Résumé

This document presents a systematic literature review analyzing machine learning (ML) models for climate change prediction and climate risk assessment, synthesizing findings from 40 peer-reviewed studies published between 2015 and 2024. The review employs thematic analysis to categorize the literature into five major clusters: deep learning architectures for temperature and precipitation forecasting, ensemble methods for extreme weather event prediction, hybrid physics-informed neural networks, spatiotemporal models for sea-level rise and glacier dynamics, and ML-based climate risk assessment frameworks for socioeconomic impact modeling. Key findings indicate that Long Short-Term Memory (LSTM) networks and Transformer-based architectures outperform traditional statistical models in long-range climate forecasting, while gradient boosting methods excel in regional risk classification. Physics-informed neural networks offer superior interpretability in data-scarce environments. The review identifies significant research gaps, including model interoperability and the integration of socioeconomic variables, and suggests future research should focus on federated learning approaches and explainable AI frameworks to enhance transparency and stakeholder trust. This comprehensive analysis contributes to understanding the methodological landscape of ML applications in climate science and highlights the potential for improved predictive accuracy and risk assessment capabilities.

Classification CIB

Mots-clés