The work described in this paper is directed at the automated estimation of population using Supervised Classification through a GIS platform applied to satellite imagery. The motivation for the work is that the collection of population data, when conducted in the traditional manner can be time consuming and inefficient. This is especially the case in rural areas that lack of sophisticated communication and transport infrastructure, especially in developing countries. The solution proposed in this paper is based on the concept of using satellite imagery in order to estimate the population. The concept includes using a small sample of satellite images to build a classifier that can then be used to predict populated areas. The main issue to be addressed is how to best use the classified result towards matching it to the number of people in the study area. In order to correspond to the task described above, the following methodology is proposing (i) an image supervised classification technique , (ii) a statistical forecasting procedure and (iii) a regional disaggregation technique which will be furtherer explained in the following sections.
With this chapter, the methodology of downloading and analyzing Sentinel2 data shall be described. The Semi-Automatic Classification Plug-in (SCP) [3] is a free open source plug-in for QGIS that is used for semi-automatic classification or supervised classification of remote sensing images as well as tools for image preprocessing, the classification post processing, and the raster calculation and is used for obtaining the imagery needed for the purposes of this study.
In order to gather the Satellite imagery, The Copernicus Open Access Hub can be used, which provides complete, free and open access to Sentinel-1, Sentinel-2 and Sentinel-3 user products. The SCP Sentinel-2 tab allows the searching and downloading of the desired imagery (Level-1C) using the Data Hub API. The search area is defined accordingly for each one of the areas of interest to include all of the respective subregions. For the purposes of this study, the date of acquisition was defined as of 2016 or later in order to make the classification process relevant to the desired results. The maximum cloud cover shall be set not to exceed 10 % to help with the process whereas bands 1, 9, and 10 shall not be included.
In this section, the steps followed when classifying the satellite images will be explained. For the purposes of this study, several compounds of Uganda and Lusaka are examined. The two regions are classified separately. The same approach is followed for both of the study regions. It is noted that multiple different satellite images might be necessary to cover the entire study area. For the Sentinel-2 images, the bands are converted to reflectance and clipped to an extend large enough to cover each area examined so as to reduce the computational time of the process. The band set which is the input image for SCP is defined and the regions of interest are created.
Iterating through the different color composites, allows completion of the classification. The defined buildings macro-class is the area of interest that will be extracted. This shall provide us with high resolution two colored pixelated images that hold information regarding the buildings of the areas. Such data shall be used for completing the methodology through the process that will be furtherer explained in the next sections.
The classification results can be validated through comparison with the original raster files as well as old shape-file layers provided by various open data sources (e.g. OpenLayers Plug-in, Geofabrik).
Following the completion of the process described above, the two classified regions are separated using polygons that describe the sub-areas of interest. For the purposes of this study, the PNG images of the clipped images (17 for Uganda, 89 for Lusaka) are extracted, each one now containing a certain number of blank pixels corresponding to the living areas of each district. Concerning their usage in the estimation algorithm that will be described, it is important to note that all extracted classified files of each study area shall correctly correspond to the respective area's land coverage ratio.
While the results of the classification process can provide a decent overview of the population distribution, the buildings information shall be used in a way that can allow to extract a population estimation. The historical Census data of the countries can be used to predict the aggregated population result that can be later split into the study areas accordingly through the classified imagery. The statistical analysis of the Census data consists of two basic steps, (i) constructing the population growth time-series and (ii) a forecasting model in order to pertain to the out of sample observation; the current year population number.
Various Census data for demographic purposes are used for both areas [12] obtained from The World Bank organization. Concerning the areas of interest of the areas of interest of the present study, the latest population data that was available for both regions of Uganda and Zambia was that of 2015.
Following the process of collecting the population data time series, we can proceed to choosing the best fitting model to forecast the 2017 population of the two countries based on previously observed values. Regarding choosing the correct forecasting technique we shall examine several error metrics (ME, MPE, MAPE, MSE, sMAPE). The trend of our time-series shall also not be overlooked and, taking into consideration the growth ratio of the examined African regions, slightly optimistic methods are suitable for the purposes of this study. The time-series dataset that was obtained by the analysis procedure described in the previous section can be used as input in the forecasting support system OMEN [10], a fully customizable web-based forecasting tool.
