Population Estimates in Informal Settlements Using Object-Based Image Analysis and 3D Modeling

By Almeida et al. , posted on August 16th, 2011 in ArticlesEarth ObservationUrban Monitoring

 
  • Facebook
  • email
  • Twitter

By Cláudia M. Almeida, Cléber G. Oliveira, Camilo D. Rennó, and Raul Q. Feitosa

Urban population has been drastically increasing in developing countries in recent decades, especially in informal settlements, which are characterized by high occupation densities. In Brazilian big cities, multistory residential buildings have also started to appear in informal urban areas. Assessing the total number of inhabitants in such places has become crucial for public policies and local management. Population estimates that do not depend on field surveys are strategic in these squatter settlements because of the safety risks to field workers. This article proposes an innovative method for estimating population in a squatter settlement of Rio de Janeiro city using not only 2D but also 3D information derived from digital elevation models obtained by orbital very high resolution images, demonstrating that remotely sensed data can provide fast and fairly accurate results that can be updated continuously in the inter-census periods.

I. INTRODUCTION

According to a 2006 UN Habitat report, at the global level, 30 percent of all urban dwellers lived in slums in 2005, a proportion that has not changed significantly since the beginning of last decade. Of these slum dwellers, 60 percent, or 581 million, are living in Asia; 20 percent, or 199 million, in Sub-Saharan Africa; and 14 percent, or 134 million, in Latin America [1]. If the share of slum dwellers remains roughly stable, however, the absolute number of their inhabitants keeps growing steadily. In the last 15 years, 283 million additional slum dwellers have joined the global urban population, and it is presently believed that the total has already surpassed one billion.

In most cities in developing countries, we clearly observe a struggle for urban space. As a result, there has been a mushrooming of high-rise residential buildings in the most recent decades, initially in central neighborhoods of big cities. Now, high-rise buildings have started to appear on the outskirts of big cities and in medium-size and small cities as well. Until a few years ago, the construction of multistory buildings was restricted to the formal city. Nevertheless, this is a phenomenon that nowadays also takes place in squatter settlements of big cities in Brazil.

Informal settlements, especially vertical ones, concentrate a great number of dwellers in high occupation densities. Population estimates that do not depend on field surveys are crucial in such settlements because of the safety risks to field workers. Moreover, conventional estimates based on either on-site sampling or head counts are time-consuming and expensive. These estimates are necessary to support local management activities, and the development and implementation of public policies and security and civil defense initiatives related to emergency and rescue services in cases of conflicts or hazards such as floods, fires and landslides.

II. POPULATION ESTIMATES

Since the end of the 1970s and the early 1980s remote sensing data have been used worldwide for the purpose of assessing the surface and the population density of human settlements at different scales. According to [2], population estimates can be performed at the local, regional and national level based on: 1) counts of individual dwelling units [3]; 2) measurement of urbanized land areas [4]; and 3) estimates derived from land-use/land-cover classification [5]. Further alternative approaches considered texture analysis [6] and residual kriging [7].

To date, all of the studies based on measurement of urbanized land areas and land-use/land-cover classification worked on a two-dimensional basis, i.e., flat habitable areas. But a clear future trend will be to embody the third dimension in population studies in those cases where multistory residential buildings are found. This is especially true in big cities of developing countries, and in Brazil, Russia, India and China, where there are commonly tens or even hundreds or thousands of high-rise residential buildings in a single town.

A couple of studies have attempted to include 3D information derived from LIDAR in population estimates for formal urban areas [8], [9]. This work, however, focuses on an innovative method to derive population estimates with the aid of an object-based image classification and 3D information obtained by digital elevation models (DSM and DTM) of IKONOS images for a vertical squatter settlement in Rio de Janeiro city.

 

Fig. 1. (a) Map of Brazil and Rio de Janeiro. (b) Satellite view of Jacarepaguá. (c) True color composition of the Rio das Pedras squatter settlement. Image acquired by IKONOS satellite on January 26, 2007.

