Abstract: Coastal wetlands are important natural resources with significant ecological value. Effectively extracting wetland vegetation information from remote sensing images and improving the classification accuracy are considerable challenges. Based on the GEE platform, this study proposes a method that integrates sliding-window-based Jeffries-Matusita (JM) distance optimization with a Stacking ensemble learning algorithm. This method overcomes the limitation of traditional JM distance, which relies solely on a single static image, and introduces a fitting equation for the mean and standard deviation of vegetation NDVI into the JM formula, thereby constructing a “feature-temporal” dual-optimization dimension. It objectively identifies the feature window that exhibits the most significant phenological differences among vegetation types and inputs the selected optimal temporal features into the Stacking ensemble learning model for classification. Taking the coastal wetland in Lianyungang as a case study, the results show that: (1) the JM-based time optimization improves average overall accuracy by 0.606% compared to using all-year imagery; (2) by incorporating radar polarization features from Sentinel-1 and slope features from the SRTM digital elevation model, the overall average accuracy increases by 1.462%; (3) the Stacking ensemble learning algorithm based on sliding window JM distance optimization achieves the highest classification accuracy, with a maximum overall accuracy of 94.390%, an average overall accuracy of 90.828%, and an average Kappa coefficient of 88.350%. These results demonstrate that the proposed method, which combines multi-source data with JM distance optimization and Stacking ensemble learning, exhibits significant potential for coastal classifying wetland vegetation.