**4.3 Change detection**

416 Remote Sensing – Applications

Fig. 6. Original GeoEye image (a), objects segmented (b), objects merged (c) (Feature

**Land use / cover classes Description** 

buildings).

Table 2. Land cover classes anticipated with object-based classification.

detailed classification of residential housing was made only in Raila village.

The objects extracted during the segmentation were then classified using Support Vector Machine (SVM) classification algorithm in an object-oriented framework along with training sets, selected by experienced user. Nine land cover classes were used all together

Buildings\_blue Residential houses with blue spectral reflectance on image. Buildings\_light Residential houses with white or bright spectral reflectance. Buildings\_brown Residential houses with brown or dark spectral reflectance.

Roads Traffic connection between villages, usually unpaved. Shadows Shadowed areas around high objects (high vegetation and

These nine urban land use/cover classes included four types of residential housing. Subclasses were choosed because of their different spectral signature inside the same land cover class (e.g. instead of selecting only class "buildings" we selected four subclasses "buildings\_blue", "buildings\_light", "buildings\_brown" and "buildings\_red", Fig. 7). This way we obtained better results than we would have using one general class only. More

Classification results were obtained as a raster image and a vector file. Vectors were exported to a single layer and later processed for the need of post-classification in ESRI ArcMap software (all polygons smaller than 2 m2 were merged with neighbouring larger

Buildings\_red Residential houses with red spectral reflectance.

Soil Areas of unvegetated soil, mudded ground. Vegetation 1 Green vegetated areas, low vegetation (grass). Vegetation 2 Green vegetated areas, high vegetation (trees).

Extraction module of ENVI EX).

(Table 2).

polygons).

Satellite data offers unique utility for monitoring and quantifying land cover change over time. Consequently, change detection has become a significant part of the remote sensing research over the last decades. The goal of remote sensing change detection is to detect the geographic location of changes, identify their type (if possible) and quantify their amount. If long term imagery time series are handled, trends can be recognised. A large number of change detection methods have been evolved and they differ in their refinement, robustness and complexity (Hall and Hay, 2003). Nowadays a three level systematisation system is proposed that differentiates change detection methods by introducing the notion of pixel, feature and object level image processing (Deer, 1998). In general change detection techniques can be grouped into two major types (Jianya et al., 2008; Coppin et al., 2004; Lu et al. 2004; Singh 1989): image differencing techniques and post-classification comparison techniques. The main difference between the two types is that image differencing methods can identify the location and the magnitude of change but can not identify the type of land use or surface changes taken place in the area. Post-classification techniques can identify the location and provide the change character. Recent advances in change detection mostly involve high resolution data and consequently object-oriented and/or multi-scale approaches, with a range of techniques to approach contextual modelling (Lang et al., 2006; Blaschke et al., 2008; Addink & Van Coillie, 2010).

#### **4.3.1 Change detection of Kibera informal settlement**

The use of VHR satellite image time series may provide a reliable approach to detect dense urban growth in detail (Hofmann et al., 2006). A generically applicable and rapid operational land cover mapping of these settlements has generally proven difficult (Netzband & Rahman, 2010). Object-based classification and land use mapping of Kibera settlement from GeoEye image (section 4.2 and 5.2) highlighted some typical problems for object delineation in slum-like areas that can be corrected only with a lot of manual work. Main difficulties are associated with informal area outer-homogenity but inner-heterogenity due to the microstructure of urban agglomeration. Object-based classification is thus very demanding in terms of methodology adaptation to informal residential areas specifics,

Object-Based Image Analysis of VHR Satellite Imagery for

change characterisation (type of change, from-to).

of the change pattern for study's specific aims.

**4.3.2 Raila village change detection** 

satellite images involved.

