Deep neural network and transfer learning for annual wetland vegetation mapping using sentinel-2 time-series data in the heterogeneous lake floodplain environment

The lake-floodplain wetlands are characterized by high biodiversity, difficult access, and significant environmental changes. Traditional remote sensing mapping methods struggle to generate consistent time-series data on wetland vegetation communities. Current research has endeavoured to address this issue through the application of deep learning methodologies. However, a significant limitation of these models is their reliance on a substantial volume of training samples, which contradicts the difficulty and high cost of obtaining samples from the lake-floodplain wetlands. Whether it is possible to construct a transferable deep learning model under small sample conditions and apply it to the mapping of lake-floodplain wetlands is an urgent issue that needs to be addressed. To solve this problem, this study first constructed a deep neural network (DNN) designed specifically for mapping complex lake-floodplain wetland vegetation under conditions of limited sample size. Subsequently, using 2021 as a reference year, a novel histogram threshold method was proposed to identify the unchanged samples for the target transfer years of 2019, 2020, 2022, and 2023. Finally, annual wetland vegetation mapping was performed in Poyang Lake using DNN and sample transfer learning (STL). The results showed that high-quality annual time-series data of wetland vegetation can be generated using the constructed DNN and STL, with all overall accuracies exceeding 80%. The histogram threshold method, which combines SAD and NDVI indicators of key phenological period, can effectively solve the problem of difficulty in determining the unchanged samples in transfer learning for heterogeneous lake wetlands. Furthermore, the performance of STL based on the constructed DNN model was significantly superior to those based on support vector machine and random forest algorithms for mapping annual wetland vegetation communities using limited training samples. This study demonstrates that the effective application of DNN and STL will be highly beneficial for long-term monitoring of vegetation in lake-floodplain wetlands, particularly where sample availability is limited.