A study conducted by Shihori Koyama et al. in collaboration with The University of Tokyo and Nagoya Institute of Technology was published in Chaos.
Accurate estimation of meteorological variables such as temperature and atmospheric pressure is essential for climate studies, disaster preparedness, and environmental monitoring. However, installation of dense observation networks is often costly or difficult because of terrain, and many locations still suffer from limited data availability. Conventional weather prediction methods often rely on large datasets and substantial computational resources to achieve high-precision predictions. In contrast, little was known about the conditions under which the desired prediction accuracy could be achieved with limited observational data and computational resources.
In this study, an Echo State Network (ESN), a lightweight machine learning model based on the reservoir computing framework, was used to investigate how accurately meteorological variables can be predicted from observations collected at geographically distant locations. Analysis conducted across multiple cities located in different climate zones—including Tokyo, New York, London, Cairo, and Canberra — demonstrated that the scope of predictability can be quantitatively assessed in each region. For example, in the analysis of the Tokyo area, it was found that temperature can be predicted effectively using data from locations 200-600 km away, and pressure can be predicted effectively using data from locations 200-400 km away. The study further demonstrated that future prediction performance can be estimated in advance by analyzing correlations in historical data. This demonstrated the potential to evaluate the expected level of accuracy before actually building a predictive model.
These findings provide a practical approach for assessing predictability in advance, even when available observational data are limited. The approach could contribute to more efficient climate modeling and may also be extended to environmental monitoring applications such as air pollution estimation.
Title: Assessing the Predictability of Meteorological Variables via Spatial Correlations Using Echo State Networks
Authors: Koyama, S., Inoue, D., Yoshida, H., Aihara, K., Tanaka, G.
Journal Name: Chaos
Published: August 7, 2026