Imaging Spectroscopy of Urban Environments
- 208 Downloads
Future spaceborne imaging spectroscopy data will offer new possibilities for mapping ecosystems globally, including urban environments. The high spectral information content of such data is expected to improve accuracies and thematic detail of maps on urban composition and urban environmental condition. This way, urgently needed information for environmental models will be provided, for example, for microclimate or hydrological models. The diverse vertical structures, highly frequent spatial change and a great variety of materials cause challenges for urban environmental mapping with Earth observation data, especially at the 30 m spatial resolution of data from future spaceborne imaging spectrometers. This paper gives an overview of the state-of-the-art in urban imaging spectroscopy considering decreasing spatial resolution, the related user requirements and existing knowledge gaps, as well as expected future directions for the work with new data sets.
KeywordsImaging spectroscopy Hyperspectral Urban Unmixing Spatial resolution Environmental Mapping and Analysis Program (EnMAP)
In 2018, 55% of the world’s population is estimated to live in cities and the numbers of urban dwellers and of mega cities will continue to increase rapidly (UN 2018). The twenty-first century is therefore referred to as the Century of the City (Seto et al. 2010). Urbanization constitutes a demographic, socioeconomic and biophysical process that substantially contributes to global environmental change (Lambin et al. 2001; Pickett et al. 2011). The expansion of impervious areas, the loss of natural environments as well as the increased landscape fragmentation along with the concentration of human consumption, production and waste generation alter biogeochemical cycles, climate regimes, hydrological systems and biodiversity at multiple scales (Grimm et al. 2008; Lawler et al. 2014). Globally, urban areas are a major source of CO2 and other greenhouse gas emissions (Kennedy et al. 2009). At the regional and local scale, cities are characterized by the urban heat island, that is, an increased air temperature (Grimmond 2007; Oke 1982), and by air pollution (Fenger 1999; Hatt et al. 2004). Urban ecosystem condition, its functioning and spatial–temporal change influence the quality of life and human health (Harlan et al. 2006; Pauleit et al. 2005; Tan et al. 2010). At the same time, contemporary urbanization has the chance to promote ecological sustainability due to increased environmental awareness and returns from innovation, productivity and efficiency (Grimm et al. 2008; Seto et al. 2010). To ensure that policy sustains benefits of urbanization and anticipates or manages the negative consequences of urban growth (UN-Habitat 2010), the interactions between socioeconomic and environmental processes in urban landscapes need to be better understood (Alberti 2005). Integrated research targeting the diverse aspects and implications of urban systems in times of rapid urbanization is of great importance to face the complexity of urban dynamics.
Remote sensing can contribute in many ways to the analysis of urban environments and provides substantial information on ecosystem characteristics (see Small et al. 2018 for a general overview on urban remote sensing). We define the term urban environment in this paper as contiguous areas of anthropogenic, artificial surfaces used for transport, commerce, production, administration or housing, plus the included or adjacent vegetation surfaces that are intensively managed and directly influenced by the artificial cover. Land cover composition as detailed as construction material abundance (Herold et al. 2003; Roessner et al. 2001), vegetation characteristics such as type, structure or condition (Alonzo et al. 2015; Degerickx et al. 2018; Delegido et al. 2014; Pontius et al. 2017), information on vertical structures and surface roughness (Zhou et al. 2005), water run-off potentials (Weng 2001) or spatial–temporal surface temperature changes (Deng and Wu 2013; Imhoff et al. 2010) can be deduced when the full spectrum of remote sensing technology is considered, including multi- and hyperspectral optical sensors, thermal sensors, microwave and lidar systems (Small et al. 2018). This way, ecologically relevant information such as spatial patterns of impervious and vegetated surface, biomass estimates and inputs for microclimatic or hydrological models can be derived (Carlson and Arthur 2000; Heldens et al. 2017; Huang et al. 2013; Ngie et al. 2014; Voogt and Oke 2003).
Given the fine-scale spatial patterns and highly frequent spectral variation of urban surfaces, the spatial resolution of remote sensing imagery becomes a key factor in mapping urban environments. Thus, the soon-to-come step from airborne to spaceborne imaging spectroscopy marks a major challenge for urban remote sensing. This overview paper reflects on the state of imaging spectroscopy and the respective potentials and challenges in the context of future spaceborne imaging spectrometers such as the Environmental Mapping and Analysis Program EnMAP (Guanter et al. 2015) or the Hyperspectral Infrared Imager HyspIRI (Hochberg et al. 2015).
