Regression-based dynamic graph convolutional network for intelligent monitoring of coastal wetlands

Abstract

Multispectral images (MSIs) with representative spectral bands and superior spatial resolution are usually considered for fine classification of wetlands; however, traditional MSI feature extraction and classification models may be suboptimal to identify differences between similar ground objects. To come up with a solution for the above difficulties, an advanced approach called a regression-based dynamic graph convolutional network (RDGCN) is proposed to complete the fine classification of wetlands. First, a dynamic GCN branch is designed to obtain the deep mutual information between samples. Second, the spatial and spectral features are fully abstracted by the dual-tunnel convolutional neural network branch. Finally, an improved standard linear regression loss is added to the graph convolution output layer to map the samples to the sparse structure representation and improve the sparsity of the output. Two improved versions of time-saving based on RDGCN are also discussed. Experiments with Sentinel 2A data in the Jiaozhou Bay and Yellow River Delta are conducted to verify the classification and mapping accuracy. The ground truth maps of the two areas are marked according to the categories of land cover after on-site investigation. The results achieved by RDGCN are superior to those of the state-of-the-art approaches.