Sentinel-1 SAR 영상과 인공지능 기법을 이용한 연안해역의 고해상도 해상풍 산출

Abstract

A study was conducted to estimate high-resolution sea wind in coastal waters using SAR satellite images and deep neural network(DNN) models. The 368 Sentinel-1 SAR images were collected that cover the coastal waters of Busan, Ulsan, and Gyeongnam in Korea from 2015 to 2020. The VV and VH polarization bands and incident angle data were obtained from the SAR images. Predicted wind data by KMA LDAPS model, observed wind data from sea buoys and water depth in the study area were matched-up with the SAR images to establish a training dataset for DNN. A total of four DNN models were constructed, those were the combinations of spatial resolutions(10m verse 100m) and wind directions(eastward wind verse northward wind). All of them went through training processes including parameter optimization. The final results of the DNN models showed correlation coefficient (CC) of 0.901 and MAE 1.467 for eastward wind, CC 0.826 and MAE 1.586 for northward wind at the spatial resolution of 10 m, CC 0.909 and MAE 1.378 for eastward wind, CC 0.827 and MAE 1.559 for northward wind at the spatial resolution of 100m. Also the same training dataset was applied to calculate sea wind by geophysical model function(GMF), for the purpose of comparing DNN method to a conventional GMF method, CMOD5.N, which are most used for wind speed retrieval with SAR image. The wind speed by CMOD5.N showedCC 0.729 and MAE 2.277. As a result, in this study, the DNN models not only show better accuracy than the popular GMF method in estimating sea wind, but also have the advantage of being able to calculate wind direction compared to the CMOD method, which can only calculate wind speed. The sea wind map of study area was implemented using QGIS with the high resolution wind data estimated by DNN models. This high resolution sea wind map could be utilized to check the business profitability of offshore wind farms. Also there found atendency that sea wind becomes stronger as it goes out to open sea with fewer islands from coastal area with many islands close to the inland. The difference in accuracy between the eastward wind DNN model and the northward wind DNN model was considered due to the low correlation between north wind and wave in southern sea area, especially in winter season when north wind dominates. For better accuracy, it is recommended to classify coastal regions and seasons along common marine meteorological characteristics, and to construct a deep neural network model for each classified case.