A deep learning system for plant disease detection and progression tracking built on the PlantVillage dataset. Given a leaf image, the model identifies the disease, classifies its stage (healthy → early → mid → late), estimates days since infection, and computes a treatment urgency score.

What It Does

Disease Classification — identifies plant disease across 38 classes (e.g. Apple Scab, Tomato Late Blight, Potato Early Blight)
Stage Classification — maps each disease to one of 4 progression stages:
- Stage 0 — Healthy
- Stage 1 — Early
- Stage 2 — Mid-stage
- Stage 3 — Late-stage
Days Estimation — regression head estimates days since infection onset
Urgency Scoring — outputs a 0–10 urgency score with action recommendations (Low / Moderate / High / Critical)
Temporal Tracking — tracks disease progression across multiple images over time and plots stage & urgency curves

Evaluation Results

Task	Metric	Score
Disease Classification	Accuracy	82.82%
Disease Classification	Weighted F1	0.8242
Stage Classification	Accuracy	80.93%
Stage Classification	Weighted F1	0.7998
Days Estimation	MAE	2.86 days
Days Estimation	RMSE	4.16 days

Usage

python

from huggingface_hub import hf_hub_download
from tensorflow import keras
import numpy as np
import json

# Download model and metadata
model_path = hf_hub_download(
    repo_id="Sharmistha-catalyst/sick-greens-plant-disease",
    filename="final_progression_model.h5"
)
metadata_path = hf_hub_download(
    repo_id="Sharmistha-catalyst/sick-greens-plant-disease",
    filename="metadata.json"
)

# Load model and metadata
model = keras.models.load_model(model_path)
with open(metadata_path) as f:
    metadata = json.load(f)

# Preprocess and predict
image = np.expand_dims(your_image_array, axis=0)  # (1, 224, 224, 3)
image = keras.applications.mobilenet_v2.preprocess_input(image)

disease_pred, stage_pred, days_pred = model.predict(image)

Dataset

PlantVillage via TensorFlow Datasets (tfds.load('plant_village'))

38 disease/healthy classes across crops including Tomato, Potato, Apple, Grape, Corn, Peach, Pepper, Strawberry, and more
Capped at ~100–102 images per class for balanced training
Split: 70% train / 15% validation / 15% test

Model Architecture

Backbone: MobileNetV2 (pretrained on ImageNet, fine-tuned)
Heads:
- Disease classification head (38-class softmax)
- Stage classification head (4-class softmax)
- Days regression head (single neuron)
Class weights applied to handle imbalance

Tech Stack

Python, TensorFlow / Keras
TensorFlow Datasets
MobileNetV2
Pandas, NumPy
Matplotlib, Seaborn, Plotly
scikit-learn
OpenCV, Pillow

Author

Sharmistha-catalyst — Hugging Face · GitHub

Sharmistha-catalyst/sick-greens-plant-disease

🌿 Sick-greens