Estimation of surface reflectance using deep neural network with KOMPSAT-3A data

The acquisition of image data from satellite is performed by the satellite’s sensor after the light from the sun is reflected in object at the surface. In this process, light passes through the earth's atmosphere twice and is affected by the scattering, absorption and reflection by the atmosphere. This effect of the atmosphere reduces the power of the sun's light entering the sensor and consequently influences image data. The process of removing this effect is called atmospheric correction. Generally, the radiative transfer model (RTM) such as the Second Simulation of a Satellite Signal in the Solar Spectrum (6S) is used in the atmospheric correction methods for surface reflectance retrieval. In general, RTM have high accuracy. But, RTM processing takes long time to perform atmospheric correction. So, several studies have applied the Look-up Table (LUT) method based on RTM. However, LUT is not an exact method due to the increment and range of input variables. In this study, we used the Deep Neural Network (DNN) method to predict surface reflectance for KOMPSAT-3A data. To Build an effective DNN model, 6S-based LUT is used as training data and the hyper-parameters have been adjusted. To evaluate the surface reflectance retrieval, we compared the surface reflectance derived of 6S RTM, 6S-based LUT and DNN methods.