The Prediction Of Finely-Grained Spatiotemporal Relative Human Population Density Distributions In China

China's transportation industry has been experiencing huge changes and the travelling frequency of citizens becomes higher and higher and more and more diversified in a short time period. The analysis and deep research on the short-term change of population density in the city-level spatial resolution are worthy of further exploration. In this study, we first used two linear regression models to build relationships between the 2010 census density, predicted 2020 census density and the Tencent density respectively to test the usability of Tencent positioning data. The Pearson's correlation coefficients r 0.58 (p < 0.01) and 0.54 (p < 0.01) demonstrates good positive correlations between the ground truth (census data) and the geographic spatiotemporal big data (Tencent positioning data), which could be used to represent the relative human population density distribution in China. Then we use the human population distribution based on the Tencent positioning dataset of China in every hour of the first 21 days, to predict the hourly distribution in the next week by seasonal autoregressive-integrated-moving-average (SARIMA) and a convolutional long short-term memory (ConvLSTM) models respectively, with the 50 x 50 km of spatial resolution. The total average of the ConvLSTM model's Root Mean Square Error (RMSE) is 139.0, while the SARIMA's one is three times greater than the value. And the coefficient of determination (R-2) values of ConvLSTM model is higher than 0.9, while the other ones are about 0.78. Comparing the two results in both time and space concludes that: the evaluation results reflected by RMSE and R-2 showed that the two models are both suitable for the prediction of Tencent density distribution in finely-grained time and space. Nevertheless, the predicted density correlated much better with the tested data at temporal and spatial scales when using ConvLSTM compared to SARIMA, and the capability of prediction in space by ConvLSTM model is more stable.