Remote Sensing of Environment 210 (2018) 217–228
Contents lists available at ScienceDirect
Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse
Mapping patterns of urban development in Ouagadougou, Burkina Faso, using machine learning regression modeling with bi-seasonal Landsat time series
T
⁎
Franz Schuga, , Akpona Okujenia, Janine Hauera,b, Patrick Hosterta,b, Jonas Ø. Nielsena,b, Sebastian van der Lindena,b a b
Geography Department, Humboldt-Universität zu Berlin, Unter den Linden 6, 10099 Berlin, Germany Integrative Research Institute on Transformations of Human-Environment Systems, Humboldt-Universität zu Berlin, Unter den Linden 6, 10099 Berlin, Germany
A R T I C LE I N FO
A B S T R A C T
Keywords: Landsat Seasonal Time series Support vector regression Unmixing Urban Informal settlements Burkina Faso
Rapid urban population growth in Sub-Saharan Western Africa has important environmental, infrastructural and social impacts. Due to the low availability of reliable urbanization data, remote sensing techniques become increasingly popular for monitoring land use change processes in that region. This study aims to quantify land cover for the Ouagadougou metropolitan area between 2002 and 2013 using a Landsat-TM/ETM+/OLI time series. We use a support vector regression approach and synthetically mixed training data. Working with biseasonal image stacks, we account for spectral variability between dry and rainy season and incorporate a new class - seasonal vegetation - that describes surfaces that are soil and vegetation during parts of the year. We produce fraction images of urban surfaces, soil, permanent vegetation and seasonal vegetation for each time step. Statistical evaluation shows that a temporally generalized, bi-seasonal model over all time steps performs equally or better than yearly or mono-seasonal models and provides reliable cover fractions. Urban fractions can be used to visualize pixel-based spatial-temporal patterns of urban densification and expansion. A simple rule set based on a seasonal vegetation to soil ratio is appropriate to delineate areas of unplanned and planned settlements and, thus, contributes to monitoring urban development on a neighborhood scale.
1. Introduction Within the last four decades, the number of city dwellers on the African continent increased from 24% to 45% of the total population. This constitutes the highest growth in urban population worldwide and the increase is likely to continue at a similar rate. It is projected that in 2030, 770 million people will be living in urban areas in Africa, compared to 83 million in 1970 (UNESDA, 2015). This demographic development might have massive environmental, infrastructural and social impacts in and around a city. Decreased water availability and quality, extreme air pollution, densification of land usage, limited access to infrastructure such as health, schooling, electricity and sanitation as well as affected ecosystem balances and services are challenges that go along with a rapid population increase, particularly in Sub-Saharan Africa (Grimm et al., 2008; Myers, 2011; Goodfellow, 2013; Li et al., 2015a; Marlier et al., 2016). Identifying urban expansion patterns and understanding the dynamics of urban settlements as well as associated consequences in Sub-
⁎
Saharan Africa have accordingly emerged as important research topics (Murray and Myers, 2007; Myers, 2011; Parnell and Pieterse, 2014). A major challenge in this respect is a lack of data, in particular reliable census data (Beauchemin and Bocquier, 2004; Satterthwaite, 2010; Potts, 2012, 2017), that can add to the understanding of the abovementioned processes and advise urban planning (Gandy, 2006; Goodfellow, 2013; Parnell et al., 2009). This is also true for this project's study site, Ouagadougou, Burkina Faso: Low availability of historic data, such as censuses, land cover maps or cadastral archives, along with a lack of urban growth governance structures and the high uncertainty of population projections in the area (Satterthwaite, 2010), makes in-situ methods of urban growth monitoring difficult. Thus, the analysis of urban densification and expansion processes in Sub-Saharan Africa is increasingly turning towards other data sources including remote sensing images (Potts, 2012; Harre et al., 2016). Multiple studies applied remote sensing techniques monitoring land use change processes in the studied area (e.g. de Jong et al., 2000; Mering et al., 2010; Gessner et al., 2015; Akintunde et al., 2016; Hou et al., 2016). Since the
Corresponding author. E-mail address:
[email protected] (F. Schug).
https://doi.org/10.1016/j.rse.2018.03.022 Received 27 November 2017; Received in revised form 13 February 2018; Accepted 16 March 2018 0034-4257/ © 2018 Elsevier Inc. All rights reserved.
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
approach can be performed with highly generalized models in a spatial and temporal context using multi-site libraries (Okujeni et al., 2016). A requirement of regression modeling is the development of continuous training information consisting of pairs of reference signatures and land cover fractions. This data is conventionally generated by the manual definition of surface class polygon compositions in high resolution reference imagery (e.g. Sexton et al., 2013; Song et al., 2016). Okujeni et al. (2013) and Okujeni et al. (2016) demonstrate that SVR training data can also be based on synthetically mixed training spectra (SVRsynthmix) generated from pure library spectra, i.e. endmembers, obtained from the imagery without the need of high resolution reference imagery. This way, the availability of multi-temporal and retrospective reference information becomes possible. In this study, we aim to quantify land cover for the city of Ouagadougou and its surroundings with Landsat Thematic Mapper (TM), Enhanced Thematic Mapper (ETM+) and Operational Land Imager (OLI) data for the period from 2002 to 2013 using five time steps. We extend onto the concept of a generalized SVRsynthmix model as suggested in Okujeni et al. (2016) by integrating series of bi-seasonal multi-spectral images. To do so, we work with stacked imagery from both phenological extremes at the end of rainy and dry season for each time step. The first study objective consists of the reliable identification of surface fractions for the study area, including seasonal vegetation into the classification scheme. Our second objective is to map urban development over time based on those fractions. Finally we wish to delineate areas with planned and unplanned development in order to better understand the spatial-temporal patterns of urban development.
