An AI-Based Computer Vision Framework for Real-Time Crop Disease Diagnosis and Precision Treatment Recommendation

Abstract

Early and accurate diagnosis of crop diseases is essential for improving agricultural productivity, minimizing crop losses, and reducing unnecessary pesticide usage. Conventional manual disease inspection is labor-intensive, subjective, and unsuitable for large-scale monitoring. This paper presents an AIbased computer vision framework for real-time crop disease diagnosis and precision treatment recommendation. The proposed framework integrates transfer learning using MobileNetV2 with image preprocessing, disease severity estimation, explainable artificial intelligence, and an intelligent recommendation engine to provide comprehensive decision support for farmers. The proposed system classifies fifteen disease categories across tomato, potato, and bell pepper crops using the PlantVillage dataset. Data augmentation and transfer learning are employed to improve classification performance while maintaining computational efficiency. To enhance model interpretability, Gradientweighted Class Activation Mapping (Grad-CAM) is incorporated to visualize disease-discriminative regions used during prediction. Furthermore, disease severity is estimated using HSV-based image analysis, and a rule-based precision recommendation engine generates treatment strategies according to disease severity, crop growth stage, treatment preference, and affected cultivation area. To support edge deployment, the trained model is converted into TensorFlow Lite format and benchmarked for lightweight inference. Experimental evaluation on the PlantVillage dataset demonstrates a classification accuracy of 92.0%, with weighted precision, recall, and F1-score of 93.0%, 92.0%, and 92.0% respectively. Post-training dynamic range quantization further reduces model size by 4.76× (13.00 MB to 2.73 MB) and improves inference latency by 3.14× (117.30 ms to 37.38 ms), while incurring negligible loss in predictive performance. A Streamlitbased web application is developed to provide real-time diagnosis through image upload and live camera capture, to-gether with explainability visualizations and precision treatment recommendations. The proposed framework demonstrates the practical potential of integrating deep learning, computer vision, and agronomic knowledge to support intelligent and sustainable precision agriculture.