Remote Sensing and Social Sensing Data Fusion for Fine-Resolution Population Mapping With a Multimodel Neural Network

Abstract

Mapping population distribution at fine spatial scales is significant and fundamental for resource utilization, assessment of city disaster, environmental regulation, and urbanization. Multisource data produced by remote and social sensing have been widely used to disaggregate census information to map population distributions at fine resolution. However, it is challenging to achieve accurate high-spatial-resolution population mapping by combining multisource data and considering geographic spatial heterogeneity. The existing approaches do not consider global and local spatial information simultaneously, resulting in low accuracy. This article proposes a multimodel fusion neural network for estimating fine-resolution population estimates from multisource data. Our approach takes into account the local spatial information and global information of each geographic unit. Specifically, a first-order space matrix of a geographic unit is used to characterize its local spatial information. We propose a multimodel neural network, which combines a convolutional neural network and a multilayer perceptron (MLP) model to estimate a fine-resolution population mapping. Using Shenzhen, China, as the experimental setting, a population distribution map was generated at a 100-m spatial resolution. The model was quantitatively validated by showing that it captured the relationship between the estimated population and the census population at the township level ($R^2=0.77$) more accurately than the WorldPop dataset ($R^2=0.51$) and the MLP-based model ($R^2=0.63$). Qualitatively, the proposed model can identify differences in population density in densely populated areas and some remote population clusters more accurately than the WorldPop population dataset.

Keywords

• Convolutional neural network (CNN)
• multimodel neural network
• population mapping
• population spatialization
• remote sensing