Seung-Teak Oh, Ji-Young Lee, Yang-Soo Kim, Jae-Hyun Lim, Journal of Internet Computing and Services, Vol. 26, No. 3, pp. 75-82, Jun. 2025
10.7472/jksii.2025.26.3.75, Full Text:  HTML
Keywords: satellite image, Natrual Light, Spectral Power Distribution (SPD), Deep Learning, CCT

Abstract

As efforts in the field of lighting to reproduce the characteristics of natural light continue, the need for spectral power distribution (SPD) of natural light has increased. SPD of natural light is essential for analyzing and reproducing the characteristics of natural light, and is generally collected through actual measurements. However, it is difficult to collect the characteristics of natural light outdoors for a certain period of time due to the need for expensive spectral measurement equipment and restrictions depending on the season and weather. For this reason, related organizations cannot continuously measure and provide SPD of natural light. Some have attempted to calculate some light characteristics such as illuminance and radiance by applying deep learning to sky or satellite data, but have not been able to calculate SPD, which is the fundamental information of natural light. In addition, there have been few attempts to calculate SPD of natural light using satellite data, which is closely related to changes in weather and natural light. Therefore, in this paper, we propose a deep learning model that calculates SPD of natural light using image data from geostationary satellites. By analyzing the satellite-provided data in the visible light wavelength band and the spectrum of actual natural light, the input factors of the deep learning model such as the attenuation effective path (AEP) of sunlight are derived and a dataset is built. After that, a CNN-DNN model that calculates the SPD of natural light by inputting the AEP and the solar elevation angle is implemented. In the performance evaluation, the suitability of the proposed model was confirmed through comparison with the existing DNN model, and the accurate natural light spectrum was calculated within the error range of MAE 0.0308 W/m2 under various weather conditions (within 20% of the outlier), and the color temperature was calculated within 5% of MAE, showing that it can be applied to related fields that require natural light characteristics.

Statistics
Cite this article