Hyperspectral Image Classification With Context-Aware Dynamic Graph Convolutional Network

Abstract

In hyperspectral image (HSI) classification, spatial context has demonstrated its significance in achieving promising performance. However, conventional spatial context-based methods simply assume that spatially neighboring pixels should correspond to the same land-cover class, so they often fail to correctly discover the contextual relations among pixels in complex situations, and thus leading to imperfect classification results on some irregular or inhomogeneous regions such as class boundaries. To address this deficiency, we develop a new HSI classification method based on the recently proposed Graph Convolutional Network (GCN), as it can flexibly encode the relations among arbitrarily structured non-Euclidean data. Different from traditional GCN, there are two novel strategies adopted by our method to further exploit the contextual relations for accurate HSI classification. First, since the receptive field of traditional GCN is often limited to fairly small neighborhood, we proposed to capture long range contextual relations in HSI by performing successive graph convolutions on a learned regioninduced graph which is transformed from the original 2D image grids. Second, we refine the graph edge weight and the connective relationships among image regions by learning the improved adjacency matrix and the 'edge filter', so that the graph can be gradually refined to adapt to the representations generated by each graph convolutional layer. Such updated graph will in turn result in accurate region representations, and vice versa. The experiments carried out on three real-world benchmark datasets demonstrate that the proposed method yields significant improvement in the classification performance when compared with some state-of-the-art approaches.

Knowledge and Geo-Object Based Graph Convolutional Network for Remote Sensing Semantic Segmentation

“…There are 16 land-cover categories involved in this scene. Similar to methods in [ 42 , 43 , 44 , 45 , 46 , 47 , 48 ], we remove 20 water absorption channels and noise channels and keep 200 channels.…”

Hyperspectral Image Classification with Localized Graph Convolutional Filtering

“…Although CNN has been successful on the domains with underlying grid-like structured data, the CNN-based methods suffer from several intrinsic drawbacks summarized by the previous studies [30,31], i.e., (1) only adapting to the regular squares regardless of the geometric changes in object regions, (2) difficulty in capturing the valuable information of class boundaries during convolving a punch of patches as the convolution kernels have fixed shape, size, and weights, (3) often take a longer training time to fit huge parameters, (4) incapable of modeling topological relations among samples whether local or nonlocal feature extraction. In this regard, the graph-based convolutional neural networks appear relatively promising to overcome the aforementioned defects and show excellent characteristics, i.e., (1) competently process the irregular image regions in the non-Euclidean (or non-grid) graph data structure, (2) multiple graph inputs can be dynamically updated and refined with multiscale neighborhood [30]. It is worth mentioning that graph representation learning represented by GCNs has received increasing attention in quantifying nonlinear features in irregular graphs converted from hyperspectral data.…”

Joint Correlation Alignment-Based Graph Neural Network for Domain Adaptation of Multitemporal Hyperspectral Remote Sensing Images

“…Wan et al [27] proposed a multiscale dynamic GNN, where the graph is dynamically updated during the training process. They further proposed a context-aware dynamic GNN that can capture the long-range contexture relations in hyperspectral data [36].…”

“…Due to the advantages of GNN [27][28][29], [34], [36], we applied GNN for feature extraction of multitemporal hyperspectral remote sensing images in this paper, which is able to effectively exploit spectral relational information in images. However, GNN can extract features but is incapable of reducing domain shift or conducting domain adaptation.…”