结合注意力机制的卫星遥感影像立体匹配

采用立体匹配技术对多视卫星遥感影像进行三维场景重建一直是摄影测量与遥感领域的核心问题。基于卷积神经网络的深度学习方法极大地促进了立体匹配技术的发展,然而其中涉及匹配困难和误匹配问题的相关研究仍然不足。为了提升卫星遥感影像不适定区域中视差估计的精度,本研究提出了一种结合注意力机制的立体匹配深度学习网络,在特征提取模块中加入注意力机制,分别从通道和空间两个维度捕获全局信息,对特征进行优化;在代价体的构建模块中构建新的代价体积,并重新设置视差的回归范围。为了验证本文方法的有效性,在US3D、WHU-Stereo两个数据集上分别与已有方法 Stereo-Net、PSM-Net进行了比较分析。结果表明,本文方法在EPE(endpointerror)和D₁两个指标上均能达到最优,取得了较好的性能,提高了立体匹配的精度,尤其在无纹理、重复纹理、遮挡及视差不连续区域表现出良好的鲁棒性。     

基于Anchor-free的交通标志检测

交通标志检测是自动驾驶中的重要研究方向,实时准确地从街景图像中检测交通标志对实现自动驾驶及智慧城市的发展具有重要意义。传统的算法基于颜色、形状特征进行检测,只能提取特定种类的交通标志,算法无法同时检测不同类型的交通标志。基于图像特征+机器学习分类器的算法需要人工设计特征,算法速度较慢。主流的基于深度学习的方法多基于先验框,在网络设计上引入了额外的超参数,且在训练过程中产生过量的冗余边界框,容易造成正负样本不平衡。本文受Anchor-free思想的启发,引用YOLO检测器直接回归物体边界框的思路,提出一种基于Anchor-free的实时交通标志检测网络AF-TSD（Anchor-free Traffic Sign Detection）。AF-TSD摒弃了先验框的设计,并引入自适应采样位置可变卷积与注意力机制,大大提高网络的特征表达能力。本文开展大量对比实验,实验结果表明本文提出的AF-TSD交通标志检测网络速度接近主流算法,但精度优于主流算法,在德国GTSDB交通标志检测数据集上取得了96.80%的精度,检测速度平均单张图片32 ms,达到实时检测的要求。     

基于地震信号反演滑坡动力学机制

以2003年千将坪滑坡事件为例,基于地震信号分析大型高速滑坡启动之后的滑体运动特性。通过国家地震台网采集因滑坡激发产生的地震信号,反演得到滑坡区域的受力-时间函数,再经由时频分析划分滑坡期间的子事件,进而推导滑体的运动参数并还原滑坡过程。结果显示,由地震信号反演所求得的滑床坡度、滑坡方向以及滑体位移等数值与现场勘踏所得数据相符;同时,通过对滑坡子事件的分析,可分辨出因对岸河堤阻挡而产生的部分滑体反倾过程,从而还原较完整的滑坡过程。     

基于特征注意力机制的RNN-Bi-LSTM船舶轨迹预测

【目的】为更准确预测船舶轨迹,基于RNN、Bi-LSTM和注意力机制,研究一种结合特征注意力机制的RNN-Bi-LSTM的船舶轨迹预测模型。【方法】基于AIS数据构建基于循环神经网络(RNN)与双向长短时记忆网络(Bi-LSTM)的混合神经网络模型,并在混合模型中加入特征注意力机制对数据特征进行权重分配,提升模型对船舶轨迹预测精度。【结果】使用实际运行的船舶AIS数据,对模型的有效性和实用性进行验证,测试集均方误差为2.751×10^-5、均方根误差为5.245×10^-3,在连续弯道预测中的均方误差为4.359×10^-6、均方根误差为2.088×10^-3。【结论】结合特征注意力机制的RNN-Bi-LSTM相较于传统的预测神经网络,船舶轨迹预测精度更高,尤其在弯道预测中也表现出较好的符合度。     

