Efficient multiscale image fusion and feature reduction for elevation data in coastal urban areas

Technological advances in the remote sensing of elevations have made it possible to improve topographic resolutions to the 1-10 meter scale and bathymetric (underwater elevations) to the 5-90 m scale. To a large extent, however, data collected at different resolutions and from different types of sensors remain largely separate and their joint information content under-exploited. For many applications, such as flood modeling, there is a vital need to combine information from these seemingly disparate data sets and extract characteristics about the surface that can aid Earth scientists and emergency planners. The research discussed in this dissertation consists of two parts that address this need.;In the first component of the work, a simplified formulation for calculating the variance of the process noise in a multiscale Kalman smoother is derived and implemented via a pruning method on the quadtree data structure. This method exploits the distribution of measurements in the finest scale to efficiently fuse multi-resolution data with different areas of spatial coverage to produce seamless single surface digital elevation models (DEMs). To further improve the accuracy of the Kalman-based fusion, a landscape-dependent measurement error is calculated for the Shuttle Radar Topography Mission (SRTM) data set, providing information on a scale intermediate to the meter-scale DEMs from airborne laser altimetry (LiDAR) and 90-m coastal DEMs created by the National Oceanic and Atmospheric Association (NOAA). Analysis of the Multiscale Kalman filter and Smoother (MKS) residuals was employed because only a single (spatially uniform) value for the SRTM measurement error is specified by the US Geological Survey (USGS). This was done by defining a LiDAR-derived hydrologic ground surface that excludes vegetation and bridges, under which flood waters can pass.;In the second component of this work, an alternative method to traditional downsampling and mesh generation (used to reduce DEM resolutions to scales at which flood and storm surge modelers can run their fluid-dynamics algorithms) is described. By recognizing that the two main DEM components in urban areas that control flood water flow (i.e. ground and buildings) exhibit different spatial frequencies, vegetation points can be filtered out, resulting in a DEM that can be further decomposed into ground and building classes. The resolution of each class can then be reduced independently, using downsampling for the ground elevations and rectangular parameterization for the building regions (so as to preserve the "channels" for water flowing between buildings even when the overall resolution is dramatically decreased). To examine the performance improvements of the new method compared to the traditional method of downsampling DEMs used by the hydrologic modeling community, two DEM-based hydrologic calculations, flow accumulation number and hydraulic water discharge rate, were calculated. The results obtained from these two calculations confirm that the new method better preserves flood model accuracy with reduced computation time as a function of DEM resolution.