A Spectral Sequence-Based Nonlocal Long Short-Term Memory Network for Hyperspectral Image Classification

Abstract

Efficient classification for hyperspectral image (HSI), which assigns each pixel of the image into a specific category, has been a critical research topic in the HSI analysis area. Under the supervised classification settings, the deep learning approaches are very useful for label prediction. However, most deep learning modeling methods cannot get the utmost out of spectral information, which is critically important for object interpretation. Consequently, a spectral sequence-based nonlocal long short-term memory (LSTM) network for HSI classification is proposed in this article. To boost the dominant role of spectral information, the LSTM network for sequential data processing is employed. Furthermore, the nonlocal diverse regions are exploited to learn contextual features for stronger discriminative ability. Meanwhile, the attention mechanism is adopted to increase the classification performance. Experiments on Salinas, Indian Pines, Pavia University, and Houston University datasets are implemented. Benefiting from the idea of processing spectral bands as sequence data and nonlocal diverse regions, spectral-spatial information is fully fused to achieve accurate pixel-level classification. It can be demonstrated that the proposed method can achieve better quantitative and qualitative classification results compared with the other state-of-the-art classification methods.