Quantification of Tree Crowns in Urban Areas Using Very High Resolution Image

Introduction

Trees constitute the most important component of green infrastructure in urban areas, providing numerous environmental, economic, and social benefits. Many studies have corroborated the benefits of planting trees in urban areas (McPherson et al. 1994; Tyrvainen et al. 2014; McHale et al. 2007; Ball 2012). Therefore, decision makers and planners, including land managers and administrators, are aware of the balance that should exist between urban infrastructure and the green space. This explain the relevance of monitoring the state of natural resources in cities. There is a need for a quick and reliable inventory of urban forest resources because it has been always difficult to update the urban green ecosystem using traditional field survey methods. Efficient urban forest management demands detailed, timely, repeatable, and spatially explicit information. In recent years, the launch of very high spatial resolution satellites (IKONOS, QuickBird, WorldView, among others), as well as the development of new classification algorithms, have initiated a new era in forest management using remote sensing technology (Plantier et al. 2006). In 2009, WorldView-2, a new satellite-borne sensor, was launched by Digital Globe. The very high spatial resolution (0.5 m in panchromatic bands and 2.0 m in multispectral bands) and four new spectral bands (Coastal, Yellow, Red Edge, and Near Infrared 2) in addition to the four standard bands (Blue, Green, Red, and Near Infrared 1) of this satellite confirm the expectation that this sensor has a high potential for tree mapping. The data provider postulates that all four new bands are strongly related to vegetation properties (Koukal et al. 2012). With the increasing availability of high spatial resolution data and the computational power to process it, remote sensing research in forestry has focused more and more on detecting and measuring individual trees as opposed to obtaining stand level statistics (Cabello-Lebric 2015). This technology provides opportunities for investigating and quantifying the structure and floristic of forests at both the stand and individual tree level (Bunting and Lucas 2006).

Although the use of high resolution satellite data is the best solution for tree quantification in urban areas, some factors, such as the increase of intra-crown spectral variance found in very high resolution (VHR) imagery and the low spectral separability between tree crowns and other vegetated surfaces in the understory (Gougeon and Leckie 2006; Hirschmugl et al. 2007; Pouliot et al. 2002) can comprise its efficiency. Furthermore, the coexistence of trees with urban infrastructures results in spatial arrangements that can complicate the interpretation of an image. In this chapter, an attempt has been made to quantify trees in urban areas using very high resolution WorldView-2 imagery using both a Normalized Difference Vegetation Index (NDVI) threshold and image segmentation techniques. A quick and accurate way to obtain inventory data on individual trees and clustered trees to replace the traditional surveying method that is less accurate and time-consuming is demonstrated. The resulting database can be used in urban planning to recognize the importance of green spaces in urban clusters. Land administrators should integrate remote sensing and geographic information system techniques in urban green space monitoring sectors, which will lead to better decision making as well as time-saving.

Methodology

19.2.1 Study Area Description

The study was conducted in Kigali city, Kicukiro District, Nyarugunga sector, Kamashashi village, which is located between 1°46’19”S to 2°05’14”S and 29°46’40'’E to 30°28’36”E. (Figure 19.1). Kigali city consists of a complex of urban infrastructures including buildings, roads, sidewalks, canals, and so on, mixed with trees distributed in different patterns (block plantations, evenly spaced, or in arbitrary spatial pattern) and other green areas.

  • 19.2.2 Data Collection and Processing
  • 19.2.2.1 Satellite Data

To perform this study, WorldView-2 satellite image was used. The WorldView-2 satellite provided high spatial resolution data in eight spectral bands, with 0.46 m and 0.5 m spatial resolutions for panchromatic band and 1.84 m to 2 m for multispectral bands at nadir and off-nadir, respectively. Images were recorded in the spectral range of 450 to 1040 nm. Further details about the sensor can be found on the WorldView-2 website (http://worldview2.digitalglobe.com/). The high spatial resolution data provided by WorldView-2 satellite enables the viewer to discriminate and map fine details like shallow reefs, individual trees, and other diverse information such as the quality of urban infrastructures, cadastral information, the health of plants, and to manage various environmental resources. Through different methodologies, WorldView-2 data provides a rapid, standardized, and objective assessment of the biophysical impact, in terms of vegetation cover and restoration interventions (Mariana et al. 2014). Photo interpretation of a time series of aerial photography is generally used to qualitatively evaluate the long-term effectiveness of restoration interventions in terms of persistency, including for recognizable structures such as terraces, grubbing patterns, revegetated areas, and so on (Rango et al. 2002).

