Abstract: In order to make full use of the existing data to improve the accuracy of remote sensing water-depth inversion of shallow water, and quickly select parameters in the machine learning model, taking Caohai Lake as the research area, geographic information(GEO) was added to the BP neural network model and random forest model(RF),and the optimal parameters of the GEO+RF model were selected by orthogonal experimental design(OED).The inverted results were compared with the multi-band log-linear model, GEO+BP neural network model and GEO+RF model.The results showed that the GEO+OED-RF model had the highest accuracy, the measured water depth-inverted water depth scatter plot had the most concentrated points, and the inversion water depth map was basically consistent with the measured water depth.It showed that the GEO+OED-RF model can effectively improve the experimental efficiency, select optimal parameters, and improve the inversion accuracy of remote sensing water-depth in shallow water areas, which can provide a method reference for remote sensing monitoring and analysis of water resources in similar areas.