Master of Science
Remote sensing can play a key role in understanding the makeup of urban forests. This thesis analyzes how high-resolution multispectral imagery, lidar point clouds, and multidate multispectral imagery allow for improved classification of London, Ontario’s urban forest. Chapter 2 uses object-based support vector machine classification (SVM) to classify five types of trees using features derived from Geoeye-1 imagery and lidar data. This results in an overall accuracy of 85.08% when features from both data sources are combined, compared with 77.73% when using only lidar features, and 71.85% when using only imagery features. Chapter 3 makes use of Planetscope and VENuS images from different seasons to classify deciduous trees, conifers, non-tree vegetation, and non-vegetation using SVM. Using multidate Planetscope images increases overall accuracy to 83.11% (8.19 percentage points more than single-date Planetscope classification), while using multidate VENuS images increases accuracy to 72.18% (2.22 percentage points higher than single-date VENuS classification).
Summary for Lay Audience
Urban trees provide numerous benefits to a city’s environment, as well as the health of its people. It is often necessary for urban planners to know the makeup of tree species in the urban forest. Trees can be identified and classified by species using remotely sensed data. This data is often imagery, but other data sources such as lidar (3D point data from laser pulses) also allow for classification. This thesis focuses on two different data sources for classifying trees. The first source is a combination high-resolution imagery and lidar data. The second contains multiple images of the same area on different days of the year.
In chapter 2, features derived from imagery and lidar, which ultimately represent the chemical and structural traits of trees, are used to classify five types of trees in London, Ontario. Object-based classification is used, meaning individual trees crowns are delineated and classified, rather than just classifying individual pixels. It is found that lidar features perform better than imagery features, resulting in more trees being classified accurately. However, combining features from both data sources results in an even higher level of accuracy.
Chapter 3 focuses on using imagery obtained on different dates, to capture seasonal changes in vegetation. Four dates are used, representing different stages of leaf development in trees. Two sensors are used, Planetscope and VENuS, which have rarely been used for multidate tree classification. Planetscope has higher-resolution, but has fewer bands, meaning it captures less detailed spectral information. VENuS has more bands but lower spatial resolution. Classification is performed on image pixels and classifies the study area into deciduous trees, conifers, non-tree vegetation and non-vegetation. Significant improvement to accuracy is found for Planetscope when using multiple dates, in particular using images from April when leaves are not present and July when leaves are fully grown. Improvement from using multiple dates is smaller when using VENuS.
Roffey, Matthew, "Vegetation and Tree Species Classification Using Multidate and High-resolution Satellite Imagery and Lidar Data" (2019). Electronic Thesis and Dissertation Repository. 6454.