Fully Convolutional Network-Based Nonlocal-Dependent Learning for Hyperspectral Image Classification

Abstract

Deep convolutional neural networks play an important role in hyperspectral image (HSI) classification tasks through hierarchical learning. Recent work based on deep learning has made great progress in exploring contextual features, with more of these approaches focusing on nonlocal contextual information. Nevertheless, the contextual information obtained by these methods still has room for improvement as they only consider the semantic level. Moreover, they ignore the importance of contextual features in the spectral domain, an important component of contextual features, especially in HSIs. This article proposes a novel HSI classification method, the nonlocal-dependent learning fully convolutional network (FCN). The network fully focuses on HSI’s spatial and spectral nonlocal contextual features by combining a context-aware module and global convolutional long short-term memory neural network (ConvLSTM) learning from shallow (spatial level) to deep (semantic level) layers. Specifically, the proposed context-aware module perceives local joint features through local and surrounding learning and refines features under a global context through global learning. To further enhance the long-range dependencies of spectral and spatial dimensions at different phases, global ConvLSTM learning is proposed to obtain multiscale fused features from spatial to semantic. Experiments on datasets with different scenes reveal that the proposed method obtains better classification performance than state-of-the-art methods.

Topics & Concepts