**4.2 Land cover classification**

414 Remote Sensing – Applications

procedures of co-registration could not be applied. The non‐linear rubber‐sheeting method was the only possibility to obtain mutually aligned images. This procedure is effective, but very time consuming, since it demands manual selection of hundreds of control points for

Fig. 4. Selecting the control points for rubber-sheeting method for image geo-referencing

technique. Geo-corrected images were evaluated through detailed visual control.

GeoEye image taken on 25th of July 2009 was selected as the reference image considering its highest spatial resolution, good matching with GPS path-network tracks and the fact it is most recent. Then QuickBird images selected for analysis were manually registered to the reference, based on cca. 1,400 manually selected control points per image and using a piecewise transformation based on triangles formed from the tie points (Fig. 4). Resampling was nearest neighbour. An average RMSE is not reported as this is local approximation

Geometrically matched GeoEye and QuickBird images were then used for OBIA. Finally three images were selected for change detection due to their best results in geo-correction phase: GE2009-07-25, QB2008-08-10 and QB2006-03-27. For change detection analysis images were resampled to uniform 1 m resolution, to be prepared for radiometric

After geometric adjustments there are still differences amongst the spectral properties of satellite images (spectral bands from the same or different images are not adjusted to each other). Hence, before pixel-based image comparisons (image differencing) the radiometric standardisation is needed. Most standardisation procedures derive from adjustments of invariant objects or from the least squares method (linear regression). The problem with the first method is that invariant areas should be verified with field measurements. Furthermore, the generally recognised invariant objects, such as deserts or light sandy beaches, do not come into play, since they cannot be found in Kibera area. The principle of

(AutoSync module of ERDAS Imagine).

standardisation.

each image.

Since GeoEye image of the whole Kibera informal settlement contains lots of information and the analysis of the total area would be too demanding in terms of computer processing, we divided image of Kibera into 12 smaller parts (according to 12 informal villages). The complete process of segmentation and classification was applied systematically for each informal village separately, with same parameter settings at each phase. This was possible due to relatively homogenous landscape over Kibera settlement. Thus we obtained 12 regional classification outputs, which were then merged in the final stage. To avoid erroneous classification on the edges of the splitted images, we applied 30 meter buffer when masking the village fragments out from the whole GeoEye image (Fig. 5).

Fig. 5. A buffer of 30 meters around the village border.

Object based classification consists of two stages. Image is first segmented into a set of segments (regions) that are considered to be homogeneous in terms of one or more spectral or spatial properties. Then follows classification where each segment is classified into belonging object class.

Supervised segmentation within software used (ENVI EX, Feature Extraction module) is defined by two segmentation parameters that influence an average size of segments: segmentation and merging. Setting different values for these two parameters causes change of size of segments, allowing for an image to be segmented at many different scales, so both parameter values influence classification results. Since structure of the Earh's surface is similar throughout the whole settlement, same general segmentation parameters were used for each of the 12 villages. Visual example of segment structures are shown on Fig.6.

While classifying Raila village we adapted segmentation parameters to best extract shapes of individual buildings inside informal settlement. Since segmentation parameters were adequate for one particular land use only (i.e. buildings), others were expected to be underor over-segmented. There is no single "optimal" scale for analysis of remote sensing images, rather there are many optimal scales that are specific to the image-objects that exist within a scale (Hay et al., 2003) and this is why using a multi-scale approach may often be preferable (Johnson and Xie, 2011). All spectral bands were used and given equal weight for image segmentation and all available attributes were calculated for all segments.

Object-Based Image Analysis of VHR Satellite Imagery for

Blaschke et al., 2008; Addink & Van Coillie, 2010).

**4.3.1 Change detection of Kibera informal settlement** 

ENVI EX).

**4.3 Change detection** 

Population Estimation in Informal Settlement Kibera-Nairobi, Kenya 417

Fig. 7. Selecting training samples for supervised classification (Feature Extraction module of

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;

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,

Fig. 6. Original GeoEye image (a), objects segmented (b), objects merged (c) (Feature Extraction module of ENVI EX).

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 (Table 2).


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

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

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 polygons).

Fig. 7. Selecting training samples for supervised classification (Feature Extraction module of ENVI EX).
