Wild rose species classification using transfer learning in convolutional neural network
|
Full Text |
Pdf
|
|
Author |
Enrico S. Lucena, Russel Kenneth P. Balingit, Alyssa Therese O. Paras, Jerahmeel Ammiel V. Mitas and Jessica S. Velasco
|
|
e-ISSN |
1819-6608 |
|
On Pages
|
33-41
|
|
Volume No. |
21
|
|
Issue No. |
1
|
|
Issue Date |
March 10, 2026
|
|
DOI |
https://doi.org/10.59018/012614
|
|
Keywords |
wild roses, plant species classification, convolutional neural networks, transfer learning, deep learning.
|
Abstract
Wild roses have gained significant importance in research, particularly in computer vision and deep learning. This study evaluates the performance of seven pre-trained convolutional neural network (CNN) models-DenseNet201, InceptionV3, MobileNet, ResNet152V2, Xception, VGG-19, and VGG-16-for classifying seven wild rose species: Burr Rose, California Rose, Dog Rose, Moyes Rose, Prickly Rose, Red-leaved Rose, and Sweet Briar. The dataset consists of 700 images, with 100 images per species, split into 80% for training and 20% for testing/validation. The models were implemented using the Keras platform and assessed based on confusion matrices, weight size, loading time, and accuracy metrics. The results indicate that DenseNet201 achieved the highest classification accuracy at 60.43%, with a weight size of 63.4 MB and a loading time of 1900.6414 seconds. In contrast, MobileNet demonstrated optimal efficiency, achieving the smallest weight size of 37.5 MB and the fastest loading time of 617.7464 seconds. Despite the challenges posed by the limited dataset and the inherent similarities between wild rose species, the findings highlight the potential of transfer learning approaches for plant species classification. This study provides a valuable foundation for future research in automated plant identification, demonstrating the feasibility of leveraging deep learning techniques to enhance species recognition. The insights gained from model comparisons, particularly in terms of accuracy and computational efficiency, can guide the selection of suitable models for similar classification tasks in botanical research and conservation efforts. Further exploration with larger and more diverse datasets may improve model performance and contribute to the development of practical applications in automated plant identification systems.
Back