Monitoring Spartina Alterniflora removal dynamics across coastal China using time series Sentinel-1 imagery

Abstract

Invasions by Spartina species have posed serious threats to coastal ecosystems worldwide. Since the introduction of Spartina alterniflora (S. alterniflora) in China in 1979, it has expanded across 68,000 ha of coastal wetlands by 2020. In 2022, the Chinese government issued the “Special Action Plan for the Prevention and Control of Spartina alterniflora (2022–2025)”, aiming for nationwide eradication by 2025. As local and provincial removal efforts progress, timely information on removal status and timing is crucial for tracking the project's progress, assessing the effectiveness of control measures, and facilitating in-depth research on S. alterniflora re-establishment mechanism. Frequent cloud cover hinders the application of optical satellite imagery in timely monitoring of S. alterniflora removal in coastal areas. The all-weather Sentinel-1 SAR sensor overcomes this limitation, offering frequent acquisitions suitable for accurate mapping of S. alterniflora removal. In this study, we present the SAR Time Series Change and change Time Detection (STS-CTD) model, a deep learning framework designed to detect S. alterniflora removal events using Sentinel-1 time series imagery, providing information on where and when S. alterniflora was removed. The model integrates Transformer encoder, multi-kernel Conv1D decoder, and Band Dropout training strategy to learn the abrupt changes in time series backscatters and radar indices caused by plant removal. We applied the model within the boundaries of national-scale S. alterniflora map and generated the first S. alterniflora removal maps across China's coastline during 2021 to 2023. Our key findings include: (1) The resultant S. alterniflora removal map achieved an overall accuracy (OA) of 97.95 % and an F1-score of 98.39 % for removal identification, and the removal timing estimation exhibited a Mean Absolute Error (MAE) of 6.77 days and a Root Mean Square Error (RMSE) of 14.68 days; (2) the multi-kernel Conv1D and Band Dropout strategy considerably improved the model performance compared to models using only Transformer encoder and conventional Dropout; (3) the STS-CTD model outperformed state-of-the-art time series analysis models, including Bi-LSTM, Bi-GRU, TCN, and InceptionTime; (4) the model demonstrated strong temporal transferability, showing promise for application in future years; (5) it effectively mitigated noise from SAR imaging and tidal inundation, though continuous inundation and incomplete removal reduced accuracy in certain areas. The STS-CTD model and the resulting national-scale maps offer an operational solution for assessing invasive species management in coastal wetlands.