Abstract: Aiming at the problem of insufficient information of spectral features for remote sensing bathymetry inversion in complex marine environments, a bathymetric retrieval model based on multispectral data and its spatial features was proposed by using three machine learning methods, multilayer perceptron (MLP), random forest (RF) and support vector machine (SVR). Taking Wuzhizhou Island as an example, the spatial location and slope were selected as spatial features, and the shallow water depth was derived by combining with the spectral features of WorldView-2 images and the in-situ measured water depth data. The results show that under the same conditions, all three machine learning algorithms obtained better inversion results, with the MLP model having the highest inversion accuracy. The model with inputting the spatial features accurately reflects the undulating features of shallow bottom topography of the study area, and it is significantly more accurate than the spectral-based model alone, with an average absolute error within 2 m. Model accuracy in shallow area at 20 m is lower than that in deep area at 20 m. Spatial features can effectively reduce the impact of complex marine environments on remote sensing bathymetry models, and spatial location information can significantly improve model accuracy.