Nepal's economy is still 25% agriculture. In the Terai — the flat southern belt that produces most of the country's rice, wheat, and lentils — crop disease can devastate entire communities. Farmers often don't recognize disease until it's too late, and agricultural extension officers are spread thin.
The Idea
KrishiSathi (meaning "Farm Friend") is a mobile app where a farmer photographs a leaf, and within seconds, gets a diagnosis and treatment recommendation — in Nepali, in simple language.
The Data Problem
Existing crop disease datasets (PlantVillage, etc.) are mostly American and European crops. Nepali rice varieties, local wheat strains, and regional growing conditions produce different visual symptoms. We needed our own data.
Data Collection: The Hard Part
We partnered with the Nepal Agricultural Research Council (NARC) and local Krishi Gyan Kendras. Over 6 months:
- Visited 45 farms across Chitwan, Nawalparasi, and Rupandehi
- Collected 10,000+ labeled photographs using a standardized smartphone protocol
- Had NARC pathologists verify labels for 15 common diseases across rice, wheat, maize, and lentils
- Augmented with synthetic variations to handle lighting and angle differences
Model Architecture
We built a lightweight CNN (MobileNetV3-based) that:
- Runs entirely on-device — no internet needed for inference
- Achieves 91% accuracy on our test set
- Processes an image in under 2 seconds on a budget Android phone
- Model size: just 12MB
The Treatment Layer
Diagnosis alone isn't useful if the farmer doesn't know what to do. We built a treatment database in collaboration with NARC:
- Organic/traditional remedies listed first
- Chemical treatments with locally available product names
- Nearest Krishi Samagri Pasal (agricultural supply shop) locations
- All content in Nepali with audio playback for low-literacy users
Impact So Far
KrishiSathi is being piloted in 12 VDCs in Chitwan district. Early feedback is emotional — farmers showing us caught-early cases that would have destroyed their harvest. We're working to expand to all 77 districts by 2027.
