Neural Network Based Quality Control of CYGNSS Wind Retrieval

Abstract

Global Navigation Satellite System – Reflectometry (GNSS-R) is a relatively new field in remote sensing that uses reflected GPS signals from the Earth’s surface to study the state of the surface geophysical parameters under observation. The CYGNSS is a first of its kind GNSS-R constellation mission launched in December 2016. It aims at providing high quality global scale GNSS-R measurements that can reliably be used for ocean science applications such as the study of ocean wind speed dynamics, tropical cyclone genesis, coupled ocean wave modelling, and assimilation into Numerical Weather Prediction models. To achieve this goal, strong quality control filters are needed to detect and remove outlier measurements. Currently, quality control of CYGNSS data products are based on fixed thresholds on various engineering, instrument, and measurement conditions. In this work we develop a Neural Network based quality control filter for automated outlier detection of CYGNSS retrieved winds. The primary merit of the proposed ML filter is its ability to better account for interactions between the individual engineering, instrument and measurement conditions than can separate thresholded flags for each one. Use of Machine Learning capabilities to capture inherent patterns in the data can create an efficient and effective mechanism to detect and remove outlier measurements. The resulting filter has a probability of outlier detection (PD) >75% and False Alarm Rate (FAR) < 20% for a wind speed range of 5 to 18 m/s. At least 75% of the outliers with wind speed errors of at least 5 m/s are removed while ~100% of the outliers with wind speed errors of at least 10 m/s are removed. This filter significantly improves data quality. The standard deviation of wind speed retrieval error is reduced from 2.6 m/s without the filter to 1.7 m/s with it over a wind speed range of 0 to 25 m/s. The design space for this filter is also analyzed in this work to characterize trade-offs between PD and FAR. Currently the filter performance is applicable only up to moderate wind speeds, as sufficient data is available only in this range to train the filter, as a way forward, more data over time can help expand the usability of this filter to higher wind speed ranges as well.