Machine Learning Classification for Plant Pathology in Nepalese Agriculture

Deployment of Convolutional Neural Networks on edge devices to identify foliar blights in tomato and potato crops in Kavre valley.
Experimental Methodology & Laboratory Protocol
Data collection of 4,500 annotated leaf images from smallholder farmers, followed by MobileNetV3 transfer learning and quantization for offline Android application inference.
Key Discoveries & Research Outcomes
- Achieved 94.2% top-1 classification accuracy across 6 major crop disease classes.
- Inference time under 65ms on low-cost quad-core ARM devices.
- Field-tested in 3 agricultural cooperatives with positive farmer feedback.
Full Publication Analysis
Foliar fungal pathogens such as early blight (Alternaria solani) and late blight (Phytophthora infestans) cause up to 40% yield losses in vegetable crops across the mid-hills of Nepal. Traditional diagnosis requires trained agronomists, who are scarce in rural agrarian wards.
Model Architecture & Quantization
We trained an optimized MobileNetV3-Small backbone using PyTorch, augmented with heavy color jitter and blur simulations to account for varying outdoor sunlight in hillside terraces. The model was converted to INT8 precision via TensorFlow Lite quantization.
Deployment on Edge Devices
The companion Android application runs offline inference locally on the smartphone, generating immediate pesticide dosage recommendations in Nepali language.
Er. Subash Karki
Faculty of Information Technology
Er. Subash Karki, Bibek Gautam, Prajwal Sharma
https://doi.org/10.1109/AGRITECH.2025.99281