Farm Guard
Crop Disease Detector
The problem Farm Guard solves
Crop diseases cause ₹50,000 crores in annual losses, yet rural farmers lack the internet connectivity and expertise for timely diagnosis. FarmGuard AI solves this "last mile" problem as an offline-first, multilingual Edge AI system that runs entirely in the browser using TensorFlow.js. Unlike cloud solutions, it requires zero internet, ensuring real-time (<200ms) detection and complete data privacy. Farmers simply point their camera to receive instant, voice-guided diagnosis and treatment plans in their local language, democratizing access to expert agricultural care.
Challenges we ran into
** Challenge 1: The "Glare" False Negativ**e
Problem: Shiny leaves (such as Cherry leaves with Powdery Mildew) reflected sunlight, causing the model to interpret white spots as "glare" rather than "disease," leading to false "Healthy" predictions.
**Solution: **We implemented a Confidence Threshold. If the model is unsure (below 60% confidence), the app prompts the user to "Move Closer" or "Adjust Lighting" rather than providing a potentially incorrect guess.
Challenge 2: Model Normalization Mismatch
Problem: The app initially provided random predictions. We discovered the model expected input values between [0, 1], but the application was sending values normalized to [-1, 1].
**Solution: **We debugged the tensor preprocessing pipeline and aligned the normalization logic strictly with the training phase parameters.
Challenge 3: Large Model Size
Problem: The full PlantVillage model exceeded 200MB, which is too large for rural mobile data connections.
Solution: We utilized MobileNetV2 (a lightweight architecture) and applied Quantization to reduce the model size to under 5MB without significant loss in accuracy.
Tracks Applied (1)