Ice Edge Verification -- Measuring the skill in our forecasts and disagreement in our observations

Abstract

Sea ice is one of the most crucial components of the polar system. Alongside its effects on the climate and weather, it also impacts human lives and operations in these regions. Climate scientists, as well as other stakeholders (such as marine industries, mining operations, governmental authorities as well as local/aboriginal population), have a high interest in getting accurate observation and prediction of sea ice presence, and its effective border – the ice edge. In this context, this thesis is focused on the verification of ice presence and ice edge, across different datasets. We do this by describing a new method of generating reference forecasts of the ice edge (as benchmark for predictions at sub-seasonal to seasonal timescales), analysing initial state errors in forecasts and analysis from ECMWF, and comparing the ice presence and ice edge difference between several observational and analysis datasets.

Operational systems generating sea ice forecasts at sub-seasonal timescales are mostly compared against simple references based solely on climatological or initial states, which can lead to a potential overestimation of their prediction skill. In chapter 1, we describe the Spatial Damped Anomaly Persistence (SDAP) method, which combines historical sea ice probability with the initial ice edge anomalies to generate probabilistic reference forecasts of the ice edge. The SDAP forecasts outperform both traditional references, as well as most dynamical forecast models from the Sub-seasonal to Seasonal (S2S) project at long lead times, establishing a more challenging benchmark for operational forecast systems.

Within the S2S database, forecasts from the European Centre for Medium- Range Weather Forecasts (ECMWF) have the highest prediction skill compared to both traditional references and SDAP forecasts, but show significant initial error. In chapter 2, we analyse this initial state issue in the ECMWF forecasts, as well as corresponding analysis from ORAS5, by measuring the errors against observations from OSISAF. We find that the initial state errors are partly due to interpolation issues, and partly systematic underestimation, especially in the summer. We show the spatial distribution of the mean bias and provide evidence that between 10 to 20% of the error can be reduced simply by subtracting the mean bias.

While forecast verification methods assume observations to be ‘the truth’, observational records also have inconsistencies that have previously only been discussed in concentration or ice extent terms. In chapter 3, we analyse differences in ice presence between several observational and analysis datasets by measuring the Integrated Ice Edge Error and bias between each pair. We find significant mismatch between observations, particularly in the summer, and identify regions where certain observations disagrees with all other. Our results show that observational records from OSISAF potentially overestimate ice presence, while those from AMSR-E/2 potentially underestimate it. In the last chapter, we discuss the results from the different studies in context of each other and mention how the inconsistencies in observations cast doubt on the forecast errors we initially measured, offering the possibility of a probabilistic observation and reemphasizing the need for accurate sea ice edge measurements and forecasts.