Refined segmentation method based on U-ASPP-Net for Arctic independent sea ice
投稿时间:2021-01-28  修订日期:2021-03-19  download
摘要点击次数: 311
全文下载次数: 8
郑付强 中国科学院上海技术物理研究所 200083
匡定波 中国科学院上海技术物理研究所 
胡勇 中国科学院上海技术物理研究所 200083
巩彩兰 中国科学院上海技术物理研究所 
黄硕 中国科学院上海技术物理研究所 
李澜 中国科学院上海技术物理研究所 
何志杰 中国科学院上海技术物理研究所 
Abstract:The accuracy of independent sea ice identification in the Arctic sea ice region has important guiding significance for ensuring navigation safety, planning Arctic routes and dynamically correcting routes. The image of FY-3 satellite MERSI II has a spatial resolution of 250m. Compared with SAR image, it has the characteristics of larger size, shorter revisit period and lower cost. As an ideal data source for dynamic identification of isolated Arctic sea ice, it can provide information about the ice conditions over large areas. To some extent, the traditional sea ice identification method based on optical remote sensing data sources are unable to distinguish the thin ice areas mixed with independent sea ice, small ice debris and ice water. Using this method, the segmentation effect on edge and small pieces of independent sea ice is poor, and independent sea ice covered by thin clouds cannot be identified. For this reason, based on the characteristics of the FY-3 MERSI II images, an algorithm of independent sea ice segmentation based on U-ASPP-Net is proposed. The algorithm introduces the Atrous Spatial Pyramid Pooling module and Atrous Depthwise Separable Convolution on the basis of U –Net to develop a new independent sea ice segmentation network U-ASPP-Net. Meanwhile, FDWloss is used as the loss function at the back end of the network. Finally, the overlap elimination strategy is used to generate the final independent sea ice segmentation map. In order to verify the accuracy and effectiveness of U-ASPP-Net, U-Net, Deeplab v3+ and partition gradient difference and bimodal threshold segmentation method are selected as control methods for experiments. The experimental results show that the independent sea-ice fine segmentation method based on U-ASPP-NET is superior to other methods in the four indexes of OA, Kappa coefficient, IOU, Dice coefficient. It has a strong ability to extract details and edges, and has a high degree of reduction to tiny sea ice. In addition, this algorithm can solve the problem of thin cloud interference that cannot be solved when extracting independent sea ice based on medium-resolution remote sensing images to a certain extent. It still has a good ability to extract sea ice under thin clouds and can provide more accurate technical support for the dynamic planning of the Arctic route.
keywords:Arctic route, FY-3 satellite, independent sea ice, U-ASPP-Net, refined segmentation
  HTML  查看/发表评论  下载PDF阅读器

《Journal of Infrared And Millimeter Waves》