Three neural network case studies in biology and natural resource management

This paper presents 3 NN case studies. In the first, fracture toughness of wood was predicted using an expanded MLP network from experimentally measured crack angle, stiffness, density and moisture content. The data is characterized by noise but the model produced physically meaningful nonlinear trends with an R2 value of 0.67. In the second study, hydraulic conductivity (K m/day) was estimated from ground water solute concentration data collected for a range of K values. Four separate NN needed to be developed for four sub-ranges of K to reduce error. In order to determine the appropriate range of K for a particular system concentration data were clustered into 4 groups using SOM. The hybrid-model was applied to an experimental aquifer and only 10% difference was found between experimental and NN estimations of K. In the third study, digital images of lamb chops were used to collect values for 102 geometric and textural variables for meat grading. Principal component analysis reduced the variables to twelve. Three- and four-layer MLP networks and discriminant function analysis (DFA) were performed on the data and the classification accuracy from 3-layer MLP was 83% and was 12% better than that from DFA.