A Transfer Learning Based CNN-GAT Framework for Hyperspectral Image Classification

Overview

This project implements a novel deep learning framework that combines Convolutional Neural Networks (CNN) with Graph Attention Networks (GAT) for effective hyperspectral image classification. The framework leverages transfer learning from pretrained CNN models to extract rich spectral-spatial features and applies graph attention to capture complex contextual relationships within the hyperspectral data. It provides an end-to-end pipeline for training, evaluation, and experimentation on benchmark hyperspectral datasets.

Key Features

• Integration of transfer learning with state-of-the-art CNN architectures for feature extraction.
• Utilization of Graph Attention Networks (GAT) to model spatial dependencies and improve classification accuracy.
• Modular and extensible code structure to easily switch CNN backbones and adjust GAT configurations.
• Training and evaluation scripts with configurable parameters.
• Pretrained models and experiment results available in the repository.

Project Structure

• MyResult/: Trained model weights and results.
• datasets/: Hyperspectral datasets for training/testing.
• models/: Implementation of CNN, GAT, and combined CNN-GAT architectures.
• scripts/: Training, evaluation, and utility scripts.

Requirements

• Python 3.7 or newer
• PyTorch
• NumPy
• Scikit-learn
• Other dependencies as specified in requirements.txt