Application of artificial neural system algorithms to image analysis of roots in soil, I. initial results

Abstract

Several back propagation artificial neural system (ANS) architectures were tested to determine their abilities to identify roots in minirhizotron images of soil. The general model had an input level, which consisted of two linear arrays 1 × 11 pixels each, a hidden level, with 7 to 11 nodes, and an output level consisting of a single node set to produce a binary root/not-root output. The inputs to the model consisted of two linear arrays, one each from a horizontal and a vertical derivative image produced from the raw image by the Savitzky-Golay algorithm. A training image was constructed by hand-editing the raw image. The back propagation model was trained by repeatedly presenting it with a set of inputs and an associated target response. The error was calculated for the output from each input/target response pair and corrections to the weighting functions were made using a gradient descent back correction algorithm. The results of this study suggest that the ANS approach has potential to identify roots in images of soil. Suggestions for improving the performance of the model are presented.