A Hyperspectral Remote Sensing Image Classification Method Based on Semi-supervised Learning

Abstract

Hyperspectral remote sensing technology is a technology that utilizes spectral imager on flying carriers such as satellites, airplanes, and drones to acquire information about the earth's surface. With the development of deep learning, the application of neural networks in hyperspectral image classification has attracted much attention. However, hyperspectral image classification faces problems such as high labeling sample costs and many redundant samples. Existing methods still have shortcomings in utilizing unlabeled samples and improving classification performance. To solve these problems, this study addresses the critical challenges in hyperspectral image classification, namely high labeling costs and spectral redundancy, by proposing a novel semi-supervised learning framework. The model integrates three key components: (1) a 3D autoencoder for joint spectral-spatial feature extraction, (2) an ECA attention mechanism for channel-wise feature enhancement through adaptive weight learning, and (3) a Siamese network with random sample pairing for supervised feature refinement. Unlike conventional approaches, our framework establishes a synergistic mechanism between unsupervised feature enhancement and supervised correction, connected through innovative skip connections. Experimental validation demonstrates superior performance, achieving 82.3% OA (κ=0.74) on PaviaU and 87.4% OA (κ=0.82) on Salinas datasets, significantly outperforming existing methods while reducing dependence on labeled samples. The results confirm the model's effectiveness in improving feature discriminability and classification accuracy for hyperspectral imagery.