A novel strategy of full-waveform light detection and ranging bathymetry based on spatial gram angle difference field conversion and deep-learning network architecture

Abstract

The airborne light detection and ranging (LiDAR) bathymetry (ALB) system is a promising and effective approach for surveying nearshore bathymetry and underwater terrain. Deep-learning techniques have been developed to reduce waveform superposition in shallow and deep areas. These methods avoid the complex transmission process of laser pulses in the water column and the intricate determination of various parameters and thresholds in traditional ALB methods. However, studies on ALB bathymetry using deep-learning techniques remain insufficient. To improve the accuracy and reliability of nearshore bathymetry, this study proposes deep-learning bathymetry fusing waveform features and spatial-angular field features (DBWSF). This method utilizes the waveform curvature to construct an energy curve, enhancing the waveform's features. Additionally, it employs a Gram angle difference field to convert the temporal waveform into a two-dimensional Gram angle difference field image, increasing the dimensions and quantity of waveform features. Finally, this method constructs a dual-path neural network with an attention mechanism to extract the water surface and bottom waveform signals precisely to achieve nearshore bathymetry. In comparison to sample data, DBWSF exhibited high bathymetric accuracy across three study areas (Ganquan Island, Lingyang Reef, and Dong Island), achieving a root mean squared error of 0.21 m. The R² in these regions were 99.6 %, 95.1 %, and 99.8 %, respectively. In the above three study areas, compared with the bathymetric results obtained using the waveform decomposition method, DBWSF was more accurate, with improvements in RMSE of 0.33, 0.27, and 0.04 m. Compared with multilayer perceptron (MLP), the corresponding accurate improvements in RMSE with DBWSF were 0.47, 0.40, and 0.25 m. Compared with other two methods, the R² value for DBWSF in the three study areas exceeded 95 % and reached a highest value of 99.8 %. The results demonstrated the bathymetric capability, reliability, and transferability of DBWSF for determining nearshore bathymetry in different water environments. The novel LiDAR bathymetric deep-learning technique can effectively and intelligently produce precise nearshore bathymetry and seafloor topography maps.