Abstract

Coral reefs are one of the most sensitive marine ecosystems, highly susceptible to climate change and human activities, especially coral bleaching. This study aims to develop an image-based classification system to assess coral health using five pre-trained Convolutional Neural Network (CNN) architectures, namely MobileNetV2, EfficientNetB0, InceptionV3, ResNet50, and DenseNet121. The dataset consists of 923 underwater images, consisting of 438 healthy coral images and 485 bleached coral images. Preprocessing techniques performed include rotation, scaling, shearing, zooming, and horizontal flipping, applied to improve the robustness of the model. The experimental results revealed that the InceptionV3 model achieved the best performance, with a validation accuracy of 92.05% and the lowest loss value of 0.195, indicating high precision and stable training behavior. DenseNet121 and ResNet50 followed closely with validation accuracies of 90.87% and 89.32%, respectively, although signs of overfitting emerged during later epochs. Meanwhile, MobileNetV2 and EfficientNetB0 showed lower accuracy but offered faster computational efficiency and training time. To improve model transparency, several interpretability techniques were used including Grad-CAM, Guided Backpropagation, Integrated Gradients, and Saliency Map. Visual analysis showed that the model consistently focused on critical coral features such as texture and color patterns when distinguishing healthy corals from bleached ones. These findings underscore the potential of deep learning in supporting automated coral reef monitoring.

Concepts :

Access to Document

SDGs

Citations by Year