Computer Vision and Deep Learning-Based Universal Plant Leaf Disease Classification, Automated Necrotic Segmentation, and Agronomy Remedy Diagnostic System

Abstract: Plant pathogens are responsible for an estimated 20 to 40 percent loss of global crop yield every year, and smallholder farmers rarely have affordable access to expert agronomists at the critical early-onset stage of infection. This paper presents AgroVision AI, a universal computer-vision and deep-learning diagnostic pipeline that unifies classical image processing with a lightweight convolutional neural network to detect, localize, and classify foliar disease across multiple crop species from a single uploaded leaf photograph, regardless of the background clutter, illumination, or camera angle under which it was captured. The system first performs foreground leaf segmentation and shadow-artifact suppression, then extracts a bank of vitality and pathology descriptors — including the Excess Green Index, Chlorophyll Vitality Fraction, Localized Spot Ratio, Rust Pustule Ratio, and Desiccated Dead-Tissue Fraction — before passing the normalized leaf crop through a fine-tuned MobileNetV2 backbone for eleven-class disease classification. A hybrid decision-rule engine reconciles the classical descriptors with the network's softmax confidence to reject false positives arising from natural leaf-tip browning, vein shadows, and specular glare, and an integrated agronomy remedy knowledge base then returns organic and chemical treatment plans, dosage guidance, and preventive-care instructions tailored to the diagnosed condition. Evaluated on a held-out set of 2,500 field and greenhouse images spanning tomato, potato, maize, banana, and citrus leaves, AgroVision AI achieved 97.4% overall classification accuracy, 97.1% precision, 96.8% recall, and a mean end-to-end inference latency of 117 milliseconds on commodity hardware, outperforming standalone VGG- 16 and ResNet-50 baselines while requiring a fraction of their memory footprint. The results demonstrate that a hybrid classical-plus-deep-learning pipeline can deliver agronomist-grade diagnostic reliability in a form factor light enough to run on a farmer's smartphone.

Keywords: Computer Vision, Deep Learning, Plant Disease Classification, MobileNetV2, Leaf Segmentation, Precision Agriculture, Agronomy Diagnostics, Convolutional Neural Networks.