Bias Correction, Anonymization, and Analysis of Smartphone Pressure Observations Using Machine Learning and Multiresolution Kriging

Abstract

With over a billion smartphones capable of measuring atmospheric pressure, a global mesoscale surface pressure network based on smartphone pressure sensors may be possible if key technical issues are solved, including collection technology, privacy, and bias correction. To overcome these challenges, a novel framework was developed for the anonym-ization and bias correction of smartphone pressure observations (SPOs) and was applied to billions of SPOs from the Weather Company (IBM). Bias correction using machine learning reduced the errors of anonymous (ANON) SPOs and uniquely identifiable (UID) SPOs by 43% and 57%, respectively. Applying multiresolution kriging, gridded analyses of bias-corrected smartphone pressure observations were made for an entire year (2018), using both anonymized (ANON) and nonanonymized (UID) observations. Pressure analyses were also generated using conventional Meteorological Assimilation Data Ingest System (MADIS) surface pressure networks. Relative to MADIS analyses, ANON and UID smartphone analyses reduced domain-average pressure errors by 21% and 31%, respectively. The performance of smartphone and MADIS pressure analyses was evaluated for two high-impact weather events: the landfall of Hurricane Michael and a long-lived mesoscale convective system. For these two events, both anonymized and nonanonymized smartphone pressure analyses better captured the spatial structure and temporal evolution of mesoscale pressure features than the MADIS analyses.