According to the current situation that multibeam filtering should be artificially eliminated, an improved cloth simulation filtering (CSF) based on point cloud segmentation and adaptive parameter adjustment was proposed. First of all, the multibeam point cloud data was segmented. Secondly, the parameters of each CSF algorithm were adjusted according to the area and standard deviation of the point cloud data. Then, the fabric filtering model was constructed to eliminate the gross error. Finally, the segmented point cloud was stitched to generate the underwater terrain point cloud data. The results show that the over-filtering defect is eliminated compared with the CSF algorithm, and the gross error elimination ratio is decreased from 1.03 to 1.01 and from 1.46 to 1.19 respectively for different regions. Compared with the manual elimination of gross error, the missed gross error can be avoided and little manual intervention is evolved.