Development and performance evaluation of a CNN model for seagrass species classification in Bintan, Indonesia

Abstract

This study presents the development and evaluation of a convolutional neural network (CNN) model for automated seagrass species classification in Bintan, Indonesia. The objective of this research is to examine how different train-validation data split ratios affect model accuracy and generalization performance. The CNN was trained under four configurations (60:40, 70:30, 80:20, and 90:10) to analyze the influence of training data volume on learning convergence and predictive capability. The results indicate that all configurations achieved high validation accuracy, with the best performance reaching 98.53% when using the 90:10 split. Evaluation on unseen data demonstrated that the 60:40 configuration provided the most consistent and reliable generalization. Performance variations were also affected by the morphological similarity between the classified species, which increases the challenge in correctly distinguishing certain classes. Overall, the findings confirm the effectiveness of CNN-based classification for supporting marine biodiversity monitoring and underline the importance of dataset composition in achieving optimal performance. Future improvements will focus on expanding data variability to enhance robustness in real-world scenarios.