Comparative Analysis of Hand-crafted Feature Descriptors and Deep Learning Architectures for Underwater Coral-reef classification

Coral reefs are considered to be the most valued ecosystems in the world due to their biodiversity. Coral-reef distribution statistics in the underwater environment impart paramount information regarding the influence of global climate change and contamination levels. Recent developments in the domain of deep learning has led to the use of architectures that help in efficient classification and provides good accuracy compared to traditional methods. In this work, the efficiency of hand-crafted feature descriptors has been compared with that of deep learning architectures like AlexNet, GoogleNet, and ResNet50. Transfer learning methods were utilized in this work rather than training the models from scratch. Experiments were performed on EILAT and RSMAS underwater coral-reef datasets to examine the efficiency of the classification models. Various performance parameters like sensitivity, specificity, precision, F1-score and accuracy are calculated to analyse the performance of the deep learning models. Experimental results suggests that the accuracy obtained with ResNet-50 architecture outperforms other models.