An ANN-based method for population Dasymetric mapping to avoid the scale heterogeneity: A case study in Hong Kong, 2016–2021

Abstract

A comprehensive understanding of population distribution is critical for assessing socio-economic issues. However, the widely used dasymetric mapping method relies on models built at a coarse administrative scale and estimates population at a fine-gridded scale. This difference in scale between the training and estimating domains results in significant heterogeneity in data distribution. To address this issue, we proposed a scale heterogeneity-avoided method based on artificial neural networks that can take population density as an independent variable and gridded properties, including remote sensing images, digital terrain models, road networks, building footprints, and land use, as dependent variables. Our experiments in Hong Kong in 2016 and 2021 showed significant advantages of the proposed method. Compared to commonly used methods, our approach demonstrated a 19.4% improvement in the root mean square error. Furthermore, the advantages of our method became more apparent at larger census units, and the accuracy of the pre-trained model for directly estimating population in other temporal phases was satisfactory. Among the geospatial data variables, land use was the most significant in accurately estimating population. Replacing land use data with random numbers led to a decrease in accuracy by over 89.0%, while other properties only resulted in decreases of 2.7% to 13.9%. We further investigated spatiotemporal changes in population distribution from 2016 to 2021, finding that population growth mainly occurred in new built-up areas, while larger population decreases occurred in old towns. Throughout the study period, the population tended to concentrate more, as the average population density increased while the median population density decreased.