The top-of-atmosphere (TOA) albedo, a key component of the earth's energy balance, can be monitored regularly by satellite observations. Compared to the previous study Song et al. (2018), this paper estimates TOA albedo by directly linking Advanced Very High Resolution Radiometer (AVHRR) narrowband reflectance with TOA broadband albedo determined by NASA's Clouds and the Earth's Radiant Energy System (CERES) instead of Moderate Resolution Imaging Spectroradiometer (MODIS). The TOA albedo product developed in this study has an increased spatial resolution, from 1° to 0.05°, and its starting year has been extended from 2000 to 1981, compared to the CERES TOA albedo product. Models of lands and oceans are established separately under different atmospheric and surface conditions using gradient boosting regression tree (GBRT) method instead of the linear regression models in the previous study. The root mean square errors (RMSEs) of the cloudy-sky, clear-sky and snow-cover models over land are 11.2%, 9.2% and 2.3%, respectively; over oceans they are 14.6%, 10.6% and 5.6%, respectively. Compared to Song et al. (2018), the improvements of the three models over land are 28.8%, 29.2% and 68.6%, respectively. Compared to the CERES product, the new product is much more accurate than that from our previous study. The global monthly mean differences of the TOA albedo obtained with the GBRT product and CERES from 2001 to 2014 are mostly less than 5%. • A novel algorithm for TOA albedo retrieval using machine learning model. • The first long-term (1981–2014) global high resolution (0.05°) TOA albedo dataset from AVHRR. • Comparison with the CERES product shows high accuracy and quality of the generated product. [ABSTRACT FROM AUTHOR]