Mapping hierarchical wetland characteristics by optical-SAR integration with collaborative spatial-spectral-temporal learning

The learning-based integration of optical and synthetic aperture radar (SAR) satellite imagery is known to be effective in promoting the accuracy of wetland land-cover classification. However, the distribution of wetland categories is characterized as spatially heterogeneous and highly dynamic. It remains a challenge to fuse the inherent characteristics of optical and SAR data by exploiting their discriminative feature representations for delineating wetland landscapes. To fully integrate the complementary information among optical and SAR data, a dual-branch deep network is proposed for mapping hierarchical wetland characteristics, which is referred to as HiWet-DBNet. Within the network, two parallel branches are designed to collaboratively learn the spatial, spectral or polarized, and temporal dependencies in the optical image and SAR image time series, respectively. Inspired by the relationships of deep and shallow features, the intra-layer features are fused across the branches to generate the multi-level wetland mapping results (i.e., general wetland land cover, and wetland vegetation types). The proposed method was tested on the Poyang Lake wetland in China using Sentinel-1 and Sentinel-2 imagery. The evaluation results show that the overall accuracy of HiWet-DBNet reaches 88.51% and 88.61% in the dry and wet seasons, which is superior to the other solutions with only a single data source or insufficient fusion of multi-modal features. For the challenging task of submerged vegetation detection, the producer’s accuracy of HiWet-DBNet is improved by 1.70% to 16.59% compared with the VBI algorithm and state-of-art deep learning-based wetland classification methods.