Enhancing Satellite LiDAR Systems: Self-Supervised 3D Tensor Denoising for the CASALS Architecture

• Daryl Tapel, Computer Engineering, University of Delaware

• Gonzalo Arce, Electrical and Computer Engineering, University of Delaware

Modern satellite LiDAR systems capture highly detailed photon-count waveforms but experience significant signal loss due to background photons and readout noise. This research addresses this issue by denoising these signals with a self-supervised neural network that can recover 3D hyperheight data cubes (HHDC) from low-resolution observations. The project focuses on data simulated by the Concurrent Artificially-Intelligent Spectrometry and Adaptive Lidar System (CASALS) forward model. This model emulates the physical sensing mechanism of the instrument, producing measurements with realistic spatial correlations and mixed Poisson-Gaussian noise. Denoising these waveforms is challenging because traditional supervised models need paired clean-noisy datasets, which are rarely available in real-world remote sensing situations. To assess these volumetric structures on the HHDC dataset, the state-of-the-art Asymmetric Pixel-Shuffle Downsampling Blind-Spot Network (AP-BSN) algorithm was used. While the original algorithm was designed to remove spatially correlated noise from standard 2D sRGB images, its structure was modified to process 3D tensors. To further improve the design for LiDAR systems, the network’s internal capacity was expanded to process vertical height bins at the same time. Pixel-downsampling strategies were also adjusted to perfectly maintain spatial resolution and eliminate stride aliasing. With this approach, the adapted 3D network demonstrated effective spatial denoising and peak intensity preservation without needing clean target image sources. This offers a strong, fully self-supervised option compared to traditional machine learning methods for recovering high-resolution 3D canopy structures in remote sensing systems.