Improved population mapping for China using remotely sensed and points-of-interest data within a random forests model

Area of Science:

• Geographic Information Systems (GIS)
• Remote Sensing
• Population Geography
• Machine Learning

Background:

• Gridded population maps are crucial for understanding human distribution and are often created by disaggregating census data using remote sensing products.
• Geospatial big data, particularly points-of-interest (POIs), offers semantic information often missing in remote sensing data, presenting opportunities for more accurate population mapping.
• Machine learning advances enable modeling complex relationships between population density and various geographic factors.

Purpose of the Study:

• To develop a high-resolution gridded population map by integrating points-of-interest (POIs) with multi-source remote sensing data.
• To assess the accuracy of the developed population map against the WorldPop dataset.
• To demonstrate the utility of geospatial big data, specifically POIs, in improving population distribution modeling.

Main Methods:

• Utilized a random forests model to combine points-of-interest (POIs) data with multi-source remote sensing data (e.g., nighttime lights, land cover).
• Disaggregated county-level 2010 census population data to a 100m x 100m grid resolution.
• Validated the resulting population map by comparing its accuracy (root mean square error) with the WorldPop dataset in major Chinese cities.

Main Results:

• The developed population map exhibited higher accuracy than the WorldPop dataset, with lower root mean square errors in key urban areas (e.g., Beijing, Shanghai).
• The new method significantly reduced the over-allocation of population in rural areas and under-allocation in urban areas observed in the WorldPop dataset.
• Demonstrated the effectiveness of incorporating points-of-interest (POIs) data in enhancing the precision of population distribution mapping.

Conclusions:

• Integrating points-of-interest (POIs) with remote sensing data and machine learning significantly improves the accuracy of gridded population mapping.
• Geospatial big data, particularly POIs, provides valuable semantic information that enhances population distribution models.
• This methodology holds promise for mapping other socioeconomic parameters with greater spatial accuracy in the future.