Using Numerical Weather Model to Predict Zenith Total Delay (ZTD) with Machine learning and Improve Tropospheric Correction for InSAR

Interferometric synthetic aperture radar is extensively used to detect the Earth's surface deformation and movement, including from earthquakes and subsidence.. Tropospheric delays, in particular due to water vapor in the lower atmosphere, significantly hinder the ability of InSAR to measure small deformation signals on large spatial and temporal scales. For example, inter-seismic deformation and slow-slip events are challenging to measure with InSAR because long wavelengths and time scales of deformation are hidden within large (>10 cm) tropospheric delays.

Because of this issue, many recent studies have investigated using various corrections for tropospheric noise in InSAR. These can be broken into several categories, including empirical functions such as topography, and the use of auxiliary data sources. Examples of the latter include interpolating Global Navigation Satellite Systems (GNSS) -derived zenith delays (e.g. Maubant et al., 2020), calculating delays using global atmospheric model data (GAM) (e.g. Bekaert et al., 2015), and machine learning approaches (e.g. Cao et al., 2020). The use of auxiliary data typically requires interpolation to the InSAR pixels and projection to the line-of-sight direction. To date, studies using linear interpolation or kriging have failed to fully remove tropospheric noise from interferograms due to the lack of high spatial resolution GAM data to interpret water vapor.

In this study, we investigate machine learning approaches to improve the tropospheric corrections based on GAMs. We develop spatial correlation models from weather model parameters (Temperature, Pressure, and Water vapor), latitude, and elevation, and train our model using GNSS interferometric delays. We use 6000 available GNSS station delays from the Nevada Geodetic Laboratory between 2017 and 2019 in the US region as the test data and the ERA-5 global weather model from European Centre for Medium-Range Weather Forecasts (ECMWF) as input features to develop spatial correlation and prediction models. We explore using Deep Neural Networks to improve the interpolation scheme and downscale the GAMs to InSAR resolution. We find an improvement over existing methods by utilizing both the GAM data as well as spatial information about the pixel location compared to the weather model data location.