基于深度学习的星载SAR海洋内波自动检测研究

Automatic Detection of Internal Waves from Space-borne SAR Images Based on Deep Learning

摘要:

海洋内波是由海水密度陡变和外力扰动所引起的一种海洋内部的波动，在合成孔径雷达 （Synthetic Aperture Radar，SAR）遥感图像上通常呈现出亮暗相间的条纹特征。基于 2014–2021 年的 390 景 Sentinel-1 SAR 海洋内波遥感图像建立训练与验证数据集，结合旋转目标检测算法，使用迁移学习方法对模型进行训练，得到基于旋转框的海洋内波自动检测模型，并将检测结果与 YOLOv8 的检测结果进行对比。研究结果表明，旋转目标检测模型较 YOLOv8 能够取得更为优异的海洋内波自动检测结果，其精确率达到 93.06%， 召回率为 90.24%，在精确率较高同时还能保证较低虚警。旋转目标检测模型为海量星载 SAR 海洋内波自动快速检测提供了一种新手段，该方法在检测海洋内波的同时还可提取内波传播方向信息，为针对性开展海洋内波动力学参数反演和过程研究提供了技术基础。

Abstract:

Internal waves are a kind of seawater fluctuation caused by steep change of seawater density and external disturbance，which are usually shown as bright and dark stripes on Synthetic Aperture Radar（SAR）remote sensing images. In this paper，a training and validation dataset is constructed based on 390 Sentinel-1 SAR internal wave remote sensing images from 2014 to 2021. Combined with the algorithm of Rotation Equivariant Detector （ReDet），the transfer learning method is used to train the model，and an automatic detection model for internal waves is obtained based on the rotating box. The detection results are compared with those from YOLOv8 model. The results show that the rotating target detection model performs better than YOLOv8 in automatic detection of internal waves，which yields an accuracy rate of 93.06% with a recall rate of 90.24% and achieves a high accuracy and a low false alarm at the same time. The rotating target detection model provides an innovative technical solution for automatic and rapid detection of internal waves among massive space-borne SAR images. The method can be used to extract the propagation direction information of internal waves，which provides a solid technical basis for dynamic parameter inversion and further process research of internal waves.