Population Estimation in Informal Settlement Kibera-Nairobi, Kenya 419

The procedure was implemented as follows. First, spectral information for a slightly coarser scale (i.e. 3 m spatial resolution) was computed for images or areas of interest. This may be accomplished with a specified neighbourhood mean value annotation. Second, change differentiation between images is performed on a coarser resolution scale and change magnitude categorisation is applied (see below for categorisation classes and their definition). Third, upper positive and negative changes are reclassified so that the mask of important changes based on the neighbourhood context characteristics is prepared. Fourth, change differentiation is calculated on the original data scale (i.e. GeoEye and QuickBird data in 1 m spatial resolution), then categorisation is applied, and finally a mask (or a mask with a buffer) of an arbitrary specified magnitude of changes (obtained in the previous step)

Normally, if there is no substantial unwanted effects (noise) due to meteorological or sensor influences in images, changes obtained from the difference image (image differentiation) are distributed normally and symmetrically, with the average at 0. Abrupt changes (objects or land cover transformations) can then be defined by thresholding the distribution tails, giving the pattern of positive and negative changes in reflectance (spectral) space. We have calculated change magnitude (transition class category) intervals for every 0.5 standard deviation. Then the criterion of 2.5 standard deviations for negative changes and 2.0 for positive changes was applied to enclose the majority of detected transformations in both real and spectral world situations. Result of such categorisation of image difference is a pattern of abrupt changes (locations of appearance, disappearance of objects), with no association to

Each output from the above automatic threshold procedure is finally refined to an arbitrary degree, depending on case study objective. For entire Kibera informal settlement changes were obtained from comparison of GeoEye 2009-07-25 and QuickBird 2006-03-27 images. Change patches smaller than 5 m2 were eliminated for the purpose of this example, mainly to reduce the impact of change artefacts belonging to small patches of rooftop renovations (Fig. 3). False changes identified due to differences in viewing angle (location of buildings and trees shadows, buildings boundaries due to different original resolution of imagery) were also removed using the results obtained from object based classification (state of land use in 2009, section 4.2). Where shadow or vegetation object class were present, change pattern was corrected in the given context. Described and implemented step-wise change pattern refinement is shown in Fig. 9, for a subset of Raila village and northern neighbourhood. In other words, with simple generalisation we could control many aspects

Results were visually examined and evaluation of change pattern characteristics (over- and under-estimations) was done throughout the area using complementary comparison of

Raila village is known to have undergone extensive development during recent years (Kibera Wikipedia, 2011; MKP, 2011). Thus temporally detailed examination was performed observing its urbanisation. Change detection procedure described in section 4.3.1 was in addition implemented for Raila village only, but for two time sequences: 2006-2008 and 2008-2009. QB 2006-03-27, 2008-08-10 and GE 2009-07-25 images were used for this example.

is overlaid in order to restrict merely the contextually supported changes.

especially when accounting for their direct relation to representation on different satellite data sources. Thus within the limited framework of this case study pixel-based approach to identify outline of urban growth was preferred. The procedure was implemented on radiometrically adjusted time series (section 4.1). GeoEye 2009-07-25 and QuickBird 2006-03- 27 images were compared for the changes over whole Kibera, and GeoEye 2009-07-25, QucikBird 2008-08-10 and 2006-03-27 images were analysed for the observation of sequential urban growth of Raila village.

The simple thresholding of difference images is a well-known method that leads to the delimitation of changes and no-changes. The advantage of this method is that it can be fully automatic. However main disadvantage is that it heavily depends on consistency of datasets. Regardless of the carefully performed data preparations certain unwanted effects remains, which may drastically burden the imagery comparisons. This data variability behaves as a detected change and may well be enclosed within identified pattern of changes, although not all of the identified differences belong to real changes (false, non-intrinsic changes). Such false effects result in over-estimation of change pattern and can cause the quantitative evaluation fail. Since this data noise originates from the pre-processing algorithms as well as the natural and technological conditions during data acquisition, it can not be completely removed with data radiometric corrections (Veljanovski & Oštir, 2011).

