Date of Award

Summer 8-30-2015

Level of Access

Open-Access Dissertation

Degree Name

Doctor of Philosophy (PhD)


Spatial Information Science and Engineering


Kate Beard-Tisdale

Second Committee Member

Neal Pettigrew

Third Committee Member

Phillippe Tissot


As sensors become increasingly compact and dependable in natural environments, spatially-distributed heterogeneous sensor network systems steadily become more pervasive. However, any environmental monitoring system must account for potential data loss due to a variety of natural and technological causes. Modeling a natural spatial region can be problematic due to spatial nonstationarities in environmental variables, and as particular regions may be subject to specific influences at different spatial scales. Relationships between processes within these regions are often ephemeral, so models designed to represent them cannot remain static. Integrating temporal factors into this model engenders further complexity.

This dissertation evaluates the use of multilayer perceptron neural network models in the context of sensor networks as a possible solution to many of these problems given their data-driven nature, their representational flexibility and straightforward fitting process. The relative importance of parameters is determined via an adaptive backpropagation training process, which converges to a best-fit model for sensing platforms to validate collected data or approximate missing readings. As conditions evolve over time such that the model can no longer adapt to changes, new models are trained to replace the old.

We demonstrate accuracy results for the MLP generally on par with those of spatial kriging, but able to integrate additional physical and temporal parameters, enabling its application to any region with a collection of available data streams. Potential uses of this model might be not only to approximate missing data in the sensor field, but also to flag potentially incorrect, unusual or atypical data returned by the sensor network. Given the potential for spatial heterogeneity in a monitored phenomenon, this dissertation further explores the benefits of partitioning a space and applying individual MLP models to these partitions. A system of neural models using both spatial and temporal parameters can be envisioned such that a spatiotemporal space partitioned by k-means is modeled by k neural models with internal weightings varying individually according to the dominant processes within the assigned region of each. Evaluated on simulated and real data on surface currents of theGulf ofMaine, partitioned models show significant improved results over single global models.