Semantic segmentation and segmentation refinement using machine learning case study: Water turbidity segmentation

Classical methods for image segmentation such as pixel thresholding, clustering, region growing, maximum likelihood have been used regularly and relied on for a long time. However, these classical methods have limitations, particularly on images where there are many overlapping pixel values between features, which is common in remote sensing images. The advent of machine learning, in particular, deep learning in computer vision and image analysis, has gained interest in the remote sensing field. Current deep learning architecture has been able to achieve high accuracy for image recognition, object detection, and segmentation. This study performed image segmentation on the coastal area with high water turbidity using Landsat-8 images. Currently, the standard tool to derive water turbidity data from Landsat-8 images is the level-2 plugin of SEADAS software. However, due to its rigorous processing method, the processing time using SEADAS Level-2 Plugin is quite long; for example, processing one Landsat-8 image took around 8 hours. As a consequence, the amount of time needed to process multiple images is increasing. Deep learning has advantages once the model trained, the inference or prediction process is quite fast. Therefore it has the potential to be used as a complementary tool to predict and segment high turbidity areas, because in deep learning. In this study, we implemented U-Net architecture with ResNet connection and used Generative-Adversarial Network (GAN) to refined segmentation results.