Computer vision method to determine the quality of Chrysanthemum cuttings through their leaf morphology

ABSTRACT : In this work, a new computer vision method is proposed to classify Chrysanthemum cuttings through their leaf morphological traits, which can optimize the selection of ideal cuttings for rooting in the plant propagation process. The implementation of the method was divided into two stages: image processing, using traditional techniques; and machine learning, through unsupervised and supervised models. In this way, to carry out the first stage, a set of 650 cuttings leaf images of the cultivar Chrysanthemum × morifolium 'Baltica' was formed, which allowed the building of a set of feature vectors corresponding to the samples. Those vectors were made up of the area, perimeter, length, and diameter features of the leaf morphological parts of veins, petiole, leaf blade, and lobes. So, leaf measured traits such as vein area, vein perimeter, vein length, vein diameter, petiole area, petiole perimeter, petiole length, petiole diameter, leaf blade area, leaf blade perimeter, leaf blade length, leaf blade diameter, upper lobe length, left upper lobe length, left lower lobe length, right upper lobe length, and right lower lobe length were defined. Then, for the second stage, the samples were labeled according to the cluster obtained when a K-means unsupervised learning model was fitted, forming the dataset. Next, different supervised learning models such as Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF) y K-Nearest Neighbors (KNN) were trained, applying the Cross-Validation (CV) method to find the respective optimal hyper-parameters that had been previously defined. Likewise, different scenarios of feature combinations were defined to evaluate the performance of the models in the training with CV. The scenarios were defined using the features of the leaf morphological traits together and separately, to find those that could determine with the highest accuracy the state of leaf development. Thus, when training with all the features, the classifiers with the best performance were found, obtaining accuracy percentages of 92% for LR, 91% for SVM and RF, and 87% for KNN. On the other hand, with the features of the morphological traits separately, the classifiers with the higher accuracy percentages were those in which lobe features were used, achieving accuracies of 80% for SVM, 78% for RF, 77% for LR, and 74% for KNN. Finally, taking into account all scenarios and model types, the best classifier was the SVM, because it obtained the best results in the scenarios with the separate morphological traits, in addition to having fewer hyper-parameters to fit compared to RF, which obtained the second best results.

• Autores: Pavas Henao, Juan Guillermo
• Fecha de publicación: 2025
• Institución: Universidad de Antioquia
• Repositorio: Repositorio UdeA
• Idioma: eng
• OAI Identifier: oai:bibliotecadigital.udea.edu.co:10495/45686
• Acceso en línea: https://hdl.handle.net/10495/45686
• Palabra clave: Computer vision; Visión por ordenador; Image processing; Proceso de imágenes; Machine learning; Aprendizaje automático; Plant propagation; Plantas - Multiplicación; Chrysanthemum; http://aims.fao.org/aos/agrovoc/c_1599; http://id.loc.gov/authorities/subjects/sh85029549; http://id.loc.gov/authorities/subjects/sh85064446; http://id.loc.gov/authorities/subjects/sh85079324; http://id.loc.gov/authorities/subjects/sh85102802; Computer vision
• Rights: embargoedAccess
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/