Integrating Deep Features with Superpixel-based Graph Attention Networks for Hyperspectral Image Classification

Hyperspectral image (HSI) classification benefits greatly from deep learning techniques such as Convolutional Neural Networks (CNNs) and Graph Neural Networks (GNNs), especially Graph Attention Networks (GATs). These methods have shown remarkable success in capturing detailed spectral and spatial information. However, a major challenge remains: effectively extracting rich and robust features from limited labeled samples, which is common in real-world HSI applications. In this paper, we introduce WFCG Transfer, a novel framework that combines the strengths of pretrained CNNs and Graph Attention Networks through a weighted feature fusion approach. By leveraging pretrained CNN architectures like ResNet, VGG, and EfficientNet, the model captures fine-grained spectral-spatial features at the pixel level. Meanwhile, a superpixel-based GAT branch models the higher-level contextual and topological relationships among regions in the image. These complementary features are adaptively integrated using a learnable weighted fusion mechanism, enabling the model to effectively balance local detail and global context. Extensive experiments conducted on three widely used real-world HSI datasets show that WFCG Transfer outperforms several current state-of-the-art classification methods, demonstrating better accuracy, robustness, and generalization. This makes the proposed framework a powerful and practical solution for HSI analysis in diverse remote sensing scenarios.