animeshakr/plant-disease-efficientnetv2s

🌿 Plant Disease Detection — EfficientNetV2S

Automated plant disease classification trained on the PlantVillage benchmark (38 classes, 54,306 images).

Live demo: 🤗 HuggingFace Space
GitHub: Animesh-Kr/Plant-Disease-Prediction

---

Model Details

Property	Value
Architecture	EfficientNetV2S
Input resolution	384 × 384
Number of classes	38
Dataset	PlantVillage (Hughes & Salathé, 2015)
Test accuracy	99.57%
Macro F1	99.48%
Top-3 accuracy	99.98%
Parameters	~21M
Training framework	TensorFlow / Keras 3

---

Evaluation Results

Model	Test Accuracy	Macro F1	Weighted F1	Top-3 Accuracy	Mean Confidence
Baseline CNN (4-block, scratch)	91.23%	89.32%	91.28%	98.74%	78.54%
EfficientNetV2S (this model)	99.57%	99.48%	99.57%	99.98%	90.55%

Interactive Visualisations

Plot	Link
3D UMAP Embeddings	Open
3D Performance Surface	Open
3D Confusion Surface	Open

Statistical Significance

• McNemar's test: p = 3.27 × 10⁻¹⁸² — improvement is astronomically significant
• Expected Calibration Error (ECE): 0.0902 — moderately well calibrated
• MC Dropout uncertainty (30 passes): flagged images are ~17 percentage points less accurate than unflagged, confirming the uncertainty flag is meaningful

---

Training Pipeline

Data

• Dataset: PlantVillage color subset — 54,306 images, 38 disease classes
• Split: Family-aware 70/15/15 train/val/test using perceptual hash deduplication
• Deduplication: pHash + Union-Find clustering prevents near-duplicate leakage
• Augmentation: RandomFlip, RandomRotation(0.08), RandomZoom(0.12), RandomTranslation(0.08), RandomContrast(0.10)

Architecture

Input (384×384×3)
    ↓
Augmentation layers
    ↓
EfficientNetV2S backbone (ImageNet pretrained, include_preprocessing=True)
    ↓  ← top 40% unfrozen during fine-tune stage
GlobalAveragePooling2D
    ↓
Dropout(0.3) → Dense(512, swish) → Dropout(0.3)
    ↓
Dense(38, softmax)

Training Strategy

• Stage 1 (Warmup): Backbone frozen, head only, lr=2e-3, 8 epochs
• Stage 2 (Fine-tune): Top 40% backbone unfrozen, lr=2e-5, 15 epochs
• Loss: CategoricalCrossEntropy + label smoothing 0.1
• Optimizer: AdamW (weight decay 1e-4)
• Precision: Mixed float16
• Batch size: 64 @ 384×384 (A100 80 GB)
• Hardware: Google Colab Pro+ A100 80 GB, 167 GB RAM

---

Supported Classes (38 total)

Plant	Diseases
Apple	Apple scab, Black rot, Cedar apple rust, Healthy
Blueberry	Healthy
Cherry	Powdery mildew, Healthy
Corn	Cercospora leaf spot, Common rust, Northern leaf blight, Healthy
Grape	Black rot, Esca (Black Measles), Leaf blight, Healthy
Orange	Haunglongbing (Citrus greening)
Peach	Bacterial spot, Healthy
Pepper	Bacterial spot, Healthy
Potato	Early blight, Late blight, Healthy
Raspberry	Healthy
Soybean	Healthy
Squash	Powdery mildew
Strawberry	Leaf scorch, Healthy
Tomato	Bacterial spot, Early blight, Late blight, Leaf mold, Septoria leaf spot, Spider mites, Target spot, Yellow leaf curl virus, Mosaic virus, Healthy

---

Usage

Option A — TFLite float16 (recommended, fast)

import tensorflow as tf
import numpy as np
import json
from PIL import Image

# Load model
interp = tf.lite.Interpreter(model_path="model_float16_quant.tflite")
interp.allocate_tensors()
ind  = interp.get_input_details()[0]
outd = interp.get_output_details()[0]

# Load labels
with open("class_indices.json") as f:
    class_map = json.load(f)   # {"0": "Apple___Apple_scab", ...}

def predict(img_path: str, top_k: int = 5):
    img = Image.open(img_path).convert("RGB").resize((384, 384))
    arr = np.expand_dims(np.array(img, dtype=np.float32), axis=0)
    interp.set_tensor(ind["index"], arr)
    interp.invoke()
    probs = interp.get_tensor(outd["index"])[0]
    top   = np.argsort(probs)[-top_k:][::-1]
    return [(class_map[str(i)], float(probs[i])) for i in top]

results = predict("leaf.jpg")
for cls, conf in results:
    print(f"{cls}: {conf*100:.1f}%")

Option B — Full Keras model

import keras
import numpy as np
import json
from PIL import Image

model     = keras.models.load_model("best_model.keras")
class_map = json.load(open("class_indices.json"))

def predict(img_path: str, top_k: int = 5):
    img  = Image.open(img_path).convert("RGB").resize((384, 384))
    arr  = np.expand_dims(np.array(img, dtype=np.float32), axis=0)
    prob = model.predict(arr, verbose=0)[0]
    top  = np.argsort(prob)[-top_k:][::-1]
    return [(class_map[str(i)], float(prob[i])) for i in top]

---

Files

File	Size	Description
model_float16_quant.tflite	~45 MB	TFLite float16 — for deployment
class_indices.json	2 KB	Label mapping {"0": "Apple___Apple_scab", ...}

---

Limitations

• Trained on controlled laboratory photographs from PlantVillage only
• Generalisation to field photographs with occlusion, variable lighting, or soil contamination is not validated
• Not intended for commercial crop management decisions

---

Citation

@misc{kumar2025plantdisease,
  author    = {Animesh Kumar},
  title     = {Plant Disease Detection using EfficientNetV2S},
  year      = {2025},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/animeshakr/plant-disease-efficientnetv2s}
}

References

• Hughes & Salathé (2015). An open access repository of images on plant health. arXiv:1511.08060
• Tan & Le (2021). EfficientNetV2: Smaller Models and Faster Training. ICML 2021
• Gal & Ghahramani (2016). Dropout as a Bayesian Approximation. ICML 2016
• Dietterich (1998). Approximate Statistical Tests for Comparing Supervised Classification Learning Algorithms. Neural Computation

---

Developed by Animesh Kumar