Improving the quality of NGS's GPS on Bench Marks with machine learning

Abstract:

The National Oceanic and Atmospheric Administration's National Geodetic Survey (NOAA's NGS) has been engaging the geodetic, survey, and geospatial communities with a crowd-sourced project called "GPS on Bench Marks." This project has been ongoing since 2014, and user participation has exploded in the last 24 months with approximately 500 submissions per month. The increasing data contributions are the direct result of NGS identifying passive control (bench marks) of geodetic importance towards improving the National Spatial Reference System (NSRS). This presentation will highlight how machine learning can be used to assess the overall quality of the GPS on bench marks dataset.

One of the ways NGS uses this data is as an independent validation for global and regional geoid models. However, there are a number of significant obstacles to ultimately arrive at a robust, validation dataset. These obstacles stem from the sheer number of bench marks available, with two independent height measurements, which are based on multiple occupations. This is compounded by observational time differences at up to decadal levels, and observations coming from thousands of different geodetic surveyors.

The general process of ‘cleaning’ the GPS on bench marks dataset has typically been done at NGS in the past for the primary purpose of including the data as constraints in a hybrid geoid model. For the latest NGS hybrid geoid model, GEOID18, this process was done much more meticulously than in prior models – resulting in a high quality dataset. However, this one-off ‘cleaning’ is not sustainable for a dataset that grows by 500+ occupations per month, but is a prime candidate for machine learning. The machine learning process is implemented in conjunction with the previous ‘cleaning’ done for GEOID18, which is used to train the machine learning model.

This presentation will highlight how machine learning is being implemented on the GPS on bench marks dataset to determine potential problems at individual bench marks; the overall performance of different machine learning models; the various model predictors that are used as input to the machine learning models and their impact on results; and the final GPS on bench marks dataset available for validation purposes.