International Journal of Science, Engineering and TechnologyAuthors: Harsh Rana, Swati, Priyanshu Vats, Aakash Kumar
Abstract: Agriculture sustains over 70% of Uttarakhand's hill-dwelling population, yet crop diseases cause estimated annual losses of ₹1,200 crore across the state (Directorate of Agriculture, Uttarakhand, 2022). Timely, accurate identification of plant diseases is critical to reducing these losses, but access to agronomists in remote hill districts such as Pithoragarh, Chamoli, and Uttarkashi remains severely limited. This paper proposes a custom Convolutional Neural Network (CNN) architecture optimised for crop disease detection under the specific agro-climatic, lighting, and device-availability conditions of Uttarakhand. The model was trained on a curated dataset of 15,850 annotated leaf images spanning six major crops — wheat, rice, tomato, potato, apple, and maize — covering 28 distinct disease classes. Images were sourced from the publicly available PlantVillage benchmark as well as original field photographs collected across five Uttarakhand districts in collaboration with local Krishi Vigyan Kendras (KVKs). The proposed architecture incorporates four convolutional blocks with Batch Normalisation, GlobalAveragePooling, and Dropout regularisation, yielding an overall classification accuracy of 96.7%, precision of 96.2%, recall of 95.9%, and F1-score of 96.0% on the held-out test set. These results outperform all evaluated baselines — SVM (78.3%), Random Forest (82.5%), VGG-16 (88.9%), ResNet-50 (91.4%), and MobileNetV2 (93.2%). The trained model was converted to TensorFlow Lite (TFLite) format and integrated into a prototype Hindi-English Android application named KrishiRakshak, which supports fully offline inference on low-end devices in under 1.5 seconds. A pilot field study with 50 farmers in Pauri Garhwal and Almora districts demonstrated an 82% correct disease identification rate using the application, compared to 47% through unaided visual inspection.
