CORAL REEF MONITORING METHODS USING DEEP LEARNING

• Main Author: Fahimna, Bima
• Format: Final Project
• Language: Indonesia
• Subjects: Geologi, hidrologi & meteorologi
• Online Access: https://digilib.itb.ac.id/gdl/view/77624
• Institution: Institut Teknologi Bandung

Coral reefs are complex marine ecosystems with high biodiversity but are easily damaged because they are vulnerable to environmental changes. The purpose of this study is to apply a deep learning model with CNN algorithm as a method of monitoring coral reefs based on coral health charts. The classification of coral reef types is divided into three classes, namely Boulder, Branching, and Table, then the classification of coral reef health levels is divided into three classes, namely Bleached, Healthy, and Partially Bleached. In this study, three CNN architectures were used, namely DenseNet169, ResNet152, and VGG19. Based on the results of the testing model, the DenseNet169 architecture provides the best performance for coral reef type identification with an accuracy of 91.33% and identification of coral reef health levels with an accuracy of 80.30%. In this study it was also found that input data in the form of low-resolution images and images with many coral reef colonies in it will give poor prediction results. Coral reef monitoring is carried out automatically using the CNN DenseNet169 model for model implementation and manually using a coral health chart. The results of manual and automatic identification show that coral reefs at five stations around the waters of Kelapa Dua Island have branching morphology types of 46%, boulders of 35%, and tables of 19% with a health level of partially bleached of 62%, bleached of 20%. , and healthy as much as 18%, this is due to natural factors, namely salinity and temperature values that are not in accordance with optimal conditions for coral reef growth, then human activity factors where each point of the station is often passed by ships.