Development of Continuous Learning, Artificial Intelligence Algorithms for Quality Control, and Autocalibration of Haptic Rain Gauge Sensors Using Near-Real-Time Precipitation Products

The advent of the Internet of Things (IoT) era has fostered previously unseen innovation in the design of low-cost sensors, capable of providing environmental data at greater spatiotemporal densities. The potential social and economic impact of these developments is significant, as measurement and data access capabilities have historically been financially and/or technically out of reach for many commercial and home users. New challenges arise, as emerging sensor technology gets deployed by users who are not familiar with calibrating and installing equipment. For this reason, in addition to the realities of utilizing low-cost sensors, data that is collected must be meticulously quality controlled (QC). Furthermore, to maximize the robustness of these datasets, instruments should be calibrated in-situ. Traditional QC is no longer capable of handling the ever-growing volume of data, the computational requirements of detailed analyses, and typically does not involve enough external datasets for proper calibration. With innovation in artificial intelligence (AI) and machine learning, in addition to more expansive datasets becoming available, a rigorous QC can now be a reality.

This presentation focuses on one such platform designed for home use, the WeatherFlow SmartWeather station. Comprised of two separate modules operating at one-minute temporal resolution: AIR (temperature, pressure, humidity, and lightning), and the focus of this presentation, SKY (wind speed and direction, insolation, and rainfall), each of which utilize innovative sensors and contain no moving parts. The rain sensor in SKY utilizes haptic sensing technology that brings the concept of a disdrometer to a home weather station. This sensor readily provides rain detection and relative intensity information, while rainfall accumulation estimates are a function of not only sensor signal but also external factors such as temperature, wind conditions, and mounting style. By combining all instrument data with high-quality, real-time precipitation products, calibrations can be achieved that provide an ever-improving level of accuracy in rainfall estimates that would otherwise not be possible.

One of the main near real-time precipitation products is the Real-Time Precipitation Gauge Quality Control (RPGQC) system. Partners Synoptic Data Corp. and MetStat, Inc. have aggregated data from over 20,000 precipitation stations across the United States and adjacent parts of Canada and Mexico. Real-time acquisition protocols coupled with a multi-sensor QC algorithm produce a dataset of 1-hour rain gauge data. This multi-sensor QC algorithm is based on surrounding gauges, radar reflectivity data, National Weather Service Stage IV gauge-adjusted radar-estimated rain data, and satellite-estimated rain data. This data source provides nearest neighbors for comparison to SKY gauges.

In addition to the RPGQC dataset used, a gauge adjusted quantitative precipitation estimate (QPE) product called MetStormLive® is integrated in the calibration of the SKY stations. This gauge adjusted QPE integrates precipitation data from a plethora of sources including the RPGQC dataset, dual-polarization radar as well as traditional low-altitude radar reflectivity, satellite precipitation estimates, and other ancillary datasets. It calibrates the radar precipitation estimates to gauge data and leverages satellite estimates to spatially interpolate between gauges in areas void of quality radar data, thus providing a seamless gridded dataset of precipitation estimates across the continental United States. Grid cell values are captured for use in the site-specific SKY gauge calibration algorithm.

By leveraging the RPGQC and MetStormLive QPE datasets in combination with AI and machine learning technologies, an adaptive continuous learning scheme is being designed and tested to auto-calibrate the SKY rain gauge sensors in near real-time. This will result in rainfall depth measurements from the SKY that are consistent with semi-professional rain gauges.