Probabilistic methods for improved change detection and prediction on sandy beaches using high resolution airborne lidar

Airborne light detection and ranging (lidar) can sample beach topography at orders of magnitude higher spatial resolutions than is practical with standard surveying methods. Data mining and pattern classification techniques offer great potential for coastal monitoring with lidar, but have been relatively unexplored. In the following research, three main contributions are presented: (1) systematic framework to mine high resolution lidar data over a beach, (2) information-theoretic approach to detect morphology indicative of erosion, (3) first research to explore modern probabilistic classifiers to model the effect of morphology on probability of erosion.;Lidar surveys were conducted over a beach on the east coast of Florida multiple times between 2003 and 2007. Through automated profile sampling, several different features are extracted from the data and segmented into binary erosion or accretion classes. Divergence measures are used to rank class separation between features. The more separation provided by a feature, the greater its potential as a morphologic indicator. Morphologic indicators can improve beach monitoring providing insight into the change dynamics and for classifying high impact zones. Deviation-from-trend performed best overall, and it is a contributing factor to anomalous erosion in the study area. Over shorter epochs, slope based features ranked high. A naive Bayes classifier is implemented to test the ability of the features on classifying erosion zones. The top features selected by divergence outperformed correlation and a median metric by approximately 5% and 3% supporting the utility of the divergence method.;To evaluate the joint effect of the features on the outcome of erosion, logistic regression is utilized. A generalized estimating equation (GEE) is applied to handle spatial correlation in the binary responses. To reduce model over fitting and address collinearity among the features, Lasso regression is employed. The ability of the classifiers to predict (classify) zones prone to erosion based solely on morphology is evaluated. Lasso GEE obtained the highest average success rate with 80% and a maximum of 86%. The logistic based classifiers substantially outperformed non-parametric naive Bayes by approximately 7%. The developed classifiers provided a powerful tool for beach characterization with lidar.