Neural networks in agriculture and natural resources: Its application to the wellhead protection area problem using GIS

Ranjan Samuel Muttiah, Purdue University

Abstract

This research has two objectives: one general and the other specific. The general objective was to find the system characteristics of agriculture and natural resources that make them amenable for study using neural networks. Neural networks are comprised of simple computational units that work together collectively to solve problems. Many agricultural and natural resources systems are characterized by decision making under conditions of uncertainty, are data intensive, and have dynamics with well defined inputs and outputs but can be difficult to describe mathematically. "Difficult" agriculture and natural resources problems in classification, time-series analysis, and heuristic optimization were solved using neural networks. The specific objective was to delineate wellhead protection areas (WHPAs) using neural networks. WHPAs are regions around water wells where human activities that might contaminate water are restricted or excluded. This research has developed a new method for the automatic demarcation of WHPAs based on numerical simulations of a non-point source model that accounts for surface factors, and saturated zone flow and transport finite element models that account for sub-surface conditions. The map area in which the simulations were performed is called the Indian Pine watershed located in Tippecanoe County, Indiana. The non-point source surface model was run to obtain nitrogen concentration in the runoff volume leaving a cell using surface elevations obtained from fractal surfaces and elevations generated from Indian Pine. The saturated zone models used aquifer data from fractal surfaces and Indian Pine to predict the draw-down and contaminant concentrations in the vicinity of the water well for various pumping and contaminant discharge rates. Land-use, erosivity index, weather conditions, transport and other coefficients were made location-specific. The results of the numerical simulations were then used to manually delineate WHPAs. The manually delineated WHPAs were learned by a cascade-correlation neural network with Gaussian hidden units. Results showed that the network accurately recalled the WHPAs used in training. When the trained neural network was tested on aquifer conditions found in the state of Vermont, the network provided significant agreement for area of the WHPAs.

Degree

Ph.D.

Advisors

Engel, Purdue University.

Subject Area

Agricultural engineering|Environmental science|Artificial intelligence|Geography

COinS