A Hyperspectral Classification Method Based on Deep Learning and Dimension Reduction for Ground Environmental Monitoring

Hyperspectral remote sensing images exhibit high dimensionality, a large volume of data, and significant redundant information. Before using deep learning methods for ground monitoring and classification, dimension reduction is often necessary. In response to the limitations of traditional principal component analysis in achieving comprehensive feature extraction, which may impact classification accuracy and computational efficiency, a combined classification method for hyperspectral remote sensing data is proposed, which uses t-Distributed Stochastic Neighbor Embedding (t-SNE) dimensionality reduction algorithm instead of traditional Principal Component Analysis (PCA) algorithm, and combines it with a deep learning network model Hybrid Spectral Network (HybridSN), to achieve accurate classification of ground cover and roof materials. Simultaneously, to further validate the dimensionality reduction effect of the t-SNE algorithm and its impact on the deep network model, for the same dataset, another deep learning network Deep Feature Fusion Network (DFFN) was set as the experimental control group. The dataset used in this article is the publicly available aerial hyperspectral remote sensing dataset University of Pavia dataset (UP), and the main process is as follows: Firstly, for the UP dataset, PCA and t-SNE algorithms are used for data dimension reduction. Subsequently, these two datasets after dimension reduction are input into HybridSN and DFFN deep learning models for classification, respectively. Finally, the accuracy of the classification results is assessed, and based on this evaluation, the effectiveness of different dimensionality reduction algorithms is compared and analyzed. The experimental results demonstrate that the HybridSN and DFFN models combined with t-SNE dimension reduction, are capable of effectively extracting hybrid species of ground objects, while pre-serving clear edge information in hyperspectral remote sensing image classification, achieving ground environment monitoring. They also exhibit superior performance in terms of overall accuracy (OA), average accuracy (AA), and Kappa coefficient.