Chris Cook M.S. defense
11 a.m. Wednesday, Oct. 14
Available only via zoom
In partial completion of the M.S. in Geosciences from the University of Alaska Fairbanks College of Natural Science and Mathematics, Department of Geosciences
Thesis title: Multi-Temporal Machine Learning Approaches for Rockfall Detection Using High-Resolution Drone-Based Imagery and Topography
Abstract: Unstable slopes, particularly in mountainous areas adjacent to roadways or other infrastructure, present a significant rockfall hazard to communities and human life. State Department of Transportation (DOTs) need to manage, prioritize, and remediate many such slopes which is difficult to do with limited resources. Remote sensing, and specifically high-resolution drone-based remote sensing, offers a safe method for imaging known hazardous corridors with great detail, but imaging alone cannot identify or help prioritize the potential hazards that may be present. Machine learning offers a pathway forward for analyzing these large datasets, but it is unclear what data and derived features, and what types of models, may be of most value for assessing rockfall.
This work seeks to answer the following relevant questions: 1. Can high-resolution drone-derived remote sensing data combined with machine learning accurately identify rockfall in surveyed transportation corridors? 2. Which combination of imaging bands and topographic features provides optimal model performance?, 3. How does the inclusion of multi-temporal versus single-temporal imagery affect detection accuracy?, and 4. How does spatial blocking and cross validation influence model performance?
Through four field data collections that captured drone-based, co-aligned, high-resolution optical, thermal, and elevation data, this work assessed the relative value of these and derived features for their effect on model performance for predicting rockfall. Both eXtreme Gradient Boosting (XGBoost) and Convolutional Neural Network (CNN) -based models were assessed for their performance in identifying areas of recent rockfall. A rockfall / no-rockfall polygon dataset was created for each time interval to use as training and testing for models, and a spatial cross validation and blocking regime was used to ensure that XGBoost and CNN models were trained on the same blocks and folds for each interval while reducing the impact of spatial autocorrelation on model performance.
Results showed that elevation and its derived topographic indices slope and roughness both held considerable importance for model accuracy, with the differenced topographic features being of primary importance. Multitemporal, 21-feature, XGBoost models achieved the best model average Receiver Operating Characteristic Area Under the Curve (ROC-AUC) of 0.92 across the three observed time intervals. Single timepoint models showed that pre-rockfall data had the least importance and had little to no predictive capabilities when used in isolation and average ROC-AUC of 0.57, suggesting that pre-rockfall imagery and topography data are not sufficient for predicting what rocks may fall in the near-term on a surveyed slope. Post-rockfall -trained models were modestly better with a ROC-AUC of 0.69, suggesting that post-rockfall imagery and topography data collection likely still would not be sufficient for the purposes of a state DOT attempting to dependably identify recent rockfall source areas. Optical and thermal features and their time interval differenced features provided models with relatively little predictive capability, with models achieving an average ROC-AUC of 0.69, suggesting that elevation and its derivatives may be sufficient for data collection for rockfall detection. Nearly all tested models showed that the inclusion of absolute elevation harmed model performance. XGBoost and CNN models showed similar capabilities in terms of model accuracy, with CNN models outperforming the XGBoost pixel-based models as patch size increases. It is likely that rockfall detection models would require multiple surveys to reliably detect rockfall, but it is evident that elevation and derived topographic features in isolation are sufficient.
Major advisors: Franz Meyer and Simon Zwieback