基于深度学习的滑坡灾害易发性分析

滑坡灾害成因机理复杂、影响因素众多,深度学习作为当前人工智能领域的热点,能够更好地模拟滑坡灾害的形成并准确预测潜在的斜坡。为了挖掘深度学习在滑坡易发性的应用潜能,本文构建了一维、二维和三维的滑坡数据表达形式,并提出3种基于卷积神经网络模型（Convolutional Neural Networks, CNN）的滑坡易发性分析处理框架：基于CNN分类器、基于CNN与逻辑回归的融合和基于CNN集成,最后以江西省铅山县为研究对象进行验证,结果表明：所有基于CNN的易发性模型都能够获得准确且可靠的滑坡易发性分析结果。其中,基于二维数据的CNN模型在所有单分类器中预测精度最高,为78.95%。此外,二维CNN特征提取能够显著提升逻辑回归的预测精度,其准确率提升7.9%。最后,异质集成策略能够大幅度提升基于CNN分类器的滑坡预测精度,其准确率提升4.35%~8.78%。     

基于双注意力残差网络的高分遥感影像道路提取模型

高分辨率遥感影像中,道路光谱信息丰富,且空间几何结构更清晰。但是,基于高分遥感影像的道路提取面临道路尺寸变化大、容易受树木、建筑物及阴影遮挡等因素影响,导致提取结果不完整。此外,高分遥感影像中同物异谱和异物同谱现象较为严重,从而影响道路提取结果连续性及细小道路信息完整性,而且难以区分道路和非道路不透水层。因此,本文提出基于双注意力残差网络的道路提取模型DARNet,利用深度编码网络,获取细粒度高阶语义信息,增强网络对细小道路的提取能力,通过嵌入串联式通道-空间双重注意力模块,获取道路特征图逐通道的全局语义信息,实现道路特征的高效表达及多尺度道路信息的深层融合,增强阴影和遮挡环境下网络模型的鲁棒性,改善道路提取细节缺失现象,实现复杂环境下高效、准确的道路自动化提取。本文在3个实验数据集对DARNet和DLinkNet、DeepLabV3+等5个对比模型进行对比试验和定量评估,结果表明,本文DARNet模型的F1分别为77.92%、67.88%和80.37%,高于对比模型。此外,定性比较表明,本文提出模型可以有效克服由于物体阴影、遮挡和高分影像光谱变化导致道路提取不准确与不完整问题,改善细...     

基于多源数据的谭家湾滑坡变形机制及稳定性评价

针对2016年5月发生于秭归县西北部的谭家湾滑坡,结合卫星遥感影像、现场勘查资料以及历史资料等多源数据,初步明确了滑坡的影响区域、特征及发生时序;综合采用钻探、槽探、物探等手段,开展室内外相关实验,明确了滑坡区的地层特性以及岩土体物理力学性质指标,通过分析该区裂缝位移及GPS数据,对该边坡的变形机制进行了探讨,并对该区稳定性进行了评价。结果表明:①谭家湾滑坡属于不规则"圈椅形"中型松散层的水库下降型滑坡,滑坡区的地表形态、地质构造及岩性等因素决定了滑坡的形成和发育,库水位和降雨的共同作用激励了滑坡的变形;②滑坡根据时序共分为3级滑体,总体呈现多次、多层、相互影响的演化特点,第三级滑体具有牵引式特征;③滑坡体内地下水位随库水位下降而下降,但下降速率缓于库水位,随之坡体内水力梯度和渗透力显著变大,此时碰到强降雨,将会导致坡体地下水赋存,岩土体软化,加剧滑坡变形,须施加必要的防护措施。④稳定性分析表明,该滑坡现处于临界稳定状态,一旦发生降雨和库水位变化,局部段可能发生失稳滑动。      