2 State-of-the-Art in Imaging Spectroscopy of Urban Environments
Field and laboratory spectroscopy allows for the most detailed spectral measurements of urban materials or urban vegetation, given the very high number of spectral bands, for example, 2150 in the case of the full-range Analytical Spectral Devices (ASD) spectroradiometer, indefinitely high spatial detail, and (at least partially) standardized illumination. Multiple spectral measurements from the field or laboratory are usually collected in spectral libraries and analyzed for characteristics that have been acquired in parallel to the measurement, for example, material composition or biophysical variables. However, only few results from urban areas using field- or laboratory-based spectral libraries are reported in the literature. Herold et al. (2004) developed a comprehensive field spectral library of more than 4500 individual spectra for the vicinity of Santa Barbara, USA. In a systematic statistical analysis, they revealed opportunities and limitations for separating urban land cover types and for identifying important wavelength regions for urban remote sensing. Ben-Dor et al. (2001) developed an urban spectral library from a collective spectral database (Price 1995) to explore the visible and near-infrared spectral characteristics of urban cover types typical for Tel-Aviv, Israel. In a follow-up study, the use of the entire visible to shortwave-infrared region based on a spectral library obtained from in situ measurements for the recognition of urban materials was illustrated (Ben-Dor 2001).
Pure image spectra that have been extracted from spectrally homogeneous regions in high spatial resolution airborne imaging spectroscopy data constitute an alternative means to develop or complement urban spectral libraries. For example, Heiden et al. (2007) developed a comprehensive field and image spectral library of more than 21,000 urban materials representative for many German cities. They explored this library to identify unique material-specific spectral features that are robust against spectral overlap between material classes and within-class variability. Findings obtained from field, laboratory or high spatial resolution image spectral libraries mainly contribute to an improved understanding of general spectral properties of urban materials, including their brightness (albedo), slope and shape of the broad continuum, absorption features or separability.
Beyond the description of urban materials, spectral measurements are also used to describe vegetation properties and the relevance of very high spectral detail has, for example, been shown in the context of mapping urban air pollution based on leaf chlorophyll content and fluorescence yield of urban street trees (Brackx et al. 2017; Van Wittenberghe et al. 2013). Even though the mixed pixel challenge is not tackled at this scale, insights into opportunities and limitations of imaging spectroscopy for urban mapping at airborne and spaceborne scales are provided.
To effectively cope with the spectral mixing without information loss, quantitative mapping approaches, that is, approaches estimating land cover fractions within pixels, are used to estimate sub-pixel fractions of urban land cover at the pixel scale. Okujeni et al. (2013) showed that the high information content of airborne imaging spectroscopy data allowed for extended VIS fraction mapping, that is, impervious, the I, into roofs and paved surfaces and vegetation, the V, into trees and low vegetation, in a support vector regression approach that used synthetically mixed training data from a 3.6-m HyMap data image endmember spectral library. Similar results for an extended VIS framework are reported by Degerickx et al. (2017) on the same data set using MESMA with different endmember selection techniques and Wetherley et al. (2017) using AVIRIS NG data at 4 m with MESMA, as well as by Chen et al. (2017) using reflective optics system imaging spectrometer (ROSIS) data at 1.3 m covering the visible to near-infrared range and a more recent sparse spectral unmixing approach.
Beyond mapping and quantifying of urban land cover, the information in high spatial and spectral resolution airborne data can be used to characterize the state and condition of urban green. Degerickx et al. (2018) employ 2-m resolution APEX data in combination with airborne lidar data to retrieve the chlorophyll content and leaf area index of individual urban trees in the city of Brussels, Belgium, using partial least squares regression. By combining these outcomes into an objective urban tree health assessment, which is shown to match the results from a traditional visual tree health assessment, the authors clearly illustrate the potential of imaging spectroscopy for urban tree health monitoring.
