Estimation of groundwater storage from seismic data using deep learning

Convolutional neural networks can provide a potential framework to characterize groundwater storage from seismic data. Estimation of key components, such as the amount of groundwater stored in an aquifer and delineate water table level, from active-source seismic data are performed in this study. The data to train, validate and test the neural networks are obtained by solving wave propagation in a coupled poroviscoelastic–elastic media. A discontinuous Galerkin method is applied to model wave propagation, whereas a deep convolutional neural network is used for the parameter estimation problem. In the numerical experiment, the primary unknowns estimated are the amount of stored groundwater and water table level, while the remaining parameters, assumed to be of less of interest, are marginalized in the convolutional neural network-based solution. Results, obtained through synthetic data, illustrate the potential of deep learning methods to extract additional aquifer information from seismic data, which otherwise would be impossible based on a set of reflection seismic sections or velocity tomograms.     

基于渐进式网络处理的低剂量Micro-CT成像方法

微米级计算机断层扫描（Micro-CT）的应用领域很广泛,在生物医学、材料科学等都有研究,而且最近几年的发展很迅速。Micro-CT图像常因辐射剂量的限制而出现噪声,所以开发一种合适的算法来抑制Micro-CT图像中噪声变得很重要。Micro-CT图像的噪声水平与扫描样本、扫描参数等参数有关,合适的噪声抑制算法应该在不同噪声水平下都有不错的性能。过去Micro-CT图像的噪声抑制算法主要是迭代重建算法,但迭代重建算法速度比较慢。深度学习方法作为近些年比较热门的图像处理方法,在临床低剂量CT图像处理上相比于传统方法效果更好、处理速度更快,有进一步在低剂量Micro-CT图像处理中应用的潜力。另外,生成对抗网络在保持图像细节上有着比卷积神经网络更好的效果,本文构建普通卷积神经网络与生成对抗网络,用于对比两者的性能差异。限制于放射源的功率,低噪声的Micro-CT数据难以获取,提出一种创新的扫描方式,可有效获取低噪声的Micro-CT数据,并对实验小鼠的进行了扫描。针对低剂量Micro-CT中较高的噪声,结合Micro-CT的成像原理,提出渐进式的低剂量Micro-CT图像处理方法,分别在解析重建前后对小鼠的Micro-CT图像进行两次处理。相比于只在断层图像上处理,渐进式方法对高噪声数据的处理效果更好。从客观指标与主观视觉效果上,分析和对比了渐进式方法与其他深度学习方法或迭代重建算法的差别。定量分析不同噪声水平下的Micro-CT图像的处理效果,分析生成对抗网络在渐进式Micro-CT图像处理中的优势与限制。渐进式Micro-CT图像处理方法生成的图像质量高、生成速度快、鲁棒性高、客观指标高,可以对进一步的高级应用如图像分割、图像三维可视化等提供帮助。      

Computation of response spectra from mining tremors using neural networks

Underground coal and copper ore exploitation in two Polish mining regions causes mining tremors and a series of other negative phenomena in the environment. Although these tremors are strictly connected with human activity, they differ considerably from other paraseismic vibrations. The moment of their occurrence is not to be foreseen likewise for earthquakes. The main problem discussed in the paper was formulated as the neural network evaluation of a relation between mining tremor energies, epicentral distances and acceleration response spectra. Back-propagation neural networks with Resilient back-propagation learning method were used. Each input vector included information about the mining tremor energy and the epicentral distance. Values of acceleration response spectrum were expected as the outputs of neural networks. Neurally evaluated spectra were compared with spectra computed on the basis of experimental data. After the network is trained and tested, it can be used for mapping of new data of mining tremor energies and epicentral distances into the spectra. Then, what is the substantial advantage of neural approach, the prediction of acceleration response spectra can be performed without recording of surface vibrations. In the light of the results, it is visible that the presented way of computation of acceleration response spectra can be peculiarly applied to prognosis of mining tremors influences on structures.     

Benchmark on the accuracy and efficiency of several neural network based phase pickers using datasets from China Seismic Network

Seismic phase pickers based on deep neural networks have been extensively used recently, demonstrating their advantages on both performance and efficiency. However, these pickers are trained with and applied to different data. A comprehensive benchmark based on a single dataset is therefore lacking. Here, using the recently released DiTing dataset, we analyzed performances of seven phase pickers with different network structures, the efficiencies are also evaluated using both CPU and GPU devices. Evaluations based on F₁-scores reveal that the recurrent neural network (RNN) and EQTransformer exhibit the best performance, likely owing to their large receptive fields. Similar performances are observed among PhaseNet (UNet), UNet++, and the lightweight phase picking network (LPPN). However, the LPPN models are the most efficient. The RNN and EQTransformer have similar speeds, which are slower than those of the LPPN and PhaseNet. UNet++ requires the most computational effort among the pickers. As all of the pickers perform well after being trained with a large-scale dataset, users may choose the one suitable for their applications. For beginners, we provide a tutorial on training and validating the pickers using the DiTing dataset. We also provide two sets of models trained using datasets with both 50 Hz and 100 Hz sampling rates for direct application by end-users. All of our models are open-source and publicly accessible.     

