README.md

metadata

language: en
license: mit
tags:
  - image-classification
  - plant-disease
  - agriculture
  - efficientnet
  - transfer-learning
  - maize
  - ghana
metrics:
  - accuracy
  - f1

Maize Disease & Pest Classifier — EfficientNet-B0

Part of the MSc thesis: "Application of Deep Learning for Low-Resource Maize Disease Diagnosis in Heterogeneous Agro-Ecological Zones"

Model details

Architecture	EfficientNet-B0
Training strategy	Two-phase transfer learning (ImageNet → maize diseases)
Test accuracy	90.94%
Weighted F1	0.9198 (validation)
Classes	7
Input size	224 × 224

Classes

0. Healthy
1. Northern Leaf Blight
2. Common Rust
3. Gray Leaf Spot
4. Ear Rot
5. Fall Armyworm
6. Stem Borer

Dataset

Combined from:

• PlantVillage (4,188 images, 4 classes) — Kaggle
• Ghana field images (406 images, 3 classes) — custom field-collected dataset

Total: 4,594 images | Class imbalance: 10.75:1

Usage

from huggingface_hub import hf_hub_download
import torch
from src.deployment.inference import MaizeClassifier
from PIL import Image

ckpt = hf_hub_download(repo_id="moro23/maize-disease-model",
                       filename="best_val_weighted_f1_0.919762_epoch_22.pth")
classifier = MaizeClassifier(ckpt, "configs/data_config.yaml")
image = Image.open("maize_leaf.jpg")
print(classifier.predict(image))

Deployment

Live demo: HuggingFace Spaces