![]() This allows one to super-resolve census counts to a finer granularity however, since the predicted weights are relative, the method cannot be used to predict population numbers for regions without census data. 18 used a CNN to map data extracted from OSM to disaggregation weights per pixel with 200 × 200 m ground sampling distance (GSD). Recent work has employed deep learning, in particular convolutional neural networks (CNNs 17) for population estimation, in an attempt to better account for spatial context. Consequently, the method’s efficacy is reduced in countries with coarse census units, as aggregation over large spatial regions induces distribution shifts between the training and test data. 6 is that administrative regions with known census counts are used as training units, which means that the input features (covariates) must be aggregated across all pixels in such a region whereas at prediction (inference) time, density values are predicted for each individual pixel, respectively feature vector. A limitation of the method proposed by Stevens et al. 6 resort to machine learning and train a random forest model to predict population density from a set of covariates like building maps and night light images. 16 propose a weighted combination of several covariates, such as land cover and proximity to roads, to compute what portion of the population to assign to each target unit. ![]() Examples of such covariates for population disaggregation are the presence of buildings 13, building counts 10, building volumes 14, and cumulative road lengths 15. Top-down approaches 6, 11 commonly use dasymetric disaggregation to redistribute the known, spatially coarse population counts for census areas on the order of many km 2 across smaller spatial units 12-for instance square blocks of size 100 × 100 m-with the help of auxiliary variables that covary with population density. The task then becomes to disaggregate that data to a much finer resolution, often a regular grid. On the contrary, top-down approaches 6, 10 rely on census data, which ensures complete coverage at the expense of much lower spatial resolution, in some cases down to a single head count per large district. A main drawback of bottom-up methods is that local surveys will necessarily remain extremely sparse and can hardly provide enough data points to scale population mapping up to the country level. Researchers have proposed different ways to locally measure population density, such as counting the (average) number of people per rooftop area 7, 8 or, if more resources are available for the local survey, specific average densities for different types of residential zones (urban-, rural-, and non-residential) 7, 9. Bottom-up methods 7, 8 start from local surveys of population density, collected at a number of sample locations, and attempt to generalize from detailed but sparse samples to the unobserved regions to cover larger areas. Generally speaking, two different approaches have been employed for population mapping 7: bottom-up and top-down. Yet, the design of effective population density models 6 that combine such data sources with low-resolution census counts remains a challenge. Remote sensing products and other openly available geographical datasets like OpenStreetMap (OSM) can serve as auxiliary, high-resolution evidence to create fine-grained population density maps 5. The problem is especially prevalent in developing countries in the global south, where humanitarian actions are more often needed yet census data availability and quality are limited. Unfortunately, census data are often only available at very coarse spatial resolution (e.g., one aggregate number for a district with hundreds or even thousands of km 2) and therefore not suitable as a basis for local planning: whether for sustainable land use and infrastructure management or for targeted disaster relief, planners need to know in more detail where the people are. Given the rapid population growth in many regions of the world 4 and the increasing rate at which populations shift in response to environmental and social changes, it is important to maintain accurate, up-to-date maps. High-resolution population maps are crucial for many planning tasks, from urban planning 1 to preparing humanitarian actions 2 and effective disaster response 3.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |