Quality assessment of potted petunia based on a probabilistic neural network classifier

Quality of ornamental plants relies on foliage and flower appearance and evaluation of such characteristics depends on people experience; therefore, more reliable scoring methods are needed. The objective of this research was to explore the usefulness of digital image segmentation with a probabilistic neural network (PNN) classifier from which quality indicators can be obtained for quantifying time-dependent ornamental characteristics. A petunia variety (Petunia multiflora cv. F1 ‘Glistering Pearls mixed’) grown in pots in the Central High Valleys of Mexico was used as model. The variety includes a mixture of genotypes with flowers in many colors. The PNN classifier was trained and tested on six color classes (four flower classes, one foliage class and one background class). From image segmentation, six quality indicators were calculated. For the test sets, the best classification scenario whose inputs were a combination of some of the channels of RGB, CIE-Lab, CIEL*uv, HSV, and CMY color models reached an overall accuracy (OA) of 98% while it was 96% when using the RGB color model alone. The quality indicators identified differences on tends of plant quality associated to flower color. White-flowered plants had larger ground cover per plant and greater ratio of flower area to total plant area, as well as a lower roundness index. Pinkflowered plants had the lowest ground cover per plant and plant perimeter, while they had the largest roundness index. This approach allowed to identify different behaviors of the genotypes based on quality indicators, and it might be used for estimating the period this species keeps its exhibition quality.