A refined Sentinel-3 sea ice classification method based on multi-waveform features and machine learning: application to the Beaufort Sea

Monitoring sea ice dynamics is essential for climate research, polar ecosystems, and maritime safety. ICESat-2 and CryoSat-2 suffer from limited spatiotemporal coverage, while Sentinel-3 provides complementary observations. However, the Sentinel-3 Sea Ice Thematic Products rely on threshold-based classification and have difficulty distinguishing gray ice from lead; moreover, unclassified measurements reduce the accuracy and spatial coverage of sea ice thickness retrieval. In this study, a refined classification method for Sentinel-3 sea ice surface types was developed by integrating multi-waveform features with machine learning algorithms, guided by Sentinel-2 imagery. Results show that a five-parameter combination consisting of Sigma0, pulse peakiness (PP), concentration, offset center of gravity width (OCOG_W), and stacked standard deviation (SSD) is the optimal scheme, with all models achieving overall accuracy (OA) above 90% and Sigma0, PP, and concentration emerging as the most important features. Among the evaluated models, the Light Gradient Boosting Machine (LGB) model achieves the best performance, with an OA of 95.6% and a Kappa coefficient of 0.904, exceeding the support vector machine (SVM) model by 1.7% and 0.039, respectively. Compared with the Sentinel-3 Sea Ice Thematic Products, the proposed method distinguishes lead, gray ice, and sea ice more accurately, thereby improving the reliability of Sentinel-3 sea ice classification results and demonstrating potential value for sea ice dynamics monitoring and related climate studies.