Mapping and monitoring urban trees in the City of Johannesburg using remote sensing techniques: the case of Randburg municipal area
Date
2021
Authors
Jombo, Simbarashe
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The precise identification and mapping of urban tree species are vital for the monitoring and management of such ecosystems. Sustainable management of trees requires reliable information about the condition and distribution of the species involved. Municipalities and other stakeholders need such information to facilitate decision making and prioritise resource allocation for effective management of the urban tree ecosystems. Urban tree managers need information on the spatial distribution, status, and changes of urban trees. The heterogeneity and complexity of the urban landscape features, however, make the tree monitoring process challenging. The high structural and spectral intra-species variability and species similarity because of phenology, shadow, and tree age differences limit tree species classification using remote sensing. Traditionally, urban tree classification and mapping were conducted through field-walking surveys which are costly, time-consuming, and labour intensive. However, the technology of remote sensing has overcome such challenges by offering repeatable, time and a cost-effective technique with the advantage of extracting chronological data of the urban trees. The main objective of this study was, therefore, to map and monitor urban trees in Randburg municipal area in the City of Johannesburg (CoJ) using field spectroscopic, WorldView-2 (WV-2), Light Detection and Ranging (LiDAR) and Moderate Resolution Imaging Spectroradiometer (MODIS) data. First, the study explored the potential of feature selection methods in the classification of urban trees using field spectroscopic data in the Randburg municipal area located in the CoJ. The effectiveness of three feature selection methods, namely, Guided Regularized Random Forest (GRRF), Partial Least Squares Discriminant Analysis (PLS-DA) and Principal Component Analysis Discriminant Analysis (PCA-DA),in selecting key wavelengths in the classification of common urban tree species was investigated and compared. The selected key wavelengths were classified using the Random Forest (RF) and Support Vector Machines (SVM) machine learning algorithms. The feature selection methods reduced the high dimensionality of the field spectroscopic data, with GRRF selecting 13 wavelengths, whilst both PCA-DA and PLS-DA selected 10 wavelengths from the entire dataset (n=1523). The results indicated that the most key wavelengths were selected from the short-wave infrared (SWIR) region, suggesting that they were the most effective in classifying the common urban tree species. The SVM classifier produced overall accuracy values of 95.3% for the GRRF, 93.3% (PLS-DA) and 86% (PCA-DA). The RF classifier results were 88.7%, 72% and 64% for the GRRF, PLS-DA and PCA-DA methods, respectively. The SVM outperformed RF in the classification of the key wavelengths. The results of this study did not show the desired spatial distribution of the classified urban trees. The results suggested using very high resolution (VHR) multispectral data would better classify, map, and show the common urban trees spatial distribution in the study area. The results were validated using ground truth data collected in the area of study. The second part of this study assessed the capability of remote sensing to map and classify urban trees using VHR multispectral data. The WV-2 imagery was used to map the common urban tree species and land use and land cover (LULC) classes in a complex urban environment of the city using pixel-based classification methods. The study tested the utility of WV-2 bands and compared the RF and SVM performance in the classification of
common urban trees and LULC classes in the Randburg municipal area. The red, red edge, near-infrared 1 (NIR1), and NIR2 bands played the most important role in classifying the common urban trees and LULC classes. The study produced accuracies of 84.2% using the RF classifier and 81.2% for the SVM classifier. The RF overall accuracy was 3% higher than SVM accuracy in the classification of the common urban trees and LULC classes. There were misclassifications in the study, and these were due to several factors, which include the “mixed pixel problem” and the reference points’ location errors. These misclassifications can be dealt with using the object-based classification method. The object-based classification is often indicated as a relatively robust classification method, and therefore, this was investigated in the third part of the study. The object-based ensemble analysis was examined to classify common urban trees using the WV-2 multispectral imagery. The study tested the object-based method's capability and tested the utility of WV-2 bands and other feature variables in the classification of common urban trees and LULC classes. The performance of the RF and SVM was compared in the object-based classification. The results indicated that the object-based classification is capable of successfully mapping urban trees using WV-2 imagery. The overall classification accuracy was 91.9% for the RF classifier and 87.3% for the SVM classifier. The overall accuracy for the RF improved from 84.2% using the pixel-based classification method to 91.9%, whilst for the SVM algorithm, it increased from 81.2% to 87.3%. The results showed an improvement in the accuracies of the object-based classification of common urban trees and other LULC classes. The near-infrared1(NIR1), NIR2 and Normalized Difference Vegetation Index (NDVI) were the most important indices for the trees' effective classification. This study shows that object-based analysis produces better results than the pixel-based approaches because it considers various aspects such as the objects’ location, shape, texture, directional pattern and area. It has been noted in previous studies that combining LiDAR data with the VHR multispectral data and vegetation indices can produce higher accuracies using object-based analysis in the classification and mapping of the urban trees. This was investigated in the fourth part of this study. The effectiveness of LiDAR point-cloud data, WV-2 bands, and vegetation indices were examined to classify and map common urban trees. The study also ranked the importance of the fused dataset (normalised Digital Surface Model (nDSM), WV-2bands and vegetation indices)in the classification of common urban trees and LULC classes. The results indicate that the fused dataset effectively maps the common urban trees and LULC classes in the study area. This was shown by high accuracies of 97% for the RF classifier and 94% for the SVM classifier. The nDSM was the most effective variable in the classification, as shown by the high Mean Decrease Accuracy (MDA) value of 0.98 and Mean Decrease in Gini (MDG) value of 0.61. The Green Normalised Difference Vegetation Index (GNDVI) was the least important variable with an MDA value of 0.18 and an MDG value of 0.01.This study provides information to the end-user on the type of data and methods that can be used in the mapping of urban tree species to produce high accuracies. An urban tree inventory and comprehensive urban tree assessment are needed to provide information on the tree distribution and benefits of particular species in an urban environment, including mitigation of air and land temperatures, filtering of gas pollutants, and mitigating consequences of the urban heat island (UHI). This suggested the need to determine the relationship between vegetation and land surface temperature (LST) in the study area. Finally, the study assessed the spatiotemporal dynamics of vegetation coverage and LST at a city-regional level. The spatial distribution of LST and NDVI was investigated in seven city-regions over five years (2001, 2005, 2010, 2015 and 2020). This study also analysed the mean variances of LST and vegetation cover over the years of study. The study also examined the relationship between LST and NDVI over the period of study. The results indicated an increase in LST in all the city-regions with the highest value of 20.1°C in city-region G, followed by 19.6°C in city-region E. The vegetation cover decreased in all the city-regions over the years of study, with the lowest NDVI value of 0.39 in city-region G, followed by city regions F (0.43) and D (0.48). There was a negative correlation between LST and NDVI values in the study area, ranging from-0.03 to -0.76. This information is useful to urban authorities to find areas that need trees to enforce policies that minimise the cutting down of trees. This study can assist municipalities and other stakeholders involved in urban planning and management of the UHI effects to mitigate local and regional thermal changes. Overall, this study showed that remote sensing techniques are applicable in the classification and mapping of trees in a heterogeneous urban area. The methodologies in this thesis can be used in the classification of other tree species in both heterogeneous and homogeneous environments, which adds knowledge to urban tree management. This is important to municipalities, researchers and other stakeholders to develop a comprehensive urban tree inventory, vital in providing sustainable urban tree management
Description
A thesis submitted to the Faculty of Science, University of the Witwatersrand, Johannesburg, in partial fulfilment of the requirements for the degree of Doctor of Philosophy (Geography and Environmental Science)