opening of the Landsat satellite archive (United States Geological Survey, 2008), the largest series of consistent space-borne earth observation data at high spatial resolution, it has been proven generally suitable for urban growth monitoring (Taubenböck et al., 2012; Sexton et al., 2013). Visually promising Very High Resolution (VHR) imagery is not sufficient for this purpose because of low and irregular image acquisition frequencies, short time series of technically consistent photographs, the lack of shortwave infrared sensors for urban surface detection and potentially large viewing angles from sensor pointing. However, general and site-specific challenges for urban remote sensing monitoring remain also for the work with archived Landsat time series. For example, large urban areas experience highly dynamic land cover changes (Lambin et al., 2003; Taubenböck et al., 2012) with several land cover types showing spectral ambiguities (Small, 2004). Particularly in semi-arid areas with mud-made buildings and dustcoated objects, urban materials and open surfaces in the surrounding areas might resemble. The resulting confusion of land cover types is a limiting factor in the delineation of settlements in urban studies. The notably high dynamics and the level of detail of urban development pose a challenge to data selection and processing. Another important challenge for the use of remote sensing approaches lies in the regional climatic conditions of Burkina Faso. Rainy seasons with heavy cloud coverage and, thus, fewer high-quality Landsat scenes lead to an unequal distribution of available imagery throughout the year. In addition, the high spectral variability of vegetated surfaces over the year with only smaller spectral variability for permanent soil and infrastructures implies seasonal differences in the spectral features of surfaces. Large areas of seasonal rain-fed vegetation and agriculture will be detected as vegetation in rainy seasons and as soil in dry seasons. That requires a general strategy of dealing with such variable spectral information and suggests the use of multi-seasonal data. In general, the use of multi-seasonal data in optical remote sensing land cover monitoring is not a new concept (e.g. Yuan et al., 2005). Performing a sensitivity analysis on Landsat imagery, Müller et al. (2015) find that seasonal information is particularly relevant for separating vegetation types in a savannah context. With regard to Burkina Faso in particular, Knauer et al. (2017) have recently shown the relevance of seasonal vegetation detection in studying agricultural processes and Zoungrana et al. (2015) suggest the use of image acquisitions from March, June, July and October for the purpose of vegetation cover analysis in Burkina Faso. With regard to urban densification, a reliable distinction of permanent soil, seasonal vegetation and permanent vegetation appears important in addition to urban infrastructures maps. This is underpinned by the fact that temporary or permanent soil coverage during expansion phases might be an indicator for increasing settlement density and ongoing urbanization. Soil, seasonal and permanent vegetation might also contribute to the detection of the type of development, i.e. if it is planned or unplanned. Although the Landsat archive offers > 40 years of mostly multiple earth observations per year, only very few studies use multi-seasonal Landsat data for detailed urban land cover mapping in semi-arid regions. Furthermore, such studies usually target at discrete urban classification schemes (e.g. Taubenböck et al., 2012; Zhu and Woodcock, 2014). Reasons for this might be manifold. The intra-annual variability of land cover and a lack of quantitative mapping approaches that lead to reliable fractional cover maps from multispectral data are among the most prominent. But the identification of urban development patterns asks for a quantification of cover fractions instead of discrete classification and requires a methodology that provides information on gradual change (MacLachlan et al., 2017) and makes use of sub-pixel information. A promising method to quantify land cover in heterogeneous urban areas is machine learning regression modeling. Machine learning algorithms generally turn out to be strong in describing multi-modal and spectrally similar classes in the urban space. Recently, it has been shown that an ensemble learning support vector regression (SVR)
2. Study area and data 2.1. Study area Our study area comprises the greater Ouagadougou metropolitan region and adjacent landscapes (Fig. 1). Counting about 1.3 million inhabitants in the last census in 2003, the capital of Burkina Faso has nearly tripled its population since 1985. With a current growth rate of 4 to 5% and a prospective population of 3.6 million in 2020 and 5.8 million in 2030 the city tends to extend its primacy in the national urban hierarchy (Harre et al., 2016; UNESDA, 2015). The reasons for this rapid urbanization are diverse, including demographic, economic and political factors (Ouédraogo, 2002; Fournet et al., 2008; Beauchemin, 2011; Nielsen and D'haen, 2015). An insufficient number of developed parcels, a lack of financial resources for their acquisition as well as land speculation have led to a massive growth of unplanned settlements in the outskirt areas. Many of these areas are marked by the establishment of uninhabited clay huts (Fournet et al., 2008), understood as an attempt to obtain registration and subsequently being attributed a parcel in case of future restructuration (Jaglin, 1995; Fourchard, 2001; Hauer et al., in press). The spatial expansion of Burkina Faso's capital Ouagadougou is shaped by two major morphologically distinguishable kinds of settlements: Formally planned settlements stand out through highly systematic infrastructural networks under municipal supervision, defined architectural specifications and documented lot ownership (Fig. 2). Unplanned settlements, however, develop beyond official planning strategies and they feature more diverse lot and building layouts and suffer from limited access to water, electricity or roads. Predominant building materials are clay, corrugated iron, concrete and, rarely, asphalt for roads. Burkina Faso is characterized by tropical summer-humid climate under the influences of a trade wind driven dry season from October to April and an Intertropical Convergence Zone driven rainy season from May to September. Whereas green vegetation largely disappears in the dry season, phenological peaks occur during the rainy season. Natural vegetation is largely shaped by rain-fed seasonal grasses, legumes and sedges as well as herbaceous woods. Cultivation encompasses maize, 218
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
Fig. 1. Used Landsat footprint (left, framed) including the study area (left, shaded). Normalized Difference Vegetation Index (NDVI) of dry and rainy season imagery for the study area in 2013/2014 (right).
Fig. 2. Aerial high resolution imagery (Google Earth) of a planned (top) neighborhood and direct comparison of planned and unplanned development (bottom) in Ouagadougou, Burkina Faso.
219
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
create a comprehensive, multi-annual image spectral library, i.e. a library containing the temporal variation of identical surfaces. We then trained multi-annual regression models using synthetic mixtures with gradually changing labels that we derived from that library of pure spectra. This multi-annual model is then applied to each of the five image pairs. Results consist of fraction images for each of the four classes in each of the five time steps, i.e. 20 fraction images in total, derived from the averaged result of all ensemble iterations (Fig. 4). 3.1. Image spectral library SVRsynthmix requires a set of spectral references dedicated to model training (spectral library). We identified four land cover types (urban surfaces, permanent vegetation, seasonal vegetation and soil) and collected spectra from each step in the Landsat time series (Fig. 5). Training spectra were respectively sampled from homogeneous surfaces based on an overlay of Landsat pixel extents and Quickbird VHR aerial imagery in Google Earth and represent bi-seasonal reflectance spectra from pure pixels within the stacked image. Ten spectra per land cover type were sampled in each of the five image stacks. Each time step contributed the same number of training spectra to an aggregated spectral library in order to encounter temporal variability, resulting in a total number of 200 spectra. Urban surfaces is the most heterogeneous class consisting of concrete and tarmac surfaces as well as corrugated iron rooftops. Seasonal vegetation, a temporally dynamic class, which resembles soil during dry season and permanent vegetation during rainy season, is supposed to contribute to understanding urban development patterns and is part of the methodological innovation of this paper (Fig. 5).