With airborne imagery at coarser spatial resolution of > 5 m ground sampling distance, classification into discrete classes has hardly been used in the urban context and quantitative mapping techniques are becoming even more important. Several authors report on the strength of imaging spectroscopy data in this context. Wetherley et al. (2017) compared results derived from AVIRIS NG at 4 m and from AVIRIS Classic at 18 m using MESMA with different image endmember library variants. The two AVIRIS sensors were flown at different altitudes and could, hence, be processed at native resolutions. For their best fraction estimates, the authors reported similar accuracies for the green vegetation, impervious and pervious (mean R2 of 0.933 at 4 m and 0.900 at 18 m). For spectrally more similar cover types in the extended VIS model, that is, turf grass, tree, paved, roof, non-photosynthetic vegetation and soil, they discovered a decrease in accuracies (mean R2 of 0.837 at 4 m and 0.760 at 18 m). Despite this decrease, these findings indicate the possibility to work with more detailed classes when adopting coarser resolution data. Okujeni et al. (2014) transferred their findings from 3.6 m data to 9 m and reported high accuracies for the work with synthetic training data when combined with support vector regression, but also other machine learning approaches. More detailed maps on urban vegetation also require quantitative mapping frameworks. Gu et al. (2015), for example, used non-metric multidimensional scaling ordination (NMDS) to map canopy traits in 11.8-m AVIRIS data. Based on these trait maps, they predicted dominant species composition of urban forest patches. However, studies at spatial scales > 5 m do not achieve the accuracies of studies with higher spatial resolution (Delegido et al. 2014) or studies including structural information from, for example, lidar sensors (Alonzo et al. 2015).
Nowadays, spaceborne imaging spectrometers can acquire data at minimal ground sampling distances of 30 m, as will be the case for EnMAP or HyspIRI (Guanter et al. 2015). So far, only few studies in urban environments have used Hyperion imagery, which comes closest to the characteristics of future systems but lacks sufficient radiometric quality in the shortwave-infrared (Kruse et al. 2003). All these studies aim for land cover fractions instead of discrete classification. Weng et al. (2008) extracted the fraction of impervious surface from Hyperion data with an RMSE of about 17% using a VIS approach that separated high and low albedo impervious surfaces as suggested by Wu and Murray (2003) based on multispectral data. Fan and Deng (2014) achieved similar accuracies for impervious surface fractions, while separating tree and non-woody vegetation in an enhanced MESMA approach. The spaceborne compact high-resolution imaging spectrometer (CHRIS) only covers the wavelength range from 410 to 1050 nm and can therefore achieve 18-m ground instantaneous field of view. It is capable of tilting to derive quasi-simultaneous image acquisitions at multiple viewing angles and has been used by various authors in an urban context (e.g., Demarchi et al. 2012a; Duca and Del Frate 2008; Licciardi and Del Frate 2011). For example, Duca and Del Frate (2008) used multi-angular acquisitions from the CHRIS instrument and improved classification accuracies of an extended VIS scheme (dark asphalt, bright asphalt, buildings, among other cover types) by almost 7% compared to a single nadir acquisition. Demarchi et al. (2012b) also showed an improvement in the accuracy of quantitative mapping of imperviousness from CHRIS data when using multi-angle instead of single, close-to-nadir imagery.
The limitations of existing spaceborne imaging spectrometers have led to an increased use of data from end-to-end sensor simulations in the context of mission preparations for EnMAP or HyspIRI (Roberts et al. 2012; Segl et al. 2012). During end-to-end simulation, an at-sensor raw digital number (DN) image is modeled for an area with known reflectance factors, assuming realistic solar illumination and atmospheric conditions. The reflectance factors may stem from high spectral and spatial airborne imagery or from forward modeling of, for example, vegetation canopies. The full radiative path from Sun to the geographic location of the area and to the sensor is simulated, followed by a simulation of the signal’s modulations within the sensor and during recording. The resulting raw DN product is afterward preprocessed to level 2a using the standard ground segment workflow. Several authors have worked on urban environments with future Earth observation imaging spectroscopy data from such simulations. Okujeni et al. (2015) used simulated EnMAP data for mapping extended VIS components. The study proved both the capability of imaging spectrometry data to quantify spectrally similar cover types like roofs and pavements, or low and high vegetation at 30 m and the superiority of the higher spectral resolution in comparison to multispectral Landsat data. This is in line with results from Roberts et al. (2012) using simulated HyspIRI data with MESMA and from Rosentreter et al. (2017) who used simulated EnMAP data and a multi-class SVR. They all conclude that mapping class fractions at the thematic detail of extended VIS from spaceborne imaging spectroscopy is worthwhile.
