Multi-input deep Neural Networks to Estimate Population from Satellite Imagery

Knowing the number and the distribution of the inhabitants on a territory, is a fundamental and necessary aspect to manage a community. The birth and mortality, employment and aging indices, calculated in relation to the population, not only have statistical or demographic value, but also make it possible to guide investments, support for employment and incentive plans for businesses. This is how census proves useful to a government and public administration in the development of a country’s economic policy. Some well-developed countries use administrative data for population estimation, or a combination of these and sample surveys. In many countries in the world, however, the census is still carried out using the traditional method of collecting data directly from the population, asked to fill in a paper or via e-mail questionnaire. Censuses of this kind require a lot of time and considerable economic resources and are therefore uncommon and often inaccurate, especially in remote and hard-to-reach areas or zona where there is a lack of institutions/organisations coordinating all the procedures. Moreover, population mapping is generally entrusted to municipalities which can carry out surveys at different times, hence introducing space-temporal heterogeneity and bias. The rapid development of the IT sector and the increasing availability of big data have led many researchers to experiment with new techniques for population estimation. In particular, the deep learning and the convolutional neural networks (CNN) could allow to obtain homogeneous population estimates without high costs, in a short time and with a good accuracy using also unstructured data, in particular satellite images, easily downloadable from the web. This article shows the results of a research conducted in two phases. In the first phase a CNN architecture Inception-ResNet-v2 has been used to provide population estimates at county and sub county level. The model converts satellite images into population density estimates. The neural network was trained using satellite images remotely sensed by Sentinel-2 in the years 2015-2016 and related to Lazio (Italy), with dimensions of 2km × 2km and a resolution of 10m. The model was first tested on the provinces of Naples and Grosseto, and subsequently on portions of the territory of Lazio. Population data provided by the National Institute of Statistics during the last ten-year census of 2011 were used to construct the target variable of the dataset and validate the results obtained. The time discrepancy between the population values and the images they refer to cannot be bridged, as the last available Istat data is, as mentioned, 2011 and Sentinel-2 was launched for the first time in June 2015. Therefore, the gap will have to be considered in the final estimates. Both regression and classification tasks were performed. In an attempt to improve these results, the use of multi-input neural networks was tested in the second phase of this research to analyse data of different nature (numeric / categorical variables, images, texts, etc.) jointly. A multi-input neural network merges in a single solution the architectures used to handle satellite images as CNN and architectures for the processing of auxiliary tabular data such as the Multi Layer Perceptron. As auxiliary information, it was deemed useful to help the network recognise the different uses to which the buildings in the images were destined. One way to obtain this information is to consider a land cover map, the Corine Land Cover (CLC). Using the map, it is possible to quantify the amount of land, in percentage terms, within each tile occupied by: built-up "Urban" type, which includes all types of residential buildings, more general "Artificial surfaces", "Agricultural areas", "Forests and semi-natural areas" and "Swamplands and Waterbodies". We created a dataset of Region of Lazio, the Province of Naples and the Province of Grosseto, with satellite tiles and CLC class percentages. Multi-input networks implemented showed a fair improvement in performance respect the first phase. In particular, it provided good results on the Grosseto dataset, while the estimates obtained on the Naples test data were still not satisfactory. The neural network showed difficulties in making correct estimates for tiles with high population density, above about 20.000 inhabitants. Focusing on the results obtained with the best performing network, it was noted that the data from Lazio and Naples, images with very similar characteristics, had very different target values: given the same number of buildings, the number of residents in Naples is much higher than in the Lazio region. This points out how this network is unable to capture the different population density that characterizes the data. We therefore sought to improve network performance by providing input information that could help the network capture the different population density. This new neural network has improved the previous ones, showing a strong generalization capacity. These studies have therefore shown that the goal of providing a tool capable of supporting the population census, exploiting now easily accessible data, such as satellite images, can be achieved. Obviously, it is essential to train the networks on a dataset that is as representative as possible. The use of auxiliary information has proved to be very useful, in fact satellite images alone can induce the network to make incorrect predictions, as the network can determine the presence of inhabited centers but is not able to distinguish the corresponding population density.