Fig. 3. Mean NDVI values for one of five selected regions of interest (ROI) in all available Landsat images for WRS-2 195/051 in the study period. Images providing consistent values for five different ROIs have been selected as dry (red) and rainy (blue) season images. Dotted lines represent the maximum and minimum NDVI within the ROI for both dry and rainy season images. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
millet, sorghum, cotton and groundnuts in rather fragmented areas with important tree and shrub population (Knauer et al., 2017; UNEP, 2010). 2.2. Imagery Our time series consisted of five time steps between 2002 and 2013 based on available Landsat TM, ETM+ and OLI data (WRS-2 195/051) with minimum cloud coverage. Low image availability in the 1990s prevented us from finding phenologically corresponding image pairs in earlier years and extending the study period further into the past. With regard to the requirements of our mapping approach (see Section 3), the integration of multispectral sensors with different spectral coverage was not possible. Each time step covers both phenological extremes, including one image recorded just after the rainy season and one during the dry season. Best suitable images were selected based on NDVI values in structurally invariant areas of seasonal vegetation (Fig. 3). NDVI is used for an overview on seasonal variation and maxima/minima in the greenness and abundance of vegetation, only. Subsequent image analyses are performed on all spectral values. All images were atmospherically corrected using the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS, Masek et al., 2006) and cloud-masked with the Fmask algorithm (Zhu and Woodcock, 2012). LEDAPS is an atmospheric correction algorithm using water vapor and air temperature inputs from the National Centers for Environmental Protection (NCEP). Fmask is an object-based cloud and cloud shadow detection routine performing cloud probability masks based on Landsat top of atmosphere reflection and brightness temperature. Water was masked using the Modified Normalized Difference Water Index that generally offers a very high overall accuracy on inland waters (Xu, 2006). All images were reduced to six optical bands. An image stack of both rainy and dry season images containing twelve bands was used for following analyses.
3.2. Ensemble Support Vector Regression with synthetic mixing SVR is a Support Vector Machine (SVM) method designed to solve regression problems. It uses non-linear kernel transformation in order to map input data into a higher dimensional feature space and, then, applies a linear regression model solving higher dimensional problems (Schölkopf and Smola, 2002). In recent years, the benefits of SVM have been proven in different remote sensing applications for both classification and fraction analysis (Pal, 2009; Mountrakis et al., 2011; Tuia et al., 2011; Okujeni et al., 2016). They turned out to be particularly useful for dealing with high spectral intra-class variability from reference input data (Waske et al., 2010; Okujeni et al., 2013). As collecting training data with gradual surface information is challenging, we used synthetic mixing to create artificial mixtures between all 200 input spectra. Compared to manual reference training fraction selection, this is a more universal and less time consuming approach. Following Okujeni et al. (2016), we used a linear mixing system with mixing increments of 20%, i.e. all pure input spectra of one class were mixed at fraction levels of 0, 20, 40, 60, 80 and 100% with all input spectra from all other classes. Mixing complexity was binary (two spectra) or ternary (three spectra) (Fig. 6). To handle very large training sample sizes, SVRsynthmix uses an ensemble framework (Okujeni et al., 2016). We used an ensemble with 20 iterations and 1,200 randomly drawn training samples each. Iteration results were averaged for the final fraction map. SVRsynthmix modeling was performed in the EnMAP-Box using the imageSVM plug-in (van der Linden et al., 2015). The latter is based on the LIBSVM algorithm (Chang and Lin, 2013), a Gaussian kernel function, and offers a fully automated grid search for tuning SVR parameters. All resulting fraction maps have been restrained to a {0,1} extent. For the purpose of comparison with the global, bi-seasonal model, the performances of a mono-seasonal dry season model, a mono-seasonal rainy season model and yearly bi-seasonal models were tested. For the mono-seasonal models, the image library was adapted according to the respective season and split in six dry season bands and six rainy season bands.
3. Methods and data analysis The quantitative land cover maps of Ouagadougou and its surroundings consist of pixel-wise regression estimates of four present land cover types and for five points in time, where rainy and dry season imagery was available in the USGS archive (Fig. 4). The regression approach, here a support vector regression, requires combinations of spectral and thematic information for model training. Starting with one Landsat image stack for each examined time step, we first extracted exemplary surface spectra from the same locations in all images to 220
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
Fig. 4. Model training and estimation for a global multi-annual regression model based on bi-seasonal Landsat image pairs.
3.4. Mapping Fraction results were used to map urban expansion during the study period as well as densification processes. For the distinction of planned and unplanned urban development, we iteratively identified a rule set or fraction index based on reference neighborhood surface fractions and with the help of expert knowledge. The hypothesis is that the composition of surface fraction shares allows conclusions on neighborhood types that we can use for urban pattern mapping. According to the study period, we only considered neighborhoods that started development in or after 2002, i.e. urban fractions in that year needed to be under a certain threshold. Reference fractions (e.g. in Fig. 8) suggested that the ratio of soil and seasonal vegetation or its change could be an indicator for planned or unplanned development. Fig. 5. Exemplary input spectra for permanent vegetation, seasonal vegetation, soil and urban surfaces.
4. Results
3.3. Validation methodology
4.1. Model evaluation
As SVRsynthmix was applied for the first time using bi-seasonal data, we statistically evaluated the general usability of the approach for such data. Reference land cover information was derived from historic Quickbird VHR imagery from Google Earth for all studied years. We applied a systematic pixel-wise approach for seven randomly chosen homogeneous neighborhoods within the Ouagadougou metropolitan area. Those neighborhoods constituted a mix of historic and recently developed quarters. Within each neighborhood, we created a grid of 25 Landsat pixels using nine of them as reference. This procedure was chosen in order to capture a representative sample of neighborhood land cover avoiding adjacent areas. We then created a grid of seven by seven reference points based on 5 m distance between each point in a 30 m Landsat pixel. Using this 5 m distance, we capture most urban features without matching grid size to objects. We determine land cover at each of the reference points at each studied time step, resulting in an overall number of 441 reference points per neighborhood and, thus, 3087 reference points per time step (Fig. 7). Reference fractions for all nine pixels of a neighborhood were averaged and compared to SVR results of the same area by means of Mean Absolute Error (MAE) values. Fig. 8 displays reference fractions for two exemplary neighborhoods and all time steps. In addition to the generalized bi-seasonal model, the yearly bi-seasonal and both mono-seasonal SVRsynthmix models were validated regardless of their contextual shortcomings in order to evaluate if the work with multi-annual libraries and the use of bi-seasonal image stacks generates disadvantages in result quality.
