Land-sea classification based on the fast feature detection model for ICESat-2 ATL03 datasets

Accurate land-sea classification of ATL03 datasets is the prerequisite for signal photons identification and higher-level data production. Current photon classification methods are mainly based on physical models and machine learning methods. Physical models necessitate a priori knowledge of the background noise rate, while machine learning methods require extensive manual intervention, which can be time-consuming. This work introduced a fast feature detection model (FFD model) for land-sea classification that eliminates the need for prior knowledge and is not constrained by time. The land-sea boundary usually experiences significant photon height changes and reflective intensity differences on both sides. The FFD model was developed to efficiently detect the inflection point of photon reflective intensity for preliminary land-sea classification. Minor classification errors were carefully corrected based on topographic geometric distribution in further refinement. The FFD model was applied to three typical land-sea boundary types to evaluate its classification accuracy. Validation of classification results with consistent Sentinel-2 images and manually tagging photons indicated that the FFD method exhibited substantial robustness to various land-sea boundary landforms. The FFD model can process nearly 105 raw photon classifications per second, with an average overall accuracy of 99.42 % for preliminary classification results. Coastal slope and surface reflection in land-sea boundary may have impact on classification accuracy. These effects were carefully eliminated in the refinement process, resulting in an average overall accuracy improvement to 99.86 %. Consequently, the FFD model represented a robust and efficient solution for automatic land-sea classification in coastal shallow water, coral reefs, and estuary areas.