Smartphone Pressure Analysis with Machine Learning and Kriging

Over one billion smartphones are now capable of measuring atmospheric pressure. Distributed worldwide, smartphones offer the potential for a surface pressure observing network of unparalleled coverage and density. The Weather Company (IBM) has begun to develop such a network by retrieving pressure observations through its smartphone app. To date, they have collected over 50 billion smartphone pressure observations (SPOs); however, the utility of these observations remains limited by poor data quality. Inaccurate metadata and smartphone sensor errors make bias-correction and quality control a requirement for for SPOs. To accomplish this task, an efficient machine learning approach to bias-correction was developed. Billions of smartphone pressure errors are predicted in advance, using a gradient-boosting model trained on past in situ pressure data and real-time smartphone sensor data. Bias-correction is performed by subtracting predicted pressure errors from SPOs. Using multi-resolution kriging, 5-km gridded analyses of bias-corrected SPOs have been generated every 5-min for the year 2018. These analyses span the continental U.S. and will be used to study mesoscale phenomena associated with surface pressure features. Smartphone pressure analyses will also be compared to kriging analyses of in situ METAR and Mesonet observations, and to model analyses from the High-Resolution Rapid Refresh model.