A Novel Deep Learning-Based Approach for Rift and Iceberg Recognition From ICESat-2 Data

• Published in: IEEE transactions on geoscience and remote sensing 2024, Vol.62, p.1-17
• Main Authors: Huang, Zhengrui, Wang, Shujie, Alley, Richard B., Li, Anqi, Parizek, Byron R.
• Format: Article
• Language: English
• Subjects: Algorithms; Altimeters; and Land Elevation Satellite-2 (ICESat-2); Antarctica; Artificial neural networks; Cloud; Deep learning; Feature extraction; Graphs; Ice; ice shelf; Ice shelves; iceberg; Icebergs; Labels; Land ice; Laser altimeters; Machine learning; Neural networks; Photons; Point cloud compression; Recognition; rift; Scene classification; Sea surface; Three dimensional models; Training
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2024.3382573

The knowledge of rifts and icebergs in Antarctica is imperative for understanding drivers and mechanisms controlling ice-shelf retreat. The description of their 3-D structural features has been challenging before the recent launch of the Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2), whose goal is to collect high-resolution elevation measurements using a photon-counting laser altimeter system. The great advancements of deep learning in multifeature characterization enable the recognition of rifts and icebergs from the Global Geolocated Photon Data (ATL03) product. Considering the insufficient 3-D information on rifts and icebergs at an extended spatiotemporal scale, we propose a novel deep learning-based approach to recognize rifts and icebergs from ATL03 data. First, ATL03 data are converted into three feature spaces, including 3-D point clouds, 2-D images, and 2-D graphs. Second, we construct a scene classification network for scene label prediction. This model builds three deep neural networks (DNNs) to separately encode three feature spaces, simultaneously extracting 3-D and 2-D morphological features and topological features from ATL03 data. These heterogeneous features are further integrated through a feature fusion layer. Finally, we implement a presegmentation algorithm to segment unlabeled ATL03 data into separate scenes and use a trained classifier to predict scene labels. Case studies on Antarctic ice shelves are conducted to validate the effectiveness of the proposed approach in terms of performance and generalization capabilities.