利用基于残差多注意力和ACON激活函数的神经网络提取建筑物

针对目前主流深度学习网络模型应用于高空间分辩率遥感影像建筑物提取存在的内部空洞、不连续以及边缘缺失与边界不规则等问题,本文在U-Net模型结构的基础上通过设计新的激活函数（ACON）、集成残差以及通道-空间与十字注意力模块,提出RMAU-Net模型。该模型中的ACON激活函数允许每个神经元自适应地激活或不激活,有利于提高模型的泛化能力和传输性能;残差模块用于拓宽网络深度并降低训练和学习的难度,获取深层次语义特征信息;通道-空间注意力模块用于增强编码段与解码段信息的关联、抑制无关背景区域的影响,提高模型的灵敏度;十字注意力模块聚合交叉路径上所有像素的上下文信息,通过循环操作捕获全局上下文信息,提高像素间的全局相关性。以Massachusetts数据集为样本的建筑物提取实验表明,在所有参与比对的7个模型中,本文提出的RMAU-Net模型交并比与F1分数2项指标最优、查准率和查全率两项指标接近最优, RMA-UNet总体效果优于同类模型。通过逐步添加每个模块来进一步验证各模块的有效性以及本文所提方法的可靠性。     

基于生成对抗网络的动水驱动型滑坡状态识别方法

动水驱动型滑坡状态识别能更有效地辅助分析滑坡形变规律, 实现滑坡状态的准确识别对深入展开动水驱动型滑坡状态研究具有重要意义。针对目前动水驱动型滑坡突变状态研究较少, 难以获得相关特征, 从而导致状态识别性能不佳的问题, 提出了一种应用于动水驱动型滑坡状态识别的生成对抗网络学习方法。本方法通过构建滑坡状态监测数据矩阵, 依据少量数据样本设计合理的生成器网络完成对滑坡状态的数据扩增并设计判别器网络实现扩增数据的筛选, 通过对抗生成网络实现对滑坡状态的分类, 达到滑坡状态识别的目的。以三峡库区白水河滑坡为研究对象, 将降雨、库水位、深部位移和地表位移等多源监测数据进行了规范化处理, 设计生成器网络和对抗器网络完成了对滑坡状态数据的扩增, 设计滑坡状态识别生成对抗网络完成了对滑坡状态的分类和识别。结果表明, 生成对抗网络对滑坡状态识别具有较高的准确率。研究结果证实本方法能够对目标区域内的动水驱动型滑坡状态进行准确识别和分类, 可直接应用于工程实际。      

基于地形特征融合的卷积神经网络滑坡识别

滑坡严重威胁着人民群众的生命财产安全。完整、准确的滑坡编录图是研究滑坡的重要资料。深度卷积神经网络方法由于众多优势而备受关注,然而卷积神经网络结构复杂,需要大量的训练样本,制约了其在滑坡制图上的发展。提出了融合地形特征的卷积神经网络建模方法。首先在遥感影像上叠加地形因子构建新的滑坡样本,然后设计提取并融合空间与光谱特征的轻量级卷积神经网络(FF-CNN),最后训练最优模型进行滑坡识别。在四川汶川地区进行的消融实验证明：在空间特征基础上融合光谱特征的FF-CNN模型滑坡识别评价指标F1分数和平均交并比(MIoU)分别提高0.020 2和0.014 4;在遥感影像上叠加地形因子后,FF-CNN模型滑坡识别评价指标F1分数和MIoU值分别提高0.066 4和0.048 2。在湖北省三峡库区和四川省都江堰市虹口乡的实验说明FF-CNN模型表现出较强的适用性和迁移能力,在滑坡识别上具有较大潜力。     

Experimental investigation on the failure mechanism of a rock landslide controlled by a steep-gentle discontinuity pair

A type of rock landslide is very common in practical engineering, whose stability is mainly controlled by the rock bridge between the steep tensile crack at the crest and the low-inclination weak discontinuities at the toe(namely, ligament is the term for the locking section in the slope). To obtain a deeper understanding into the failure process of this kind of landslide, twenty-four physical slope models containing a steep-gentle discontinuity pair(a steep crack in the upper part and a low-inclination discontinuity in the lower part) were tested by applying vertical loads at the crests. The results indicate that the inclination angle of the ligament(θ) has great influence on the failure and stability of this type of rock slope. With the change of θ, three failure patterns(five subtypes) concerning the tested slopes can be observed, i.e., tensile failure of the ligament(Type 1), tension-shear failure of the ligament(Type 2) and two-stage failure of the main body(Type 3). The failure process of each failure mode presents five stages in terms of crack development, vertical load, horizontal/vertical displacements and strains in the ligaments. The specific range of the ligament angle between different failure patterns is summarized. The discussion on the failure resistances and ductility of different failure patterns, and the guiding significances of the experimental findings to the stability evaluation and the reinforcement were conducted.     

