Airborne bathymetric lidar has become an essential tool in coastline surveys, waterway safety assessments, and marine resource management. However, laser absorption and scattering in seawater significantly weaken echo signals as water depth increases, rendering traditional waveform processing methods ineffective in detecting weak seabed echoes. Consequently, enhancing the bathymetric detection capability of airborne lidar is of considerable importance.

To address this limitation, the study introduces a deep learning-based method for inverse bathymetry. The proposed approach converts lidar waveforms into images, enabling the utilization of adjacent frame waveform information, which facilitates the extraction of weak seabed echoes compared to traditional single-waveform processing. In terms of data preprocessing, a linear approximation method is employed to rapidly eliminate water scattering effects, thereby reducing variability in echo waveforms of different water qualities. This simplifies model training by eliminating the need to account for variations in echo characteristics caused by differing water qualities. For the neural network model, the U-Net architecture is adopted, incorporating the spatial convolutional neural network (SCNN) module to enhance feature fusion and improve semantic segmentation performance. The waveform data is transformed into an image-based input by converting each temporal waveform moment into a point cloud, which is subsequently projected onto a plane.

To validate the proposed method, two datasets, A and B, were constructed for model training and testing. Results demonstrate that the linear approximation-based preprocessing effectively mitigates variability of different water qualities. Furthermore, tests conducted on measured datasets from various regions and moments show that the model significantly outperforms traditional waveform processing methods. Specifically, results for datasets from Dazhou Island and Dacheng Wan in Zhangzhou reveal improvements in the correct prediction rate and maximum detectable water depth. These findings underscore the superior capability of the model in detecting weak seafloor echo signals and its robustness to diverse water qualities.

The method proposed in this study for water depth inversion, which involves converting waveforms into images and applying semantic segmentation through deep learning networks, significantly enhances the depth detection capability of airborne bathymetric lidar. The preprocessing approach, employing linear approximation to remove water scattering, ensures that the model can effectively handle echo signals from diverse water qualities. The conversion of waveforms into image-based inputs facilitates the improved recognition of weak seabed signals. Data processing results from different moments and regions demonstrate the model robustness to varying water qualities, with substantial improvements observed in both the accuracy and maximum depth of detection compared to traditional waveform processing methods. However, the approach has limitations in scenarios where surface and seafloor signals are mixed. The detection of seafloor signals in shallow water will be a key focus of future research.