crop-disease-detection

Crop Disease Detection

Description

End-to-end crop disease detection skill supporting classification (which disease), detection (where is the disease), and segmentation (disease region) at three granularity levels, with multiple SOTA models and training strategies built in.

When to Use

• User wants to do plant/crop disease detection, classification, or recognition research
• Need to train or fine-tune disease detection models
• Need to compare different model performances on disease detection tasks
• Need to implement cross-domain (lab to field) disease detection transfer
• Need to perform zero-shot or few-shot disease identification with novel pathogens
• Need to deploy lightweight disease detection on edge devices

Prerequisites

pip install torch torchvision ultralytics transformers segment-anything timm open-clip-torch albumentations

Supported Datasets

Dataset	Task	Classes	Images	Environment
PlantVillage	Classification	38	54,305	Controlled lab
PlantDoc	Detection, classification	27	2,596	Real field

Quick Start

from agri_cv_research.models import YOLODetector, ViTClassifier
from agri_cv_research.datasets import PlantVillageDataset

dataset = PlantVillageDataset(root="./data/plantvillage", split="train", download=True)

# Classification with Vision Transformer
model = ViTClassifier(backbone="vit_base_patch16_224", num_classes=38)
model.train(dataset, epochs=50, batch_size=32)

# Evaluate on test set
metrics = model.evaluate(dataset, split="test")
print(f"Accuracy: {metrics['accuracy']:.4f}")

# Predict on a single image
prediction = model.predict(test_image)
print(f"Predicted class: {prediction['class']}")

Code Examples

Multi-Model Comparison Experiment

Compare YOLODetector, ViTClassifier, CLIPFewShot, and SAMSegmentor on the same dataset:

from agri_cv_research.models import (
    YOLODetector,
    ViTClassifier,
    CLIPFewShot,
    SAMSegmentor,
)
from agri_cv_research.datasets import PlantVillageDataset

train_ds = PlantVillageDataset(root="./data/plantvillage", split="train", download=True)
test_ds = PlantVillageDataset(root="./data/plantvillage", split="test", download=True)

models = {
    "YOLOv8-m (detect)": YOLODetector(model_size="m", task="detect", num_classes=38),
    "ViT-B/16": ViTClassifier(backbone="vit_base_patch16_224", num_classes=38),
    "CLIP (16-shot)": CLIPFewShot(backbone="ViT-L/14", num_classes=38, shots=16),
    "EfficientNet-B4": ViTClassifier(backbone="efficientnet_b4", num_classes=38),
}

results = {}
for name, model in models.items():
    print(f"\n--- Training {name} ---")
    model.train(train_ds, epochs=30, batch_size=32)
    results[name] = model.evaluate(test_ds)
    print(f"{name} → Accuracy: {results[name]['accuracy']:.4f}")

# Print comparison table
print("\n=== Model Comparison ===")
for name, metrics in results.items():
    print(f"{name:25s} | Acc: {metrics['accuracy']:.4f}")