<![CDATA[The popularity of aquascaping has increased significantly in recent years. However, beginners often face difficulties in identifying aquatic plant species due to their highly similar visual characteristics, which may lead to improper plant care. This study evaluates and compares the performance of three Convolutional Neural Network (CNN) architectures, namely MobileNetV3 Large, ResNet18, and EfficientNet-B0, for classifying six aquascape plant species: Anubias, Bucephalandra, Cryptocoryne Wendtii, Floaters, Hornwort, and Vallisneria Spiralis. The dataset consists of 1,998 images resized to 224 × 224 pixels and enhanced through data augmentation techniques, including rotation, horizontal flip, color jitter, and Gaussian blur, to improve model generalization. The models were trained using the PyTorch framework with transfer learning, fine-tuning based on ImageNet pretrained weights, the AdamW optimizer, class weighting, and an early stopping strategy. Experimental results show that ResNet18 achieved the highest test accuracy of 92.7%, followed by EfficientNet-B0 with 90.3% and MobileNetV3 Large with 88.7%. These findings indicate that the residual learning architecture of ResNet18 is particularly effective for aquatic plant classification on the proposed dataset, while MobileNetV3 Large remains a suitable alternative for deployment on resource-constrained devices.]]>
Copyrights © 2026