A cross-validation competition between the methods Naive, MAPA [9], ETS [8], Theta [5], ARIMA [7], SES, Holt and Damped [6] is performed through the platform. The rolling origin evaluation procedure [11] is used with the validation window matching the forecasting horizon and the performance of each method is evaluated based on the minimization of the squared errors. Thus, the 'best fitted' model is selected, which in this case was the autoregressive integrated moving average (ARIMA) model for both Uganda and Zambia; suitable for low frequency and short horizons.
The processes described above provide with an arithmetic output regarding Uganda's and Zambia's total population. Such information is used to proceed to the following step which is a top-down division of the countries' population in order to come to a result about each specific area of interest in the given shapefiles. It is however important to note that the subareas of interest might not add up to the entire countries' extend and therefore a straightforward division of the entire population to the acreage of the living areas (as obtained from the classification process) might not be possible. Because of this difficulty, a different approach shall be examined. In order to split the total population into the different areas, a weight variable will be calculated for each pixel corresponding to living areas in the classified images. To achieve this, we gather Census data about housing and population in a small subregion of each one of the countries of interest and estimated the multiplier for each pixel.
Uganda For the purposes of this study, we shall first examine the areas of Uganda. Regarding the region of Katwe Lake, according to Census data obtained from the Uganda Bureau of Statistics[13], the region had a total population of about 23,559 people in 2014. The quotient of the compound's population and Uganda's total population is estimated to be equal to ∼ 0.00063368 in 2014. Assuming that the population density of the region has not changed significantly, it is estimated that the compound's population for the year of 2017 shall equal to ∼ 25, 970. We proceed to examine the classification result extracted by the methodology that was described in the previous section. We shall proceed to sum the pixels of the living areas (rgb(255,255,255)) in the region; The above process gives us a result of 2,242 total pixels that represent living areas in the compound of Lake Katwe. Thus, we calculate that the density multiplier of each pixel that represents living area in the compounds of Lusaka, shall equal to ∼ 11.5834.
Lusaka The same process is followed for the regions of Lusaka, where the district of Kanyama is examined. According to Census data from 2010 [14], the Kanyama compound had a total population of about 366,170 people in 2010 leading us to the conclusion that the quotient of the compound's population and Zambia's total population equaled to ∼ 0.026703 in that year. Assuming that the population density of the region has not changed significantly, the compound's population for the year of 2017 is estimated to be equal to ∼ 388, 002. We then proceed to examine the classification results for Lusaka; it is noted that the Kanyama compound consists of three polygon shapes: (i) Old Kanyama (ZMB TW0029), (ii) New Kanyama (ZMB TW0050) and (iii) Kanyama West (ZMB TW0078). The aggregated result for those three subregions can be examined. We shall proceed to sum the pixels of the living areas (rgb(255,255,255)) in the three districts using the same methodology. It is estimated that a result of total 64,199 pixels represents living areas in the compound of Kanyama. Therefore we calculate that the density multiplier of each pixel that represents living area in our Lusaka study area shall equal to ∼ 7.02352.
After calculating the weight variables for both regions as described above, the clipped pixelated images along with the multipliers can be used as input to calculate the population of each region. For each compound, each white pixel is multiplied by the estimated weight variable of the respective study area (11.5834 for Uganda -7.02352 for Lusaka). The outputs provide with the final dataset containing the population estimations for all the subregions of each country. It is important to note that the accuracy of the procedure described, depends largely on the classification output. A well classified image similar to the one obtained from the regions of Uganda during this study, can improve the results whereas a poorly classified region similar to the one obtained from Lusaka shall result to a dataset that differs greatly from its actual population.
4 Notes and Comments.
The results obtained from the described methodology shall provide us with the final dataset containing information on the population data for each one of the study areas. However, the same procedure can be followed for estimating the dwellings of the regions of interest by appropriately parameterizing the census data input along with the forecasting model described in the previous sections.