| Knowledge of the flow rates in rivers and major streams is critical for understanding and effectively managing watersheds because it bears on such things as predicting flood vulnerability and fluvial erosion of the landscape, as well as mitigating the damaging effects of floods and waterborne pollution. Spatially dense networks of in situ streamflow measurements are generally nonexistent because of the high cost of installing, calibrating, and maintaining the necessary equipment. A framework is therefore needed to predict streamflow from other available data. The relevant data are often of different modalities, however, such as images (e.g., digital topographic maps), time-series (e.g., an in situ streamflow guage), and spatio-temporal (e.g., daily radar-based precipitation estimates), and the estimation of parameters that exhibit complex spatio-temporal dependencies from such multi-modal measurements poses a significant challenge. Most data fusion algorithms explored to date work with data belonging to a single mode or "family" possessing one-to-one sample mappings, such as images or time series data, and cannot be directly applied here. In this work, a simple yet efficient probabilistic framework based on Bayesian networks (BNs) is developed that provides a mechanism to incorporate spatial, temporal and spatio-temporal features. Graph topologies are generated based on physical models characterizing the underlying random process.;Solving high-dimensional estimation problems over large data sets leads to computational demands that are often impractical to run even on high-end General Purpose Processors. This is true for the framework proposed in this work as well. FPGA-based reconfigurable computing has been successfully used to accelerate computationally intensive problems in a wide variety of scientific domains to achieve speedup over traditional software implementations. However, this potential benefit is quite often not fully realized be cause creating efficient FPGA designs is generally carried out in a laborious, case-specific manner requiring a great amount of redundant time and effort. Algorithm decomposition, performance prediction, and design reuse for parallel implementation all need to be performed in an efficient and structured manner in order to develop successful FPGA designs in a productive way. To address these challenges, an approach for pattern-based decomposition of algorithms for FPGA design and development is proposed.;The major contributions of this work include (1) development of a scalable approach to fuse spatial, temporal and spatio-temporal watershed data via a BN to estimate streamflows; (2) prediction of uncertainties in streamflow estimates as a function of location (space), condition (time) and the particular mix of sensors; and (3) accelerating algorithm execution on FPGAs using pattern-based decomposition with significant increases in productivity via design reusability. (Full text of this dissertation may be available via the University of Florida Libraries web site. Please check http://www.uflib.ufl.edu/etd.html)... |
