Data-Driven Sparse Sensor Selection for Observing-Network Optimization and Its Impact on Data Assimilation

Abstract This study primarily focuses on the optimized planning of observational networks, using data assimilation and numerical weather prediction (NWP) as evaluation frameworks. The target observations are ground-based Global Navigation Satellite System (GNSS) precipitable water vapor (PWV) retrievals currently operated in Japan. Using a greedy sparse sensor placement algorithm, 500 operational GNSS observation sites were ranked according to their contribution to minimizing the domain-averaged PWV reconstruction error over a ten-year period, with respect to the NWP analysis fields used as initial condition in Japan’s operational NWP system. The ranked observation sites were divided into three groups—top, middle, and bottom 100—and each group was assimilated separately into the data assimilation system to evaluate its impact. Assimilating the top 100 sites yielded better results than assimilating the middle or bottom 100 sites. Significant improvements were found in the PWV analysis field, mid- to lower-tropospheric humidity, and temperature throughout the troposphere. Although the magnitude of improvement decreased with forecast lead time, assimilation of the top-ranked sites led to statistically significant gains in PWV forecasts and maintained significant improvements in humidity and temperature at multiple vertical levels. These results demonstrate that assimilation of objectively selected, high-ranked observations consistently enhances forecast skill compared with lower-ranked observations. Previous studies have rarely examined whether objectively ranked sparse subsets of observations can produce meaningful impacts within operational systems. The novelty of this study lies in linking climatology-based sparse sensor ranking with an operational data assimilation framework.