HSI Classification Method Based on U-Nets and GCN-CNN in the Background of Artificial Intelligence

Abstract

Hyperspectral images typically have large-sized features and fuse a large amount of spatial and spectral data, increasing the complexity of feature selection and effective mining. In addition, its dimensional redundancy and feature intersection further exacerbate the interpretability problem of the model, limiting the overall improvement of classification performance. Therefore, this paper proposes an HSI classification model built on U-Nets and a graph convolutional network. This model utilizes multi-scale superpixel segmentation to enhance the flexibility of spatial structure modeling and achieves synchronous extraction of spatial topological relationships between land features from multiple scales through a multi-scale graph convolution architecture. The experiment showed that the proposed model achieved an F1 value of 96.7% on the comprehensive datasets (Indian Pines, Pavia University, Salinas, and GRSS 2013), demonstrating good robustness and generalization ability. Regardless of whether under interference conditions, the average classification entropy and average mutual information between categories of the proposed model were significantly lower than those of comparative models. Under the condition of random loss of some bands, the average classification entropy and average mutual information value between categories of the research model were 0.28 and 0.79, and 0.31 and 0.77 under Gaussian noise interference. The research model has strong discriminative ability in hyperspectral image classification tasks and effectively deals with complex scenes such as noise interference and data loss.