The MAE for mapped fractions of four land cover types was calculated for model predictions from the global model. Resulting error values were compared to those from the yearly bi-seasonal models as well as the mono-seasonal dry and rainy season models in order to assess the validity and accuracy of the suggested approach before interpreting spatial-temporal patterns. In dry and rainy season models, seasonal vegetation cannot be mapped as a class. The error assessment shows that the performance of a global, bi-seasonal model is constantly equal or superior to the other models (Table 1). Performances of global dry and rainy season models vary, with better results for the dry season model in all three classes. In the yearly, bi-seasonal model, seasonal vegetation appears as a new class, while the model provides equally good results for urban and permanent vegetation compared to the global dry season model. Soil achieves a weaker mapping accuracy. The global, bi-seasonal model only shows a slight decrease in MAE for soil compared to the dry season model, while achieving highest values for the three other classes. The results from the error assessment are confirmed by the distribution of reference and prediction values in scatter plots and by the slope and intercept of linear model fits to these values (Fig. 9). Scatter plots provide insights into biases in predictions from our global model, overestimating selected areas in terms of urban surfaces and soil and underestimating seasonal vegetation by an intercept of −0.12. Permanent vegetation is more sensitive to changes due to a low value range and higher point clustering leading to a rather arbitrary slope result.
221
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
Fig. 7. Validation scheme. Grid of nine validated Landsat pixels, each with 7 ∗ 7 reference points. This scheme is applied on seven neighborhoods in five different years (7 points ∗ 7 points ∗ 9 LS pixels ∗ 7 neighborhoods ∗ 5 years = 15.435 reference points in total).
Fig. 6. Exemplary binary mixture of Seasonal Vegetation with Soil, Permanent Vegetation and Urban Surfaces in fraction portions of 0/20/40/60/80/100%.
4.2. Mapping urban expansion Fig. 8. Reference fractions for Tampouy non-lôti (left) and Tanghin Karpala (right) for all time steps. Values indicate class fractions from 441 reference points, averaged over nine Landsat pixels respectively.
Visual inspection of fraction maps shows a general growth of the greater Ouagadougou area in all cardinal directions (Fig. 10). We observe a general spread of urban surfaces from 2002 to 2013, represented by a) an increase of urban fraction in the urban core and selected suburban areas and b) new areas featuring urban fraction in the outskirt areas. Along with that comes a belt of extensive bare soil surface surrounding built-up areas. High fractions of bare soil do rarely occur within the city core, but mainly constitute lots and street blocks prepared for development along the urban fringe. Most of the city's surroundings are covered with seasonal vegetation on cultivated or unused land. Within the city, seasonal vegetation fractions are particularly high on fallow ground. Permanent vegetation at fractions above 10% is only observed on few irrigated areas and around water reservoirs and does not show any obvious trend between 2002 and 2013 (black areas in Fig. 10).
Table 1 Class-wise averaged MAE comparison between the applied models. Best values in bold.
222
Class
Global dry seas. model
Global rainy seas. model
Yearly biseas. model
Global biseas. model
Urban Soil Seasonal veget. Permanent veget.
0.13 0.08 – 0.05
0.14 0.15 – 0.16
0.13 0.14 0.16 0.05
0.10 0.10 0.13 0.04
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
Fig. 9. Class-wise global, bi-seasonal model validation results. One point represents the averaged fractions within one reference area in one year. The confidence level (grey shadowing) is 0.95 and should be viewed with caution because of the sample size.
Fig. 10. Fraction development of urban surfaces, seasonal vegetation and soil from 2002 to 2013.
relevant, but moderate share of urban surfaces in 2007 (between 20% and 40%) and an increased share of urban surfaces (above 40%) in 2013 (Fig. 12).
Fig. 11 describes the temporal dimension of urban expansion on a pixel level. Pixel color relates to the first time step when the urban fraction is > 20%, under the condition that subsequent fraction values do not drop below 10%. Using this simple rule set, we can visualize neighborhoods relating to their year of first emergence encountering effects of noise and restructuring. We see that some neighborhoods can be clearly linked to a period of development, suggesting a spatially focussed pattern of expansion. Whereas in total, 98.6 km2 are covered by relevant urban fraction (i.e. > 20%) in the Ouagadougou metropolitan area in 2002, that number reached 180.2 km2 in 2013 and, thus, nearly doubled. Working with class cover fractions, we can also map urban densification. Between 2007 and 2013, 29.79 km2 of the metropolitan area underwent processes of urban densification, which is 1/6 of the total urban area in 2013. For this example, we observe pixels that have a
4.3. Unplanned and planned urban development Areas of planned and unplanned development over time can also be represented on a pixel level. For planned and unplanned urbanization pattern detection, pixels were classified as planned development when
fracseasonalvegetation2013 /fracsoil2013 > 2 /3 and as unplanned development when
fracseasonalvegetation2013 /fracsoil2013 ≤ 2 / 3 both under the conditions that 223
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
Fig. 11. Urban surface cover by year of first observation from 2002 to 2013. Shading describes the urban surfaces fraction in 2013. Reading example: Dark orange pixels are surfaces that experienced urban development (> 20% urban fraction) in 2011 and a fast development until 2013. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)
fracurban2002 < 0.3
5. Discussion
and
Ouagadougou experienced rapid spatial expansion and densification between 2002 and 2013. This can be clearly monitored based on archived Landsat data. Using a global, bi-seasonal SVRsynthmix model allows for a clear delineation of spatial-temporal changes in urban land cover and provides more thematic information than mono-seasonal models while being more accurate or at least similarly accurate (Table 1). The method of support vector regression using synthetic mixing suggested by Okujeni et al. (2013) and Okujeni et al. (2016) is capable of dealing with complex bi-seasonal information in order to generate accurate maps of urbanization processes in an area with particular challenges with regard to development patterns and climate. Resulting land cover fractions are highly accurate considering surface fraction identification and estimation in an urban environment. With the help of seasonal vegetation as a class, our proposed method accounts for intra-annual spectral surface variability. It provides information on urban expansion patterns and distinguishes planned from unplanned settlement types. Considering urban expansion (Fig. 11), we can identify densification in the urban core and selected suburban areas and an increase in area covered by urban surfaces in the outskirt areas. Resulting time series information makes it possible to temporally attribute neighborhood growth. Regression models are able to temporally allocate densification processes (Fig. 12) and gradual development, and outmatch classifications that are limited to assigning a winning class, only. Fraction results
fracurbant > fracurbant − 1 ∗0.9 This rule set is based on the iterative finding that the seasonal vegetation/soil ratio contributes to describing planned and unplanned areas. Neighborhoods with mostly planned urbanization processes appear clustered, whereas those with unplanned development are scattered (Fig. 12). In unplanned settlements, the increase in urban fractions is comparatively fast. In Château d'Eau de Tampouy, a neighborhood identified as unplanned and developed after 2002, seasonal vegetation is cleared in favor of open soil between 2002 and 2007 and replaced by urban fraction in the years after, while the overall level of soil fractions stays above 45% (Figs. 13A & 14A). In planned settlements, fraction development is generally slower. Tanghin Karpala (B) shows less dense urban structures and lower urban fractions, whereas soil and seasonal vegetation fractions show equal shares (Figs. 13B & 14B). Similar fraction patterns for planned development can be observed in the neighborhood of Kouritenga (Figs. 13C & 14C), with the exception that the eastern part has some unplanned urban structures (Figs. 13D & 14D), is slightly older and features urban surfaces in 2002 already. Among planned areas, long-established neighborhoods generally show a higher fraction of permanent vegetation. 224
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
Fig. 12. Densification of urban class fractions between 2007 (20% < urban fraction < 40%) and 2013 (urban fraction > 40%).
