Abstract: [Purposes]Coastal wetlands play a crucial role in maintaining coastal stability,and accurate monitoring of these ecosystems is of great importance.Synthetic aperture radar (SAR)imagery,with its all-weather and time-series acquisition capability,provides strong support for wetland monitoring in both temporal and spatial dimensions.However,feature redundancy limit the performance of classification models.[Methods]To address this issue,this study uses 30 Sentinel-1 SAR images acquired over the Yellow River Delta wetlands in 2020 as the data source.An improved minimum redundancy maximum relevance with cross-validation feature selection strategy is proposed to identify the optimal feature subset for machine-learning classifiers,and Yellow River Delta wetlands maps from 2020 to 2024 are generated.[Findings]The results show that the wetland classification accuracy of the proposed method under extreme gradient boosting reached 89.66%,outperforming feature selection methods such as random forest,recursive feature elimination,and principal component analysis.The Spartina alterniflora in the Yellow River Delta was extensively removed in 2023 and completely eradicated by 2024.Under the new pattern,tidal flats have been continuously transformed into native salt marsh vegetation such as Phragmites australis and Suaeda salsa.[Conclusions]The proposed method provides a new technical reference for wetland monitoring using SAR imagery．

Key words: synthetic aperture radar, coastal wetland classification, improved minimum redundancy maximum relevance algorithm, machine learning, feature selection