Application of machine learning for lead detection from ICESat-2 geolocated photon data (ATL03)

Abstract:

ICESat-2 laser altimeter has a very good spatial resolution of 17 m footprint with spacing 0.7 m. This outstanding spatial resolution can make it possible to find narrow leads, and eventually estimate sea ice freeboard accurately. However, the sea ice products of ICESat-2 (ATL07 and ATL10) employs 150-photon aggregation to detect leads (open water) and calculate sea ice freeboard, which can undermine the performance of freeboard estimation in two respects: (1) irregular spatial resolution (> 10 m) and (2) contamination of the height segment by mixed surface types in the 150-photon aggregation. Therefore, in order to maintain the original spatial resolution of the ICESat-2 photon data and improve the lead detection accuracy, we attempt to detect leads directly from the ICESat-2 geolocated photon product (ATL03). Various machine learning (ML) techniques are applied for this purpose: including artificial neural network (ANN), deep neural network (DNN), and long short term memory network (LSTM). As a ground truth for training and testing the ML models, we use Sentinel-2 true color (RGB) images that are coincident with the ATL03 tracks. Compared to the leads detected by the ATL10 product, our ML algorithms show better performances in detecting leads with a better spatial resolution (2 m). By detecting leads more accurately with fine resolution, these ML techniques can contribute to the accurate estimation of sea ice freeboard and thickness over the polar region.