Poster Session 2, Tuesday, October 4, 10:40–12:40
GOCI Super-resolution via GOCI-II and U-Net deep learning model
Super-resolution is the process of creating high-resolution images from low-resolution images. It is an important class of image processing techniques in image processing and computer vision. In recent years, deep learning techniques have been applied to tackle super-resolution tasks, ranging from convolutional neural network based method to recent promising approaches using generative adversarial net. This study aims to improve spatial resolution of the Geostationary Ocean Color Imager (GOCI) with a spatial resolution of 500 m operated from June 2010 to March 2021 using GOCI-II data with a spatial resolution of 250 m and U-Net deep learning model. For this, we used the hourly GOCI-II remote-sensing reflectance (Rrs) product from December 2020 to March 2021, the period when the images were acquired at the same time as GOCI. The study area includes the coast of the Korean Peninsula, which has a complex coastal system. The U-Net model was trained and tested using a total of 679 images acquired over 103 days. To determine the optimal patch size, we repeatedly performed training by generating various patch pairs of GOCI and GOCI-II from 32 x 32 to 256 x 256. As a result, we found that the U-Net model is well trained from patch sizes below 64 x 64, and that the spatial distribution of marine environments occurring nearshore is more detailed in high-resolution super-resolved GOCI image than in the original GOCI image. This result will be useful because it may be applied to super-resolution for another satellite images.
Soo Mee Kim, Korea Institute of Ocean Science and Technology; Korea Maritime and Ocean University, Korea, [email protected]
Young-Heon Jo, Pusan National University, Korea, [email protected]