neighborhoods representing exemplary development patterns, but is also reflected in the dynamic development of the urban soil belt, preceding new built-up areas. Looking at neighborhoods that undergo a transformation from a rather unplanned organisation to a plotted and parcelled lot arrangement, we see that those processes usually take place when the unplanned settlements are not yet at their densest stage. Visual analysis shows that neighborhood re-organisation is in general systematic. Fraction ratios of seasonal vegetation and soil adapt towards the above described pattern of planned areas. The slight constant underestimation of soil in the applied model could be due to an underestimation of soil in the used validation imagery, since we had to perform validation on dry season images only. Accordingly, pixels referred to as seasonal vegetation might still have a relatively large share of soil in the bi-seasonal signal. The imperfect temporal overlap of Landsat imagery and VHR reference imagery could also account for some of the observed errors in such a highly dynamic urban area. Results from this study confirm the value of multi-seasonal information for mapping land cover in Western Africa (e.g. Zoungrana et al., 2015) and the challenges of high spectral similarities, in particular in vegetated areas (Knauer et al., 2017). The temporally generalized bi-seasonal model is highly universal due to its thematic and methodological quality. It is capable of encountering spectral variations of land surfaces over the year and meets the requirement of their
offer a high level of detail on a sub-pixel level and make it possible to study fine-scale processes. We further state that fraction patterns between unplanned and planned urban development considerably differ, as exemplified for several areas across the city (Fig. 13). Both types of settlement are distinguished through their swift increase in urban fractions over time and their seasonal vegetation to soil ratio. Unplanned settlements grow fast and feature a relatively high seasonal vegetation to soil ratio (about > 2/3), whereas planned settlements often feature urban fraction growth rates of < 20% and a much lower seasonal vegetation to soil ratio. At the beginning of neighborhood development, areas of seasonal vegetation are cleared for subsequent construction. In the case of unplanned development, those are rapidly being replaced by urban surfaces. If no or only sparse development occurs, those areas might be recovered by seasonal vegetation in planned neighborhoods, whereas in unplanned neighborhoods, open space is either kept clear of vegetation for further construction purposes or used for human activity, such as waste disposal, soil quarrying or markets. A generally less dense construction in planned areas, such as Tanghin Karpala (Figs. 13B and 14B), leads to a higher overall presence of seasonal shrubs and grasses compared to soil surfaces in those neighborhoods. On a micro scale, we also see that a sudden change in soil fractions might be an indicator for future urban surface coverage. That can be visualized in selected 225
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
Fig. 13. Unplanned and planned urban patterns in metropolitan Ouagadougou (top left), Tampouy and Kouritenga (top right) colored corresponding to above mentioned rule set.
Fig. 14. Google Earth satellite imagery (top) and historic fraction composition of the selected areas A, B, C and D (bottom).
synthetic mixtures, so that SVR accounts for urban spectral variability with a relatively small number of library spectra and hence relatively small efforts by the user. Okujeni et al. (2013) use 23 urban surface spectra on much more heterogeneous 3.6 m hyperspectral data in Berlin and prove the robustness of SVR. Combining SVRsynthmix and bi-seasonal input data allows the reliable derivation of urban cover, soil and vegetation cover fractions, i.e. components similar to the Vegetation – Impervious Surfaces – Soil (V-I-S) framework according to Ridd (1995), yet allowing for a more thematically oriented and mixed urban fraction.
representation in spectral libraries as highlighted in Dudley et al. (2015). In the Ouagadougou area, the diversity in spectral training data within classes over the year is high and training spectra from one year are likely to be beneficial for the representation of the feature space in the same region in other years. SVRsynthmix, as a machine learning approach, uses the advantage of a single spectral library across years and encounters that issue. We achieve good results when using spectra from only ten locations within the urban area. The amount of spectra used for training multiplies with the number of years and the number of 226
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
respective year. The transferability of the method to regions with seasonal land cover features but different climatic conditions is also unclear.
Moreover, the additional seasonal vegetation cover might be regarded as a temporal expansion of the two-dimensional mixing space as, for example, described by Small (2005). Such an additional dimension representing seasonal change in semi-arid regions appears useful with regard to studies addressing the spectral mixing space and makes SVRsynthmix, in principle, transferable to any land cover research with a seasonal component. Overall, the integration of intra-annual spectral variation provides valuable information for mapping urban expansion processes. The Landsat data archive offers the image density needed for the purpose of this study from 2002 on; the global, bi-seasonal SVRsynthmix model produces fractions as accurate as required for urban growth mapping. Multi-temporal information on fractions allows for delineating settlement types. Results from this approach go clearly beyond discrete indicators for urban settlements by providing quantifications for specific urbanization patterns representing the characters of urban neighborhoods and, thus, outperforms previous studies with manual digitizing (Fournet et al., 2008) or discrete classification for Ouagadougou (de Jong et al., 2000) and other cities in developing or emerging countries (Griffiths et al., 2010; El Garouani et al., 2017; Li et al., 2015b; Hu et al., 2016), or even those using different sensor types and data input (Taubenböck et al., 2012; Kumar et al., 2015; Benza et al., 2016). Our methodology is applicable to smaller scales than those determining neighborhood typology based on high resolution imagery or structural analysis only (Kuffer et al., 2014; Taubenböck and Kraff, 2014). The results might serve urban planning by providing an early alert system for highly dynamic urban areas and potentially contribute to more anticipatory planning