Intercomparison of Machine-Learning Methods for Estimating Surface Shortwave and Photosynthetically Active Radiation

Abstract

Satellite-derived estimates of downward surface shortwave radiation (SSR) and photosynthetically active radiation (PAR) are a part of the surface radiation budget, an essential climate variable (ECV) required by climate and vegetation models. Ground measurements are insufficient for generating long-term, global measurements of surface radiation, primarily due to spatial limitations; however, remotely sensed Earth observations offer freely available, multi-day, global coverage of radiance that can be used to derive SSR and PAR estimates. Satellite-derived SSR and PAR estimates are generated by computing the radiative transfer inversion of top-of-atmosphere (TOA) measurements, and require ancillary data on the atmospheric condition. To reduce computational costs, often the radiative transfer calculations are done offline and large look-up tables (LUTs) are generated to derive estimates more quickly. Recently studies have begun exploring the use of machine-learning techniques, such as neural networks, to try to improve computational efficiency. Here, nine machine-learning methods were tested to model SSR and PAR using minimal input data from the Moderate Resolution Imaging Spectrometer (MODIS) observations at 1 km spatial resolution. The aim was to reduce the input data requirements to create the most robust model possible. The bootstrap aggregated decision tree (Bagged Tree), Gaussian Process Regression, and Neural Network yielded the best results with minimal training data requirements: an R 2 of 0.77, 0.78, and 0.78 respectively, a bias of 0 ± 6, 0 ± 6, and 0 ± 5 W / m 2 , and an RMSE of 140 ± 7, 135 ± 8, and 138 ± 7 W / m 2 , respectively, for all-sky condition total surface shortwave radiation and viewing angles less than 55°. Viewing angles above 55° were excluded because the residual analysis showed exponential error growth above 55°. A simple, robust model for estimating SSR and PAR using machine-learning methods is useful for a variety of climate system studies. Future studies may focus on developing high temporal resolution direct and diffuse estimates of SSR and PAR as most current models estimate only total SSR or PAR.