Content area
Climate change significantly impacts global socioeconomic systems, highlighting an urgent need for accurate and reliable climate predictions. Recently, deep learning techniques have shown considerable potential in climate science due to their strong capabilities in processing complex datasets. However, several critical challenges remain, including limited spatial resolution of climate data, inadequate utilization of climate teleconnection information, and unexplored numerical forecasting capabilities of Large Language Models (LLMs). Focusing explicitly on climate teleconnections, this thesis systematically addresses these challenges through three interconnected research tasks.
First, we propose a Temporal-aware Implicit Neural Representation Interpolation method, enabling flexible reconstruction of high-resolution climate data at arbitrary spatial scales. By effectively integrating temporal information and inherent physical characteristics of climate data, the proposed method significantly improves data quality and model generalization capabilities, thus providing essential data support for subsequent teleconnection-based studies.
Second, to effectively leverage the cross-variable, cross-region, and cross-scale information embedded within climate teleconnections, we introduce three innovative information-sharing strategies: (1) a Vision Transformer model integrating interseasonal and interannual spatiotemporal information, significantly improving ENSO (Niño3.4) prediction accuracy and lead times; (2) a multitask deep learning framework that captures complex spatial and temporal dependencies among different climate variables and regions, considerably enhancing precipitation forecasting across the contiguous United States; and (3) a knowledge distillation approach leveraging the ClimaX foundation model, efficiently transferring large-scale information into compact models, improving prediction accuracy and physical interpretability.
Finally, we systematically explore and evaluate the numerical forecasting capabilities of GPT-4o, specifically examining its performance in teleconnection-driven short-term (15-day) and long-term (12-month) rainfall predictions. By analyzing its predictive reliability and limitations under various scenarios, we clarify the potential and constraints of LLMs in climate forecasting, providing valuable directions for their future integration with domain-specific expert models.
Collectively, this thesis significantly advances the theoretical understanding and methodological innovation of deep learning for climate teleconnections, providing robust tools and valuable insights for future climate prediction research.
