WHAT & WHY
WHAT IT IS
An AI-powered computer vision diagnostic tool that scans crop leaf photos to identify fungal, bacterial, and viral plant pathogens with 96.4% accuracy.
TARGET AUDIENCE
Agronomists, agricultural researchers, and farmers requiring instantaneous field crop diagnostic analysis.
PROBLEM SOLVED
Eliminates multi-day lab testing delays by providing instant field diagnostics and treatment recommendations.
PAIN POINT & CONTEXT
Crop leaf pathogens spread rapidly, causing severe agricultural yield losses when diagnostic lab reports take days to process.
PROJECT GOALS
Train a deep CNN vision model with >95% classification accuracy across 38 pathogen classes.
Build a FastAPI inference server responding in sub-200ms.
Create an intuitive web scanner interface usable on field mobile devices.
USER INSIGHTS
INSPIRATION & BENCHMARKS
PlantVillage dataset research and precision agriculture automation.
KEY USER INSIGHTS
Field agronomists require offline-capable image capture and high-confidence heatmaps showing leaf damage regions.
Trained a ResNet-50 Convolutional Neural Network fine-tuned on 54,000 leaf images, wrapped in FastAPI with a Next.js 15 UI.
INSTANT INFERENCE SCANNER
Drag-and-drop or camera snap diagnostic scanner with sub-200ms response.
GRAD-CAM HEATMAP VISUALIZATION
Highlights exact visual pixel regions driving the AI model's diagnosis.
TREATMENT RECOMMENDATION ENGINE
Provides actionable chemical and organic treatment steps based on pathogen severity.
DATA PIPELINE & FLOW
Next.js frontend sends image payload to Python FastAPI backend running PyTorch model inference with Grad-CAM heatmap generation.
BUILD TIMELINE
Curated 54,000+ augmented leaf images across 38 disease categories.
Trained ResNet-50 CNN model achieving 96.4% validation accuracy.
Engineered sub-200ms inference server and visual heatmap generator.
Built mobile-responsive Next.js frontend with live camera scanner.
CHALLENGES & SOLUTIONS
Overfitting on specific field lighting conditions.
Applied aggressive Albumentations image transforms (color jitter, random shadows, rotation).
Slow GPU inference cold starts on cloud API instances.
Containerized model runtime with ONNX Runtime CPU/GPU execution provider.
SYSTEM OUTCOMES
Achieved 96.4% test set classification accuracy across 38 pathogen categories.
Sub-180ms total inference response time.
Successfully evaluated on over 1,500 real field crop samples.
LEARNINGS & ROADMAP
KEY LEARNINGS
Deep learning models in production require visual interpretability (Grad-CAM) to gain domain expert trust.
FUTURE ROADMAP (V2.0)
- Edge deployment via ONNX Mobile for 100% offline field scanning.
- Drone telemetry integration for large-scale aerial crop scanning.
- Multi-spectral imagery support.
