Скачать или смотреть Caleb Robinson, A Deep Learning Approach for Population Estimation from Satellite Imagery

In previous literature, questions of "where" people live and "how many" people live there have been studied independently. Population estimates are made for different administrative areas (e.g. US counties) based on cohort components or other methods; however, the locations of the people inside of the administrative area is largely unknown. Similarly, dasymetric modelling techniques for disaggregating existing population counts within administrative areas have been studied; however, there is no ground truth data with which to validate these methods. In addition, there are often large gaps between census counts in many countries. In the US, a national census is taken every 10 years, and it has only recently been supplemented with annual, non-comprehensive surveys, i.e. the American Community Survey. In some countries several decades can pass without a new census. In general, taking a census is extremely expensive and requires a great deal of organization and time. Knowing where people live and how many people live in a place are fundamental aspects of community planning at every scale. Considering this, we aim to create high resolution gridded population estimates using only satellite imagery. This jointly answers the questions of where people live, and how many people live there. Specifically, we train deep convolutional neural networks to estimate population in the US at a ~1km^2 resolution using a concatenation of Landsat 7 and Nighttime Light satellite imagery. To train our networks we disaggregate population counts at the census tract level to use as targets. We validate our model on held out data by aggregating individual grid cell estimates at the county level then comparing to the ground truth values.