Results Presentation and Discussion

Digital Image Classification

The land use/land cover map provides a comprehensive data set in terms of the overall landscape across the study area. The Kamashashi cell is covered by 10 land use categories (classes): agriculture land occupies 3%, bare soil 1%, built-up land 17%, crop land 43%, grassland or shrub 6%, wetland 10%, open ground 4%, trees area о 11% and water bodies 2% of the total area (Figure 19.3).

Methodological flow chart

FIGURE 19.2 Methodological flow chart.

Vegetation Mapping and Tree Crown Delineation

Based on NDVI thresholds, vegetated area and nonvegetated area were separated. According to Meera et al. (2016), NDVI values vary between -1 and 1, where negative values are nonvegetation and postivie values represent vegetation. The NDVI for vegetation generally ranges from 0.3 to 0.8, with larger values representing “greener” surfaces (Lachowski 1996). A threshold of 0.41 was used to mask out nonvegetated areas from the NDVI image leaving trees and grassland/shrub (Figure 19.4). The spectral reflectance of vegetation is completely different from the reflectance properties of the background material (i.e., water, soil, and settlements). This unique character of the vegetation spectrum makes it possible to separate vegetation from background material with

Land cover map of the Kamashashi cell

FIGURE 19.3 Land cover map of the Kamashashi cell.

NDVI masking for vegetation and nonvegetation areas

FIGURE 19.4 NDVI masking for vegetation and nonvegetation areas.

remotely sensed multispectral data that at least includes NIR and red region reflectance. NDVI is a good index for distinguishing vegetation and nonvegetation covers (Holme et al. 2008). In the Kamashaski cell, the nonvegetated area covers 215.73 hectares, while the vegetated area covers 583.27 hectares.

To separate tree cover and low vegetation, a threshold of (0.7 < threshold > 0.41) was achieved. This was based on image texture analysis, which is considered to be a measure of the spatial variation of image tone or intensity (Haralick 1979). This is useful to identify grassland areas since the spectral variance of grassland objects is smaller than that of tree crowns in very high resolution imagery. Finally, tree canopies (Figure 19.6) were also masked out from the NDVI image. A total of

864,002 trees were mapped and counted through tree canopies in the Kamashashi cell (Figure 19.5).

Image Segmentation Process

Image segmentation was introduced to extract neighborhood information and, preserving homogeneity throughout the satellite image, segmentation algorithms were used to subdivide the entire image at the pixel level (Baatz and Schape 2000; Benz et al. 2004). Segmentation algorithms ideally generate image objects that match the target objects; the shape value was 0.2 and the compactness was 0.8. The threshold of the NIR band was used for masking tree crown with values of 0.41 and 0.68. The features on the resulting image were then classified as trees

Urban Tree Canopies Classification and Counting

According to Jennings et al. (1999), the degree of tree density is expected in percentages. From different NDVI thresholds, this study classified trees canopies as follows: Very High Forest, Canopy Density, High Forest Canopy Density, Moderate Forest Canopy Density, Low Forest Canopy Density, and Very Low Forest Canopy Density (Figure 19.6). The Table 19.1 indicates different

Tree canopies

FIGURE 19.5 Tree canopies.

Tree canopies isolated as polygons

FIGURE 19.6 Tree canopies isolated as polygons.

statistics for each category. The tree canopy cover occupies 49.40 hectares out of 799 hectares in the Kamashashi cell (Table 19.1).

Through the vectorization process, individual trees were delineated and counted. A total of 806,484 trees were found in the Kamashashi cell.

Accuracy Assessment

To assess the accuracy of the study, a total number of 500 ground control points (tree locations) were collected and grouped according to the size of tree crowns. These were then compared to the results obtained from both the NDVI threshold and image segmentation results. The overall accuracy was 99.70% and kappa statistics was 0.9963 (Table 19.2).

TABLE 19.1

Percentage of Tree Canopy Density


Canopy Density (%)

NDVI Threshold

Very High Trees Canopy Density


0.70 > trshld > 1

High Trees Canopy Density


0.60 > trshld > 0.70

Moderate Trees Canopy Density


0.50 > trshld > 0.60

Low Trees Canopy Density


0.46 > trshld > 0.50

Very' Low Trees Canopy Density


0.41 > trshld > 0.46

TABLE 19.2

Accuracy Assessment Table


Class Names

Producer Accuracy (%)

User Accuracy (%)


Group 1




Group 2




Group 3




Group 4




Group 5




Group 6




Group 7




Group 8




Group 9




Group 10



Tree quantification was successfully achieved using pan-sharpened WorldView-2 data since the experiments provided the results with high accuracy (99.70% overall accuracy) in identifications of tree crown using digital images processing techniques. Both NDVI and segmentation-based methods provided reliable results for individual trees mapping. Counting from complex mixed infrastructures in urban areas using very high resolution satellite images despite a certain number of factors that can comprise the satellite image analysis in urban areas, such as shadow of tall buildings, clouds, and calibration and radiometric errors. Compared to existing traditional methods that are time-consuming, less accurate, and labor-intensive, the present method constitutes a quick and accurate way to obtain inventory data on individual trees and clustered trees as well. Such a database can be used optimally in urban planning, recognizing the importance of green spaces in the Kigali Master Plan implementation. The data set can also include and incorporate other factors to predict or model land resources management processes.

< Prev   CONTENTS   Source   Next >