Deep Computer Vision and Transfer Learning Approach for Multi-Class Mango Leaf Disease Detection and Segmentation

Accurate and timely identification of mango leaf diseases is essential for minimizing yield losses and supporting sustainable agricultural practices. This study presents a deep computer vision–based framework for multi-class mango leaf disease detection and instance segmentation using transfer learning. The proposed approach employs the YOLOv12 architecture to jointly perform object detection and pixel-level segmentation across seven mango leaf classes, including both diseased and healthy samples. A curated dataset of 3,479 annotated images was augmented and partitioned for training, validation, and testing to enhance model generalization. Experimental results demonstrate rapid and stable convergence, with balanced optimization across classification, localization, and segmentation tasks. The model achieved near-perfect performance, reaching mAP@0.5 scores of 1.0 for bounding box detection and 0.99 for mask segmentation, alongside consistently high F1 scores and 100% recall across all disease categories. Qualitative evaluations on unseen images further confirm the model’s robustness under varying lighting conditions, orientations, and disease severity levels. These findings highlight the effectiveness of the proposed framework as a reliable and scalable solution for automated mango leaf disease detection, offering strong potential for real-world deployment in precision agriculture and intelligent crop monitoring systems.