wisamWR/Tomato_Leaf_Diseases-CNN_Comparative_Analysis

Tomato Leaf Disease Classifier (EfficientNetV2B0)

Model klasifikasi gambar buat deteksi penyakit daun tomat. Base: EfficientNetV2B0 (transfer learning, ImageNet weights), head custom (GlobalAveragePooling → Dense 128 → Dropout 0.5 → Softmax 10 kelas). Dipilih sbg model terbaik dari perbandingan 14 arsitektur CNN (VGG, ResNet, DenseNet, MobileNet, EfficientNet, Xception, NASNet).

Model Details

• Architecture: EfficientNetV2B0 (frozen base, ImageNet pretrained) + custom classification head
• Input: RGB image, resize ke 224x224, preprocess pakai tf.keras.applications.efficientnet_v2.preprocess_input
• Output: 10 kelas (softmax)
• Framework: TensorFlow / Keras (.h5 format)
• Training: 10 epoch, batch size 64, Adam optimizer, mixed precision, augmentation (flip, brightness, contrast, saturation, rotation), oversampling kelas minoritas + class weights buat handle imbalance

Classes

Tomato_Bacterial_spot
Tomato_Early_blight
Tomato_Late_blight
Tomato_Leaf_Mold
Tomato_Septoria_leaf_spot
Tomato_Spider_mites_Two_spotted_spider_mite
Tomato_Target_Spot
Tomato_Tomato_Yellow_Leaf_Curl_Virus
Tomato_Tomato_mosaic_virus
Tomato_Healthy

Performance (validation set)

Metric	Score
Accuracy	97.38%
Precision	0.9702
Recall	0.9631
F1-Score	0.9664

Best di antara 14 model yang dibandingkan (VGG16/19, ResNet50/101/152/50V2, EfficientNet B0/B3/V2B0/V2B3/V2S, DenseNet121/169/201, Xception, NASNetMobile/Large, MobileNetV3 Small/Large).

Dataset

PlantVillage Tomato Leaf Dataset — gambar daun tomat terbagi ke 10 kelas (9 penyakit + sehat). Split stratified 80/20 train-validation, oversampling kelas minoritas di set training.

Usage

import tensorflow as tf
import numpy as np
from PIL import Image
from huggingface_hub import hf_hub_download

model_path = hf_hub_download(repo_id="<username>/<repo-name>", filename="EfficientNetV2B0_best.h5")
model = tf.keras.models.load_model(model_path)

CLASS_NAMES = [
    "Tomato_Bacterial_spot", "Tomato_Early_blight", "Tomato_Late_blight",
    "Tomato_Leaf_Mold", "Tomato_Septoria_leaf_spot",
    "Tomato_Spider_mites_Two_spotted_spider_mite", "Tomato_Target_Spot",
    "Tomato_Tomato_Yellow_Leaf_Curl_Virus", "Tomato_Tomato_mosaic_virus", "Tomato_Healthy",
]

image = Image.open("daun_tomat.jpg").convert("RGB").resize((224, 224))
arr = np.expand_dims(tf.keras.applications.efficientnet_v2.preprocess_input(np.array(image, dtype=np.float32)), axis=0)
preds = model.predict(arr)[0]
print(CLASS_NAMES[int(np.argmax(preds))], float(np.max(preds)))

Demo Streamlit ada di app.py.

Limitations

• Dilatih cuma 10 epoch dgn base model frozen — belum fine-tuned penuh.
• Performa bisa turun di gambar dgn kondisi pencahayaan/background beda dari dataset PlantVillage (yg umumnya background bersih studio).
• Bukan pengganti diagnosis ahli pertanian/agronomis.

Training Notebook

Detail eksperimen, EDA, dan perbandingan semua arsitektur ada di tomato-leaf-diseases-cnn-comparative-analysis.ipynb.