aimlin9/agrosense

AI-powered crop disease detection for smallholder farmers in Ghana. MobileNetV2 transfer learning, 97.3% validation accuracy on PlantVillage. # 🌱 AgroSense **AI-powered crop disease advisory for smallholder farmers in West Africa.** A farmer photographs a diseased crop. Within seconds the system identifies the disease using a fine-tuned MobileNetV2 model, generates farmer-friendly treatment advice via Gemini, and stores everything for follow-up. Built mobile-first, with an SMS fallback planned for farmers without smartphones. --- ## 📊 The Model Two-phase transfer learning on MobileNetV2, fine-tuned on PlantVillage (54,305 images, 38 disease classes, 14 crops). Trained on a free Kaggle P100 GPU. | Phase | What happened | Best val accuracy | |---|---|---| | **1** — frozen base | Train classification head only, lr = 1e-3, 10 epochs | **94.6%** | | **2** — fine-tune | Unfreeze top 30 layers, lr = 1e-5 (100× smaller), 10 epochs | **97.28%** | Final model: **2.7 MB TFLite file** with default quantization. Negligible accuracy loss vs the Keras original. ### Training curves The dashed vertical line marks the Phase 1 → Phase 2 transition. Notice the brief training-loss spike at epoch 11 (the "fine-tuning shock" as previously-frozen weights start receiving gradients), followed by recovery and a steady climb to 97.28%. ### Predictions on held-out validation images 12 out of 12 correct on a randomly sampled batch from the validation set. Most predictions at 99–100% confidence. The model expresses appropriate uncertainty on visually ambiguous cases — e.g., the Tomato Septoria leaf spot at 53.9% — which is exactly the behavior we want for safety-critical advice. --- ## 🏗️ Architecture ``` ┌─────────────────┐ ┌──────────────────────────────────────┐ │ Mobile / SMS │────▶│ FastAPI Backend │ │ (planned) │ │ │ └─────────────────┘ │ ┌─────────────────────────────────┐ │ │ │ POST /api/diagnose │ │ │ └────────┬────────────────────────┘ │ │ │ │ │ ▼ …