Evaluation of Machine Learning Algorithms for the Classification of Lake Ice and Open Water from Sentinel-3 SAR Altimetry Waveforms

Lakes cover a significant fraction of the landscape in many northern countries. They play a key role in regulating weather and climate and also have a significant impact on northern communities since the presence (or absence), extent and thickness of lake ice affect transportation (ice roads), food availability, recreational activities, and tourism in wintertime. The drastic decline in in-situ observations of lake ice phenology (i.e., freeze-up and break-up dates and ice cover duration) and lake ice thickness globally over the last three decades make remote sensing technology a viable means for monitoring lake ice conditions. Although satellite radar altimetry has been used in various cryospheric and hydrological studies, little work has been conducted on lake ice compared to, for example, sea ice and the estimation of lake water levels. This study was carried out using Sentinel-3A/B SAR altimetry data acquired over three ice seasons (2018-2019, 2019-2020 and 2020-2021) at 11 large lakes across the Northern Hemisphere. We explored the information provided by radar waveforms to discriminate between open water, first (young) ice, growing ice and melting ice using machine learning models. 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 (LTTP), early tail to peak power (ETTP) and the maximum value of the echo power. Four machine learning algorithms including Random Forest (RF), Gradient Boosting Trees (GBT), K Nearest Neighbour (KNN) and Support Vector Machine (SVM) classifiers were tested to assess their capability in classifying the lake surfaces across all years. Manual class labelling based on Sentinel-3 Synthetic Aperture Radar Altimeter (SRAL) waveforms and complementary satellite data (Sentinel-1 imaging SAR data, Sentinel-2 Multispectral Instrument (MSI) Level 1C data, and MODIS Aqua/Terra data) was performed to create training and test samples for the classifiers. Accuracies greater than 95% were achieved across all classifiers using a 4-parameter combination (Sigma0, PP, OCOG Width, and LEW). Amongst all waveform parameters, Sigma0, OCOG width and PP were found to be the most important parameters for discriminating between lake ice 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 and have faster processing speeds. Additionally, consistent results (>93.7% accuracy in all classifiers) achieved on the accuracy assessment carried out for each lake 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 has potential for application on smaller lakes too since SAR mode data (~300 m along-track resolution) is used in the study