Estimating the daily mean blue-sky land surface albedo on the Tibetan Plateau using convolutional neural network

Abstract

ABSTRACT Surface albedo is a critical indicator of Earth’s ability to reflect solar radiation, influencing energy balance and climate dynamics. Research on blue-sky albedo on the Tibetan Plateau has often focused on clear-sky conditions, overlooking seasonal snow cover and variable weather. This paper uses convolutional neural networks (CNNs) to estimate blue-sky surface albedo on the plateau. The residual neural network 18 (ResNet18) model outperforms other deep learning architectures, such as long short-term memory (LSTM) networks, significantly improving the correlation between in situ observations and remote sensing data. Albedo simulation accuracy is notably higher than moderate resolution imaging spectroradiometer (MODIS) products, especially in areas with homogeneous surface properties. Surface albedo variation rates show that albedo changes have not been significant, influenced by factors such as summer vegetation greenness, high-altitude glacier mass balance reduction, and winter snow cover fluctuations. These results highlight regional differences in albedo change trends on the Tibetan Plateau. This study demonstrates the effectiveness of CNNs in simulating surface albedo, offering valuable insights into albedo changes and establishing a framework for refining albedo estimation methodologies.