Regional September Sea Ice Forecasting with Complex Networks and Gaussian Processes

Lawrence

2021

The Cryosphere

Self Cite

Abstract. Observations of sea ice freeboard from satellite radar altimeters are crucial in the derivation of sea ice thickness estimates, which in turn provide information on sea ice forecasts, volume budgets, and productivity rates. Current spatio-temporal resolution of radar freeboard is limited as 30 d are required in order to generate pan-Arctic coverage from CryoSat-2 and 27 d are required from Sentinel-3 satellites. This therefore hinders our ability to understand physical processes that drive sea ice thickness variability on sub-monthly timescales. In this study we exploit the consistency between CryoSat-2, Sentinel-3A, and Sentinel-3B radar freeboards in order to produce daily gridded pan-Arctic freeboard estimates between December 2018 and April 2019. We use the Bayesian inference approach of Gaussian process regression to learn functional mappings between radar freeboard observations in space and time and to subsequently retrieve pan-Arctic freeboard as well as uncertainty estimates. We also employ an empirical Bayesian approach towards learning the free (hyper)parameters of the model, which allows us to derive daily estimates related to radar freeboard spatial and temporal correlation length scales. The estimated daily radar freeboard predictions are, on average across the 2018–2019 season, equivalent to CryoSat-2 and Sentinel-3 freeboards to within 1 mm (standard deviations <6 cm), and cross-validation experiments show that errors in predictions are, on average, ≤ 4 mm across the same period. We also demonstrate the improved temporal variability of a pan-Arctic daily product by comparing time series of the predicted freeboards, with 31 d running means from CryoSat-2 and Sentinel-3 freeboards, across nine sectors of the Arctic, as well as making comparisons with daily ERA5 snowfall data. Pearson correlations between daily radar freeboard anomalies and snowfall are as high as +0.52 over first-year ice and +0.41 over multi-year ice, suggesting that the estimated daily fields are able to capture real physical radar freeboard variability at sub-weekly timescales.

Section: Discussionsupporting

confidence: 54%

Section: Introductionmentioning

confidence: 99%

A Bayesian approach towards daily pan-Arctic sea ice freeboard estimates from combined CryoSat-2 and Sentinel-3 satellite observations

Lawrence

2021

The Cryosphere

Self Cite

“…Of course, an investigation into the drivers of the temporal variability in the CS2S3 field would be an additional way to validate the product, although this goes beyond the scope of our study. We conclude here that the Gaussian Process Regression method is an extremely robust tool for modelling a wide range of statistical problems, from interpolation of geo-spatial data sets, as presented here and in other works (Le Traon et al 1997;Ricker et al 2017), to time series forecasting (Rasmussen and Williams 2006;Sun et al 2014;Gregory et al 2020). The Gaussian assumption holds well in many environmental applications, and the fact that the covariance structure can take any form, so long as K is a symmetric positive semi-definite matrix, means that the model can be tailored very specifically to the problem at hand.…”

Section: Discussionsupporting

confidence: 54%

Response to Dr. Alessandro Di Bella (RC1)

2021

Self Cite

Observations of sea ice freeboard from satellite radar altimeters are crucial in the derivation of sea ice thickness estimates, which in turn inform on sea ice forecasts, volume budgets, and productivity rates. Current spatio-temporal resolution of radar freeboard is limited as 30 days are required in order to generate pan-Arctic coverage from CryoSat-2, or 27 days from Sentinel-3 satellites. This therefore hinders our ability to understand physical processes that drive sea ice thickness variability on sub-monthly time scales. In this study we exploit the consistency between CryoSat-2, Sentinel-3A and Sentinel-3B radar freeboards in order to produce daily gridded pan-Arctic freeboard estimates between December 2018 and April 2019. We use the Bayesian inference approach of Gaussian Process Regression to learn functional mappings between radar freeboard observations in space and time, and to subsequently retrieve pan-Arctic freeboard, as well as uncertainty estimates.The estimated daily fields are, on average across the 2018-2019 season, equivalent to CryoSat-2 and Sentinel-3 freeboards to within 2 mm (standard deviations < 5 cm), and cross-validation experiments show that errors in predictions are, on average, ≤ 3 mm across the same period. We also demonstrate the improved temporal variability of a pan-Arctic daily product by comparing time series of the predicted freeboards, with time series from CryoSat-2 and Sentinel-3 freeboards, across 9 sectors of the Arctic. The mean of predicted and CryoSat-2 or Sentinel-3 time series are generally consistent to within 3 mm, except for the Canadian Archipelago and Greenland, Iceland and Norwegian Seas, which show discrepancies greater than 1 cm due, in part, to biases between CryoSat-2 and Sentinel-3 observations in these locations.

“…Similar to previous works (Fountalis et al 2014(Fountalis et al , 2015Gregory et al 2020), our method here is based on complex networks;…”

Section: Introductionmentioning

confidence: 98%

Network connectivity between the winter Arctic Oscillation and summer sea ice in CMIP6 models and observations

Stroeve

2022

Preprint

Self Cite

Abstract. The indirect effect of winter Arctic Oscillation (AO) events on the proceeding summer Arctic sea ice extent suggests an inherent winter-to-summer mechanism for sea ice predictability. On the other hand, operational regional summer sea ice forecasts in a large number of coupled climate models show a considerable drop in predictive skill for forecasts initialised prior to the date of melt onset in spring, suggesting that some drivers of sea ice variability on longer time scales may not be well represented in these models. To this end, we introduce an unsupervised learning approach based on cluster analysis and complex networks to establish how well the latest generation of coupled climate models participating in phase 6 of the World Climate Research Programme Coupled Model Intercomparison Project (CMIP6) are able to reflect the spatio-temporal patterns of variability in northern-hemisphere winter sea-level pressure and Arctic summer sea ice concentration over the period 1979–2020, relative to ERA5 atmospheric reanalysis and satellite-derived sea ice observations respectively. Two specific global metrics are introduced as ways to compare patterns of variability between models and observations/reanalysis: the Adjusted Rand Index – a method for comparing spatial patterns of variability, and a network distance metric – a method for comparing the degree of connectivity between two geographic regions. We find that CMIP6 models generally reflect the spatial pattern of variability of the AO relatively well, although over-estimate the magnitude of sea-level pressure variability over the north-western Pacific Ocean, and under-estimate the variability over the north Africa and southern Europe. They also under-estimate the importance of regions such as the Beaufort, East Siberian and Laptev seas in explaining pan-Arctic summer sea ice area variability, which we hypothesise is due to regional biases in sea ice thickness. Finally, observations show that historically, winter AO events (negatively) covary strongly with summer sea ice concentration in the eastern Pacific sector of the Arctic, although now under a thinning ice regime, both the eastern and western Pacific sectors exhibit similar behaviour. CMIP6 models however do not show this transition on average, which may hinder their ability to make skilful seasonal to inter-annual predictions of summer sea ice.