T4 GPUs on Kaggle) using an initial dataset of 363 original images (121 per class: early blight, late blight, and healthy leaves), expanded to 920 images through a realistic data augmentation pipeline (including rotation, flipping, shear, noise, and exposure variation). Thanks to early stopping, training converged efficiently within 43 epochs, achieving outstanding performance: precision = 0,9854, mAP50 = 0,9950, and recall = 1,0000. These metrics demonstrate high sensitivity and specificity, even with limited data. In conclusion, the proposed system represents a viable, robust, and scalable solution for integration into precision agriculture applications, enabling timely diagnoses and reducing reliance on manual inspections.