processes in regions where research and planning often operate after the fact (Caldeira, 2016). If data quality and density permit, the proposed method can provide a history of spatial expansion in all semi-arid areas that lack well-documented archives and, possibly, experience short-term developments with potential linkage to specific events (e.g. policy changes or political transitions). Future research should focus on the aspect of method transferability on multiple levels. Due to its bi-seasonal character, requirements for image selection are more challenging than using one point in time for processing. Knowing that regional image availability varies, finding adequate image dates with consistent phenology is crucial. In this study, high cloud coverage during rainy seasons plus a lack of Landsat TM data in the mid-1990 and ETM+ with SLC-off imagery in most of the 2000s led to a general deficit of available images during phenological peaks during that period. The prerequisite of image pairs and spectral consistency throughout seasons and time steps considerably reduces data availability for this study. For the purpose of a better identification of temporal developments, future studies ought to work on denser time series, e.g. through image compositing (Griffiths et al., 2013) or virtual sensor constellations between Landsat and Sentinel 2a and 2b (Wulder et al., 2015) and the further integration of multi-sensor data. Sentinel 2, in particular, appears promising with regard to an annual availability of image pairs ideally representing seasonal differences. Resulting annual time series are expected to lead to more robust interpretations of temporal trajectories of land cover fractions, due to a more reliable identification of dates with minimum and maximum vegetation coverage and annual availability of such bi-seasonal image pairs. In the case that appropriate imagery is available, bi-seasonal SVRsynthmix should be applicable to any semi-arid environment that features similar land cover characteristics. Using location-specific spectral libraries should produce similar results in other areas than Ouagadougou, whereas the spatial transferability of bi-seasonal libraries still requires research. Vegetation and urban development patterns might have different characteristics across regions which raises the question, if rule-based mapping for planned and unplanned development patterns also applies for new study areas. Moreover, research should clarify if spectral library transferability is feasible not only in a spatial, but also in a temporal dimension, i.e. if land cover can be mapped without integrating any spectral training data from the
6. Conclusion Results from this study show that a support vector regression method with synthetic spectral mixing and bi-seasonal series of Landsat data is suitable for mapping urban extent and land cover fractions, and this way increase process understanding in a rapidly growing WestAfrican city on a small and large scale. Accounting for intra-annual land cover variability is advantageous in order to draw conclusions on potential future urban development. We can identify specific land cover compositions for different types of urban patterns that point towards adaptive planning recommendations. Our study adds a seasonal dimension that is non-existent in comparable urbanization studies in remote sensing and monitors developments on a smaller spatial scale than social science research could. We provide and extend a strong and universal approach for quantitative (urban) land cover mapping with little demand for manual referencing, suitable for mono- and multiseasonal studies using spectral data. The applied methodology contributes to global urban mapping by offering a reliable way to account for climatic particularities. It is potentially transferable to other areas that meet the requirements of image availability and might be even more promising in the future when integrating additional data sources such as Sentinel or using high performance computation environments such as Google Earth Engine. Acknowledgements The authors would like to thank Mamadou Kabré from the IGB (Institut Géographique du Burkina) for his support with the identification of the neighborhoods referred to in this paper. Clara Sichau helped labeling reference data. J. Hauer, J. Nielsen and S. van der Linden were supported by Humboldt-Universität zu Berlin as part of its institutional strategy within the German Excellence Initiative. Stiftung HumboldtUniversität and the Federal Ministry of Education and Research granted scholarships (Deutschlandstipendium) for J. Hauer and F. Schug. This research further contributes to the Landsat Science Team 2012-2017 (United States Geological Survey, 2017). References Akintunde, J.A., Adzandeh, E.A., Fabiyi, O.O., 2016. Spatio-temporal pattern of urban growth in Jos Metropolis, Nigeria. Remote Sens. Appl. Soc. Environ. 4, 44–54. http:// dx.doi.org/10.1016/j.rsase.2016.04.003. Beauchemin, C., 2011. Rural-urban migration in West Africa. Towards a reversal? Migration trends and economic situation in Burkina Faso and Côte d'Ivoire. Popul. Space Place 17, 47–72. http://dx.doi.org/10.1002/psp.573. Beauchemin, C., Bocquier, P., 2004. Migration and urbanisation in francophone West Africa. An overview of the recent empirical evidence. Urban Stud. 41, 2245–2272. http://dx.doi.org/10.1080/0042098042000268447. Benza, M., Weeks, J.R., Stow, D.A., López-Carr, D., Clarke, K.C., 2016. A pattern-based definition of urban context using remote sensing and GIS. Remote Sens. Environ. 183, 250–264. http://dx.doi.org/10.1016/j.rse.2016.06.011. Caldeira, T.P.R., 2016. Peripheral urbanization. Autoconstruction, transversal logics, and politics in cities of the global south. Environ. Plann. D Soc. Space 35, 3–20. http://dx. doi.org/10.1177/0263775816658479. Chang, C.-C., Lin, C.-J., 2013. LIBSVM: A Library for Support Vector Machines. https:// www.csie.ntu.edu.tw/~cjlin/papers/libsvm.pdf. Dudley, K.L., Dennison, P.E., Roth, K.L., Roberts, D.A., Coates, A.R., 2015. A multi-temporal spectral library approach for mapping vegetation species across spatial and temporal phenological gradients. Remote Sens. Environ. 167, 121–134. http://dx. doi.org/10.1016/j.rse.2015.05.004. El Garouani, A., Mulla, D.J., El Garouani, S., Knight, J., 2017. Analysis of urban growth and sprawl from remote sensing data. Case of Fez, Morocco. Int. J. Sustain. Built Environ. 6, 160–169. http://dx.doi.org/10.1016/j.ijsbe.2017.02.003. Fourchard, L., 2001. De la ville coloniale à la cour africaine. Espaces, pouvoirs et sociétés à Ouagadougou et à Bobo-Dioulasso (Haute-Volta) fin XIXe siècle - 1960. L'Harmattan, Paris. Fournet, F., Meunier-Nikiema, A., Salem, G., 2008. Ouagadougou (1850–2004). (s.l.: IRD Éditions).