To overcome this drawback a contextual multi-level change detection approach was applied that can efficiently treat most of the unwanted differences and suppress sensor related noise (Veljanovski, 2008). Taking into account the neighbourhood and change information by joining two spatial scales (Fig. 8), approach reduces amount of small size false differences.

Fig. 8. Change detection approach takes into account the neighbourhood and change information by joining different spatial scales.

The model is based on focal information logic that gives averaged change information in a slightly reduced spatial scale – within the specified neighbourhood. It is based on the fact that in a larger geographic area (e.g. a 3 x 3 pixel window or 3 m spatial resolution for resampled VHR data) the information of the changes will tend to level abrupt change information if a small spatial scale change is present, and will show the averaged difference if the majority of pixels in the observed window are subjected to change. Computing the piecewise change information between two time-successive data sets provides valuable information regarding the location and numeric change value derived from contextual information within the specified neighbourhood.

especially when accounting for their direct relation to representation on different satellite data sources. Thus within the limited framework of this case study pixel-based approach to identify outline of urban growth was preferred. The procedure was implemented on radiometrically adjusted time series (section 4.1). GeoEye 2009-07-25 and QuickBird 2006-03- 27 images were compared for the changes over whole Kibera, and GeoEye 2009-07-25, QucikBird 2008-08-10 and 2006-03-27 images were analysed for the observation of sequential

The simple thresholding of difference images is a well-known method that leads to the delimitation of changes and no-changes. The advantage of this method is that it can be fully automatic. However main disadvantage is that it heavily depends on consistency of datasets. Regardless of the carefully performed data preparations certain unwanted effects remains, which may drastically burden the imagery comparisons. This data variability behaves as a detected change and may well be enclosed within identified pattern of changes, although not all of the identified differences belong to real changes (false, non-intrinsic changes). Such false effects result in over-estimation of change pattern and can cause the quantitative evaluation fail. Since this data noise originates from the pre-processing algorithms as well as the natural and technological conditions during data acquisition, it can not be completely removed with data radiometric corrections (Veljanovski & Oštir, 2011).

To overcome this drawback a contextual multi-level change detection approach was applied that can efficiently treat most of the unwanted differences and suppress sensor related noise (Veljanovski, 2008). Taking into account the neighbourhood and change information by joining two spatial scales (Fig. 8), approach reduces amount of small size false differences.

Fig. 8. Change detection approach takes into account the neighbourhood and change

The model is based on focal information logic that gives averaged change information in a slightly reduced spatial scale – within the specified neighbourhood. It is based on the fact that in a larger geographic area (e.g. a 3 x 3 pixel window or 3 m spatial resolution for resampled VHR data) the information of the changes will tend to level abrupt change information if a small spatial scale change is present, and will show the averaged difference if the majority of pixels in the observed window are subjected to change. Computing the piecewise change information between two time-successive data sets provides valuable information regarding the location and numeric change value derived from contextual

information by joining different spatial scales.

information within the specified neighbourhood.

urban growth of Raila village.

The procedure was implemented as follows. First, spectral information for a slightly coarser scale (i.e. 3 m spatial resolution) was computed for images or areas of interest. This may be accomplished with a specified neighbourhood mean value annotation. Second, change differentiation between images is performed on a coarser resolution scale and change magnitude categorisation is applied (see below for categorisation classes and their definition). Third, upper positive and negative changes are reclassified so that the mask of important changes based on the neighbourhood context characteristics is prepared. Fourth, change differentiation is calculated on the original data scale (i.e. GeoEye and QuickBird data in 1 m spatial resolution), then categorisation is applied, and finally a mask (or a mask with a buffer) of an arbitrary specified magnitude of changes (obtained in the previous step) is overlaid in order to restrict merely the contextually supported changes.

