Can time-series Sentinel images be used to properly identify wetland plant communities?

Abstract

Understanding the distribution of wetland plant communities is critical to biodiversity conservation and wetland habitat sustainable management, especially for migratory birds. However, limited road accessibility and low spectral discriminability make the mapping of wetland plant communities inadequate for wetland health assessment, necessitating the improvement of classification methods. In this study, we proposed a random forest classifier that combined multi-source remote sensing features for wetland plant community classification and evaluated this method for the Momoge Ramsar wetland site (MRWS) in China. The major result of this work was that the phenological and time-series features based on Sentinel-2 images were the most valuable discriminators for wetland plant community classification in the MRWS. The SAR_sum extracted from Sentinel-1 images also had high importance for classification. Moreover, the spatial pattern of different wetland plant communities was revealed, and the resultant classification map had a high overall accuracy (91.3%) and Kappa coefficient (0.90). And the six important features prompted the classification accuracy to reach 84.8%. In 2020, the total coverage area of natural wetland plant communities in MRWS reached 628.5 km2 (42.1%), of which Carex meyeriana distributed the most widely, followed by Suaeda glauca, Phragmites australis, Typha orientalis, and Scirpus triquater. The findings of this study can provide scientific decision-making support for the protection and management of migratory birds and plants in the MRWS. The proposed method employed freely available Sentinel 1/2 satellite images and fully automated programs on Google Earth Engine, and has guiding significance for large-scale and long-time-series wetland classification.

Keywords

• wetland plant community
• google earth engine
• random forest classifier
• phenological feature
• sentinel time-series image