Thesis Format
Monograph
Degree
Doctor of Philosophy
Program
Civil and Environmental Engineering
Supervisor
Kopp, Gregory A
Abstract
The use of forests as a damage indicator for forensic tornado analysis has been underutilized in the past. As such, better utilizing forensic tree damage would aid the assessment of tornado risk in Canada. Previous studies using forests have been performed in an ad hoc manner. There are several emerging methods for analysis of tornado damage in forests involving different approaches including the Box Method in Canada. These approaches still need development to be used systematically with many aspects still rooted in expert judgment or that are performed manually with subjective processing of data.
The goal of this work is to create a framework for systematically analyzing tornado damage through forests and to start the process for a fully automated methodology. This includes examining and assessing the different methods used, manually examine tornado tracks with these methods, determine the needs for automated processing of the initial data set, and develop an approach for the Box Method.
The current approaches used for forest tornado analysis were examined and compared and the strengths and weaknesses identified. Manual analysis was performed with the Box Method. Issues with how subjective the analysis can be were examined in-depth. The rigourous analysis yielded solutions with how to handle the subjective analysis previously used. Many of the areas that are problematic from an objective standpoint are the edges of the determined tornado path. These aspects include defining where these bounds are, how to handle areas with scattered treefall and irregular patterns of treefall damage. These issues lead to different solutions in identifying the centrelines of the tornado and the width of damage. Using different areas of the damage will lead to different tornado intensities so properly defining or identifying this issue is important.
Using the raw imagery, an AI algorithm was used to identify treefall and perform pre-processing of data for further analysis. This included identifying the treefall and treefall direction. A framework for an automated Box Method was proposed with details of the proposed algorithm and the possible areas which will need improvement.
This work focuses on the Box Method, with the intention of automating the Box Method. To better examine this method, the analysis was performed manually for the Box Method. For this analysis, tracks were selected to examine, and the treefall was identified. In practice, identification of trees has been performed by estimation, whereas a computer-driven analysis focuses on exact measurements, whether it be individual tree segments or selected areas of treefall. The manual analysis underwent a more thorough examination of the tornado track, while providing insight into how expert opinion related decision points influence the analysis affecting the identified tornado track, the treefall and observed degree of damage. This more rigourous analysis was compared to the operational examination, observing which portions of the analysis may be less well defined for a computer driven analysis, relying on expert opinion to refine and adjust. These comparisons were observed and attempts at creating operational computer-driven alternatives were made.
With a computer-driven analysis in mind, an algorithm was trained to determine the initial identified trees (masks) for tornado analysis. This began with the training and identification of treefall throughout forests and the treefall direction. The models were trained based on several training sets. The reliability of these data sets and flaws of the artificial intelligence learning model were identified. Furthermore, a starting model for determining the treefall based on the treefall direction and location was created. These models were developed to their initial stages and will need future development. The current models require more inputs as the issues faced with a fully automated procedure were discovered to be more nuanced and complex.
With the raw imagery data, the algorithm for automatically analyzing the Box Method could be considered. From the treefall data, image processing of the treefall is necessary to manipulate the image and observe the pixels in a manner that can be appropriately used in this method. The algorithm was broken down into individual sections to conduct the analysis and the individual portions were used to analyze portions of tornado tracks. With this algorithm, the framework for a fully autonomous Box Method can be performed with refinement.
This work started the development of analysis of tornado tracks through forests, addressing some of the initial issues from gathering raw imagery. Different analysis methods were considered and examined and development of a framework for the Box Method that could be utilized by other analysis method was created.
Summary for Lay Audience
Tornadoes cause a considerable amount of damage in North America. Tornado counts are incorrect particularly because of missing tornadoes in forests. The tornadoes occurring through forests are harder to identify and analyze, leading to difficulties in determining tornado occurrence and intensity. Trees in these areas may be better used to examine these tornadoes.
There are currently difficulties in using trees to determine the wind speed. Forest damage is often examined on a case-by-case basis. In more recent years, methods have been created to examine treefall damage for tornadoes. These methods still have challenges including their dependence on expert opinion and need to be manually processed.
This thesis examines different methods and attempts to develop approaches to automate these in a more systematic way. The belief is that a program can better analyze the treefall and estimate the wind speeds, this will benefit how researchers analyze tornadoes and better understand tornado risks.
To perform this analysis, treefall was examined manually. The manual analysis had several portions where expert opinion is relied on. These portions of the analysis had differing opinions from expert to expert. As such, solutions that compromise how these experts see the damage were developed.
Automation of treefall was the next problem identified. The important characteristics of treefall were found, and an artificial intelligence was developed and trained to find this treefall. Other aspects that would be useful for automation were explored and considered.
From this, a framework of how the Box Method would be automated was created. This framework focused on the individual steps and removed as much of the expert judgment as possible.
Overall, the automation frameworks were created for forest analysis. The work examined different pieces of how tornado analysis is performed and used these together to create an automated method that can be used in a consistent way.
Recommended Citation
Hong, Emilio S., "Development of Automated Analysis Methods for Tornado Damage to Trees in Forests" (2024). Electronic Thesis and Dissertation Repository. 10343.
https://ir.lib.uwo.ca/etd/10343