3 User and Observational Requirements
The literature overview in the previous section clearly shows that imaging spectroscopy data is mainly used for the identification of materials and the estimation of their associated land cover. Based on the reported state-of-the-art, the following user and observational requirements can be derived.
The requirements in terms of temporal resolution strongly depend on the thematic classes to be studied. On the one hand, individual images are usually sufficient to map anthropogenic surfaces. Ideally, the acquisition date and time minimize the share of shadow and the degree of reflectance anisotropies. The combination of leaf-on and leaf-off data from the same year, on the other hand, can help quantifying deciduous tree cover or surface cover underneath canopies. The advantage of multiple images per year is even higher when mapping tree species: information from various dates to better describe phenology stages will increase separability and improve characterization of vegetation structure and vigor (e.g., using multispectral data refer to Pontius et al. 2017; Tigges et al. 2013; Wirion et al. 2017). Denser time series of acquisitions may help characterize phenology and the influences of, for example, neighboring artificial surfaces on microclimate models.
Radiometric resolution requirements are not as high as, for example, in aquatic applications (see Giardino et al. 2018, this issue). However, the analysis of images with a high amount of shadow may benefit from high radiometric resolution and better signal-to-noise ratios in order to better differentiate land cover in shaded areas (van der Linden and Hostert 2009).
Looking at spectral resolution requirements, the benefits of the high spectral information content have been demonstrated in many studies. Particularly, when classification schemes aim for a high level of detail, such as material composition, additional spectral information leads to significantly higher accuracies at all spatial scales (Herold et al. 2003, Okujeni et al. 2015). The description of vegetation characteristics in urban environments also benefits from high spectral resolution (Alonzo et al. 2014; Delegido et al. 2014; Launeau et al. 2017).
Therefore, there is evidence that data from future spaceborne spectrometers will contribute relevant information to studies at 30-m scale and for subsequent analyses, for example, environmental models, at such spatial scales, despite the loss of spatial detail compared to airborne sensors. The final accuracy and level of detail that can be achieved with actual data cannot be predicted based on simulated data from end-to-end simulation. The reflectance surfaces underlying such simulations always include artefacts following the complex surface geometry of the urban environment, for example, through object displacement (van der Linden and Hostert 2009), or illumination anisotropies from large view angles in the original airborne data (Schiefer et al. 2006).
In comparison with other environments, the requirements for mapping and monitoring urban areas with Earth observation data may to some extent be similar to those from geological applications (e.g., van der Meer et al. 2012), where material identification and fraction estimates prevail. Only relatively large and homogeneous urban vegetation surfaces, for example, larger forest patches, may require similar approaches as forest mapping outside urban areas (see Hill et al., under review, this issue).
4 Challenges and Future Directions
Nowadays, spaceborne multispectral imagery is available from finer than 1 m spatial resolution for the analysis of urban environments. In the near future, imaging spectroscopy data will be acquired at 30 m resolution with sensors such as EnMAP and HyspIRI, as well as the Hyperspectral Imager Suite (Matsunaga et al. 2017) and PRISMA (Guarini et al. 2017). These sensors enable repeated observations almost everywhere on the globe under more standardized acquisition conditions and with operational preprocessing. Still, the selection of data sets by users working in urban environments will always depend on the intended level of spatial detail. Given the technical trade-off in spectral and spatial resolution, spaceborne imaging spectroscopy data alone will hence not automatically become number one choice, for example, by urban planners, due to the small size of most urban objects.
To incorporate the potential of higher spectral resolution, data fusion techniques can be expected to play a pivotal role in fostering the use of such data. For the fusion with secondary optical data, approaches based on Sentinel-2 Multispectral Imager (MSI) data appear promising (Yokoya et al. 2016). The relevance of increased spatial detail in combination with high spectral information content for urban mapping has already been shown using CHRIS data (Demarchi et al. 2012b).
The benefits of including structural information especially from lidar are mentioned in Sect. 2. With the Carnegie Airborne Observatory-2 (Asner et al. 2012), similar multi-sensor constellations have been successfully implemented for airborne observation. Such structural information may be used at the original 30 m or even higher resolutions by downscaling the spectral signal as done by Alonzo et al. (2014) for analyzing single urban trees. With detailed digital surface models being available for many urban areas around the world as well as products from microwave sensors such as TanDEM-X, more fusion approaches become feasible, once imaging spectroscopy is available at higher frequencies. Attempts to achieve this in urban environments with multispectral data exist (Zhu et al. 2012).
