| Abstract [eng] |
Gridded population maps are essential tools for resource allocation, urban planning, and disaster response. Traditionally, methods for creating gridded population maps often simply redistribute people based on settlement patterns and use a large amount of derived data, such as distances to various objects. As a result, maps produced by these models can be quite inaccurate and are not always available due to data limitations. This study addresses this problem by proposing a model, based on CNN and U-Net architectures, that can produce gridded population maps with 10 meter resolution, which relies solely on satellite imagery. This approach enables the generation of such maps for any location and time period where satellite data is available. The study also introduces a new metric for evaluating the accuracy of population density maps, ΔAvg.%, which shows the percentage difference between census based and model generated gridded population maps across different population density intervals. |