III. STUDY AREA

 

The investigated squatter settlement, which is named Rio das Pedras, is located in Jacarepaguá, a western sector of Rio de Janeiro city. The site is situated on marshlands, close to the Pedras river (Fig. 1). The dwellers themselves executed the filling works for occupying the area, which experienced two occupation booms: in the mid-1960s and in the early 1980s. Multistory buildings started to appear in the mid-1990s.

IV. CONCEPTS AND METHODS

A. DSM and DTM Generation

In order to cope with the third dimension of the study area, a digital surface model (DSM) was initially built using a stereo pair of IKONOS panchromatic images, acquired on January 26, 2007. The software Geomatica 10.0.3 executed the following processing operations: 1) collection of tie points (TPs); 2) estimation of the model parameters based on the image metadata and stereo pair orientation; 3) generation of epipolar images for the whole scene; 4) calculation of parallaxes through stereo-correlation; and 5) DSM generation.

The IKONOS RPC model provides a functional relationship from the object space to the image space [10]. The RPC functional model consists in a ratio of two cubic functions of the object space coordinates. Separate rational functions usually relate normalized (scaled and offset) line and sample coordinates (xijyij) to normalized latitude, longitude and ellipsoidal height (ø, λH). The RPC model is given as [10]:


 
 
 
 
 
 
 
where xijyij are image coordinates; ø, λH are latitude, longitude and height; and the polynomial (k = 1, 2, 3, 4) has the form below:

P(ø, λH) = C1 + C2λC3. ø + C4H + C5λ

C6λH + C7. ø. H + C8λ2 + C9. ø2C10H2

+ C11. ø. λH + C12λ3 + C13λ .ø2 + C14λH2       (3)

C15λ2. ø + C16. ø3 + C17. ø. H2 + C18λ2H

C19. ø2H + C20H3

 

Fig. 2. (a) Left epipolar image and (b) right epipolar image, both oriented to the East. (c) DSM for the whole scene, oriented to the North (UTM/WGS84). The study area is located close to the bottom right corner.

The epipolar images (Fig. 2a and 2b) were generated and oriented as a function of IKONOS orbit. After the stereo pair orientation and the generation of epipolar images, the parallaxes were automatically calculated. This task was accomplished by means of search and correlation windows that locate the tie points (a total of 35) in both images. The correlation measure employed by the Geomatica OrthoEngine algorithm is the normalized cross-correlation coefficient. The generated DSM (Fig. 2c) was used for orthorectifying the panchromatic and the multispectral bands, which were later pansharpened using the principal components method.

 

The digital terrain model (DTM) was generated by means of an OrthoEngine filtering operation employing the tie points collected on the ground. For the special purpose of this work, the elevation model actually used in the calculations resulted from the subtraction between the DSM and the DTM, containing information on the buildings’ heights and their distribution over the study area.

B. Object-Based Image Analysis

 

Fig. 3. Schematic view of the semantic network showing the land-cover classes found in the study area belonging to Levels 2 and 1. Level 3 only differentiates blocks, street network and water streams.

Object-Based Image Analysis (OBIA) belongs to the major field of expert systems, a subfield of artificial intelligence, and employs modeling strategies based on thematic or specific knowledge inserted by an expert user. The basic principles of OBIA comprise multiresolution segmentation; hierarchical semantic networks meant for expert knowledge modeling and storage; and fuzzy logic for handling uncertainties inherent in the input data. OBIA was used for discriminating residential from non-residential areas according to the land-cover classification and to the average target size. A three-level semantic network was conceived based on heuristics, in which a GIS layer containing the city streets network was segmented and classified at Level 3. A discrimination among vegetation, shadow, sandy soil, asphalt and built-up areas was carried out and assigned to Level 2. A more refined classification of urban land cover was finally executed at Level 1, in all cases considering spectral and geometric information (Fig. 3)

 

C. Population Estimates

The 37 census districts contained within Rio das Pedras were used to calculate the projected area of the habitable surface in this squatter settlement. For each of the districts, all areas corresponding to classes related to the residential use were summed up (e.g., asbestos, dark concrete, etc.), according to the OBIA classification results. Big targets not related to the residential use, shadow on streets, clay bare soil, water streams, metallic roofs and vegetation were excluded from the calculation.

