A Deep Learning–Based Automated Cloud Amount Estimation Method Using Sky Imager Images

Cloud amount is a key meteorological variable that directly affects surface solar radiation and precipitation variability. However, cloud amount observations from the Korea Meteorological Administration’s Automated Surface Observing System (ASOS) rely on subjective judgments by human observers and suffer from low temporal resolution. Satellite-based cloud products also face limitations in adequately capturing point-scale cloud variability due to spatial resolution constraints.
To address these limitations, this study developed an automated cloud amount estimation system by integrating a sky imager with deep learning techniques. A dataset was constructed using sky imager images collected from January 2020 to July 2025 and corresponding concurrent ASOS cloud amount observations. A CNN-based classification model and a U-Net–based segmentation model were independently developed. The CNN model estimates cloud amount at the image level, while the U-Net model performs pixel-level cloud segmentation using cloud masks generated by a normalized Red–Blue Ratio (nRBR) algorithm as ground truth data.
Validation results show that the CNN model achieved a correlation coefficient (R) of 0.95 and an RMSE of 1.27 when compared with ASOS observations, while the U-Net model achieved a cloud detection accuracy of approximately 0.97, demonstrating stable reproduction of cloud distributions. The proposed system enables rapid cloud amount estimation from high–temporal-resolution continuous observations and suggests its potential applicability to photovoltaic power forecasting as well as agricultural and meteorological applications.