Denoising and Classification of ICESat-2 Photon Point Cloud based on Convolutional Neural Network

ICESat-2 (Ice, Cloud, and land Elevation Satellite-2) launched by NASA (National Aeronautics and Space Administration) in 2018 is a laser altitude measurement satellite. The advanced topographic laser altimeter system (ATLAS) instrument on-board ICESat-2 employs a micro-pulse and multi-beam photon counting laser altimeter system with low energy consumption, high detection sensitivity, and high repetition rates, and thus greatly improves the sampling density in the along-track distance. However, it introduces a significant number of solar noise photons in the raw data. How to effectively remove the noise photons and classify the signal photons into ground photons and canopy photons is critical for subsequent applications such as the estimation of terrain elevation and forest height, and it has been a hot and challenging topic in the current research. In this paper, a denoising and classification algorithm based on convolutional neural network was proposed. The convolutional neural network has made a series of breakthrough research results in the fields of image classification, object detection, semantic segmentation, and so on. To remove obvious noise photons, the photons were first divided into grids in the along-track distance and elevation direction, and the rough signal photons were gridded into pictures. Then, the convolutional neural network was employed to perform the final denoising and classification. Finally, the proposed algorithm was tested with the airborne LiDAR datasets, including DSM (Digital Surface Model) and DTM (Digital Terrain Model), and was further compared with ATL08 (land and vegetation height) products. Experimental results show that our proposed algorithm can remove noise photons effectively in bare land and forest areas. Moreover, this algorithm can simultaneously remove noise photons and classify signal photons into ground photons and canopy photons in forest areas. The R2 and RMSE values of the retrieved ground surface in the bare land areas were 1.0 and 0.72 m, respectively. In the forest areas, the R 2 of the estimated ground surface and canopy surface were 1.0 and 0.70 with the RMSE values of 1.11 m and 4.99 m, respectively. The reason for this result may be that it is difficult for photons to penetrate the forest canopy and reach the ground surface in forest areas, which causes the RMSE value of the forest area to be larger than that of the bare land area. In this paper, the deep learning algorithm was used to realize the denoising and classification of photon counting data, and good results were achieved in bare land and forest areas, which provides a reference for subsequent photon counting LiDAR data processing. 2021, Science Press. All right reserved.