 

Fig. 4. Object-based classification of urban land cover in the Rio das Pedras squatter settlement, conducted at Level 1 of the semantic network.

The population of Rio city in 2000 was compared to the estimated population in 2006 (produced by the Brazilian Institute for Geography and Statistics [IBGE]), and was assessed by field interviews at sampled dwellings. The IKONOS stereo pair was acquired in January of 2007, very close to the end of 2006, when the estimates were released, so 2006 was selected as the reference year. The percentage increase in population for the whole city was applied to each of the districts.

 

The population density was calculated according to data supplied from thePereira Passos Institute (IPP) in Rio as a function of the inhabitants of the settlement divided by its area. It was assumed that the density remained constant from 2000 to 2006. And finally, a height factor derived from the DSM was generally applied to all of the districts. This factor results from the division of the volume related to the residential use by the average headroom for the settlement, set as 2.1 m. A height factor of 1.3 m was employed, corresponding to a 30 percent increase in the orthogonally projected residential area.

In order to evaluate the estimation results statistically, a hypothesis test was employed, using a paired t-test for means. The null hypothesis (H0: μR – μE = 0) states that there are no significant difference between the means of the reference (μR) and the estimated data set (μE), and hypothesis one (H1: μR – μE ≠ 0) states the opposite.

V. RESULTS

The final classification of urban land cover for the study area conducted at Level 1 of the semantic network with its respective legend is presented in Fig. 4.

Table showing STATISTICAL ACCURACY OF THE POPULATION ESTIMATESThe 2000 Census reported the population in Rio city at 5,613,897, and IBGE estimated the 2006 population at 6,132,652. This increase percentage of 9.24 percent was applied to the 2000 population of each census district in the squatter settlement so as to generate the 2006 reference data. In this way, the total population in the settlement in 2006 was 43,342 inhabitants, derived from the total sum of population for each district, and the population estimated with the OBIA classification and the IKONOS digital height model was 43,295 inhabitants. Table 1 presents the results of the statistical accuracy tests.

VI. DISCUSSION

The statistical tests showed a very good Pearson correlation coefficient (ρ) and a fairly good coefficient of determination (R2). Most importantly, the pvalue result surpassed 5 percent, which means that the reference and the estimated data sets present no significant difference between their means.

Although the digital height model resulting from the subtraction between the DSM and DTM provided accurate estimates, 3D models with finer elevation accuracy would be desirable, especially when dealing with dwelling units of reduced size, as observed in squatter settlements. In such cases, the height factor could be customized for each census district.

The reference data could be refined at the informal settlement level, in the sense that the projected population rate adopted would concern not the city of Rio de Janeiro as a whole, but specifically the Rio das Pedras settlement.

VII. CONCLUSION

The association of OBIA and 3D information derived from a digital height model from IKONOS proved to be effective for the purpose of population estimates. The main advantage of conducting population estimates by means of remotely sensed data lies in the fact that they offer fast and fairly accurate results, and can be continuously updated in the inter-census periods.

The authors acknowledge that this is a work in progress and will continue to tackle the many open questions both in the DSM/DTM generation and in the classification process. As directions for future work, the authors plan to carry out a trend analysis and a statistical validation for the digital surface model, execute accuracy tests for the object-based image analysis and explore 3-D models with higher elevation accuracy, all in support of developing more consistent methods for population estimates.

