Boosting Classifiers for Weed Seeds Identification

The identification and classification of seeds are of major technical and economical importance in the agricultural industry. To automate these activities, like in ocular inspection one should consider seed size, shape, color and texture, which can be obtained from seed images. In this work we complement previous studies on the discriminating power of these characteristics for the unique identification of seeds of 57 weed species. In particular, we discuss the possibility of improving the naïve Bayes and artificial neural network classifiers previously developed in order to avoid the use of color features as classification parameters. Morphological and textural seed characteristics can be obtained from black and white images, which are easier to process and require a cheaper hardware than color ones. To this end, we boost the classification methods by means of the AdaBoost.M1 technique, and compare the results obtained with the performance achieved when using color images. We conclude that boosting the naïve Bayes and neural classifiers does not fully compensate the discriminating power of color features. However, the improvement in classification accuracy might be enough to make the classifier still acceptable in practical applications.