Estimating the Magnitude of OMPS Nadir Mapper Geolocation Anomalies using Machine Learning

Estimating Magnitude of OMPS Nadir Mapper Geolocation Anomalies using Machine Learning

Warren Porter, Banghua Yan and Ninghai Sun,

(To be submitted to the 105th AMS Annual Meeting
Topic: Automated Handling of On-Orbit Satellite Anomaly using Artificial Intelligence)

Abstract

Since October 2011, the STAR Integrated Calibration/Validation System (ICVS) has monitored the quality of Science Data Records (SDR) of observations from the Ozone Monitoring Suite Nadir Mapper (OMPS NM), flying onboard the Suomi National Polar-orbiting Partnership (SNPP) and Joint Polar Satellite System (JPSS)-1 (NOAA-20) and JPSS-2 (NOAA-21). In particular, ICVS monitors the geolocation quality of this data using Visible Infrared Imaging Radiometer Suite (VIIRS) M1 band data as a reference. Recent errors with the GPS clock mechanism aboard SNPP have degraded geolocation results multiple times in 2024, which affect geolocation accuracies of both VIIRS and OMPS NM. Accurate geolocation is vital to the useful interpretation and utilization of OMPS NM SDRs and geolocation errors can lead to further errors in products dependent on OMPS NM SDRs. Existing geolocation monitoring methods fail to detect geolocation errors that are common to both VIIRS and OMPS NM; therefore, an alternative approach is needed to monitor geolocation accuracy of OMPS SDR data, unburdened from dependency on the geolocation accuracy of VIIRS onboard the same satellite.

This study seeks to leverage the correlation between OMPS NM Radiance observations in a small area of interest (AOI) from different platforms to construct a generalized linear model (GLM) to estimate the magnitude of geolocation errors from decreases in the correlation coefficient. This GLM is trained using synthetic errors inserted into nominal OMPS NM SDRs of varying magnitude that cause the correlation coefficient to decrease proportionately.

Computation of these correlation coefficients requires each observation to be paired up with the nearest neighboring observation from another OMPS NM sensor (VIIRS sensor) housed on another JPSS platform or an observation from an earlier day from the same JPSS platform, which is assumed to not be experiencing any geolocation errors. Identification of nearest neighboring observations is done using a k-d tree, an unsupervised machine learning technique.

This study compares the performance of this technique for various AOIs. AOIs that perform best because they are more sensitive to smaller geolocation errors feature few clouds and strong heterogeneity, such as coastal deserts in the Middle East and South America. This technique also performs better for estimating errors with S-NPP OMPS NM because there is only a 50-minute difference between its observations and the corresponding observations from NOAA-20 and NOAA-21. On the other hand, the 100 minutes separating NOAA-20 from NOAA-21 observations leads to a lower overall correlation coefficient, which can disguise decreases in correlation due to geolocation errors.

The quality of the analysis is further increased by developing and applying new quality control (QC) criteria in the training dataset. Eliminating observations matched with other observations more than 20 km away is an import QC method as this pairs of nearest neighbors demonstrate poor correlations with each other because the observations are so far apart. Dropping them from the training dataset increases the overall correlation whenever the artificial perturbations are less than the distance threshold.

Disclaimer: the presentation contents are solely the opinions of the authors and do not constitute a statement of policy, decision, or position on behalf of NOAA or the U. S. Government.