BACKPROPAGATION NEURAL NETWORK DESIGN AND EVALUATION FOR CLASSIFYING WEED SPECIES USING COLOR IMAGE TEXTURE

Abstract

Color co-occurrence method (CCM) texture statistics were used as input variables for a backpropagation (BP) neural network weed classification model. Thirty-three unique CCM texture statistic inputs were generated for 40 images per class, within a six class data set. The following six classes were studied: giant foxtail, large crabgrass, common lambsquarter, velvetleaf, ivyleaf morningglory, and clear soil surface. The texture data was used to build six different input variable models for the BP network, consisting of various combinations of hue, saturation, and intensity (HSI) color texture statistics. The study evaluated classification accuracy as a function of network topology, and training parameter selection. In addition, training cycle requirements and training repeatability were studied. The BP topology evaluation consisted of a series of tests on symmetrical two hidden-layer network, a test of constant complexity topologies, and tapered topology networks. The best symmetrical BP network achieved a 94.7% classification accuracy for a model consisting of 11 inputs, five nodes at each of the two hidden layers and six output nodes (11 ×5 ×5 ×6 BP network). A tapered topology ( 11 ×12 ×6 ×6 BP network) out performed all other BP topologies with an overall accuracy of 96.7% and individual class accuracies of 90.0% or higher.