A squeeze-and-excitation channel attention module was incorporated into a U-Net-based convolutional neural network to develop a deep learning model, named CrossWaterNet, for water body identification from multi-source multispectral remote sensing images. The model was trained using a dataset comprising 140000 labeled water body samples. To address the systematic jagged-edge errors in water body boundaries caused by pixel discreteness in remote sensing images, a water boundary smoothing and reconstruction method based on Gaussian filtering and B-spline curve fitting was further proposed. Taking the three types of surface water bodies, including rivers, reservoirs and lakes, as the objects, the improved normalized difference water body index method, the OmniWaterMask model and the DeepWaterMapV2 model were selected for comparative analysis. The results show that the CrossWaterNet model is more sensitive to key spectral bands and exhibits robust generalization performance across different optical sensors, enabling it to accurately identify various types of surface water bodies, including rivers, reservoirs, and lakes. Moreover, the proposed boundary smoothing and reconstruction method effectively mitigates the systematic errors introduced by pixel discreteness, thereby improving the stability and consistency of boundary representation. Keywords: water body identification; remote sensing technology; deep learning; multi-source multispectral remote sensing images; smoothing and reconstruction 〖FL