227
Remote Sensing of Environment 210 (2018) 217–228
F. Schug et al.
Humboldt-Universität zu Berlin. Okujeni, A., van der Linden, S., Tits, L., Somers, B., Hostert, P., 2013. Support vector regression and synthetically mixed training data for quantifying urban land cover. Remote Sens. Environ. 137, 184–197. http://dx.doi.org/10.1016/j.rse.2013.06.007. Okujeni, A., van der Linden, S., Suess, S., Hostert, P., 2016. Ensemble learning from synthetically mixed training data for quantifying urban land cover with support vector regression. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 10, 1640–1650. http://dx.doi.org/10.1109/JSTARS.2016.2634859. Ouédraogo, M., 2002. Land Tenure and Rural Development in Burkina Faso. Issues and Strategies. http://pubs.iied.org/pdfs/9183IIED.pdf. Pal, M., 2009. Kernel methods in remote sensing. A review. ISH J. Hydraul. Eng. 15, 194–215. http://dx.doi.org/10.1080/09715010.2009.10514975. Parnell, S., Pieterse, E. (Eds.), 2014. Africa's Urban Revolution. Zed Books, London. Parnell, S., Pieterse, E., Watson, V., 2009. Planning for cities in the global South: an African research agenda for sustainable human settlements/shaken, shrinking, hot, impoverished and informal. Emerging research agendas in planning. Prog. Plan. 72, 233–241. http://dx.doi.org/10.1016/j.progress.2009.09.001. Potts, D., 2012. Challenging the myths of urban dynamics in sub-Saharan Africa. The evidence from Nigeria. World Dev. 40, 1382–1393. http://dx.doi.org/10.1016/j. worlddev.2011.12.004. Potts, D., 2017. Urban data and definitions in sub-Saharan Africa. Mismatches between the pace of urbanisation and employment and livelihood change. Urban Stud. 15, 004209801771268. http://dx.doi.org/10.1177/0042098017712689. Ridd, M.K., 1995. Exploring a V-I-S (vegetation-impervious surface-soil) model for urban ecosystem analysis through remote sensing. Comparative anatomy for cities. Int. J. Remote Sens. 16, 2165–2185. http://dx.doi.org/10.1080/01431169508954549. Satterthwaite, D., 2010. Urban Myths and the Mis-use of Data That Underpin Them. Helsinki, WIDER. Schölkopf, B., Smola, A.J., 2002. Learning with Kernels. Support Vector Machines, Regularization, Optimization, and Beyond. MIT Press, Cambridge, Mass. Sexton, J.O., Song, X.-P., Huang, C., Channan, S., Baker, M.E., Townshend, J.R., 2013. Urban growth of the Washington, D.C.–Baltimore, MD metropolitan region from 1984 to 2010 by annual, Landsat-based estimates of impervious cover. Remote Sens. Environ. 129, 42–53. http://dx.doi.org/10.1016/j.rse.2012.10.025. Small, C., 2004. The Landsat ETM+ spectral mixing space. Remote Sens. Environ. 93, 1–17. http://dx.doi.org/10.1016/j.rse.2004.06.007. Small, C., 2005. A global analysis of urban reflectance. Int. J. Remote Sens. 26, 661–681. http://dx.doi.org/10.1080/01431160310001654950. Song, X.-P., Sexton, J.O., Huang, C., Channan, S., Townshend, J.R., 2016. Characterizing the magnitude, timing and duration of urban growth from time series of Landsatbased estimates of impervious cover. Remote Sens. Environ. 175, 1–13. http://dx.doi. org/10.1016/j.rse.2015.12.027. Taubenböck, H., Kraff, N.J., 2014. The physical face of slums. A structural comparison of slums in Mumbai, India, based on remotely sensed data. J. Housing Built Environ. 29, 15–38. http://dx.doi.org/10.1007/s10901-013-9333-x. Taubenböck, H., Esch, T., Felbier, A., Wiesner, M., Roth, A., Dech, S., 2012. Monitoring urbanization in mega cities from space. Remote Sens. Environ. 117, 162–176. http:// dx.doi.org/10.1016/j.rse.2011.09.015. Tuia, D., Verrelst, J., Alonso, L., Perez-Cruz, F., Camps-Valls, G., 2011. Multioutput support vector regression for remote sensing biophysical parameter estimation. IEEE Geosci. Remote Sens. Lett. 8, 804–808. http://dx.doi.org/10.1109/LGRS.2011. 2109934. United Nations Environment Programme, 2010. Africa Water Atlas. Nairobi. https://na. unep.net/atlas/africaWater/downloads/africa_water_atlas.pdf. United Nations: Department of Economic and Social Affairs, 2015. World Urbanization Prospects. The 2014 Revision. New York. https://esa.un.org/unpd/wup/ Publications/Files/WUP2014-Report.pdf. United States Geological Survey, 2008. Opening the Landsat archive. U.S. Department of the Interior. https://pubs.usgs.gov/fs/2008/3091/pdf/fs2008-3091.pdf. United States Geological Survey, 2017. Landsat Science Team 2012–2017. U.S. Department of the Interior. https://landsat.usgs.gov/2012-2017-science-team. Waske, B., van der Linden, S., Benediktsson, J.A., Rabe, A., Hostert, P., 2010. Sensitivity of support vector machines to random feature selection in classification of hyperspectral data. IEEE Trans. Geosci. Remote Sens. 48, 2880–2889. http://dx.doi.org/ 10.1109/TGRS.2010.2041784. Wulder, M.A., Hilker, T., White, J.C., Coops, N.C., Masek, J.G., Pflugmacher, D., Crevier, Y., 2015. Virtual constellations for global terrestrial monitoring. Remote Sens. Environ. 170, 62–76. http://dx.doi.org/10.1016/j.rse.2015.09.001. Xu, H., 2006. Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery. Int. J. Remote Sens. 27, 3025–3033. http://dx.doi.org/10.1080/01431160600589179. Yuan, F., Bauer, M.E., Heinert, N.J., Holden, G.R., 2005. Multi-level land cover mapping of the twin cities (Minnesota) metropolitan area with multi-seasonal Landsat TM/ ETM+ data. Geocarto Int. 20, 5–13. http://dx.doi.org/10.1080/ 10106040508542340. Zhu, Z., Woodcock, C.E., 2012. Object-based cloud and cloud shadow detection in Landsat imagery. Remote Sens. Environ. 118, 83–94. http://dx.doi.org/10.1016/j. rse.2011.10.028. Zhu, Z., Woodcock, C.E., 2014. Continuous change detection and classification of land cover using all available Landsat data. Remote Sens. Environ. 144, 152–171. http:// dx.doi.org/10.1016/j.rse.2014.01.011. Zoungrana, B., Conrad, C., Amekudzi, L., Thiel, M., Da, E., Forkuor, G., Löw, F., 2015. Multi-temporal Landsat images and ancillary data for land use/cover change (LULCC) detection in the southwest of Burkina Faso, West Africa. Remote Sens. 7, 12076–12102. http://dx.doi.org/10.3390/rs70912076.
