Active bathymetric simultaneous localization and mapping (SLAM) technologies can yield accurate and robust navigational results for long-range autonomous underwater vehicles (AUVs), by spurring vehicles to areas with complex terrain actively. However, most underwater terrain tends to be smooth under the scouring and erosion of ocean currents, making it quite challenging to calculate the feature-richness level in active bathymetric SLAM methods by extracting point, line, or even plane features. This paper proposed a both efficient and accurate terrain saliency quantification method for active bathymetric SLAM. In this work, we train PointNet-based PointNetMLP to estimate the uncertainty of bathymetric point cloud registration which can quantify the terrain saliency of the corresponding bathymetric point cloud. And the performance of the proposed method was tested using both play-back and simulation experiments.