Normally, if there is no substantial unwanted effects (noise) due to meteorological or sensor influences in images, changes obtained from the difference image (image differentiation) are distributed normally and symmetrically, with the average at 0. Abrupt changes (objects or land cover transformations) can then be defined by thresholding the distribution tails, giving the pattern of positive and negative changes in reflectance (spectral) space. We have calculated change magnitude (transition class category) intervals for every 0.5 standard deviation. Then the criterion of 2.5 standard deviations for negative changes and 2.0 for positive changes was applied to enclose the majority of detected transformations in both real and spectral world situations. Result of such categorisation of image difference is a pattern of abrupt changes (locations of appearance, disappearance of objects), with no association to change characterisation (type of change, from-to).

Each output from the above automatic threshold procedure is finally refined to an arbitrary degree, depending on case study objective. For entire Kibera informal settlement changes were obtained from comparison of GeoEye 2009-07-25 and QuickBird 2006-03-27 images. Change patches smaller than 5 m2 were eliminated for the purpose of this example, mainly to reduce the impact of change artefacts belonging to small patches of rooftop renovations (Fig. 3). False changes identified due to differences in viewing angle (location of buildings and trees shadows, buildings boundaries due to different original resolution of imagery) were also removed using the results obtained from object based classification (state of land use in 2009, section 4.2). Where shadow or vegetation object class were present, change pattern was corrected in the given context. Described and implemented step-wise change pattern refinement is shown in Fig. 9, for a subset of Raila village and northern neighbourhood. In other words, with simple generalisation we could control many aspects of the change pattern for study's specific aims.

Results were visually examined and evaluation of change pattern characteristics (over- and under-estimations) was done throughout the area using complementary comparison of satellite images involved.

#### **4.3.2 Raila village change detection**

Raila village is known to have undergone extensive development during recent years (Kibera Wikipedia, 2011; MKP, 2011). Thus temporally detailed examination was performed observing its urbanisation. Change detection procedure described in section 4.3.1 was in addition implemented for Raila village only, but for two time sequences: 2006-2008 and 2008-2009. QB 2006-03-27, 2008-08-10 and GE 2009-07-25 images were used for this example.

Object-Based Image Analysis of VHR Satellite Imagery for

needed.

**5. Results** 

respectively.

**5.1 Kibera** 

**5.1.1 Kibera land cover map 2009** 

Population Estimation in Informal Settlement Kibera-Nairobi, Kenya 421

Nonetheless, no estimation so far guessed by the MKP, or the UN, or the Government of Kenya or by other actors can be taken for granted and does not represent the real dimension of the population of Kibera. In general, no estimation can be proved nor refuted until an exhaustive census will be taken throughout the whole slum (Kibera Wikipedia, 2011).

Because population is not directly related to land cover surface reflectance, population estimation is still a challenging task based purely on remote sensing spectral signatures. Although population is not directly measurable on the remote sensing images this technology may provide good approximation of population estimation by measurement of visible variables, e.g. the number of residential buildings and/or the area of build-up zone (Zhang, 2003). There exist many studies using different approaches on remote sensing data for population estimation. Studies date from the early 1970s onward, where air photos were utilized for manual counts of dwelling units. There are three most used methods of population estimation by remote sensing: residence count method, area (density) method and regression model method (Zhang, 2003). Residence count method was mostly done in first period of studies on this topic on the western urban environment (Horton, 1974, Barrett & Curtis, 1986). Area density method was used by H.H. Wang (1990), F.Z. Wang (1990), P. Sautton (1998), Langford et al. (1994), Z.J. Lin (2001) and others. Regression model method is currently also often used (Galeon, 2010, Dengsheng et al., 2006, Zhang, 2003). With each type of method some ancillary field survey data are

Considering the above situation and the fact that for Kibera we lack other potential sociogeographic data (elsewhere applied to predict population with regression technique), we decided to assess the population on residential land cover class information obtained from object-based classification with density per area method solely. For each village a total area of buildings was calculated and different occupation scenarios (i.e. persons/living area)

With object-based (contextual) classification performed on GeoEye image with Feature Extraction module of ENVI EX, it was possible to obtain accurate land cover map and following this, total residential area of Kibera slum and its divisions (villages) with very high accuracy. From this data, those related to build-up areas, were used for population estimation. With multi-level contextual change detection implemented in Erdas Imagine, it was possible to obtain representative change pattern reflecting where in informal settlement intensive urbanisation processes have taken place. Results are presented in the following order: land cover mapping, change pattern identification and population estimation, for Kibera informal settlement (section 5.1) and Raila village (section 5.2),

Object based classification and post-classification on GeoEye 2009-07-25 image was performed for each village in Kibera informal settlement separately. Finally individual

results were joined in a land cover map of Kibera informal settlement (Fig. 10).

were tested to observe the range of possible population fluctuation.

