Annual mapping of Spartina alterniflora with deep learning and spectral-phenological features from 2017 to 2021 in the mainland of China

Abstract

ABSTRACT Spartina alterniflora (S. alterniflora) expanded continuously in the coastal zone of the mainland in China, which caused serious ecological problems. Currently, there are several studies on large-scale time-series mappings of S. alterniflora but the time interval of these mappings is over three years. Consequently, these studies fail to capture the rapid dynamics of S. alterniflora. Leveraging the temporal transferability of DeepLabv3+, this study annotated 2020 Sentinel-2 data as training data to obtain the optimal model. Subsequently, we applied this model to predict Sentinel-2 data for other years (2017, 2018, 2019 and 2021), respectively. Ultimately, we produced accurate time-series maps of S. alterniflora from 2017 to 2021 in mainland China (China mainland S. alterniflora, CMSA). Meanwhile, it also confirmed the high transferability of the DeepLabv3+ model in the time-series mapping of S. alterniflora. The obtained annual time-series maps have revealed a more detailed account of how S. alterniflora changes over time that previous studies had failed to capture. It has enriched our understanding of the dynamic changes in S. alterniflora. The mapping results revealed that S. alterniflora continued to grow along the mainland coast from 2017 to 2021 but at a slower growth rate. In some local regions, we discovered: (1) S. alterniflora populations can naturally shrink under certain conditions, such as in the Dandou Sea; (2) the area of S. alterniflora has increased slightly after decreasing for several years (e.g. Yueqing Bay). However, in some other regions, there is no observed regrowth after removal (e.g. Dongtan on Chongming Island and Chuandong Port).