Enhanced Retrieval of Spring Snow Depth Over Pan‐Arctic Sea Ice From AMSR2 Observations for 2013–2023 Using Multi‐Model Machine Learning

Abstract

Snow on sea ice modulates the Earth's radiation budget and sea ice mass balance, yet satellite retrievals of Arctic snow depth remain subject to substantial algorithmic uncertainty. We develop a multi‐model machine learning (ML) framework for retrieving spring snow depth over Arctic sea ice from Advanced Microwave Scanning Radiometer 2 (AMSR2). Three structurally distinct ML models are trained with high‐resolution airborne measurements from the Alfred Wegener Institute IceBird campaigns in April 2017 and 2019. Evaluation using a composite airborne data set from NASA's Operation IceBridge for spring 2013–2015 yields correlation coefficients of 0.66–0.70 and mean errors of 0.1–0.3 cm across the three models. A performance‐weighted fusion scheme is then introduced to combine the three ML models into a merged framework (AMSR2‐Merged) and generate a daily snow depth record for March–April from 2013 to 2023. Independent validation against ground‐based observations during the MOSAiC expedition shows a mean error of 0.16 cm for AMSR2‐Merged, with error reductions of 50%–98% relative to established empirical algorithms and 16%–88% relative to individual ML models. Over the study period, regionally averaged spring snow depths range from ∼15 to 23 cm in the Barents Sea to ∼23–25 cm in the central Arctic Ocean, with the largest interannual variability reaching ∼5 cm. Notably, the Arctic‐wide spring snow depth is ∼10 ± 5 cm lower than that in the Warren climatology (1954–1991). The uncertainty of AMSR2‐Merged ranges from ∼2.5 to 6 cm and peaks in the marginal ice zones, where AMSR2 data and inter‐model spread are the dominant sources.