基于改进DeepLabV3+模型的海冰提取方法——以北极格陵兰海为例  

An extraction method for sea ice based on improved DeepLabV3+model:Taking the Arctic Greenland Sea as an example

作　　者：孙士昌 王志勇[1,2] 李振今 张保敬 田康 赵相禹 Sun Shichang;Wang Zhiyong;Li Zhenjin;Zhang Baojing;Tian Kang;Zhao Xiangyu(College of Geodesy and Geomatics,Shandong University of Science and Technology,Qingdao 266590,China;National Demonstration Center for Experimental Surveying and Mapping Education(Shandong University of Science and Technology),Qingdao 266590,China)

机构地区：[1]山东科技大学测绘与空间信息学院,山东青岛266590 [2]测绘工程国家级实验教学示范中心(山东科技大学),山东青岛266590

出　　处：《海洋学报》2024年第8期131-142,共12页

基　　金：国家自然科学基金(41876202)。

摘　　要：海冰是全球气候变化的指示剂,北极海冰的变化关系到全球变暖、海平面上升等。针对传统语义分割模型对海冰进行提取时存在细节提取不精确、提取速度慢等问题,构建了一种改进DeepLabV3+的海冰提取方法。首先,将主干网络Xception替换为MobileNetV2,在保证海冰提取精度的同时大幅度降低模型参数量,节约时间;其次,将ASPP改进为DenseASPP,在进行海冰的多尺度特征提取时进一步扩大感受野,获得更为密集的特征;最后,引入坐标注意力机制,同时强化关注通道和空间上的特征,加强海冰边缘细节信息提取。选取北极格陵兰海为实验区,通过对该海域2020–2022年间冬季的10景Sentinel-1A双极化SAR影像进行处理、标注之后形成数据集进行实验,对比U-Net、PSPNet和DeepLabV3+等经典模型。结果表明:本文方法的m IoU达到了88.46%,mPA达到了94.16%。相较于传统DeepLabV3+,mIoU提高了2.35%,mPA提高了2.90%,参数量和GFLOPs分别减少了45.08 M和106.01 G,同时训练模型时间和提取海冰时间分别减少了68%和30%。对比U-Net、PSPNet等模型,同样取得了最优结果。与其他模型相比,本文新构建的模型对海冰特征的学习能力更强,能获取更多海冰细节信息,并大幅度节约用时,能够为研究全球变暖环境下的海冰退化监测问题提供技术支持。Sea ice is an indicator of global climate change,and the change of Arctic sea ice is related to global warming and sea level rise.Aiming at the problems such as inaccuracy and slow speed of extracting details from sea ice by traditional semantic segmentation model,an improved DeepLabV3+sea ice extraction method was constructed.Firstly,we replaced the Xception backbone network with MobileNetV2,which significantly reduces the network’s parameter count and save time while maintaining the accuracy of sea ice extraction.Secondly,we enhanced the ASPP module to DenseASPP,further expanding the receptive field during multi-scale feature extraction for sea ice,resulting in denser features.Lastly,we introduced a coordinate attention mechanism to strengthen the focus on both channel and spatial features,enhancing the extraction of fine edge details in sea ice.The Greenland Sea in the Arctic is selected as the experimental area,and 10 Sentinel-1A dual-polarization SAR images from the winter of 2020 to 2022 in the sea area are processed and labeled to form a data set for the experiment,we compared our method with classic models such as U-Net,PSPNet and DeepLabV3+.The results showed that our method achieved anmIoU of 88.46%and an mPA of 94.16%.Compared to the traditional DeepLabV3+,mIoU increased by 2.35%,mPA increased by 2.90%,and the parameter count and GFLOPs decreased 45.08 M and 106.01 G,respectively.Meanwhile,the training time and sea ice extraction time decreased by 68%and 30%,respectively.Compared to U-Net、PSPNet and other models,the optimal results are also obtained.Compared with other models,the new model constructed in this paper has a stronger learning ability about sea ice characteristics,can obtain more detailed information of sea ice and greatly saves time,and can provide technical support for the study of sea ice degradation monitoring under global warming environment.

关 键 词：海冰提取 深度学习 MobileNetV2 DenseASPP 坐标注意力 

分 类 号：P731.15[天文地球—海洋科学]