Multiview ConvLSTM Based on Autoencoder for Hyperspectral Sparse Unmixing

Abstract

Hyperspectral sparse unmixing is an crucial preprocessing technique. In recent years, deep learning-based methods have gained increasing attention in spectral unmixing, with particular emphasis on the performance of unsupervised autoencoder (AE). To fully exploit both spectral and spatial information in hyperspectral bands for unmixing, this study explores a framework for multi-view spectral and spatial information unmixing based on autoencoder (AE). We incorporate multi-view spectral information by utilizing spectral partitioning and introduce a ConvLSTM encoder that leverages recurrent neural networks (RNNs) to synergistically exploit multi-view spectral and spatial information for more effective unmixing. The experimental results on synthetic datasets confirm the superiority of the proposed method in achieving excellent unmixing performance.