Abstract: Accurate identification of seabed substrate types is essential for understanding the distribution of benthic marine communities and for planning the sustainable development of marine resources. Machine learning algorithms are an effective means of identifying substrate types. Given the limitations of substrate classification of single acoustic data of islands and reefs, fusing multispectral remote sensing data provides a new approach to overcome these limitations. This study proposes proposes a seabed classification method based on feature selection and Genetic Algorithm Extreme Gradient Boosting (GA-XGBoost), fusing multispectral remote sensing data and multibeam data from multiple sources. Firstly, the WorldView-2 multispectral data and multibeam data are preprocessed to unify the geographic coordinate system and align the spatial resolution. Then, the spectral features of the multispectral image, the topographic features of the bathymetry data, and the texture features of the backscatter intensity are extracted to form an 18-dimensional feature parameters. The 12-dimensional optimal feature subsets are then selected from these 18-dimensional features based on the XGBoost (Extreme Gradient Boosting) algorithm combined with forward stepwise feature selection. Next, the GA-XGBoost classification model is established, and the model was trained and tested using single-source and multi-source data respectively. It is then compared with the BPNN (Back Propagation Neural Network), GA-BPNN (Genetic Algorithm-Back Propagation Neural Network) and XGBoost classification algorithms. Finally, the optimal GA-XGBoost classification model was applied to classify and visualize the substrate in the entire study area. Experimental results show that the overall accuracy of the method in seabed substrate classification IS 91.23%, with a Kappa coefficient of 0.87 and an F1 score of 0.911 8, which is significantly better than the single data source input and the comparison algorithm. This indicates that GA-XGBoost can provide a new way for the rapid and accurate classification of seabed substrate.