An Improved Damage Identification Method Based on Mode Information

This paper presents a new algorithm to predict locations and severities of damage in structures by changing modal parameters. An existing algorithm of damage detection is reviewed and the new algorithm is formulated to improve the accuracy of damage locating and severity estimation by eliminating the ＂erratic assumptions and limits in the existing algorithm. The damage prediction accuracy is numerically assessed for each algorithm when applied to a two-dimensional frame structure for which pre-damage and post-damage modal parameters are available for only a few modes of vibration. The analysis results illustrate the improved accuracy of the new algorithm when compared to the existing algorithm.     

Preliminary studies on the dynamic prediction method of rainfall-triggered landslide

Rainfall-triggered landslides have posed significant threats to human lives and property each year in China. This paper proposed a meteorologicalgeotechnical early warning system GRAPES-LFM(GRAPES: Global and Regional Assimilation and Pr Ediction System; LFM: Landslide Forecast Model),basing on the GRAPES model and the landslide predicting model TRIGRS(Transient Rainfall Infiltration and Grid-based Regional Slope-Stability Model) for predicting rainfall-triggered landslides.This integrated system is evaluated in Dehua County,Fujian Province, where typhoon Bilis triggered widespread landslides in July 2006. The GRAPES model runs in 5 km×5 km horizontal resolution, and the initial fields and lateral boundaries are provided by NCEP(National Centers for Environmental Prediction) FNL(Final) Operational Global Analysis data. Quantitative precipitation forecasting products of the GRAPES model are downscaled to 25 m×25 m horizontal resolution by bilinear interpolation to drive the TRIGRS model. Results show that the observed areas locate in the high risk areas, and the GRAPES-LFM model could capture about 74% of the historical landslides with the rainfall intense 30mm/h. Meanwhile, this paper illustrates the relationship between the factor of safety(FS) and different rainfall patterns. GRAPES-LFM model enables us to further develop a regional, early warning dynamic prediction tool of rainfall-induced landslides.     

Formation mechanism of a rainfall triggered complex landslide in southwest China

This case study is about a landslide that occurred after 4 days of heavy rainfall,in the morning of June 29,2012,in Cengong County,Guizhou Province of China,geographical coordinated 108°20′-109°03′E,27°09′-27°32′N,with an estimated volume of 3.3×106 m3.To fully investigate the landslide process and formation mechanism,detailed geotechnical and geophysical investigations were performed including borehole drilling,sampling,and laboratory tests coupled with monitoring of displacement.Also,a combined seepage-slope stability modeling was performed to study the behavior of the landslide.After the heavy rainfall event,the sliding process started in this area.The landslide development can be divided into different parts.The man-made fill area,spatially distributed in the south side of the landslide area with low elevations,slid first along the interface between the slope debris and the strongly weathered bedrock roughly in the EW direction.Consequently,due to severe lateral shear disturbance,the slope in the main sliding zone slid next towards the SW direction,along the sliding surface developed within the strongly weathered calcareous shale formation located at a depth of 25-35 m.This means it was a rainfall triggered deep-seated landslide.Finally,retrogressive failure of a number of upstream blocks occurred,which moved in more than one direction.The initial failure of the man-made fill area was the‘engine’of the whole instability framework.This artificial material with low permeability,piled up in the accumulation area of surface and sub-surface and destroyed the drainage capacity of the groundwater.The numerical modeling results agreed with the analysis results obtained from the laboratory and field investigations.A conceptual model is given to illustrate the formation mechanism and development process of the landslide.     

