UVM Theses and Dissertations
Format:
Print
Author:
Pearce, Andrea Rebecca
Dept./Program:
Civil and Environmental Engineering
Year:
2011
Degree:
PhD
Abstract:
Combining computational and human neural networks created a collaborative process that cultivated a positive feedback loop between computational and subject experts. The original expert-driven science hypotheses necessitated modifications and development of new computational algorithms, and led to more efficient and better-informed researchers. We demonstrate with two environmental field (groundwater and surface-water) applications containing physical, chemical and biological data by using a modified Self Organizing Map (SOM) in tandem with human experts to identify distinct groups within these environmental data.
The SOM is an artificial neural network (ANN) commonly used as a K-means clustering method. The method outperforms many traditional clustering methods on noisy datasets (e.g. high dispersion, outliers, non-uniform cluster densities); and is appropriate when combining the multiple correlated and auto-correlated data associated with most hydrochemical research. We modify the existing algorithm to allow weighting of the input variables according to their relative importance, and by adding a post-processing radial basis function to estimate group membership between measurement locations autocorrelated in space.
In addition, we use a nonparametric multivariate analysis of variance (MANOVA) in tandem with the SOM to identify an optimal number of clusters. Through an iterative, unsupervised training phase, the SOM self-organizes the input data using a distance llletric that measures similarities between input variables. The resulting output map (U-matrix), and maps of individual input variables (component planes) may be visualized in any number of dimensions (we use 2-D maps) for communicating the dataset organization (clustering) to science experts.
Our first application involves subsurface microbiological community composition data from landfill leachate contaminated groundwater. The SOM/MANOVA methodology is able to distinguish between tiers of contamination in this multi-contaminant environment using expert knowledge to guide data preprocessing and weight the input variables. Results show a composite delineation representative of overall groundwater contamination at the landfill based only on microbiological information. Our second application combines chemical, biological and physical data from the water column and underlying lake sediments in Missisquoi Bay, Lake Champlain, Vermont, USA. This bay is plagued by cyanobacteria blooms, which often produce cyanotoxin.
The SOM methodology, in tandem with scientific experts, identifies a significant division our dataset separating samples from one location with and without a cyanobacteria bloom. Cyanobacteria cell counts were not used as input in this analysis. This research corroborates a previously suggested hypothesis that cyanobacteria growth is supported by a set of necessary, but insufficient, conditions and requires a trigger to initiate, propagate and sustain the bloom. Results from these exploratory analyses may help develop monitoring strategies in Lake Champlain, specifically guiding the spatial and temporal sampling frequency necessary to identify the environmental drivers of the blooms.
The SOM is an artificial neural network (ANN) commonly used as a K-means clustering method. The method outperforms many traditional clustering methods on noisy datasets (e.g. high dispersion, outliers, non-uniform cluster densities); and is appropriate when combining the multiple correlated and auto-correlated data associated with most hydrochemical research. We modify the existing algorithm to allow weighting of the input variables according to their relative importance, and by adding a post-processing radial basis function to estimate group membership between measurement locations autocorrelated in space.
In addition, we use a nonparametric multivariate analysis of variance (MANOVA) in tandem with the SOM to identify an optimal number of clusters. Through an iterative, unsupervised training phase, the SOM self-organizes the input data using a distance llletric that measures similarities between input variables. The resulting output map (U-matrix), and maps of individual input variables (component planes) may be visualized in any number of dimensions (we use 2-D maps) for communicating the dataset organization (clustering) to science experts.
Our first application involves subsurface microbiological community composition data from landfill leachate contaminated groundwater. The SOM/MANOVA methodology is able to distinguish between tiers of contamination in this multi-contaminant environment using expert knowledge to guide data preprocessing and weight the input variables. Results show a composite delineation representative of overall groundwater contamination at the landfill based only on microbiological information. Our second application combines chemical, biological and physical data from the water column and underlying lake sediments in Missisquoi Bay, Lake Champlain, Vermont, USA. This bay is plagued by cyanobacteria blooms, which often produce cyanotoxin.
The SOM methodology, in tandem with scientific experts, identifies a significant division our dataset separating samples from one location with and without a cyanobacteria bloom. Cyanobacteria cell counts were not used as input in this analysis. This research corroborates a previously suggested hypothesis that cyanobacteria growth is supported by a set of necessary, but insufficient, conditions and requires a trigger to initiate, propagate and sustain the bloom. Results from these exploratory analyses may help develop monitoring strategies in Lake Champlain, specifically guiding the spatial and temporal sampling frequency necessary to identify the environmental drivers of the blooms.