Loading...
Rockfall cluster filtering for lidar-monitored rock slopes using machine learning
Emmons, Edward Bennett
Emmons, Edward Bennett
Citations
Altmetric:
Advisor
Editor
Date
Date Issued
2024
Date Submitted
Keywords
Collections
Research Projects
Organizational Units
Journal Issue
Embargo Expires
Abstract
With the increased availability of lidar and photogrammetry in recent years, common rockfall monitoring workflows output large quantities of potential rockfall, including regions of erroneous apparent slope change that need to be filtered out as part of developing a rockfall database. Manual separation of true rockfalls from erroneous change clusters is extremely time intensive, and machine-learning-based filters are not effective enough to act as a true replacement for human decision-making. However, using classification models, the most ‘obvious’ false samples can be removed from the change clusters required for manual evaluation, reducing the number of clusters for review and saving dozens of hours of work for a given monitoring site. This study assesses the performance of pre-existing and novel features used to discern true rockfall from erroneous change clusters, and uses oversampling techniques to address the significant class imbalance present in rockfall databases that hinder machine learning model performance. Results indicate that models based on existing geometric features tend to perform approximately as well as models relying on larger numbers of more complex features. It was also found that addressing class imbalance with oversampling is one of the most effective ways to improve classifier performance. Creating synthetic true rockfall samples for the classification model to alter the class ratio from ~1:100 to 1:1 true-to-false samples significantly reduces the bias of classification models and in turn drastically improves performance. In contrast, despite being used for imbalanced data problems, cost matrices are not were found to not be well suited for addressing class imbalance in the context of rockfall classification problems. Finally, this study evaluated the potential for model generalization between monitoring sites and demonstrated that while generalization performance is site-specific, there is potential for using a preliminary filter based on a generalized training data set supplemented with limited site-specific training. Combinations of two and three training sites (separate from the testing site) were able to perform almost as well as the site-specific classifier for one of the testing sites considered, although in other cases, substantial site-specific data needed to be added for training to achieve performance approaching that of a fully site-specific model. These findings highlight the importance of feature engineering and addressing class imbalances to improve rockfall classification, and confirm that preliminary filtering models can effectively reduce manual effort, even with limited site-specific training.
Associated Publications
Rights
Copyright of the original work is retained by the author.