ABSTRACT
Although data-driven methods including deep neural network (DNN) were introduced, there was not enough assessment about spatial characteristics when using limited ground observation as reference. This work aimed to interpret the feasibility of several machine learning approaches to assess the spatial distribution of solar radiation on Earth based on the Communication, Ocean, and Meteorological Satellite (COMS) Meteorological Imager (MI) geostationary satellite. Four data-driven models were selected (artificial neural network (ANN), random forest (RF), support vector regression (SVR), and DNN), to compare their accuracy and spatial estimating performance. Moreover, we used a physical model to probe the ability of data-driven methods, implementing hold-out and k-fold cross-validation approaches based on pyranometers located in South Korea. The results of analysis showed the RF had the highest accuracy in predicting performance, although the difference between RF and the second-best technique (DNN) was insignificant. Temporal variations in root mean square error (RMSE) were dependent on the number of data samples, while the physical model showed relatively less sensitivity. Nevertheless, DNN and RF showed less variability in RMSE than the others. To examine spatial estimation performance, we mapped solar radiation over South Korea for each model. The data-driven models accurately simulated the observed cloud pattern spatially, whereas the physical model failed to do because of cloud mask errors. These exhibited different spatial retrieval performances according to their own training approaches. Overall analysis showed that deeper layers of networks approaches (RF and DNN), could best simulate the challenging spatial pattern of thin clouds when using satellite multispectral data.
