Urban Population Distribution Mapping with Multisource Geospatial Data Based on Zonal Strategy

Abstract

Mapping population distribution at fine resolutions with high accuracy is crucial to urban planning and management. This paper takes Guangzhou city as the study area, illustrates the gridded population distribution map by using machine learning methods based on zoning strategy with multisource geospatial data such as night light remote sensing data, point of interest data, land use data, and so on. The street-level accuracy evaluation results show that the proposed approach achieved good overall accuracy, with determinant coefficient (R2) being 0.713 and root mean square error (RMSE) being 5512.9. Meanwhile, the goodness of fit for single linear regression (LR) model and random forest (RF) regression model are 0.0039 and 0.605, respectively. For dense area, the accuracy of the random forest model is better than the linear regression model, while for sparse area, the accuracy of the linear regression model is better than the random forest model. The results indicated that the proposed method has great potential in fine-scale population mapping. Therefore, it is advised that the zonal modeling strategy should be the primary choice for solving regional differences in the population distribution mapping research.

Comparison of Twelve Machine Learning Regression Methods for Spatial Decomposition of Demographic Data Using Multisource Geospatial Data: An Experiment in Guangzhou City, China

“…A reasonable inference is that due to the fact that the differences in the socio-economic environment within the region were ignored, a single or global KNN model may have led to unreasonable estimations in the model output. According to previous studies, the zonal strategy can effectively solve the shortcomings of the global model by using secondary partition modeling [23,63]. Therefore, the zonal strategy should be given priority when dealing with the shortcomings of the global model.…”

“…Therefore, determining the appropriate cell size is the first step in the spatial decomposition of demographic data. According to a previous study [23], the cell's area should be close to 10% of the smallest subdistrict are in the overall study area. Hence, the cell size of 150 m was chosen for our study.…”

“…The relative error (RE) is the ratio of the absolute error of a measurement to the measurement being taken [23].…”

“…The typical global gridded population datasets mainly include the Gridded Population of the World (GPW) [19], the Global Rural-Urban Mapping Project (GRUMP) [2], LandScan Global [20], Global Human Settlement Population Grid datasets (GHS-POP) [21], and Worldpop [22]. However, most of these datasets are designed for extensive coverage and free accessibility for the undeveloped regions in the world; their accuracy on a local scale, such as that of a city or subdistrict, still need to be analyzed [23,24].…”

High-Precision Population Spatialization in Metropolises Based on Ensemble Learning: A Case Study of Beijing, China

“…Meanwhile, since the outbreak of COVID-19 in late 2019, recurrent outbreaks have occurred due to the high population densities and the frequent population movements in metropolitan areas [3][4][5]. Therefore, understanding the accurate spatial distribution of population is of great significance for public health, urban development planning and disaster assessment management, especially in complex metropolises [6][7][8][9][10][11].…”