Deep learning assisted exponential waveform decomposition for bathymetric LiDAR

The processing of bathymetric LiDAR waveforms is an important task, as it provides range and radiometric information to determine the precise location of water surface and bottom, and other characteristics like amplitude. The exponential waveform decomposition proved to be an effective algorithm for bathymetric LiDAR waveforms processing, however, it heavily relies on the high-quality initial estimates of the model parameters. This paper proposes to make use of deep learning to obtain the initial values directly from the input received waveforms without any hand-crafted features and prior-knowledges. Additionally, to provide training samples, we presents a method to create the synthetic bathymetric LiDAR waveforms by simulating of the backscatter cross function returned from water bodies. Two networks with different sensitivities of weak signals were trained by these synthetic waveforms, and used to estimate the initial values of the model parameters, a least square optimization follows up to obtain the final waveform decomposition result. This deep learning assisted exponential waveform decomposition method is applied to the real waveforms acquired by RIEGL VQ-840-G. The results show that estimations with the help of deep learning is less influenced by the intermediate peaks backscattered from objects and particles in water, producing a cleaner point cloud with less isolated points below water surface than the original exponential waveform decomposition. Moreover, the proposed sensitive DL-XDC is even able to detect some very weak bottom returns with low SNR.