Tensorflow Grapevine Disease Detection

Description

This project develops a mobile application for detecting diseases on grapevines using a Deep Learning model. The implementation leverages TensorFlow and Keras to build a CNN-based classifier for identifying three common diseases: Black Rot, ESA (Net Blight), and Leaf Blight.

📁 Dataset

The dataset originates from Kaggle, containing 9,027 images of grapevine leaves. The diseases are categorized as:

Black Rot
ESCA (Net Blight)
Leaf Blight

The dataset is well-balanced with a slight overrepresentation of ESCA and Black Rot. All images are in .jpeg format with dimensions 256x256 pixels.

📊 Model Architecture Selection

We evaluated pre-trained models from keras applications to balance accuracy, model size, and inference speed. The selection criteria included:

Maximize accuracy
Minimize size (1/size)
Maximize CPU speed (1/CPU Time)

The score formula used for selection was:
Score = \frac{Accuracy}{Size . CPU Time}

Top Models (Score > 0.05):

MobileNetV2 (Smallest: 14 MB, High Accuracy: 77%)
MobileNet (Fastest: 22.6 ms)
NASNetMobile
EfficientNetB0

Conclusion:
MobileNetV2 was chosen for its optimal balance between accuracy, size, and speed.

🍇 Grapevine Diseases

Key Diseases:

Black Rot
ESCA (Net Blight)
Leaf Blight

🤖 Model Structure

Architecture:

# Auto Stop
early_stopping = EarlyStopping(monitor="val_loss", min_delta=0.2, patience=10)

# Model
model = Sequential()
model.add(tf.keras.applications.MobileNetV2(
    input_shape=(IMG_HEIGHT, IMG_WIDTH, CHANNELS),
    include_top=False, 
    weights='imagenet'
))
model.add(tf.keras.layers.GlobalAveragePooling2D())
model.add(tf.keras.layers.Dense(100, activation='relu'))
model.add(tf.keras.layers.Dense(100, activation='relu'))
model.add(tf.keras.layers.Dense(NUM_CLASSES, activation='softmax'))

optimizer = tf.keras.optimizers.Adam(learning_rate=LEARNING_RATE)

model.compile(
    optimizer=optimizer,
    loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
    metrics=['accuracy']
)

Parameters:

Total params: 7.12M (27.17 MB)
Trainable params: 2.36M (9.01 MB)
Non-trainable params: 34.11K (133.25 KB)

🛠️ Training Details

Batch Size: 32
Epochs: 100 (reduced to 25 via early stopping)
Data Augmentation: Not used (insufficient improvement in accuracy)
Normalization: Pixel values normalized to [0, 1]

📊 Results

Performance:

Validation Accuracy: ~99.9%
Confusion Matrix Analysis:
- Model biased toward ESCA and Healthy classes.
- Suspected causes:
  1. Original dataset imbalance
  2. Similar visual features across diseases