REFERENCES

[1] United Nations Habitat. (2006). State of the World´s Cities 2006/7. [Online]. Available:www.unhabitat.org/pmss/getElectronicVersion.asp?nr=2101&alt=1

[2] J. R. Jensen, Remote Sensing of the Environment: An Earth Resource Perspective, 2nd ed. Upper Saddle River, NJ: Prentice Hall, 2007, 592 p.

[3] C. P. Lo, “Automated population dwelling unit estimation from high-resolution satellite images: A GIS approach,” Int. J. Remote Sens., vol. 16, pp. 17-34, 1995.

[4] A. J. Tatem, A. Noor, and S. Hay, “Defining approaches to settlement mapping for public health management in Kenya using medium spatial resolution satellite imagery,” Remote Sens. Environ., vol. 93, pp. 42-52, 2004.

[5] F. M. Henderson and Z. G. Xia, “SAR applications in human settlement detection, population estimation and urban land use pattern analysis: A status report,” IEEE Trans. Geosci. Remote Sens., vol. 35, pp. 79–85, Jan. 1997.

[6] X. Liu, K. Clarke, and M. Herold, “Population density and image texture: A comparison study,” Photogramm. Eng. Rem. S., vol. 72, pp. 187-196, 2006.

[7] X. H. Liu, P. C. Kyriakidis, and M. F. Goodchild, “Population-density estimation using regression and area-to-point residual kriging,” Int. J. Geogr. Inf. Sci., vol. 22, pp. 431-447, 2008.

[8] J. Im, Z. Lu, L. Quackenbush, and K. Halligan, “Population estimation based on residential building volume using LIDAR remote sensing,” Proc. 55th Annu. Meet. Assoc. Am. Geogr., Washington, D.C., 2010.

[9] F. Qiu, H. Sridharan, and Y. Chun, “Spatial autoregressive model for population estimation at the census block level using LIDAR-derived building volume information,” Cartogr. Geogr. Inf. Sci., July, 2010.

[10] J. Grodecki, “IKONOS stereo feature extraction – RPC approach,” in Proc. ASPRS Conf. 2001, St. Louis, MO, 2001.

Cláudia M. Almeida obtained her B.Sc. in Town Planning at the University of São Paulo, Brazil, her M.Sc. in Infrastructure Planning from the University of Stuttgart, Germany and her Ph.D. in Remote Sensing from both the Centre for Advanced Spatial Analysis of the University College London (CASA-UCL), United Kingdom, and the National Institute for Space Research (INPE), São José dos Campos, Brazil.

Since 2004, she has been working as a Researcher at INPE. Her research interests include remote sensing, high spatial resolution sensors, fuzzy classification, spatial dynamic modeling and cellular automata.

Cléber G. Oliveira received his B.Sc. in Cartographic Engineering at São Paulo State University – UNESP, in Presidente Prudente, SP, Brazil, his M.Sc. and his Ph.D. in Remote Sensing at the National Institute for Space Research (INPE), São José dos Campos, Brazil.

His research interests include SAR images processing and classification, stereoscopy, radargrammetry and polarimetry.

Camilo D. Rennó obtained his B.Sc. in Agronomy Engineering at São Paulo State University – UNESP, in Presidente Prudente, SP, Brazil, his M.Sc. and his Ph.D. in Remote Sensing at the National Institute for Space Research (INPE), São José dos Campos, Brazil.

Since 2002, he has been working as a Researcher at INPE. His research interests include digital images processing, applied statistics, hydrological modeling and environmental studies related to the Amazon forest.

Raul Q. Feitosa received his B.Sc. in Electronic Engineering from the Aeronautics Technological Institute – ITA, in São José dos Campos, Brazil, and his Ph.D. in Computer Science from the Friedrich Alexander University Erlangen (FAUEN), Nürnberg, Germany.

Since March 1993, he has been working as a Professor at the Catholic University of Rio de Janeiro, Brazil. His research interests include digital images processing, high spatial resolution sensors, remote sensing, cognitive methods and computer visualization.

 
Advertisements