Research on Fast Acquisition and Effective Recovery Method of Glacier Velocity Field from SAR

申玉涵

张奎

硕士

重庆大学

SAR偏移量追踪;;冰川流速场;;克里金插值;;深度神经网络;;CUDA并行计算

冰川运动是全球气候变化的重要组成部分之一,对冰川运动进行监测能够为研究气候变化、海平面变化等提供直观的数据。冰川流速场作为冰川运动中最重要的监测指标,是研究冰川物质分布和冰川运动的基础。利用合成孔径雷达干涉测量技术能够对冰川运动进行全天时、全天候、大面积监测,同时借助偏移量追踪技术便可以获取高精度的冰川流速场。但是,冰川区域监测范围较大,此时应用偏移量追踪技术会产生计算量过高的问题,不能完成对冰川流速场的快速获取,降低了偏移量追踪技术的适用性。因此为了解决此问题,需要研究偏移量追踪技术的高性能实现。同时,由于在冰川区域得到的合成孔径雷达图像之间不能保持良好的相干性,导致在某些失相干严重的区域不能获取准确的冰川流速,进而无法得到全面、连续的冰川流速场。因此为了得到完整的冰川流速场,需要研究对流速缺失部分的有效恢复方法。鉴于以上问题,本文对冰川流速场的快速获取和有效恢复进行了相应的研究,具体工作和贡献主要有以下三点:1)基于CUDA(Compute Unified Device Architecture)完成了偏移量追踪技术的高性能实现。首先,阐述了合成孔径雷达及其干涉测量技术的基本原理,并说明了利用合成孔径雷达干涉测量技术获取冰川流速场的局限性,从而引出偏移量追踪技术。然后,详细分析了基于互相关法的偏移量追踪技术,针对其应用在冰川流速场获取场景中出现的计算量过大的问题,提出了一种高性能实现方案。最后,利用CPU(Central Processing Unit)的内存管理策略以及GPU(Graphics Processing Unit)的高性能计算,借助CUDA对偏移量追踪技术进行了高效实现,完成了对冰川流速场的快速获取,也提高了偏移量追踪技术在冰川流速场监测中的适用性;2)建立了基于克里金法的冰川流速场恢复模型。首先,介绍了普通克里金法和协同克里金法的建模过程和验证方法。然后,分别选取合适的区域化变量,建立相应的半方差函数模型,完成了两种克里金插值方法对冰川流速场的恢复。最后,对两种方法的恢复结果进行了定性、定量的比较,并对克里金法恢复冰川流速场的效果进行了分析;3)建立了基于深度神经网络的冰川流速场恢复模型。首先,根据深度神经网络的特点,建立了一种深度神经网络恢复冰川流速场模型,确定了相应的超参数以及训练方法,对冰川流速场失相干严重的部分进行了恢复。然后,利用冰川地区的梯度信息作为辅助,提高了恢复结果的全面性、连续性。最后,通过将携梯度信息的神经网络与两种克里金插值及优化前深度神经网络的结果进行对比分析,验证了深度神经网络对冰川流速场恢复的有效性,也证明了加入梯度信息作为辅助变量能够提高对冰川流速场恢复的效果。

Glacier movement is one of the important components of global climate change.It can provide intuitive research data for climate change and sea level change by monitoring glacier movement.The glacier velocity field,as the most important monitoring indicator in glacier movement,is the basis for studying glacier material distribution and glacier movement.Synthetic Aperture Radar Interferometry enables full-day,all-weather,and large-area monitoring of glacier motion,and high-precision glacier velocity fields can be obtained with offset tracking technology.However,due to the excessive monitoring range of the glacier area,the calculation of the offset tracking technology is too complicated.Therefore,the high-performance implementation of the offset tracking technology must be implemented to quickly obtain the glacier velocity field.At the same time,due to the decoherence of SAR images obtained from glacial areas,it is impossible to restore a comprehensive and continuous glacier velocity field in some areas.Therefore,it is necessary to explore an effective recovery method for the missing portion of the flow rate to obtain a complete glacier velocity field.In view of the above problems,this thesis studies the rapid acquisition and effective recovery of glacier velocity field.The main work and contributions of this thesis mainly include the following three points:1)This thesis completed the high-performance implementation of the offset tracking technology based on CUDA(Compute Unified Device Architecture).First of all,the basic principles of synthetic aperture radar and its interferometry technology are expounded,and the limitation of obtaining the velocity field of glacier by synthetic aperture radar interferometry is illustrated,and the offset tracking technology is introduced.Then,the offset tracking technology based on cross-correlation method is analyzed in detail,and a high-performance implementation scheme is proposed for the problem of excessive computational complexity in the scene of glacier velocity field acquisition.Finally,this thesis uses the memory management strategy of the CPU(Central Processing Unit)and the high-performance computing of the GPU(Graphics Processing Unit),the offset tracking technology is efficiently implemented by CUDA,and the rapid acquisition of the glacier velocity field is completed.It also improves the applicability of offset tracking technology in glacier velocity field monitoring.2)This thesis established a model of glacier velocity field recovery based on Kriging method.Above all,the principle of ordinary Kriging interpolation and cooperative Kriging interpolation are introduced.Then,this thesis selected the appropriate regionalization variables,established the corresponding semi-variance function model,and completed the restoration of the glacier velocity field by two Kriging interpolation methods.Finally,qualitative and quantitative comparisons were made between the recovery results of the two methods,and the effect of the Kriging method on recovering the velocity field of the glacier was analyzed.3)This thesis established a model of glacier velocity field recovery based on deep neural network.Firstly,according to the characteristics of deep neural network,a deep neural network is used to recover the glacier velocity field model,the corresponding hyperparameters and training methods were determined,and the part of the glacier velocity field with serious coherence was restored.Then,this thesis used the gradient information of the glacial area as an assist,the comprehensiveness and continuity of the recovery results are improved.Finally,by comparing with the results of two Kriging interpolation and pre-optimal depth neural networks,the role of deep neural network in the recovery of glacier velocity field is verified.