Fasal Doctor Case Study — Offline ML Crop Disease Scan

Project Metrics

• Tech Stack: Flutter, TensorFlow Lite, PyTorch (model training), Python.
• Model Accuracy: 92.4% diagnostic confidence on regional Punjab crops (Wheat, Cotton, Rice).
• Data Model: On-device SQLite advisory database compiled from PAU recommendations.

The Rural Agriculture Challenge

In remote farming communities across Muktsar Sahib, Bathinda, and Faridkot, cellular signals are frequently unstable or absent. When a farmer spots a fungal infection or pest damage on a wheat leaf, sending a high-resolution image to a remote cloud API server is impossible due to network timeouts and expensive mobile data charges. This delays vital diagnoses, allowing crop diseases like wheat yellow rust or cotton leaf blight to spread rapidly and destroy entire harvests. Fasal Doctor eliminates this issue by performing 100% of its computational diagnostics directly on the user's mobile device processor.

Model Quantization & Mobile Performance

To run a convolutional neural network on a smartphone without causing CPU overheating or excessive battery drain, I designed an optimized machine learning pipeline.

• INT8 Quantization Flow: I trained a custom MobileNetV2 model on the PlantVillage dataset supplemented with over 3,000 images of regional crop diseases. I then applied Post-Training Quantization (PTQ) to convert weights from 32-bit floats to 8-bit integers, compressing the model size from 54MB to only 11.2MB with less than 0.8% loss in accuracy.
• PAU Chemical Advisory Engine: Once a disease matches a classification confidence threshold of >85%, the app queries a local SQLite database populated with Punjab Agricultural University (PAU) advisory standards. The app displays the exact chemical formulas (e.g., Propiconazole 25 EC), volume calculations, and safety rules in Punjabi and English.

Multi-Threaded Flutter Isolates

Running image inference inside a mobile app requires careful resource management. If the camera stream is processed on the main UI thread, the application's interface will lag or freeze, ruining the user experience. To solve this, I wrote code that sends camera frames to a background Flutter Isolate. The background thread handles image scaling, normalization, and inference computation before passing the classification results back to the main UI.

Real-World Field Impact

By deploying this solution locally, farmers receive immediate biological data without spending money on continuous cellular data plans or cloud services. The key advantages of this on-device approach include:

• Zero Server Costs: Since all diagnostics run on-device, hosting costs do not scale with user volume.
• Sub-2 Second Latency: Inference operates in under 2 seconds, regardless of signal strength in deep fields.
• Empowered Farming: Access to accurate PAU advisories allows farmers to buy the correct pesticide dosage, reducing chemical runoff.

Offline Diagnostics Accuracy & Data Calibration

A major challenge in deploying agricultural computer vision models is the variance in lighting conditions. Sunlight angles, cloud shadows, and dust on camera lenses alter image properties, causing false diagnoses. To stabilize accuracy, the app runs local image processing filters before inference. The captured camera frame is dynamically adjusted for brightness, contrast, and histogram equalization, ensuring the leaf is clearly distinguished from the background soil before passing the matrix to the interpreter.