Abstract: To address the problems of tropospheric delay errors, deconvolution errors and the large amount of time and disk space required to process data over a large area in InSAR during data processing,in this paper, the atmospheric correction data of 134 Sentinel-1 lifting rail images of Kunming city from September 16, 2016 to May 5, 2021 are processed by LiCSBAS and synthetic aperture radar general atmospheric correction online service product, and the subsidence information of the main urban area of Kunming city is obtained.On this basis, five typical land surface subsidence areas are obtained and their temporal and spatial distribution characteristics are analyzed. Then, deep forest and long term memory network models are used to predict the time series values, and absolute error (ε), root mean square error (RMSE) and nash coefficient (NSE) are introduced to evaluate the models. Both the deep forest and LSTM prediction models are within 4 mm, RMSE values are 0.70 and 3.01, and NSE values are 0.92 and 0.81, respectively.The results show that the deep forest prediction model has a good effect. The urban surface monitoring and prediction method combined with LiCSBAS and machine learning model can provide a reference for future land subsidence monitoring and disaster warning.

Key words: tropospheric delay error, LiCSBAS, deconvolution error, deep forest, time series prediction