Machine learning based classification of lake ice and open water from Sentinel-3 SAR altimetry waveforms

The aim of the study was to evaluate, for the first time, the capability of different machine learning (ML) algorithms in classifying along-track lake surface conditions (open water and ice types) across ice seasons (freeze-up, ice growth and break-up periods) from Sentinel-3 A/B synthetic aperture radar altimeter (SRAL) data. To achieve this goal, over 107,500 radar waveforms extracted from 11 large lakes across the Northern Hemisphere and three ice seasons (2018–2021) were manually labelled using complementary satellite data (Sentinel-1 imaging Synthetic Aperture Radar (SAR), Sentinel-2 Multispectral Instrument (MSI) Level 1C, and MODIS Aqua/Terra data) for the training and testing of the ML algorithms in discriminating between open water, young (thin) ice, growing ice and melting ice. The four ML algorithms tested include Random Forest (RF), Gradient Boosting Trees (GBT), K Nearest Neighbor (KNN) and Support Vector Machine (SVM). To characterize the waveforms, seven waveform parameters were derived: Leading Edge Width (LEW), Offset Center of Gravity (OCOG) Width, Pulse Peakiness (PP), backscatter coefficient (Sigma0), late tail to peak power (LTPP), early tail to peak power (ETPP) and the maximum value of the echo power (Max). Accuracies >95% were achieved across all classifiers using a 4-parameter combination (Sigma0, PP, OCOG Width, and LEW). Among all waveform parameters, Sigma0, OCOG width and PP were found to be the most important parameters for discriminating between lake ice types and open water. Despite showing comparable classification performances in the overall classification, RF and KNN are found to be a better fit for global lake ice mapping as both are less sensitive to their internal hyperparameters. Additionally, consistent results (>93.7% accuracy in all classifiers) achieved on the accuracy assessment carried out for each lake (out-of-sample testing) revealed the strength of the classifiers for spatial transferability. Implementation of RF and KNN could be valuable in a pre-or post-processing step for identifying lake surface conditions under which the retrieval of water level and ice thickness may be limited or not possible and, therefore, inform algorithms currently used for the generation of operational or research products. While the research focused on 11 of the largest lakes of the Northern Hemisphere, the classification approach presented herein has potential for application on smaller lakes too since data in SAR mode (∼300 m along-track resolution) are used.