Regional climate projections using a deep-learning-based model-ranking and downscaling framework: application to...

Area of Science:

• Climate Science
• Machine Learning
• Data Science

Background:

• Accurate regional climate projections require high-resolution downscaling of Global Climate Models (GCMs).
• Existing methods for evaluating and downscaling GCMs can be improved through advanced computational techniques.

Purpose of the Study:

• To develop and evaluate a deep-learning framework for ranking Coupled Model Intercomparison Project Phase 6 (CMIP6) GCMs.
• To refine climate projections using advanced deep learning models for high-resolution downscaling.
• To assess model performance across different Köppen-Geiger climate zones and seasons.

Main Methods:

• A Deep Learning-TOPSIS (DL-TOPSIS) mechanism was employed to rank 32 CMIP6 models based on nine performance criteria.
• Five climate zones (Tropical, Arid, Temperate, Continental, Polar) were analyzed across four seasons.
• Four deep learning models (Vision Transformer (ViT), GeoSTANet, CNN-LSTM, ConvLSTM) were used for downscaling outputs to 0.1° resolution.

Main Results:

• NorESM2-LM, GISS-E2-1-G, and HadGEM3-GC31-LL were identified as top-performing GCMs, outperforming others.
• GeoSTANet demonstrated the highest accuracy in capturing temperature extremes (TXx, TNn), with RMSE = 1.57°C, KGE = 0.89, NSE = 0.85, and r = 0.92.
• Transformer-based downscaling models generally outperformed conventional deep learning methods, though ViT struggled with fine-scale fluctuations.

Conclusions:

• The developed multi-criteria ranking framework effectively improves GCM selection for regional climate studies.
• Transformer-based deep learning models offer superior performance for high-resolution climate data downscaling.
• This scalable framework enhances climate projections, supporting climate impact assessments and adaptation planning.