This study established a random forest regression model (RFRM) using terrain factors, climatic and river factors, distances to the capitals of provinces, prefectures (Fu, in Chinese Pinyin), and counties as independent variables to predict the population density. Then, using the RFRM, we explicitly reconstructed the spatial distribution of the population density of Gansu Province, China, in 1820 and 2000, at a resolution of 10 by 10 km. By comparing the explicit reconstruction with census data at the township level from 2000, we found that the RFRM-based approach mostly reproduced the spatial variability in the population density, with a determination coefficient (R 2 ) of 0.82, a positive reduction of error (RE, 0.72) and a coefficient of efficiency (CE) of 0.65. The RFRM-based reconstructions show that the population of Gansu Province in 1820 was mostly distributed in the Lanzhou, Gongchang, Pingliang, Qinzhou, Qingyang, and Ningxia prefecture. The macro-spatial pattern of the population density in 2000 kept approximately similar with that in 1820. However, fine differences could be found. The 79.92% of the population growth of Gansu Province from 1820 to 2000 occurred in areas lower than 2500 m. As a result, the population weighting in the areas above 2500 m was~9% in 1820 while it was greater than 14% in 2000. Moreover, in comparison to 1820, the population density intensified in Lanzhou, Xining, Yinchuan, Baiyin, Linxia, and Tianshui, while it weakened in Gongchang, Qingyang, Ganzhou, and Suzhou.Sustainability 2020, 12, 1231 2 of 16 of population with regular grid cells and can be used to reveal the pattern of population growth and migration [5]. Furthermore, the gridded historical population datasets are widely used in the historical reconstruction of land use and land cover change (LUCC), such as the conversion from woodland to cropland, which is conducive to the quantitative estimations of carbon emissions in historical periods [6][7][8]. Therefore, there has been a large demand to determine the explicit spatial distribution of population.To date, there are a large number of global-and national-scale gridded population datasets including the Gridded Population of the World (GPW), Global Rural-Urban Mapping Project (GRUMP), WorldPop datasets, and China 1 km Gridded Population (CnPop) datasets [9]. These datasets played critical roles in resource allocation and management [10], climate change research [11,12], disease risk assessment [13], and other fields. These existing studies mostly focused on modern times; however, there are a few population gridded datasets for historical periods. This may be partly explained by the lack of documented historical census data.Overall, the modelling approaches of most population gridded datasets can be divided into two categories: a spatial interpolation (SITP) approach and multi-factor integration (MFI) approach. The SITP approach is based on geo-statistics. Under the SITP approach, population density is represented as a function of location, i.e., the X-coordinate and ...