LAUSR

Wetland is one of the most productive resources on earth, and it provides vital habitats for several unique species of flora and fauna. Over the last decade, mapping and monitoring wetlands by utilizing deep learning (DL) models and remote sensing data gained popularity due to the importance of wetland preservation. In general, DL-based methods have shown astonishing achievement in wetland classification, but some practical issues, such as limited training samples, still need to be addressed. Moreover, the performance of most of the DL approaches is decreased when moderate-resolution images with few features are used as input data. One solution to breaking the performance bottleneck of a single model is to fuse two or more of them. To this end, we strive to investigate and develop a multimodel DL algorithm for wetland classification based on the combination of a graph convolutional network (GCN) and a shallow convolutional neural network (CNN), which is called the Wet-GC algorithm hereinafter. In doing this, moderate-resolution Sentinel-1 (S1) synthetic aperture radar (SAR) and Sentinel-2 (S2) multispectral optical imagery are fed into the GCN and CNN models, respectively. As we know from the literature, the synergistic use of S1 SAR and S2 optical imagery can be used to extract different types of wetland features and increase the class discrimination possibility. Hence, wetland mapping by jointly using GCN and CNN has the ability to boost the wetland classification task. Findings indicate that the efficiency of Wet-GC with an overall accuracy (OA) of 88.68% outperforms the results obtained from random forest (OA = 84.88%), support vector machine (OA = 82.86%), extreme gradient boosting (OA = 86.55%), and ResNet50 (OA = 86.93%). The outcomes reveal that the Wet-GC architecture proposed in this article has an excellent capability to be applied over large areas with minimal need for training samples and can perform acceptably in supporting regional wetland mapping.