In this article, we present ECWL exposure assessments that address this problem by employing CoastalDEM, a new DEM developed using a neural network to perform nonlinear, nonparametric regression analysis of SRTM error. This model incorporates 23 variables, including population and vegetation indices, and was trained using lidar-derived elevation data in the US as ground truth.
Which section suggests "trained on data from areas with climates and landscapes wildly different than where the model is applied"?
EDIT: I see it now in "Methods". Here's the LIDAR dataset they used, Digital Coast coastal Lidar by NOAA: https://coast.noaa.gov/digitalcoast/data/coastallidar.html