Classifying Invasive Alien Species in the Philippines Using Convolutional Neural Networks

• Main Authors: Aliño, Rey, Fernandez, Proceso L, Jr, Diesmos, Arvin
• Format: text
• Published: Archīum Ateneo 2023
• Subjects: Alien species Philippines; Biodiversity; Convolutional neural networks; Deep learning; Invasive; Life Sciences; Physical Sciences and Mathematics; Physics
• Online Access: https://archium.ateneo.edu/physics-faculty-pubs/162; https://jestec.taylors.edu.my/Special%20Issue%20ICITE2022/ICITE_2022_07.pdf
• Institution: Ateneo De Manila University

The proliferation of Invasive Alien Species (IAS) in the Philippines is a major threat to its biodiversity. Towards reducing such threat, deep learning technology can be applied to collect taxonomic information which may then assist in strategies and plans to fight IAS. This study presents implementations of Resnet18, MobileNetV2 and GoogLeNet, three known convolutional neural network (CNN) models, previously used for other deep learning tasks, for classifying twenty-four (24) IAS in the Philippines (PH). In this interdisciplinary study, a dataset of 2,581 images of 24 invasive species was first collected. The initial images were obtained from the ASEAN Centre for Biodiversity (ACB) and supplemented by images from the International Union for Conservation of Nature (IUCN) database, the Global Biodiversity Information Facility (GBIF) and Google Images. The images were pre-processed and then used to train the three CNN models to classify the 24 invasive species. We used five-fold cross validation to evaluate the performance of our models. Precision, recall, f1-score and overall accuracy metrics were recorded and showed that the three models can accurately classify the twenty-four IAS PH in our dataset. The top performing model, ResNet18, achieved a 90.8% average accuracy while MobileNetV2 and GoogLeNet achieved average accuracies of 87.4% and 87%, respectively. While ResNet18 had higher average accuracy than the other two models, a one-way analysis of variance test of the accuracies of the three models across the five-fold training and validation, however, showed no statistically significant difference.