Assessing the predictability of meteorological variables via spatial correlations using echo state networks

Collecting spatially dense meteorological observations is essential for accurate climate modeling and prediction. However, such observations are often costly and difficult to obtain. This limitation motivates the prediction of meteorological variables at locations where direct observations are sparse or unavailable, using limited data from distant observation points. In this study, we address a fundamental question: How does the predictability of meteorological variables depend on the spatial distance between the prediction target and the observation point? To cope with limited data, we employ the echo state network, which is a lightweight machine-learning model within the reservoir computing framework. Predictive models are constructed separately for near-surface temperature and atmospheric pressure using time-series data of meteorological variables obtained from a high-quality climate reanalysis dataset. Our analyses not only confirm the qualitative property that prediction accuracy degrades as the distance between the target point and observation point increases but also provide quantitative estimates of the geographical ranges over which predictions can be made within an acceptable margin of error. These results suggest that prior analysis of spatial correlations in limited observational data can be used to estimate meteorological predictability in advance, thereby contributing to the development of climate modeling with improved efficiency in data collection and computation.