新铺下二台滑坡变形机制及中长期预报模型

以奉节新铺下二台滑坡为例,基于GPS位移监测数据、裂缝数据、降雨量及库水位等多源数据,总结分析了大型古滑坡的复活规律,引入滑坡中长期预报模型,实现了以季度或月份为时间单位的跨水文年滑坡位移预测,并通过岩土体蠕变压缩模型,验证了推移式滑坡后缘裂缝形成机理。结果表明：(1)降雨是下二台滑坡变形的主导因素,滑坡变形使得滑体产生裂缝并成为降雨入渗通道,加剧了岩体破碎与软弱层软化,降低了滑坡稳定性,集中持续降雨可使滑坡失稳破坏;(2)通过模型预测值与地表监测数据的比较,将年降雨量作为滑坡中长期预报模型中的主控因素具有实际可操作性且有助于提高滑坡中长预报精度;(3)推移式滑坡后缘裂缝由滑坡推移式位移和岩土体压缩形成,引入蠕变压缩模型计算的裂缝宽度并和监测数据的比较说明,蠕变压缩模型非常适合该类边坡,同时应用岩土体蠕变压缩模型反推得到岩土体平均变形模量,判断岩体破碎程度,可以为滑坡稳定性分析及后续工程治理提供参考。     

基于物理模型试验的杆塔基础滑坡防护措施效果研究

大量穿越山地丘陵区的高压输电线路杆塔基础常位于滑坡灾害高易发斜坡地段,施加适当防护措施提高其稳定性,是保障输电线路持续安全运行的关键。为研究不同防护措施对杆塔基础滑坡的防护效果,以湖北省巴东县燕子滑坡为地质原型,设计制作物理试验模型,分别开展了极端降雨条件下滑坡在无防护、施加抗滑桩与格构护坡时的物理模型试验,从试验角度揭示了滑坡变形破坏特征与不同防护措施的防护效果。试验结果表明：在2种极端降雨工况(50,100 mm/h)下,无防护的滑坡体历经了坡表冲刷、裂缝扩展、局部垮塌变形与整体滑动的演化过程;抗滑桩措施对滑坡整体的防护效果显著,滑坡整体处于稳定状态,杆塔基础变形较小,杆塔倾斜率满足规范,但坡表会出现冲刷垮塌现象;格构护坡措施能有效减少坡面冲刷和坡脚垮塌风险,但在持续强降雨条件下对杆塔基础的整体稳固作用稍弱。物理模型试验结果与滑坡历史变形和实际治理效果吻合,试验结论可为类似杆塔基础滑坡的破坏机理研究与防护工程设计提供借鉴。     

Long runout mechanism of the Shenzhen 2015 landslide: insights from a two-phase flow viewpoint

A catastrophic landslide occurred at Hongao dumpsite in Guangming New District of Shenzhen, South China, on December 20, 2015. An estimated total volume of 2.73×106 m3 of construction spoils was mobilized during this event. The landslide traveled a long distance on a low-relief terrain. The affected area was approximately 1100 m in length and 630 m in width. This landslide made 33 buildings destroyed, 73 people died and 4 people lost. Due to the special dumping history and other factors, soil in this landfill is of high initial water content. To identify the major factors that attribute to the long runout character, a two-phase flow model of Iverson and George was used to simulate the dynamics of this landslide. The influence of initial hydraulic permeability, initial dilatancy, and earth pressure coefficient was examined through numerical simulations. We found that pore pressure has the most significant effect on the dynamic characteristics of Shenzhen landslides. Average pore pressure ratio ofthe whole basal surface was used to evaluate the degree of liquefaction for the sliding material. The evolution and influence factors of this ratio were analyzed based on the computational results. An exponential function was proposed to fit the evolution curve of the average pore pressure ratio, which can be used as a reasonable and simplified evaluation of the pore pressure. This fitting function can be utilized to improve the single-phase flow model.     

Fracture property identification method based on shrinkage factor particle swarm optimization

In the multi-wave and multi-component seismic exploration,shear-wave will be split into fast wave and slow wave,when it propagates in anisotropic media. Then the authors can predict polarization direction and density of crack and detect the development status of cracks underground according to shear-wave splitting phenomenon. The technology plays an important role and shows great potential in crack reservoir detection. In this study,the improved particle swarm optimization algorithm based on shrinkage factor is combined with the Pearson correlation coefficient method to obtain the fracture azimuth angle and density. The experimental results show that the modified method can improve the convergence rate,accuracy,anti-noise performance and computational efficiency.