Çiftçi et al. (2026) Deep Learning-based Seasonal Forecasting Over K-means-derived Climate Zones in Türkiye

Identification

• Journal: Earth Systems and Environment
• Year: 2026
• Date: 2026-02-11
• Authors: Nida Doğan Çiftçi, Sefer Baday, Ahmet Duran Şahin
• DOI: 10.1007/s41748-026-01038-1

Research Groups

• Faculty of Aeronautics and Astronautics, Department of Meteorological Engineering, Istanbul Technical University, Istanbul, Türkiye
• Department of Informatics Applications, Institute of Informatics, Istanbul Technical University, Istanbul, Türkiye

Short Summary

This study developed an integrated framework using K-means clustering, Principal Component Analysis (PCA), and Long Short-Term Memory (LSTM) deep learning to redefine climate zones and enhance seasonal climate forecasting in Türkiye, demonstrating that cluster-based forecasts significantly reduce errors compared to aggregated approaches.

Objective

• To develop an integrated approach that redefines climate zones and enhances seasonal forecasts through multivariate clustering, climate extremes (using ETCCDI indices), and deep learning in Türkiye, aiming to improve predictive accuracy and regional specificity for climate risk assessment.

Study Configuration

• Spatial Scale: Türkiye, utilizing daily observations from 82 ground-based meteorological stations.
• Temporal Scale: Daily observations from 1993–2022 (30 years) for clustering and PCA, with seasonal forecasts generated for the year 2023.

Methodology and Data

• Models used: K-means clustering, Principal Component Analysis (PCA), Long Short-Term Memory (LSTM) deep learning model, and Linear Regression (LR) as a baseline for comparison.
• Data sources: Daily observations of minimum temperature (TN, in degrees Celsius), maximum temperature (TX, in degrees Celsius), and total precipitation (RR, in millimetres) from 82 ground-based meteorological stations operated by the Turkish State Meteorological Service (TSMS). These data were quality-controlled and homogenized using the climatol R package. 29 climate extreme indices (ETCCDI) were calculated from these daily observations.

Main Results

• K-means clustering, particularly with k=10, provided the most detailed and meteorologically consistent climate zones for Türkiye, effectively balancing regional gradients and local heterogeneity.
• PCA of the 29 ETCCDI indices revealed that precipitation extremes (e.g., PRCPTOT, R95P, RX1DAY) primarily dominated the first principal component (PC1), explaining 41.2% of the total variance. Temperature extremes (e.g., TX90P, TN10P, TXx, TNn) loaded on both PC1 and PC2 (explaining 23.7% of the variance), indicating distinct climatic drivers.
• LSTM-based seasonal forecasts demonstrated clear spatial and seasonal asymmetries in predictive skill:
  • Minimum Temperature (TN): Root Mean Squared Error (RMSE) was lowest in summer (approximately 1-2 °C), moderate in spring and autumn (approximately 2-3 °C), and highest in winter (up to 4-5 °C in cold clusters like C8). Relative errors were notably high in cold clusters (e.g., over 100% in C8 during winter).
  • Maximum Temperature (TX): RMSE generally ranged from 2-3 °C, with the lowest errors in summer (1.5-2.5 °C) and the highest in spring (approaching 5 °C in C8). Relative errors were mostly below 15-20%.
  • Total Precipitation (RR): RMSE showed broader ranges, typically 50-70 mm in winter (exceeding 120 mm in the wettest cluster C9). The most skillful season was summer, with RMSE values of 10-20 mm in some clusters. Relative errors ranged from 20-30% in the wettest clusters (C9) to over 60-80% in drier clusters (C6-C7).
• The integrated k-means–PCA–LSTM framework demonstrated that cluster-based forecasts significantly reduced errors compared to aggregated approaches, providing a more robust basis for climate risk assessment.

Contributions

• Developed and validated an integrated framework for climate zone redefinition and enhanced seasonal forecasting in Türkiye, combining K-means clustering, PCA of ETCCDI indices, and LSTM deep learning.
• Provided a statistically robust and physically interpretable climate zoning framework for Türkiye, capturing both large-scale climatic gradients and localized microclimatic features.
• Demonstrated that training deep learning models within objectively defined climate zones significantly improves seasonal forecast accuracy and reliability compared to national-scale or aggregated approaches.
• Offered one of the first applications of a deep learning framework for seasonal forecasting tailored to Türkiye’s redefined climate zones, providing actionable information for climate risk assessment and adaptation strategies.

Funding

Not applicable.

Citation

@article{Çiftçi2026Deep,
  author = {Çiftçi, Nida Doğan and Baday, Sefer and Şahin, Ahmet Duran},
  title = {Deep Learning-based Seasonal Forecasting Over K-means-derived Climate Zones in Türkiye},
  journal = {Earth Systems and Environment},
  year = {2026},
  doi = {10.1007/s41748-026-01038-1},
  url = {https://doi.org/10.1007/s41748-026-01038-1}
}