AN IMPROVED DESIGN OF SOLAR SURFACE RADIATION PARAMETERIZATION USING SATELLITE DATA

Abstract

This thesis examines solar energy system variables, including normalized reflectance obtained by a bidirectional reflectance distribution function (BRDF) model and insolation estimated by a neural network (NN) method. We analyzed MTSAT-1R and SPOT/VEGETATION images of Northeast Asia taken between January and December 2006. One research objective was to retrieve normalized reflectance using a kernel-driven BRDF model based on that of Roujean (1992). Using various statistical analyses, we tested the semi-empirical model results against Moderate-Resolution Imaging Spectroradiometer (MODIS) surface land cover types. The estimated normalized reflectance at a standard solar-target-sensor geometry was input to the NN model under cloudy conditions. Solar surface insolation (SSI) indicates how much solar radiance reaches the Earth’s surface at a specified location during the daytime and is a critical parameter for climate change estimation and numerical weather prediction. We calculated SSI by two different methods using MTSAT-1R data for Northeast Asia. For estimating incident solar insolation using remotely sensed data, the simplified physical model was generally considered the best method. This not only used geostationary satellite data to calculate SSI with a physical model, but also developed a new SSI calculation approach using an NN model to obtain more accurate results. The physical model of Kawamura used in this study was designed for SSI retrieval under various atmospheric conditions. To improve the atmospheric transmittance due to ozone while maintaining a practical atmospheric status, we substituted constant ozone values (from Ozone Monitoring Instrument [OMI] ozone data) as a function of latitude. Each transmittance was optimized for Northeast Asia using MTSAT-1R data. When conducting a NN simulation with remotely sensed data and ancillary data, the various potential input parameters should be examined in regard to their effect on the accuracy of output results. Thus, before adding input parameters to the NN, we explored the possible parameters in regard to network efficiency and the elimination of collinearity. An NN model with one hidden layer was then used to simulate SSI using early-stop and Levenberg-Marquardt back propagation (LM-BP) methods. We separated the NN architecture into two parts, one to deal with cloudy conditions and the other to deal with a clear sky; this step was conducted because different processes occur under complicated physical characteristics of clouds. The NN-model SSI estimates were compared with ground-based pyranometer measurements and showed better agreement with ground measurements than did estimates by conventional methods, especially under the clear sky condition.