Abstract

This document details a project that employs deep learning techniques, specifically Convolutional Neural Networks (CNNs), to accurately classify plant diseases from image data. The implementation compares a custom-built CNN against a transfer learning model using VGG16.

Introduction 🌾

The early and accurate detection of plant diseases is crucial for maintaining crop health and ensuring food security. This project addresses this challenge by leveraging deep learning to automate the identification of 38 different classes of plant diseases from images. By building and evaluating two distinct CNN architectures, the project demonstrates a robust methodology for creating effective classification models.

Methodology

Dataset

The project utilizes the "New Plant Diseases Dataset (Augmented)", a comprehensive collection of images for this task. The dataset is structured as follows:

• Training Set: 70,295 images
• Validation Set: 17,572 images
• Total Classes: 38 (including healthy plant categories)

Data Preprocessing and Augmentation

To prepare the data for the models, the following steps were taken using ImageDataGenerator from TensorFlow/Keras:

• Image Resizing: All images were resized to a uniform input size of 224x224 pixels.
• Normalization: Pixel values were scaled from the [0, 255] range to [0, 1] to aid in model convergence.
• Data Augmentation: To improve model generalization and prevent overfitting, the training dataset was augmented in real-time with the following transformations:
  • Rotation Range: 20 degrees
  • Width/Height Shift Range: 20%
  • Horizontal Flips

Model Architectures

Two distinct models were developed and compared:

1. Custom CNN: A sequential model was constructed from scratch, featuring:

   • Multiple Conv2D layers with relu activation and L2 regularization (0.001).
   • MaxPooling2D layers for down-sampling feature maps.
   • BatchNormalization to stabilize and accelerate training.
   • A Flatten layer followed by Dense layers for classification.
   • Dropout (0.5) to reduce overfitting.

2. VGG16 (Transfer Learning): A more advanced model leveraging the VGG16 architecture, pre-trained on ImageNet.

   • Base Model: The convolutional base of VGG16 was used as a fixed feature extractor (base_model.trainable = False).
   • Custom Head: A GlobalAveragePooling2D layer was added, followed by Dense layers with BatchNormalization and Dropout (0.4) to adapt the model to the specific plant disease classes.

Experiments and Implementation ⚙️

• Frameworks: The models were built and trained using TensorFlow with the Keras API.
• Optimizer: The Adam optimizer was used with a learning rate of 0.001.
• Loss Function: categorical_crossentropy was employed, suitable for multi-class classification.
• Training Epochs: The models were trained for 15 epochs.
• Callbacks:
  • EarlyStopping: Monitored validation loss and halted training after 10 epochs with no improvement, preventing overfitting.
  • ModelCheckpoint: Saved the best-performing model weights based on validation performance.

The training was configured to utilize available GPUs for accelerated computation.

Results and Comparison 📈

The performance of both models was evaluated on the validation set using accuracy, precision, and recall. The final metrics from the training history are summarized below:

The results indicate that the VGG16 transfer learning model outperformed the custom CNN across all metrics on both training and validation data, demonstrating the power of leveraging pre-trained weights for computer vision tasks.

Conclusion

This project successfully implements and validates deep learning models for plant disease classification. The superior performance of the VGG16 model highlights the effectiveness of transfer learning.

Future enhancements could include:

• Experimenting with other architectures like ResNet or EfficientNet for potentially even better performance.
• Fine-tuning hyperparameters such as learning rate and batch size.
• Deploying the trained model into a user-friendly web or mobile application for real-world use by farmers and botanists.

This work is licensed under the MIT License, encouraging further development and use by the community.