Sepaspour et al. (2025) Future climate prediction and projection: A systematic review of classical and advanced methodologies

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-09-27
• Authors: Reza Sepaspour, Pouria Hajikarimi, Fereidoon Moghadas Nejad
• DOI: 10.1007/s00704-025-05795-3

Research Groups

• Department of Civil & Environmental Engineering, Amirkabir University of Technology (Tehran Polytechnic), Tehran, Iran

Short Summary

This systematic review analyzes 4,276 studies (2014–2024) on climate variable forecasting methodologies, including classical, machine learning, deep learning, hybrid, and General Circulation Models (GCMs), to identify trends, gaps, and comparative effectiveness. It concludes that integrating machine learning and deep learning with high-resolution GCM outputs will be crucial for future climate forecasting.

Objective

• To systematically review and critically evaluate classical, machine learning, deep learning, hybrid, and General Circulation Model (GCM) methodologies for climate variable forecasting, analyzing their applications, advantages, limitations, and recent trends over the past decade (2014-2024).

Study Configuration

• Spatial Scale: Global (reviewing studies worldwide).
• Temporal Scale: Studies published between 2014 and 2024.

Methodology and Data

• Models used:
  • Classical statistical models: Autoregressive Integrated Moving Average (ARIMA), Seasonal ARIMA (SARIMA), Prophet model, Exponential Smoothing (SES, Holt-Winters), Box-Jenkins, Theta (θ), Seasonal-Trend Decomposition of Time Series (STL), STL with Locally Weighted Regression (STL-LOESS).
  • Machine Learning (ML): Artificial Neural Networks (ANN), Random Forest (RF), Decision Trees (DT), Gradient Boosting (GB), Support Vector Machines (SVM), Regression Models (RM).
  • Deep Learning (DL): Recurrent Neural Networks (RNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), Convolutional Neural Networks (CNN).
  • Hybrid models: Combinations such as CNN-BiLSTM, Wavelet-SVM, EMD-LSTM, Hybrid WRF-CNN, LSTM-ARIMA, SARIMA-SVM, GRU-CNN.
  • Physics-based models: General Circulation Models (GCMs) from Coupled Model Intercomparison Project Phase 5 (CMIP5) and Phase 6 (CMIP6).
• Data sources: 4,276 peer-reviewed studies identified from Scopus and Google Scholar databases.

Main Results

• A total of 4,276 peer-reviewed studies published between 2014 and 2024 were systematically reviewed.
• Classical models (ARIMA, SARIMA) remain widely applied due to simplicity and interpretability but are limited in handling nonlinear dynamics. The Prophet model is a growing modern alternative.
• Machine learning methods (ANN, SVM, RF, DT, GB, RM) demonstrate superior capability in capturing complex spatiotemporal dependencies, with Regression Models (46%) and ANNs (31%) being the most frequently used.
• Deep learning approaches (LSTM, GRU, CNN, RNN) show advanced performance for complex spatiotemporal data, with LSTM (57%) being the most widely adopted deep learning method.
• Hybrid models, which integrate multiple techniques, increasingly combine strengths for enhanced accuracy and robustness in complex forecasting tasks.
• General Circulation Models (GCMs), particularly those developed under CMIP5 and CMIP6 initiatives, provide indispensable long-term climate projections. CMIP6 models, incorporating Shared Socioeconomic Pathways (SSPs), show increased adoption due to improved precision and comprehensive scenario integration.
• Precipitation and temperature are the most frequently forecasted climate variables across all methodologies.
• The review emphasizes that the integration of machine learning/deep learning with high-resolution GCM outputs, coupled with advances in high-performance computing, will be central to the next generation of climate forecasting.

Contributions

• Provides a comprehensive systematic review of a wide spectrum of climate variable forecasting methodologies (classical, ML, DL, hybrid, GCMs) over the past decade (2014-2024).
• Identifies recent trends, methodological gaps, and the comparative effectiveness of different approaches in climate forecasting.
• Highlights the increasing shift towards advanced machine learning and deep learning techniques, and the growing adoption of CMIP6 models for long-term projections.
• Emphasizes the critical future role of integrating data-driven ML/DL models with physics-based GCM outputs for enhanced climate forecasting.
• Offers a critical evaluation of the advantages and limitations of each method, providing guidance for method selection based on problem complexity, time horizon, and data availability.

Funding

Not specified in the paper.

Citation

@article{Sepaspour2025Future,
  author = {Sepaspour, Reza and Hajikarimi, Pouria and Nejad, Fereidoon Moghadas},
  title = {Future climate prediction and projection: A systematic review of classical and advanced methodologies},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05795-3},
  url = {https://doi.org/10.1007/s00704-025-05795-3}
}