**4.2.1 Legend**

The legend distinguishes four categories: the built-up, the non built-up (vegetation and bare soil), water and clouds.

Some classes are difficult to discriminate using only spectral characteristics, especially so in countries in sub-Saharan Africa. The spectral confusions are numerous, for example, the fields are easily confused with the built-up. Production facilities and services, and the buildings for residential use in some places have the same spectral signature as the sand and burned areas. To overcome these problems, we have enriched the description of spectral regions of texture parameters (see 4.2.3).

### **4.2.2 Selection of training and validation areas**

Training and validation areas were selected based on a visual interpretation of SPOT images supported by a consultation of Google Earth and the plan of the city of Kinshasa, and edited by Aquaterra Kin Art in 1997, ensuring changes due to differences between dates of these documents. 68 areas were selected in common areas of the SPOT images. To ensure an equivalent content of classes on each date, only areas unchanged between 1995 and 2005 were selected. The sample was divided into two, 34 areas for training and 34 for validation.

#### **4.2.3 Choice of attributes**

The attributes used in the classifications were chosen on the basis of visual interpretation. The regions are described in terms of spectral averages in each spectral band and the NDVI and the textural point of view, by the standard deviations on the green and red bands, and two textural parameters of Haralick (1973), such as homogeneity and entropy of the panchromatic band.

### **4.2.4 Segmentation and classification**

eCognition was used to perform segmentation and classification. This software can simultaneously use a variety of data, panchromatic and multispectral images or vector data bases, and can create multiple levels of segmentation using a hierarchical approach.

The segmentation algorithm is the "multiresolution segmentation." According to the "Definiens Developer 7 User Guide" (2007), this algorithm merges the pixels into segments of image by minimizing the average heterogeneity and maximizing their respective homogeneities. It can do the same with image segments from a previous segment. The procedure iteratively merges the pixels or segments, as long as the maximum threshold of heterogeneity is not exceeded. Homogeneity is defined as a combination of spectral properties and form. The spectral homogeneity is based on the standard deviation of the distribution of the colour and consistency of form is based on the deviation from a compact or smooth (Cantou et al., 2006). The procedure can be influenced by the scale factor that limits the size of the resulting segments. The segmentation was performed on the image of 2005 spectral bands of green, red and near infrared respectively, giving them a weight of 2, 1 and 1. The scale parameter was chosen by trial and ,error and set at 20 with the shape parameter 0.1 (0.5 for compactness and 0.5 for smoothing).

The algorithm for supervised classification of the nearest neighbour was used. It ranks the regions according to their proximity to areas of statistical training.

### **4.2.5 Validation**

466 Remote Sensing – Applications

Given the uneven quality of SPOT images and the strong texture of the buildings, they were

The legend distinguishes four categories: the built-up, the non built-up (vegetation and bare

Some classes are difficult to discriminate using only spectral characteristics, especially so in countries in sub-Saharan Africa. The spectral confusions are numerous, for example, the fields are easily confused with the built-up. Production facilities and services, and the buildings for residential use in some places have the same spectral signature as the sand and burned areas. To overcome these problems, we have enriched the description of spectral

Training and validation areas were selected based on a visual interpretation of SPOT images supported by a consultation of Google Earth and the plan of the city of Kinshasa, and edited by Aquaterra Kin Art in 1997, ensuring changes due to differences between dates of these documents. 68 areas were selected in common areas of the SPOT images. To ensure an equivalent content of classes on each date, only areas unchanged between 1995 and 2005 were selected. The sample was divided into two, 34 areas for training and 34 for validation.

The attributes used in the classifications were chosen on the basis of visual interpretation. The regions are described in terms of spectral averages in each spectral band and the NDVI and the textural point of view, by the standard deviations on the green and red bands, and two textural parameters of Haralick (1973), such as homogeneity and entropy of the

eCognition was used to perform segmentation and classification. This software can simultaneously use a variety of data, panchromatic and multispectral images or vector data

The segmentation algorithm is the "multiresolution segmentation." According to the "Definiens Developer 7 User Guide" (2007), this algorithm merges the pixels into segments of image by minimizing the average heterogeneity and maximizing their respective homogeneities. It can do the same with image segments from a previous segment. The procedure iteratively merges the pixels or segments, as long as the maximum threshold of heterogeneity is not exceeded. Homogeneity is defined as a combination of spectral properties and form. The spectral homogeneity is based on the standard deviation of the distribution of the colour and consistency of form is based on the deviation from a compact or smooth (Cantou et al., 2006). The procedure can be influenced by the scale factor that

bases, and can create multiple levels of segmentation using a hierarchical approach.

classified by a supervised method and object-oriented software using eCognition.

**4.2 Land use classification** 

**4.2.1 Legend** 

soil), water and clouds.

**4.2.3 Choice of attributes** 

panchromatic band.

regions of texture parameters (see 4.2.3).

**4.2.4 Segmentation and classification** 

**4.2.2 Selection of training and validation areas** 

The classifications are evaluated by comparing 34 areas of validation within the matrix of confusion. Indices are calculated to assess the quality of results (Richards, 1993):


The overall accuracy is good (> 80%) obtained for the different classifications (Table 2). The Kappa coefficient is only acceptable for the classification of 1995 (85%) and 2005 (92%). The classification of 2000 has a poor Kappa (64%) caused by the fog that covers the southwest of the city. This result will therefore not be used subsequently.


Table 2. Classification accuracy

Extensive field visits conducted in late January 2009 to the end March 2009 in the extension zones of Kinshasa to understand the factors of urbanization has revealed the existence of different confusions and omissions in the class "building". For example, here are some for the image of 1995 and 2005. They are located in Figure 1 and identified in Table 3.


Table 3. Confusion and omission errors for the class built-up in 1995 and 2005

The Mapping of the Urban Growth of Kinshasa (DRC)

These errors will not be included in the analysis.

building when the matrix changes can be quantified.

The layout of the lines of roads and railways digitized,

The location change was analysed using:

Maps and plans of the city of Kinshasa,

The digital elevation model and slope map.

March 31, 1995 and KJ 4 096-358 July 01, 2005)

**5. Results** 

**5.1 Urban growth** 

Population data,

Through High Resolution Remote Sensing Between 1995 and 2005 469

The results obtained at different dates are generalized by removing polygons classified as "built"

The superposition of classes "built" on two successive dates can map the evolution of the

Field visits conducted in-depth from the end of January 2009 through to March 2009 in

The map resulting from the comparison of land use classifications in 1995 and 2005 shows that the extensions of the city is concentrated in the southwest and northeast of Kinshasa (Figure 2).

Fig. 2. Evolution of urban sprawl between 1995 and 2005 (Sources: Images Spot KJ 3 096-358

the extension zones of Kinshasa to understand the factors of urbanization,

with an area less than 1km² and the inclusion of less than 1 km² within the urban sprawl.

Fig. 1. (a and b): Location of misclassification (Source: Delbart and Wolff, 2002 for the map of municipal boundaries)

These errors will not be included in the analysis.

The results obtained at different dates are generalized by removing polygons classified as "built" with an area less than 1km² and the inclusion of less than 1 km² within the urban sprawl.

The superposition of classes "built" on two successive dates can map the evolution of the building when the matrix changes can be quantified.