2006-2009 3-band difference image for Kibera subset (Raila village). Change pattern over 2006-2009 3-band difference image.

multi-level threshold procedure.

false changes due to shadows differences. renovations.

small patches due to rooftop

Fig. 9. Overview of change pattern intermediate results through implemented processing steps.

#### **4.4 Population estimation**

Population statistics gives very important information for understanding of modern society. Demographic research is one of the main research directions of social science for a better understanding of the interactions between population growth and social, economic and environmental conditions. The collection of population data depends mainly on the census, which is labour-intensive, time-consuming and demands high financial resources. The 2009 Kenya Population and Housing Census reported Kibera's population to be 170,070 (Karanja, 2010). This report was far from the belief of that time that Kibera slum was of the biggest informal urban settlements in the world. Several actors had provided and published over the years increasing estimations of the size of its population, most of them stating that it was the largest slum in Africa with the population exceeding 1 million. According to Davis (2006), a well known expert on urban slums, Kibera had a population of about 800,000 people. International Housing Coalition (IHC, 2007) talked about more than half a million people. UN-Habitat (2004) had released several estimations ranging between 350,000 and 1 million people. These statistics mainly come from analyses of aerial images of the area. IRIN (2006) estimated a population density of 2000 residents per hectare. In 2008 an independent team of researchers began a door-by-door survey named Map Kibera Project (MKP, 2011). A trained team of locals, after having developed an ad-hoc surveying methodology, has so far gathered census data of over 15,000 people and completed the mapping of 5000 structures, services (public toilets, schools), and infrastructures (drainage system, water and electricity supply) in the Kianda village. Considering data collected for Kianda village, the population of the whole Kibera slum can be estimated between 235,000 and 270,000 people.

Nonetheless, no estimation so far guessed by the MKP, or the UN, or the Government of Kenya or by other actors can be taken for granted and does not represent the real dimension of the population of Kibera. In general, no estimation can be proved nor refuted until an exhaustive census will be taken throughout the whole slum (Kibera Wikipedia, 2011).

Because population is not directly related to land cover surface reflectance, population estimation is still a challenging task based purely on remote sensing spectral signatures. Although population is not directly measurable on the remote sensing images this technology may provide good approximation of population estimation by measurement of visible variables, e.g. the number of residential buildings and/or the area of build-up zone (Zhang, 2003). There exist many studies using different approaches on remote sensing data for population estimation. Studies date from the early 1970s onward, where air photos were utilized for manual counts of dwelling units. There are three most used methods of population estimation by remote sensing: residence count method, area (density) method and regression model method (Zhang, 2003). Residence count method was mostly done in first period of studies on this topic on the western urban environment (Horton, 1974, Barrett & Curtis, 1986). Area density method was used by H.H. Wang (1990), F.Z. Wang (1990), P. Sautton (1998), Langford et al. (1994), Z.J. Lin (2001) and others. Regression model method is currently also often used (Galeon, 2010, Dengsheng et al., 2006, Zhang, 2003). With each type of method some ancillary field survey data are needed.

Considering the above situation and the fact that for Kibera we lack other potential sociogeographic data (elsewhere applied to predict population with regression technique), we decided to assess the population on residential land cover class information obtained from object-based classification with density per area method solely. For each village a total area of buildings was calculated and different occupation scenarios (i.e. persons/living area) were tested to observe the range of possible population fluctuation.
