Learning Surrogates of a Radiative Transfer Model for the Sentinel 5P Satellite

Abstract

Surrogate models approximate the predictions of other models. The motivation for learning surrogate models can come from computational concerns, when the predictions of the original model are computationally expensive to obtain. In contrast, the surrogate models are computationally efficient.

In this paper, we propose a framework for machine learning of surrogate models, which operate on the same input and output spaces as their original models. Instead of learning direct mappings from the input to the output space (and vice versa), we first assess the intrinsic dimensionality of the input and output spaces and reduce it appropriately, by using PCA and autoencoders. Predictive models are learned on the reduced spaces by the use of neural networks and their predictions are mapped to the original spaces.

We apply the framework to learn a surrogate model for a complex radiative transfer model RemoTeC, designed and built at SRON in the Netherlands. The original model predicts shortwave infrared (SWIR) spectra, for a given state vector of atmospheric parameters, representative of any geo-location that the Sentinel 5P satellite may encounter. The results indicate a low dimensionality of both the input and the output space and are accurate in both the forward and reverse direction.

J. Brence and J. Tanevski—These authors contributed equally.