Gandy, M., 2006. Planning, anti-planning and the infrastructure crisis facing metropolitan Lagos. Urban Stud. 43, 371–396. http://dx.doi.org/10.1080/00420980500406751. Gessner, U., Machwitz, M., Esch, T., Tillack, A., Naeimi, V., Kuenzer, C., Dech, S., 2015. Multi-sensor mapping of West African land cover using MODIS, ASAR and TanDEMX/TerraSAR-X data. Remote Sens. Environ. 164, 282–297. http://dx.doi.org/10. 1016/j.rse.2015.03.029. Goodfellow, T., 2013. Planning and development regulation amid rapid urban growth. Explaining divergent trajectories in Africa. Geoforum 48, 83–93. http://dx.doi.org/ 10.1016/j.geoforum.2013.04.007. Griffiths, P., Hostert, P., Gruebner, O., van der Linden, S., 2010. Mapping megacity growth with multi-sensor data. Remote Sens. Environ. 114, 426–439. http://dx.doi. org/10.1016/j.rse.2009.09.012. Griffiths, P., van der Linden, S., Kuemmerle, T., Hostert, P., 2013. A pixel-based Landsat compositing algorithm for large area land cover mapping. IEEE J. Sel. Top. Appl. Earth Observ. Remote Sens. 6, 2088–2101. http://dx.doi.org/10.1109/jstars.2012. 2228167. Grimm, N.B., Faeth, S.H., Golubiewski, N.E., Redman, C.L., Wu, J., Bai, X., Briggs, J.M., 2008. Global change and the ecology of cities. Science (New York, N.Y.) 319, 756–760. http://dx.doi.org/10.1126/science.1150195. Harre, D., Moriconi-Ebrard, F., Heinrigs, P., 2016. Urbanisation Dynamics in West Africa 1950–2010. OECD Publishing. Hauer, J., Nielsen, J.Ø., Niewöhner, J., 2018. Landscapes of Hoping – Urban Expansion and Emerging Futures in Ouagadougou, Burkina Faso. Anthropological Theory. (doi: in press). Hou, H., Estoque, R.C., Murayama, Y., 2016. Spatiotemporal analysis of urban growth in three African capital cities. A grid-cell-based analysis using remote sensing data. J. Afr. Earth Sci. 123, 381–391. http://dx.doi.org/10.1016/j.jafrearsci.2016.08.014. Hu, T., Yang, J., Li, X., Gong, P., 2016. Mapping urban land use by using Landsat images and open social data. Remote Sens. 8, 151. http://dx.doi.org/10.3390/rs8020151. Jaglin, S., 1995. Gestion urbaine partagée à Ouagadougou. Pouvoirs et périphéries (1983–1991). Editions Karthala, Paris. de Jong, S.M., Bagre, A., van Teeffelen, P.B.M., van Deursen, W.P.A., 2000. Monitoring trends in urban growth and surveying city quarters in Ouagadougou, Burkina Faso using SPOT-XS. Geocarto Int. 15, 63–70. http://dx.doi.org/10.1080/ 10106049908542154. Knauer, K., Gessner, U., Fensholt, R., Forkuor, G., Kuenzer, C., 2017. Monitoring agricultural expansion in Burkina Faso over 14 years with 30 m resolution time series. The role of population growth and implications for the environment. Remote Sens. 9, 132. http://dx.doi.org/10.3390/rs9020132. Kuffer, M., Barros, J., Sliuzas, R.V., 2014. The development of a morphological unplanned settlement index using very-high-resolution (VHR) imagery. Comput. Environ. Urban. Syst. 48, 138–152. http://dx.doi.org/10.1016/j.compenvurbsys.2014.07.012. Kumar, U., Milesi, C., Nemani, R.R., Basu, S., 2015. Multi-sensor multi-resolution image fusion for improved vegetation and urban area classification. ISPRS - Int. Arch. Photogramm. Remote. Sens. Spat. Inf. Sci. XL-7/W4, 51–58. http://dx.doi.org/10. 5194/isprsarchives-XL-7-W4-51-2015. Lambin, E.F., Geist, H.J., Lepers, E., 2003. Dynamics of land use and land cover change in tropical regions. Annu. Rev. Environ. Resour. 28, 205–241. http://dx.doi.org/10. 1146/annurev.energy.28.050302.105459. Li, E., Endter-Wada, J., Li, S., 2015a. Characterizing and contextualizing the water challenges of megacities. JAWRA J. Am. Water Res. Assoc. 51, 589–613. http://dx. doi.org/10.1111/1752-1688.12310. Li, X., Gong, P., Liang, L., 2015b. A 30-year (1984–2013) record of annual urban dynamics of Beijing City derived from Landsat data. Remote Sens. Environ. 166, 78–90. http://dx.doi.org/10.1016/j.rse.2015.06.007. van der Linden, S., Rabe, A., Held, M., Jakimow, B., Leitão, P., Okujeni, A., Schwieder, M., Suess, S., Hostert, P., 2015. The EnMAP-box—A toolbox and application programming interface for EnMAP data processing. Remote Sens. 7, 11249–11266. http://dx. doi.org/10.3390/rs70911249. MacLachlan, A., Roberts, G., Biggs, E., Boruff, B., 2017. Subpixel land-cover classification for improved urban area estimates using Landsat. Int. J. Remote Sens. 38, 5763–5792. http://dx.doi.org/10.1080/01431161.2017.1346403. Marlier, M.E., Jina, A.S., Kinney, P.L., DeFries, R.S., 2016. Extreme air pollution in global megacities. Curr. Clim. Chang. Rep. 2, 15–27. http://dx.doi.org/10.1007/s40641016-0032-z. Masek, J.G., Vermote, E.F., Saleous, N.E., Wolfe, R., Hall, F.G., Huemmrich, K.F., Gao, F., Kutler, J., Lim, T.-K., 2006. A Landsat surface reflectance dataset for North America, 1990–2000. IEEE Geosci. Remote Sens. Lett. 3, 68–72. http://dx.doi.org/10.1109/ LGRS.2005.857030. Mering, C., Baro, J., Upegui, E., 2010. Retrieving urban areas on Google Earth images. Application to towns of West Africa. Int. J. Remote Sens. 31, 5867–5877. http://dx. doi.org/10.1080/01431161.2010.512311. Mountrakis, G., Im, J., Ogole, C., 2011. Support vector machines in remote sensing. A review. ISPRS J. Photogramm. Remote Sens. 66, 247–259. http://dx.doi.org/10. 1016/j.isprsjprs.2010.11.001. Müller, H., Rufin, P., Griffiths, P., Barros Siqueira, A.J., Hostert, P., 2015. Mining dense Landsat time series for separating cropland and pasture in a heterogeneous Brazilian savanna landscape. Remote Sens. Environ. 156, 490–499. http://dx.doi.org/10. 1016/j.rse.2014.10.014. Murray, M.J., Myers, G.A., 2007. Cities in Contemporary Africa. Palgrave Macmillan, Basingstoke. Myers, G., 2011. (I)n(f)ormal cities. In: Myers, G. (Ed.), African Cities. Alternative Visions of Urban Theory and Practice, pp. 70–103 (New York, USA). Nielsen, J.Ø., D'haen, S.A.L., 2015. Discussing Rural-to-urban Migration Reversal in Contemporary Sub-Saharan Africa. The Case of Ouagadougou, Burkina Faso.
228