Seasonal forecasting and trend analysis in redefinedclimate zones: A deep learning and climate indicesapproach

Abstract

Improving seasonal forecasting and long-term trend assessment requires a data-driven redefinition of climate zones that explicitly incorporates climate extremes and regional heterogeneity, particularly in climatically complex regions such as Türkiye. Climate variability and extremes over Türkiye exhibit strong spatial contrasts due to the combined influence of large-scale atmospheric circulation, complex topography, and regional-scale processes. Conventional climate classifications based primarily on mean conditions often fail to fully represent this heterogeneity, limiting their usefulness for climate change assessment and seasonal prediction. This thesis addresses these challenges through a sequence of interrelated studies that progressively examine observed climate extremes, redefine climate regions using datadriven approaches, and assessthe skill of both machine learning–based and operational seasonal forecasting systems within a regionalized framework. The first component of the thesis investigates long-term changes in temperature and precipitation extremes across Türkiye over the period 1960–2019. To capture potential regime shifts associated with accelerated warming during the late twentieth century, the analysis period is divided into two sequential sub-periods: 1960–1989 and 1990– 2019. Daily maximum, minimum, and mean temperature data, together with total precipitation, are analyzed for the four conventional seasons. The evaluation is conducted within a regional framework based on seven widely used climatological climate regions over Türkiye, allowing regional-scale differences in climate variability and extremes to be explicitly assessed. Probability Density Functions (PDFs) are employed to examine shifts in the distributions of mean temperature and precipitation between the two sub-periods, providing insight into changes in central tendency, variability, and distributional shape. In addition, a suite of extreme and relative climate indices recommended by the Expert Team on Climate Change Detection and Indices (ETCCDI) is used to quantify changes in temperature and precipitation extremes across the identified climate regions. This combined distribution-based and index-based approach enables a comprehensive assessment of both gradual climate change signals and changes in extreme behaviour. The results reveal a clear and spatially coherent warming signal across Türkiye, with temperature-related extreme indices exhibiting widespread and robust upward trends. Indices associated with minimum temperatures show particularly pronounced increases, highlighting the strong contribution of nighttime and cold-season warming to the overall climate signal. The TXn index, representing the monthly minimum of daily maximum temperature, exhibits an increase of approximately 2 °C per decade, with most stations showing a consistent rise between the two sub-periods. These findings demonstrate that temperature-related extremes constitute the most robust and dominant manifestation of climate change across Türkiye. In contrast, precipitation extremes display substantially greater spatial heterogeneity and do not exhibit a uniform national-scale trend. While annual total precipitation and extreme precipitation indices such as R95p and R99p increase in some regions—most notably parts of the Black Sea region—they decrease or remain unchanged in Mediterranean, continental, and transitional climate zones. Overall, the first study establishes that warming-related extremes represent the most consistent climate change signal, whereas precipitation extremes remain regionally differentiated and inherently less predictable. Recognizing the limitations of conventional climate classifications in representing this spatial complexity, the second component of the thesis introduces a data-driven framework to redefine climate regions and enhance seasonal climate prediction. New climate zones are objectively identified using k-means clustering, an unsupervised machine learning method, applied to multivariate daily observations of minimum and maximum temperature and total precipitation from 82 stations covering the period 1993–2022. A range of cluster solutions (k = 2–15) is evaluated, with a 10-cluster configuration identified as providing the most effective balance between large-scale xxvii climatic gradients and local heterogeneity. To characterize the climatic behaviour of the newly defined regions, climate extremes are quantified using ETCCDI indices, and Principal Component Analysis (PCA) with biplot visualization is applied to identify the dominant drivers of variability. The PCA results indicate that precipitation extremes dominate the leading mode of variability, while temperature extremes contribute to both primary and secondary components, reflecting their widespread but spatially differentiated influence. These findings confirm that climate extremes, rather than mean climate variables alone, play a central role in defining regional climatic characteristics. Building upon this regionalization, seasonal forecasts for temperature and precipitation are generated using a Long Short-Term Memory (LSTM) deep learning model trained separately within each cluster. This cluster-based forecasting approach reveals pronounced spatial and seasonal asymmetries in forecast skill. Temperature forecasts exhibit the highest accuracy during summer, with root mean square errors typically between 1 and 2 °C, while errors increase during winter due to enhanced atmospheric variability. Precipitation forecasts show larger errors overall, with skill varying strongly across regions and seasons, and the largest uncertainties occurring during winter and in wetter clusters. Importantly, the cluster-based approach consistently outperforms aggregated, domain-wide forecasting methods, demonstrating the value of regionally tailored prediction frameworks. The third component of the thesis extends this regionalized perspective to the evaluation of an operational seasonal forecasting system. The seasonal forecast skill of the European Centre for Medium-Range Weather Forecasts (ECMWF) Seasonal Forecast System Version 5 (SEAS5) 51-member ensemble is assessed for temperature and precipitation over Türkiye for the period 1993–2022 using the previously derived cluster-based framework. Seasonal forecasts are evaluated at lead times of one, two, and three months (lead-1, lead-2, and lead-3), and seasonal values are derived consistently from these lead-time configurations. Forecast performance is evaluated using a comprehensive set of deterministic and probabilistic metrics, including correlation coefficients, discrimination measures, Brier Skill Scores, and Receiver Operating Characteristic (ROC) Skill Scores. Lead-time sensitivity is also explicitly evaluated to assess the robustness of seasonal forecast skill across different lead configurations. The results indicate that seasonal temperature forecasts from SEAS5 exhibit relatively high and stable skill across all clusters and seasons, with limited sensitivity to lead time. In contrast, precipitation forecasts display substantially weaker and more spatially heterogeneous skill. In many regions, particularly those influenced by complex topography and coastal processes, precipitation forecast skill approaches climatological performance, highlighting the inherent challenges of predicting precipitation at seasonal timescales. Although large-scale circulation patterns such as the NAO influence seasonal climate variability over Türkiye, their persistence is only weakly translated into meaningful local-scale forecast skill by the SEAS5 system. This limitation reflects the dominant role of regional-scale processes, internal atmospheric variability, and topographic effects in shaping precipitation predictability. Overall, skill magnitudes remain below levels required for deterministic applications, particularly for precipitation, underscoring the inherently probabilistic nature of seasonal forecasts and the importance of uncertainty-aware interpretation. Forecast skill exhibits a strong seasonal dependence, with the highest discrimination and probabilistic skill occurring during spring and summer, when large-scale circulation variability is comparatively reduced. In addition to conventional fixed seasons, forecast skill is further evaluated using overlapping three-month moving seasons, including JAS, ASO, SON, OND, NDJ, DJF, JFM, FMA, MAM, AMJ, MJJ, and JJA. This moving-season framework allows for a more continuous assessment of seasonal predictability and provides improved insight into transitions between climatological seasons. An additional assessment of drought predictability using the Standardized Precipitation Index (SPI) shows that meaningful skill is primarily achieved at seasonal accumulation scales (SPI-3), whereas short-term drought predictability (SPI-1) remains limited across most regions. These results indicate that SEAS5 can provide actionable information mainly for slowly evolving drought conditions rather than short-term hydrological extremes. Taken together, the findings of this thesis demonstrate that the integrated use of observed climate extremes, data-driven regionalization, deep learning–based forecasting approaches, and operational seasonal prediction systems provides a more realistic and informative framework for climate analysis over Türkiye. By explicitly accounting for regional heterogeneity and the limited transferability of large-scale circulation signals to local-scale predictability, the thesis highlights both the capabilities and the limitations of current seasonal forecasting systems.