With the launch of the ICESat-2 satellite in October 2018, hydrographers obtained access to a vast amount of data that contain information about shallow bathymetry. Such data are especially valuable in the areas where lidar surveys have not or cannot be conducted. However, these data, consisting of geolocated photons (photon events or PEs) reflected from sea or ocean floor are noisy and require careful processing for extraction of reliable bathymetric information. Several algorithms have been proposed for automation of this labor-intensive and time-consuming task. In this paper, we propose a novel approach for such processing that has shown a promising high degree of accuracy (>90%) and often outperforms human experts.

Our approach utilizes the machine learning algorithm, specifically, Light Gradient-Boosting Machine (LightGBM). The human expert processes ICESat-2 tracks by marking bathymetric PEs that will be used to train the model. All PEs in the dataset are then converted to vectors of sectorial counts of PEs that describe the spatial distribution of PEs in the immediate vicinity. (The neighborhood and sectors are defined by sectioned nested ellipses.) These vectors are used by LightGBM to train a model for a binary classification (bathymetric or non-bathymetric PE). Two datasets from different geographic locations were used for investigation of algorithm efficiency and robustness. F1-scores for 20% test samples for both datasets exceeded 98%. For cross-validation (the model trained on one dataset was applied to the other dataset), the F1-score was 94%. Further investigation suggests the method is relatively insensitive to ellipse size and shape and the number of sectors employed. This work suggests that it may be possible to train a model that is effective for most ICESat-2 tracks.