J2.2 Assessment of Machine Learning Ensembles for Modeling DEM Uncertainty in Marshes with Terrestrial Laser Scanning

Wednesday, 9 January 2019: 3:15 PM
North 124B (Phoenix Convention Center - West and North Buildings)
Chuyen Nguyen, Texas A&M University−Corpus Christi, Corpus Christi, TX; and M. J. Starek, P. Tissot, and X. Cai

Precise and accurate measurement of marsh surface elevations is a necessary component in studying resiliency to relative changes in sea level or episodic events. Terrestrial laser scanning (TLS) is a lidar-based approach for dense 3D measurements. The main objective of this research is to evaluate and compare two machine-learning (ML) methods for modeling bias and uncertainty in Digital Elevation Models (DEMs) of different types of marsh surfaces derived from TLS. The two ML methods explored are: Artificial Neural Network (ANN) and Support Vector Machine (SVM). These approaches are compared when estimating DEMs by predicting the difference between TLS and RTK measurements. The spatial variability of the modeled elevations uncertainty is estimated as part of the model calibration based on the TLS point cloud data. Local point cloud attributes are rasterized by using radial bins based on geometric characteristics of the point cloud: point density, surface roughness, Principle Component Analysis (PCA) curvature 1, PCA curvature 2, PCA curvature 3. These features are then used as predictors for a neural network and SVM where the predicted outcome is the estimated DEM measurement difference at a given grid cell. A set of RTK GNSS data points (> 600) were collected as ground truth to train the model. Data points were collected in exposed and vegetated terrain. Ensemble model were calibrated while randomly selecting 75% of the data for model training and 25% percent for independent performance testing for each ensemble member. A good ANN model topology was selected after varying successively the number of hidden neurons; a good SVM model was also selected after varying successively the kernel functions. Each selected model was evaluated via bootstrapping with increasing number of RTK points to assess learning and prediction performance relative to the density of RTK measurements. The final output is a statistical representation of measurement bias and uncertainty on a cell-by-cell basis thereby providing a means to correct for DEM error and propagate spatially varying change detection uncertainty. The developed framework is generalizable to other terrain types and survey methods that generate 3D point cloud data such as airborne lidar or structure-from-motion photogrammetry.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner