The current study was conducted in 2013 to identify the seeds of three species of Amaranthus, Amaranthus viridis L., Amaranthus retroflexus L. and Amaranthus albus L., by using the artificial neural network (ANN) and canonical discriminant analysis (CDA) methods. To begin with, photographs were taken of the seeds and 13 morphological characteristics of each seed extracted as predictor variables. Backward regression was used to find the most influential variables and seven variables were derived. Thus, predictor variables were divided into two sets of 13 and seven morphological characteristics. The results showed that the recognition accuracy of the ANN made using 13 and seven predictor variables was 81.1 and 80.3%, respectively. Meanwhile, recognition accuracy of the CDA using the seven and 13 predictor variables was 74.0 and 75.7%, respectively. Therefore, in comparison to CDA, ANN showed higher identification accuracy; however, the difference was not statistically significant. Identification accuracy for A. retroflexus was higher using the CDA method than ANN, while the ANN method had higher recognition accuracy for A. viridis than CDA. In addition, use of 13 predictor variables yielded a greater identification accuracy than seven. The results of the current study showed that using seed morphological characteristics extracted by computer vision could be effective for reliable identification of the similar seeds of Amaranthus species.

Giant ragweed exhibits a high degree of polymorphism among individual plants in seed size, shape, spininess, and color. These features may play an important role in giant ragweed seed survival and predation avoidance; however, they are difficult to evaluate because of lack of quantification methods. A computer imaging technique was developed for describing and classifying giant ragweed seeds using digital images of the seed top and side views. Seed samples collected from 20 different giant ragweed plants (classes) were mounted and digitally scanned. Quantitative features were extracted from the seed images, including color, width, height, area, and seed perimeter. A polygon (convex hull) of the seed image based on the seed outline was constructed, from which spininess indices were developed. Fisher's linear discriminant with normalized nearest neighbor classification was used to classify randomly selected images of individual seeds according to class (maternal origin), using the extracted features as a database. The best classification rate achieved was 99%, with 138 out of 140 seeds correctly matched using data from both the top and side views. Seed features were easily extracted and varied from 1.2- to 4.5-fold among classes. Area and perimeter measurements varied least within classes but varied most among classes, suggesting that these features discriminate effectively among seeds from different plants in giant ragweed. Convex hull area : seed area ratio, using the seed top view images, was the best index of seed spininess, aligning well with visual assessment and providing greatest discrimination among classes. This experiment shows that in the case of giant ragweed, seeds from different plants are distinguishable in an objective and quantitative manner. This imaging technique can be applied to identification of seeds from different species and to studies on variable seed morphology within a species.