摘要
卫星图像海冰检测对灾害预警和海上交通安全具有重要意义。目前在卫星遥感海冰自动检测业务化运行中,多以SAR数据为主,光学图像相对较少。光学图像海冰检测的困难在于难以区分海冰、云和高泥沙浓度海水,尤其在我国渤海。文章将深度学习语义分割模型UNet与UNet++成功应用于我国海洋卫星HY-1C/D的海岸带成像仪光学图像渤海海冰检测,克服了海冰、云、海水难以区分的困难,并在UNet与UNet++基础上首次提出输入了海冰、云和海水的概率分布的语义分割模型UNet-PM和UNet++-PM。研究表明,输入概率图的模型均比其对应的原生模型的性能有明显提高,效果最好的UNet-PM模型海冰分割的召回率、精度和F1测度分别达到0.905、0.823和0.862,有效提高了检测效果。
Detecting sea ice using satellite images is of significant importance for disaster warning and maritime safety.Currently,in the operational automatic detection of sea ice using satellite remote sensing,SAR data is predominantly used,while optical images are relatively scarce.The challenge in detecting sea ice using optical images lies in distinguishing between sea ice,clouds,and high sediment concentration seawater,especially in China’s Bohai Sea.This paper successfully applies the deep learning semantic segmentation models UNet and UNet++to the detection of sea ice in Bohai Sea using optical images from the coastal zone imager of China’s HY-1C/D ocean satellites,overcoming the difficulty of distinguishing between sea ice,clouds,and seawater.Additionally,we propose,for the first time,the semantic segmentation models UNet-PM and UNet++-PM,which incorporate the probability distributions of sea ice,clouds,and seawater into the UNet and UNet++models.Research shows that models incorporating probability maps significantly improve performance compared to their original counterparts.The best-performing UNet-PM model achieves recall,precision,and F1 scores of 0.905,0.823,and 0.862 for sea ice segmentation,respectively,effectively enhancing detection performance.
作者
李斌
曾侃
LI Bin;ZENG Kan(College of Marine Technology,Ocean University of China,Qingdao,Shandong 266000,China)
出处
《遥感信息》
北大核心
2025年第2期124-130,共7页
Remote Sensing Information
