UVM Theses and Dissertations
Format:
Print
Author:
Besaw, Lance E.
Dept./Program:
Civil and Environmental Engineering
Year:
2009
Degree:
PhD
Abstract:
In this dissertation, we develop, modify and apply artificial neural networks (ANNs) to solve complex problems in surface and subsurface hydrology. Parametric statistical and physics-based models are widely used for conditional simulation, classifying, clustering and mining data as well as state-variable, parameter and time-series estimation. However, many environmental datasets comprise enormous amounts of multiple data types (continuous, nominal and categorical), represente different spatio-temporal scales, and do not conform to parametric assumptions. In addition, resources are not typically available to develop and calibrate appropriate physics-based (or other traditional) models. Given the abundance of information collected about the Earth and its human-induced changes, data-driven ANNs provide computational tools to understand and improve the way we make decisions about natural resources. Four different ANNs are developed and applied to data fi-om real-world environmental engineering problems.
A modified counterpropagation network (CPN) is used to assimilate subsurface biogeochemical data, estimate spatially auto-correlated variables and produce equiprobable conditional simulations for subsurface characterization. The classification and clustering abilities of the CPN and an ANN known as the Kohonen Self-organizing Map (SOM) are used to classify stream geomorphic condition and susceptibility to channel adjustment. Recurrent CPN and generalized regression neural network (GFNN) algorithms were developed to forecast streamflow in ungauged streams using publically available climate and discharge records for the Winooski River basin, Vermont. Finally, a supervised spiking neural network (SNN) was combined with an evolutionary strategy to explore the algorithm's ability to forecast streamflow. In all applications, the ANNs prove comparable or superior to traditional methods (where direct comparisons are applicable). This work demonstrates the usefulness of ANNs for nonlinear input-output function approximation, pattern recognition/classification and data clustering in surface and subsurface hydrology applications.
A modified counterpropagation network (CPN) is used to assimilate subsurface biogeochemical data, estimate spatially auto-correlated variables and produce equiprobable conditional simulations for subsurface characterization. The classification and clustering abilities of the CPN and an ANN known as the Kohonen Self-organizing Map (SOM) are used to classify stream geomorphic condition and susceptibility to channel adjustment. Recurrent CPN and generalized regression neural network (GFNN) algorithms were developed to forecast streamflow in ungauged streams using publically available climate and discharge records for the Winooski River basin, Vermont. Finally, a supervised spiking neural network (SNN) was combined with an evolutionary strategy to explore the algorithm's ability to forecast streamflow. In all applications, the ANNs prove comparable or superior to traditional methods (where direct comparisons are applicable). This work demonstrates the usefulness of ANNs for nonlinear input-output function approximation, pattern recognition/classification and data clustering in surface and subsurface hydrology applications.