Gopina/maize-disease-classifier

🌿 Crop Disease Classifier — T2.1

Edge-AI crop disease detection for offline diagnostics in rural Africa.
Trained for the AIMS KTT Hackathon · Tier 2 · Challenge 1.

Model Description

A fine-tuned MobileNetV3-Small backbone that classifies maize, cassava, and bean leaf images into 5 classes. Exported to ONNX (FP32 + INT8) for CPU-only edge deployment — no GPU, no internet required at inference.

File	Size	Format
model.onnx	3.78 MB	ONNX FP32
model_int8.onnx	~1.2 MB	ONNX INT8 (static quantization)

Classes

Index	Label	Disease
0	healthy	No disease
1	maize_rust	Puccinia sorghi — common maize rust
2	maize_blight	Exserohilum turcicum — northern leaf blight
3	cassava_mosaic	Cassava Mosaic Virus (CMV)
4	bean_spot	Phaeoisariopsis griseola — angular leaf spot

Performance

Split	Macro F1
Clean test set	1.00
Field robustness (blur + JPEG + brightness jitter)	1.00
Robustness drop	0.00 pp (target: < 12 pp) ✅

Usage

import cv2, numpy as np, onnxruntime as ort

session    = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])
input_name = session.get_inputs()[0].name
CLASSES    = ["healthy", "maize_rust", "maize_blight", "cassava_mosaic", "bean_spot"]

img = cv2.imread("leaf.jpg")
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
img = cv2.resize(img, (224, 224)).astype(np.float32)[np.newaxis] / 255.0

probs = session.run(None, {input_name: img})[0][0]
print(CLASSES[probs.argmax()], f"{probs.max():.2%}")

Training

• Backbone: MobileNetV3-Small (ImageNet pretrained)
• Phase 1: Head training — 20 epochs, lr=1e-3
• Phase 2: Fine-tuning top 20 layers — 10 epochs, lr=1e-4
• Data: 300 images/class from PlantVillage + Cassava Leaf Disease (TFDS)
• Augmentation: RandomFlip, RandomRotation, RandomZoom, RandomBrightness, RandomContrast

Intended Use

Designed for deployment via Village Agent relay networks in Rwanda and DRC. Supports offline diagnosis with USSD/SMS fallback for feature-phone users. See the GitHub repo for the full system design including USSD templates in Kinyarwanda + French.

Limitations

• bean_spot class uses augmented synthetic data (no real Bean Angular Leaf Spot images found in TFDS) — real-world performance on this class may be lower.
• Trained on 224×224 lab-condition images; field performance depends on photo quality.
• Not a substitute for agronomist advice in ambiguous cases.