Deep learning for spatiotemporal forecasting in Earth system science: a review

Abstract

ABSTRACT Deep learning (DL) has demonstrated strong potential in addressing key challenges in spatiotemporal forecasting across various Earth system science (ESS) domains. This review examines 69 studies applying DL to forecasting tasks within climate modeling and weather prediction, disaster management, air quality modeling, hydrological modeling, renewable energy forecasting, oceanography, and environmental monitoring. We summarize commonly used DL architectures for spatiotemporal forecasting in ESS, key technical innovations, and the latest advancements in spatiotemporal predictive applications. While DL architectures have proven capable of handling spatiotemporal data, challenges remain in tackling the complexities specific to ESS, such as complex spatiotemporal data, scale dependencies, model interpretability, and integration of physical knowledge. Recent innovations demonstrate growing efforts to integrate physical knowledge, improve model explainability, adapt DL architectures for domain-specific needs, and quantify uncertainties. Finally, this review highlights key future directions, including (1) developing more interpretable hybrid models that synergize DL and traditional physical approaches, (2) extending model generalizability through techniques like domain adaptation and transfer learning, and (3) advancing methods for uncertainty quantification and missing data handling.