Mapping herbaceous wetlands using combined phenological and hydrological features from time-series Sentinel-1/2 imagery

Abstract

The herbaceous marshes are important habitats for many endangered waterfowls. Recognizing the distribution extents of the herbaceous marshes accurately is necessary to assess the biodiversity conservation strategies. However, the classification results of herbaceous marshes differ significantly using single date imagery because of the different vegetation phenological rhythm and hydrological condition. A deep learning algorithm (temporal convolutional neural network (TempCNN)) and the CVHIs dataset were used to map wetlands in the Zhalong National Nature Reserve in China. The CVHIs dataset was constructed based on the hydrological and phenological characteristics of different wetland vegetation types from time-series Sentinel-1 and Sentinel-2 images. The results showed the following. (1) The proposed method was stable and scalable and resulted in OAs of 92.69%, 89.18%, and 88.61% and kappa coefficients of 0.91, 0.87, and 0.86 in 2019, 2020, and 2021, respectively. (2) The crucial phenological periods to distinguish between herbaceous marshes and meadows were June, July, and August, and the optimal CVHIs corresponded to the phenological stages of the wetlands vegetation. (3) The optimal feature variables and its derivation time were selected from the CHVIs based on the TempCNN algorithm, which mitigated the impacts of seasonal variability of vegetation and hydrological conditions on the classification accuracies.