For studies aiming at monitoring temporal developments or mapping entire regions as well as for comparative studies on urban areas at continental or even global scale, future spaceborne imaging spectrometers will introduce major advantages with regard to data availability: (1) at least for selected regions, multiple observations per month become possible and (2) data will be globally available under stable sensor characteristics and with standardized preprocessing workflows. These changes offer several opportunities for the work with imaging spectroscopy data. Yet the suite of analysis approaches will have to be adapted at some points, before being able to make ideal use of the new wealth of data. With regard to multi-temporal data sets, for example, time-series analysis has to move beyond the analysis of simple, mostly vegetation-related indices that may also be calculated from multispectral data. Such indices are better derived from instruments such as Sentinel-2 with significantly higher repetition rate and spatial detail. Instead, new spectral indices or feature extractions from waveband regions only present in imaging spectroscopy data are required (Leitão et al. 2015). This way, temporal changes may be incorporated in the analysis of urban areas to observe urban growth, densification or changes in composition.
When the same sensor and preprocessing are used for comparative analysis, the transfer of models, spectral libraries and even expert knowledge, for example, on characteristic spectral features as described in Heiden et al. (2007) will offer new ways for developing biophysically based semiempirical or purely empirical mapping models.
This will, however, require joint activities, for example, the generation of urban spectral libraries including image spectra from cities in all regions that shall be monitored. First attempts in this direction are already being made for foliar traits (e.g., Singh et al. 2015) and may in similar ways be extended for urban areas. The idea of using synthetic mixtures between different surface spectra from spectral libraries (Okujeni et al. 2017) to map surface fractions in imagery of different spectral resolution has been proven useful (Priem et al. 2016, 2017). Similarly, Wetherley et al. (2017) used spectra from images at different spatial resolutions and this way improved results at all scales. In addition, more work is needed to adapt spectral libraries to new image data sets, by pruning techniques (Degerickx et al. 2017) or extension of libraries by automated integration of unknown surface material spectra (Jilge et al. 2017).
Finally, the production of more detailed and accurate maps requires new frameworks that allow several application fields to make use of the same high-quality maps derived from urban imaging spectroscopy data. In the context of land cover mapping, advances were made regarding land cover classification scheme standardization. Most of urban planning activities are based on urban spatial units such as land parcels, urban blocks, urban structure types, urban biotopes or administrative units. However, the nomenclature of such units is highly complex and lack standardization and consistency. This led, for example, to the concept of Functional Urban Areas in the Urban Atlas of the European Copernicus Services (http://land.copernicus.eu/local/urban-atlas). Environmental modelers need more physical-based parameters such as abundance of impervious areas, the building density and the abundance, structure and condition of vegetation. Over the past years, the concept of local climate zone mapping, for example, has become widely recognized even beyond pure climate science. It characterizes regions of uniform surface cover, structure, material and human activity (Bechtel et al. 2015; Stewart and Oke 2012). Based on spaceborne urban surface material mapping, any urban spatial unit can be characterized and thus can support the standardized physical description of cities.
The future availability of spaceborne imaging spectroscopy data will offer room for new analysis approaches of urban environments, yet efforts at several ends are needed for an operational use of the new data.
The authors are grateful to the editors of the special issue and the organizers of the ISSI workshop on Exploring the Earth’s Ecosystems at a Global Scale with Imaging Spectroscopy Data in Bern, Switzerland, in November 2016, where the present paper was framed. The work of the authors was supported by the German Federal Ministry of Economic Affairs and Energy in the framework of the EnMAP Core Science Team (FKZ 50EE1622) and by the Belgian Science Policy Office in the framework of the Stereo III Project UrbanEARS (SR/00/307).
