Research on the extraction method of coastal wetlands based on sentinel-2 data

Wetlands play an important role in ecological health and sustainable development, and dynamic monitoring of their spatial distribution is crucial for developing management and conservation measures. The types of coastal wetlands are complex and diverse, natural and artificial wetlands are easily confused, making precise classification more difficult. The coastal wetland of Chongming Island in China, which has diverse types and unique and complex ecological and hydrological characteristics, was deliberately chosen as a challenging case study. The objective of this study was to research effective method of fine classification of coastal wetlands, by constructing feature variables and proposing strategies for multi-level selection and fusion of feature variables. Sentinel-2 data with rich spectral information and high spatial resolution was be used. In this study, firstly, the classification effect of characteristic variables such as vegetation index, water body index, red edge index, and texture index were evaluated. Focusing on the "different objects with same spectra" of the humid planning land and farm growing ponds, the spectral characteristics of them were analyzed and a "water-rich soil index (WRSI)" was established. Subsequently, correlation analysis and J-M distance method were used to multi-level selection for the feature variables and four sets of features combination schemes were established. Finally, random forest (RF) was applied to classify coastal wetlands using different feature combination schemes, and the accuracy of different schemes was compared and verified. The results show the following: 1)Texture features have a promoting effect on improving classification accuracy. The constructed "water rich soil index"(WRSI) has the effectively contribution to identification and classification of farm growing ponds and humid planned land, improving the overall classification accuracy by 6.52%. 2)By multi-level selecting and fusion of feature variable sets, both accuracy and efficiency for classification are improved. For different features combination schemes, the classification accuracy is up to 90.03% by integrating spectral features, spectral index, texture index, and WRSI. This study evaluates the potential of Sentinel-2 data in coastal wetland classification, constructs effective feature parameters, and provides a new idea for wetland information extraction. The resulting classification map can be used for sustainable management, ecological assessment and conservation of the coastal wetland.