Semantic segmentation of seagrass habitat from drone imagery based on deep learning: A comparative study

Abstract

Listen

Translate article

In this study, the utilization of drone images and deep learning to monitor the seagrass habitat, which is important in the marine ecosystem, is evaluated. Two experiments were conducted to compare the effect of image normalization and the performance of deep learning models in semantic segmentation with drone optical images acquired for the alpine habitats in coastal waters. Z-score and Min-Max normalization techniques were used to examine the effect of image normalization, and U-Net, SegNet, PSPNet, and DeepLab v3+ were used to compare the performance of the deep learning models. As a result, Min-Max normalization demonstrated outstanding performance for optical images, and Z-score normalization for black and white images. Regardless of the normalization of the image, the performance of the models was ranked in the order of U-Net, PSPNet, SegNet, and DeepLab v3+. Although the latest model, DeepLab v3+, was expected to have excellent performance, in fact, U-Net, having a relatively simple structure and a small number of parameters, showed the best performance. As the accuracy of semantic results seems to depend on the deep learning models and normalization methods, an experiment to determine an appropriate normalization method and deep learning model should be preceded for the semantic segmentation of high-resolution optical images in coastal waters.