Performance Comparison of Three Artificial Neural Network Algorithms in Identifying Seeds of Twenty Weed Species

This study was conducted to investigate the efficiency of three artificial neural network algorithms employed to identify seeds of 20 weed species from their scanned images. A total of 15 features related to seed shape and size were extracted from the images using an image-editing program these image-extracted data were fed as inputs into three neural networks of Multilayer perceptron (MLP), RBF/GRNN/PNN Network and Generalized Feed Forward (GFF) neural network employed for seed identification purposes. RBF/GRNN/PNN network is a combined network of Radial Basic Function (RBF), General Regression Neural Network (GRNN) and Probabilistic Neural Network (PNN). After the training stage, each network was tested. The results of testing stage indicated that Generalized Feed Forward network had the highest identification accuracy (90%). This network was able to identify 8 out of twenty species by 100% accuracy. The least seed identification accuracy, using this network, was 52%. The accuracy of RBF/GRNN/PNN network was found to be 61% and this network could accurately identify only 4 species with 100% accuracy. The least precision percentage using this network was zero. The Multilayer perceptron network with 71% identification accuracy had an intermediate efficiency among the three networks. The overall results showed that GFF had the highest efficiency in identifying the studied weed seeds among the three networks.