Hyperspectral LiDAR saturated signal waveform recovery for ranging based on neural networks

Hyperspectral LiDAR (HSL) is an advanced active detection system that can simultaneously acquire the three-dimensional information and spectral information of a range of targets. However, high-reflectivity targets often encounter to signal saturation, which introduces errors in both time-of-flight (ToF) ranging and hyperspectral reflectance calculation. Such saturation is a joint result of the spectrum of the supercontinuum laser source and the spectral response of the photodetector, and it causes an unbalanced detection performance of different spectral channels of the HSL. This paper proposed a neural network-based waveform recovery framework that addresses saturation-induced distortions in HSL systems to achieve precision ranging. What we believe to be a newly constructed fitting function was designed to simulate saturated waveforms through signal amplification, amplitude truncation, and temporal broadening. Subsequently, A dataset combining simulated and experimental data was built for training. Finally, we implemented a comparative evaluation of three neural network architectures-convolutional neural networks (CNN), radial basis function neural networks (RBFNN), and generalized regression neural networks (GRNN)-were evaluated for waveform recovery and ranging accuracy. RBFNN achieved the best balance between performance and efficiency. The results show that our method can obtain accuracy ranging for high-reflectivity targets with an average mean squared error of 60.623, an average coefficient of determination of 0.9935, and an average correlation coefficient of 0.9991. These results demonstrate that our method can solve the saturation recovery problem for HSL ranging in high-reflectivity signal acquisition.