Salt marshes are one of the world's most valuable and vulnerable ecosystems. The accurate and timely monitoring of the distribution and composition of salt marsh vegetation is crucial. With the increasing number of archived multi-source images, the time-series remote sensing approach could play an important role in monitoring coastal environments. However, effective construction and application of the time series over coastal areas remains challenging because satellite observations are severely affected by cloud weather. Here, we constructed a pixel-level time series by intercalibrating the Landsat images from different sensors. Based on the time series, the XGBoost algorithm was introduced for salt marsh vegetation classification. The feasibility and stability for the classification using the pixel-level time-series and XGBoost algorithm (PTSXGB) were evaluated. Five types of salt marsh vegetation from the 3 sites in the Yangtze River Delta, China, were classified. The results demonstrated that (a) the intercalibration for the Landsat images from different sensors is necessary for increasing the number of available observations and reducing the differences among spectral reflectances. (b) The salt marsh vegetation classification using PTSXGB achieved a favorable performance, with an overall accuracy of 81.37 ± 2.66%. The classification was especially excellent for the widespread Spartina alterniflora and Scirpus mariqueter . (c) Compared with the classifications using single images, the classifications using PTSXGB were more stable for different periods, with the mean absolute difference in the overall accuracy less than 3.90%. Therefore, PTSXGB is expected to monitor salt marsh vegetation's long-term dynamics, facilitating effective ecological conservation for the coastal areas.