基于深度学习的全极化SAR影像冰川边界识别

Identification of glaciers using fully polarimetric SAR data based on deep-learning

冰川识别对于周边地区水资源与气候变化监测具有重要意义。全极化SAR影像包含地物表面散射、偶次散射、体散射、统计特性等丰富的特征,而深度学习能够充分挖掘影像信息,因此使用全极化SAR影像结合深度学习能够得到精确的冰川识别效果。本文基于喜马拉雅山脉西端ALOS2-PALSAR全极化影像,使用VGG16特征提取网络与全卷积神经网络模型U-net相结合的VGG16-unet对冰川进行识别。采用的特征包括极化相干矩阵对角线元素、Freeman-Durden、H/A/α、Pauli、VanZyl、Yamaguchi这5种极化分解参数共计19种特征。为了充分利用影像信息,对这些特征进行分析与组合,并比较它们之间的冰川识别精度,以选取最佳特征。由于冰川与非冰川的地形具有明显差异,因此将DEM、坡度、局部入射角等作为辅助特征与极化特征结合。通过对比不同极化特征分类精度得出,基于物理特性的Pauli、Freeman-Durden、VanZyl、Yamaguchi特征分类的精度较高,其中Pauli特征分类的精度最高,整体精度(OA)达到92.54%,平均用户交并比(mIoU)达到78.78%。加入地形数据后整体精度(OA)提升至94.34%,平均用户交并比(mIoU)提升至82.35%。为了进一步提高冰川的识别精度,提出了一种基于单波段特征整体精度(OA)及召回率(Recall)筛选出的SDV(表面散射、偶次散射、体散射)特征交叉组合方式,结果显示,该组合整体精度(OA)达到94.98%,用户交并比(mIoU)达到85.67%,比Pauli特征分类精度分别高出0.64%和3.32%。上述结果表明,选择最佳的特征组合方式并结合深度学习在提升冰川识别精度中具有重要的作用。

Glacier identification is important for monitoring water resources and climate change in surrounding areas.Although optical images have achieved high accuracy in glacier boundary identification,optical images are affected by cloud cover,and reproducing information under the clouds is difficult.Fully polarized SAR images contain rich features,and deep learning can fully exploit image information.Therefore,using fully polarized SAR images combined with deep learning can compensate for the lack of optical images and obtain accurate glacier recognition results.In this paper,VGG16-unet(VGG16 combined with U-net)is used to identify glaciers based on ALOS2-PALSAR fully polarized images of the western part of the Himalayas.The features include the diagonal elements of the polarization coherence matrix,Freeman-Durden,H/A/a,Pauli,VanZyl,and Yamaguchi polarization decomposition parameters totaling 19 features.To make full use of the image information,these features are analyzed and combined,and the glacier recognition accuracies are compared to select the best features.Given evident differences between glacier and nonglacier topography,elevation,slope,and local incidence angle are combined with polarization features as auxiliary features.Comparing the classification accuracy of different polarization features reveals the accuracy of Pauli,Freeman-Durden,VanZyl,and Yamaguchi features based on physical characteristics is higher,among which Pauli features have the highest accuracy with an Overall Accuracy(OA)of 92.54%and an average user intersection ratio(mloU)of 78.78%.The OA is improved to 94.34%,and the mloU is improved to 82.35%after adding the topographic data.In order to improve the recognition accuracy of glaciers further,a feature crosscombination approach is proposed,and results show the OA of the combination reaches 94.98%,and the mloU reaches 85.67%,which are 0.64%and 3.32%higher than the classification accuracy of Pauli features,respectively.Selecting the best feature combination method and combining with deep learning plays an important role in improving the accuracy of glacier recognition,and the use of neural networks combined with fully polarized SAR images can effectively compensate for the shortcomings of optical images in glacier boundary identification.