19.2.2.2 Image Processing

Image fusion as a process of merging several images acquired with different spatial and spectral resolution at the same time together to form a single image to enhance the information extraction

Study area location

FIGURE 19.1 Study area location.

  • (Sarup and Singhai, 2010) was performed to enhance the quality of the WorldView-2 image. Panchromatic and multispectral images of the WorldView-2 satellite were fused using the principal component (PCA) technique. In this process, the first principal component of low resolution data is replaced by high resolution data (Shamshad et al. 2004). The PCA is the easiest and most useful of an eigen vector-based multivariate analysis, as it helps to reveal the internal structure of data in an unbiased way (Priya and Chudasama, 2015). The PCA technique helps to retain the spectral characteristics of the multispectral imagery to enhance the spatial resolution of the fused image.
  • 19.2.2.3 Shadow Enhancement

Based on the direction of the sun, the shape and position of shadows can serve as additional information in tree shape detection (Hung et al. 2006). The green cover of the study area is composed of a mixture of mature trees, small trees, and bushes. Therefore, shadow was used to distinguish the mature trees from low-lying vegetation (shrubs). The first principal component analysis was performed to enhance the shadow and the output (decorrelated image) was classified using an imageclustering algorithm. The maximum iterations were limited to 10 with two classes (i.e., shadow and nonshadowed areas). Due to less spectral variability in the shadow and water body, NDVI was calculated using the original image of WorldView-2 to extract water bodies (pixel values > -1 and < 0), and this was used to mask out water body areas from the classified image.

  • 19.2.2.4 Tree Crown Extraction
  • 19.2.2.4.1 NDVI Approaches

The tree crown identification process is affected enormously by spectral separability of tree crown pixels with respect to their background. Therefore, it is better to remove all other features that hinder the correct identification and delineation of tree crowns using NDVI. The use of NDVI is very important because it is insensitive to intra-crown shadow variations resulting from the sun elevation angle and physical structure of trees (Ardila, et ah, 2012). This property is generally exploited to extract tree crowns with circular and compact shape. NDVI was computed using Near Infrared and Red bands in the WorldView-2 image, and then a threshold (0.5 < threshold > 0.22) that differentiated trees from other features was generated. The threshold was used to extract crown areas from the pan-sharpened image of WorldView-2; the output image has pixel values of mature trees as well as the low-lying shrubs. It was observed that the low-lying shrubs had approximately the same NDVI values as the mature trees and crowns were subjected to supervised classification (maximum likelihood). Maximum likelihood assumes that the distribution of a class sample is normal. While working through a maximum likelihood algorithm, it is necessary to have well-defined training areas and pure signatures to acquire an expected result. Knowledge of the data, and of the classes desired, is required before classification (Jensen 1996). The image was classified in seven classes or categories according to their spectral reflectance. For this study, at least 20 training sites were collected for each class for classification. The number of classes was further reduced to six to eliminate some low-lying shrubs. Thereafter, the final output of classification was subjected to vectorization to extract tree crowns. The vector was then used to quantify trees.

19.2.2.4.2 Segmentation

The optimal segmentation parameters depend on the scale and nature of the tree crowns to be detected, which differ considerably between coniferous and deciduous trees (Ardila et al. 2012). By multiresolution segmentation, the image was segmented to evaluate tree crown polygons using several levels of detail and adapted to shape and compactness parameters until the crowns and segmented polygons become almost congruent in order to get individual trees and tree clusters w'ith similar shapes and spectral properties.

19.2.2.4.3 Vectorization

Vector and raster are two types of spatial data structures used in a geographic information system (GIS). With the development of GIS and remote sensing (RS) technologies, it is easy to rapidly convert raster to vector data and establish topological relations among vectorized polygons is becoming a bottleneck in data integration between GIS and RS (Chen and Zhao, 2005). Therefore, based on the previous process, a vectorization method was proposed to classify, segment, and texture RS raster data quickly in order to automatically establish topological relations.

19.2.2.5 Accuracy Assessment

A field survey was conducted using a handheld GPS to collect ground truth points, which were used in accuracy assessment and validation. Accuracy assessment is a general term used to compare the classification to geographical data providing the assumptions are true, so as to determine the accuracy of the classification process (Madhura and Suganthi 2015) (Figure 19.2).

 
Source
< Prev   CONTENTS   Source   Next >