Early Detection of Crop Diseases Using Deep Learning Models

Abstract

Crop diseases pose a serious threat to global food security by significantly reducing agricultural yield and economic stability. Although recent advances in deep learning have enabled automated crop disease detection with high accuracy, most existing systems focus primarily on leaf-based classification and fail to provide disease severity assessment, which is crucial for real-world agricultural decision-making. This paper presents a deep learning powered system for early detection of crop diseases that integrates HSV color-space segmentation with convolutional neural networks (CNNs) to detect diseases affecting both leaves and fruits while quantitatively estimating disease severity. The proposed system enhances lesion visibility through HSV-based pre-processing and isolates infected regions prior to classification using a transfer-learning-based EfficientNet-B4 architecture. Disease severity is estimated by calculating the percentage of infected tissue and categorizing it into mild (1–10%), moderate (11–25%), and severe (>25%) levels. Extensive data augmentation techniques are employed to improve robustness under varying illumination and real-field conditions. Experimental evaluation conducted on benchmark datasets and real-field images achieved classification accuracies of 99.18% for leaf diseases, 96.47% for fruit diseases, and 83.33% under field conditions. A web-based interface enables real-time disease diagnosis with visual lesion interpretation and severity reporting. The results demonstrate that the proposed system is accurate, interpretable, and suitable for deployment in precision agriculture for early disease management.