Abstract: Aiming at the problems of low precision and algorithm robustness in the fine classification of wetland vegetation in large river-connected lakes, based on remote sensing cloud platform GEE and Sentinel-2 images, this paper studied the optimization scheme of vegetation classification in Poyang Lake wetland by different training sample quantity, simultaneous phase characteristics data and machine learning classification algorithms.The results showed that:(1)With the increase of the training samples number, the classification accuracy of vegetation types increased first and then stabilized.When the number of training samples of different vegetation types reached 550,the classification accuracy reached the peak stable state.(2) The classification accuracy of data sets with different phase characteristics was significantly different, specifically, monthly time series data set > dry season data set > four seasons data set > single time phase.The overall accuracy of monthly time series data set was the highest, and the overall accuracy and kappa coefficient were 82% and 0.79,respectively.(3) Different remote sensing classification algorithms could obtain different accuracy of classification results.RF classification accuracy was the highest, followed by SVM and CART.(4) When the number of training samples of different vegetation types reached 550,the Sentinel-2 time sequence image and RF algorithm could be used to obtain the best classification results.This study can be a reference for the fine classification of Poyang Lake wetland and provide technical support for its protection.