Point-DCA-Enhanced Deep Learning Network for Photon Signal Classification in UAV-Based Photon-Counting LiDAR Bathymetry

A photon-counting LiDAR system deployed on uncrewed aerial vehicles (UAVs) enables highly mobile, high-efficiency, high-resolution shallow-water bathymetry. Photon-counting LiDAR technology can significantly enhance underwater weak signal detection capabilities but concomitantly introduces substantial noise sources, severely interfering with the extraction of target photon signals; this necessitates denoising while precisely discriminating between water surface and bottom photons for refraction correction. To address these challenges, this study proposes a dual cross-attention enhanced PointNet++ network (Point-DCA) for classifying noise, water column, bottom, and vegetation photons in bathymetric data. By capturing cross-channel and hierarchical spatial information through its DCA mechanism, Point-DCA overcomes the limited feature fusion capacity of PointNet++’s native skip connections, significantly improving classification performance. To validate the framework, a purpose-built photon-classification dataset was developed through raw photon data decoding, manual annotation, and computation of neighborhood point features to validate the framework. Experimental results showed that, without additional features, Point-DCA achieved a 4.4% higher overall accuracy (OA), 12.8% improvement in mean F1-score (mF1), and 17.3% increase in mean intersection over union (mIoU) compared to PointNet++. After incorporating neighborhood point features, Point-DCA improved the OA to 97.8%, mF1 to 95.7%, and mIoU to 91.9%. Tests across four water-depth scenarios validated the efficacy of Point-DCA in end-to-end photon point cloud denoising and classification.