As the decline in Arctic sea ice extent makes this region more accessible, the need is increasing for effective seasonal sea ice forecasting to facilitate operational planning. Recently, coupled global climate models (CGCMs) have been used to address the need for effective sea ice forecasting on seasonal time scales. This thesis assesses the operational utility of the Canadian Seasonal to Interannual Prediction System (CanSIPS) for seasonal sea ice forecasting. This assessment consists of two separate studies. The first uses hindcasting to analyze the skill of two versions of CanSIPS, as well as an intermediate version, on the pan-Arctic as well as regional scales. This approach allows for an overall assessment of the system's skill in addition to providing insight with regards to the features in each version which improved that skill. This study finds that the use of a new initialization procedure for sea ice concentration and thickness improved forecast skill on the pan-Arctic scale as well as in the Central Arctic, Barents Sea, Laptev Sea, and Sea of Okhotsk. This study also shows that the substitution of one of the constituent models in the system improved forecast skill on the pan-Arctic scale as well as in the GIN, Barents, Kara, East Siberian, Chukchi, Bering, and Beaufort Seas. Overall, the new version of CanSIPS was found to be generally more skillful than previous versions. The second study conducts a potential predictability experiment on CanCM4, the constituent CGCM common to all versions of CanSIPS considered in this study. This study follows the methodology introduced by \cite{Bushuk2018} which allows for a more complete assessment of the dependency of potential predictability on initialization month than previous studies and for comparisons to be made between potential predictability and operational skill. This analysis is again done on both the pan-Arctic and regional scale. The findings of this experiment show that CanCM4 has relatively low potential predictability relative to other models and explains results previously presented in a multi-model study by \cite{Day2016}. Further, the characteristics of CanCM4's potential predictability share similarities with other models including greater predictability at longer lead times for winter target months than summer target months, greater predictability in the Atlantic sector than the Pacific sector, and the presence of the spring predictability barrier on the pan-Arctic scale as well as in several regions. The comparison of operational skill to potential predictability provides a general overview of the ``skill gap" which may be closed with improvements in initialization procedures and model physics. This comparison does, however, come with some caveats due to differences in the statistical characteristics of the perfect model and the climate system it represents. Together, the operational skill assessment of different versions of CanSIPS and the potential predictability experiment conducted on one of its constituent models, CanCM4, demonstrate that while room for improvement exists, the recent development of this forecast system has clearly increased its operational utility as a seasonal sea ice forecasting tool. / Graduate

Arctic sea ice has undergone a dramatic transformation in recent decades, including a substantial reduction in sea ice extent in summer months. Such changes, combined with relatively recent advancements in seasonal (1-12 months) to decadal forecasting, have prompted a rapidly-growing body of research on forecasting Arctic sea ice on seasonal timescales. These forecasts are anticipated to benefit a vast array of end-users whose activities are dependent on Arctic sea ice conditions. The research goal of this thesis is to address fundamental challenges pertaining to seasonal forecasts of Arcitc sea ice, with a particular focus placed on improving operational sea ice forecasts in the Canadian Seasonal to Interannual Prediction System (CanSIPS). Seasonal forecasts are strongly dependent on the accuracy of observations used as initial condition inputs. A key challenge initializing Arctic sea ice is the sparse availability of Arctic sea ice thickness (SIT) observations. I present on the development of three statistical models that can be used for estimating Arctic SIT in real time for sea ice forecast initialization. The three statistical models are shown to vary in their ability to capture the recent thinning of sea ice, as well as their ability to capture interannual variations in SIT anomalies; however, each of the models is shown to dramatically improve the representation of SIT compared to the climatological SIT estimates used to initialize CanSIPS. I conduct a thorough assessment of sea ice hindcast skill using the Canadian Climate Model, version 3 (one of two models used in CanSIPS), in which the dependence of hindcast skill on SIT initialization is investigated. From this assessment, it can be concluded that all three statistical models are able to estimate SIT sufficiently to improve hindcast skill relative to the climatological initialization. However, the accuracy with which the initialization fields represent both the thinning of the ice pack over time and interannual variability impacts predictive skill for pan-Arctic sea ice area (SIA) and regional sea ice concentration (SIC), with the most robust improvements obtained with two statistical models that adequately represent both processes. The final goal of this thesis is to improve the quantification of uncertainty in seasonal forecasts of regional Arctic sea ice coverage. Information regarding forecast uncertainty is crucial for end-users who want to quantify the risk associated with trusting a particular forecast. I develop statistical post-processing methodology for improving probabilistic forecasts of Arctic SIC. The first of these improvements is intended to reduce sampling uncertainty by fitting ensemble SIC forecasts to a parametric probability distribution, namely the zero- and one- inflated beta (BEINF) distribution. It is shown that overall, probabilistic forecast skill is improved using the parametric distribution relative to a simpler count-based approach; however, model biases can degrade this skill improvement. The second of these improvements is the introduction of a novel calibration method, called trend-adjusted quantile mapping (TAQM), that explicitly accounts for SIC trends and is specifically designed for the BEINF distribution. It is shown that applying TAQM greatly reduces model errors, and results in probabilistic forecast skill that generally surpasses that of a climatological reference forecast, and to some degree that of a trend-adjusted climatological reference forecast, particularly at shorter lead times. / Graduate