- Ben-Dor E (2001) Imaging spectrometry for urban applications. In: van der Meer FD, De Jong SM (eds) Imaging spectrometry—basic principles and prospective applications. Remote sensing and digital image processing, vol 4. Springer, Dordrecht, pp 243–281Google Scholar
- Ben-Dor E, Levin N, Saaroni H (2001) A spectral based recognition of the urban environment using the visible and near-infrared spectral region (0.4–1.1 µm). A case study over Tel-Aviv, Israel. Int J Remote Sens 22:2193–2218Google Scholar
- Chen F, Wang K, Van de Voorde T, Tang TF (2017) Mapping urban land cover from high spatial resolution hyperspectral data: an approach based on simultaneously unmixing similar pixels with jointly sparse spectral mixture analysis. Remote Sens Environ 196:324–342. https://doi.org/10.1016/j.rse.2017.05.014 CrossRefGoogle Scholar
- Damm A (2008) Hyperspektrale Fernerkundung zur Ableitung pflanzenphysiologischer Parameter von Stadtbäumen – Strahlungstransfermodellierung für Berliner Kastanienbestände. Dissertation, Humboldt-Universität zu BerlinGoogle Scholar
- Demarchi L, Canters F, Cariou C, Licciardi G, Chan JC-W (2014) Assessing the performance of two unsupervised dimensionality reduction techniques on hyperspectral APEX data for high resolution urban land-cover mapping. ISPRS J Photogramm Remote Sens 87:166–179. https://doi.org/10.1016/j.isprsjprs.2013.10.012 CrossRefGoogle Scholar
- Fan F, Deng Y (2014) Enhancing endmember selection in multiple endmember spectral mixture analysis (MESMA) for urban impervious surface area mapping using spectral angle and spectral distance parameters. Int J Appl Earth Obs Geoinf 33:290–301. https://doi.org/10.1016/j.jag.2014.06.011 CrossRefGoogle Scholar
- Guarini R, Loizzo R, Longo F, Mari S, Scopa T, Varacalli G (2017) Overview of the Prisma space and ground segment and its hyperspectral products. Paper presented at the proceedings of 2017 ieee international geoscience and remote sensing symposium, July 23–28, 2017, Fort Worth, TexasGoogle Scholar
- Herold M, Schiefer S, Hostert P, Roberts DA (2007) Applying imaging spectrometry in urban areas. In: Quattrochi DA, Weng QH (eds) Urban remote sensing. CRC Press Inc., Boca Raton, pp 137–161Google Scholar
- Hill J, Buddenbaum H, Townsend PA Imaging spectroscopy of forest ecosystems. Surv Geophys (under review)Google Scholar
- Matsunaga T et al (2017) Current status of hyperspectral imager suite (HISUI) onboard International Space Station (ISS). Paper presented at the proceedings of 2017 IEEE international geoscience and remote sensing symposium, July 23–28, 2017, Fort Worth, TexasGoogle Scholar
- Priem F, Okujeni A, van der Linden S, Canters F (2016) Use of multispectral satellite imagery and hyperspectral endmember libraries for urban land cover mapping at the metropolitan scale. In: SPIE remote sensing. SPIE, p 100080K. https://doi.org/10.1117/12.2240929
- Priem F, Canters F, Okujeni A, van der Linden S (2017) Optimizing mixed spectra generation for regression-based unmixing of land cover in urban areas. In: 2017 joint urban remote sensing event (JURSE), 6–8 March 2017, pp 1–4. https://doi.org/10.1109/jurse.2017.7924554
- Roberts DA, Quattrochi DA, Hulley GC, Hook SJ, Green RO (2012) Synergies between VSWIR and TIR data for the urban environment: an evaluation of the potential for the hyperspectral infrared imager (HyspIRI) decadal survey mission. Remote Sens Environ 117:83–101. https://doi.org/10.1016/j.rse.2011.07.021 CrossRefGoogle Scholar
- Seto KC, Sanchez-Rodriguez R, Fragkias M (2010) The new geography of contemporary urbanization and the environment. Ann Rev Environ Resour 35:167–194. https://doi.org/10.1146/annurev-environ-100809-125336 CrossRefGoogle Scholar
- Small C, Okujeni A, van der Linden S, Waske B (2018) 6.07—remote sensing of urban environments A2—Liang, Shunlin. In: Comprehensive remote sensing. Elsevier, Oxford, pp 96–127. https://doi.org/10.1016/B978-0-12-409548-9.10380-X
- UN (2018) World urbanization prospects. The 2018 revision. https://esa.un.org/unpd/wup/Publications/Files/WUP2018-KeyFacts.pdf. Accessed 7 June 2018
- UN-Habitat (2010) State of the world's cities 2010/2011 - cities for all: bridging the urban divide state of the world's cities reports. UN-Habitat, 224 pGoogle Scholar