Angiosperm species classification using transfer learning in convolutional neural networks
|
Full Text |
Pdf
|
|
Author |
Jessica S. Velasco, Jannabelle J. Tiguara, Abhrei Mikael D. Torres, Marinel B. Valencia, Jovencio V. Merin and Jomer V. Catipon
|
|
e-ISSN |
1819-6608 |
|
On Pages
|
139-147
|
|
Volume No. |
20
|
|
Issue No. |
3
|
|
Issue Date |
March 21, 2025
|
|
DOI |
https://doi.org/10.59018/022526
|
|
Keywords |
angiosperm classification, deep learning, convolutional neural networks, transfer learning, python.
|
Abstract
The species composition provides the fundamental individual biological characteristics of an environment and ecosystems. The ability to identify species of individual plants or trees over an assessment plot and a forest stand is necessary for automatic mapping of plant distribution, biological diversity, stand structure, and even for diagnosing the dynamics of a forest stand. The advantage of developing plant mapping techniques utilizing image classification is the ability to identify signs of climate change based on plant phenology. Due to the large number of sequential layers and input data, deep learning takes a long time to train. Due to its capacity to do parallel processing, a powerful computer with a Graphic Processing Unit is the best choice for the training process. To classify angiosperm species, this study used photos from the University of Oxford dataset to represent 10 different angiosperm or flowering plant species. The 800 images used for the training and testing of various convolutional network models are made up of angiosperm species like bluebell, cowslip, crocus, daisy, dandelion, fritillary, iris, snowdrop, sunflower, and windflower. This study employed pre-trained weights from several convolutional neural network models, such as DenseNet 201, InceptionV3, MobileNet, ResNet152V2, ResNet50, and Xception, to classify the angiosperm species. The ResNet50 model had the lowest accuracy, 91.87%, while the Xception model had the greatest accuracy, 97.5%.
Back