基于改进U-Net卷积神经网络的储层预测

传统的U-Net卷积神经网络大多存在深层网络梯度消失的问题。本文在U-Net卷积神经网络中加入残差模块,提出了一种改进U-Net卷积神经网络。残差模块保证了U-Net卷积神经网络在误差反向传播过程中梯度的存在,在一定程度上可以缓解梯度消失的问题。最后将改进U-Net卷积神经网络应用于实际储层预测中,实际数据测试结果表明基于改进U-Net卷积神经网络在岩性识别以及“甜点”预测上均能取得较好的效果。      

Fracture-frequency prediction from borehole wireline logs using artificial neural networks

Borehole-wall imaging is currently the most reliable means of mapping discontinuities within boreholes. As these imaging techniques are expensive and thus not always included in a logging run, a method of predicting fracture frequency directly from traditional logging tool responses would be very useful and cost effective. Artificial neural networks (ANNs) show great potential in this area. ANNs are computational systems that attempt to mimic natural biological neural networks. They have the ability to recognize patterns and develop their own generalizations about a given data set. Neural networks are trained on data sets for which the solution is known and tested on data not previously seen in order to validate the network result. We show that artificial neural networks, due to their pattern recognition capabilities, are able to assess the signal strength of fracture-related heterogeneity in a borehole log and thus fracture frequency within a borehole. A combination of wireline logs (neutron porosity, bulk density, P-sonic, S-sonic, deep resistivity and shallow resistivity) were used as input parameters to the ANN. Fracture frequency calculated from borehole televiewer data was used as the single output parameter. The ANN was trained using a back-propagation algorithm with a momentum learning function. In addition to fracture frequency within a single borehole, an ANN trained on a subset of boreholes in an area could be used for prediction over the entire set of boreholes, thus allowing the lateral correlation of fracture zones.     

基于深度降噪自编码神经网络的中国大陆地壳厚度反演

本文采用基于数据驱动的深度降噪自编码网络构建了瑞雷面波群速度、相速度频散特性与地壳厚度的正反演函数关系,并利用最新频散模型反演了中国大陆的地壳厚度.对于神经网络架构体系的评价,除了考虑传统意义上的测试误差、训练误差之外,本文还用已知物理原理的正演结果与网络预测结果进行比较;在设计网络构架时,同时考虑地球模型和面波频散的...     

Identification and classification of multiple reflections with self-organizing maps

Artificial neural networks can be used effectively to identify and classify multiple events in a seismic data set. We use a specialized neural network, a self-organizing map (SOM), that tries to establish rules for the characterization of the physical problem. Selected seismic data attributes from CMP gathers are used as input patterns, such that the SOM arranges the data to form clusters in an abstract space. We show with synthetic and real data how the SOM can identify and classify primaries and multiples, and how it can classify the various types of multiple corresponding to a certain generating mechanism in the subsurface.     

Real‐time recurrent learning neural network for stream‐flow forecasting

Various types of neural networks have been proposed in previous papers for applications in hydrological events. However, most of these applied neural networks are classified as static neural networks, which are based on batch processes that update action only after the whole training data set has been presented. The time variate characteristics in hydrological processes have not been modelled well. In this paper, we present an alternative approach using an artificial neural network, termed real‐time recurrent learning (RTRL) for stream‐flow forecasting. To define the properties of the RTRL algorithm, we first compare the predictive ability of RTRL with least‐square estimated autoregressive integrated moving average models on several synthetic time‐series. Our results demonstrate that the RTRL network has a learning capacity with high efficiency and is an adequate model for time‐series prediction. We also investigated the RTRL network by using the rainfall–runoff data of the Da‐Chia River in Taiwan. The results show that RTRL can be applied with high accuracy to the study of real‐time stream‐flow forecasting networks. Copyright © 2002 John Wiley & Sons, Ltd.     

New method of generating spectrum compatible accelerograms using neural networks

A new method is proposed for generating artificial earthquake accelerograms from response spectra. This method uses the learning capabilities of neural networks to developed the knowledge of the inverse mapping from the response spectra to earthquake accelerogram. In the proposed method the neural networks learn the inverse mapping directly from the actual recorded earthquake accelerograms and their response spectra. A two-stage approach is used. In the first stage, a replicator neural network is used as a data compression tool. The replicator neural network compresses the vector of the discrete Fourier spectra of the accelerograms to vectors of much smaller dimension. In the second stage, a multi-layer feed-forward neural network learns to relate the response spectrum to the compressed Fourier spectrum. A simple example is presented, in which only 30 accelerograms are used to train the two-stage neural networks. This example demonstrates how the method works and shows its potential. © 1998 John Wiley & Sons, Ltd.     

Using Artificial Neural Networks to Predict the Presence of Overpressured Zones in the Anadarko Basin,Oklahoma

Many sedimentary basins throughout the world exhibit areas with abnormal pore-fluid pressures (higher or lower than normal or hydrostatic pressure). Predicting pore pressure and other parameters (depth, extension, magnitude, etc.) in such areas are challenging tasks. The compressional acoustic (sonic) log (DT) is often used as a predictor because it responds to changes in porosity or compaction produced by abnormal pore-fluid pressures. Unfortunately, the sonic log is not commonly recorded in most oil and/or gas wells. We propose using an artificial neural network to synthesize sonic logs by identifying the mathematical dependency between DT and the commonly available logs, such as normalized gamma ray (GR) and deep resistivity logs (REID). The artificial neural network process can be divided into three steps: (1) Supervised training of the neural network; (2) confirmation and validation of the model by blind-testing the results in wells that contain both the predictor (GR, REID) and the target values (DT) used in the supervised training; and 3) applying the predictive model to all wells containing the required predictor data and verifying the accuracy of the synthetic DT data by comparing the back-predicted synthetic predictor curves (GRNN, REIDNN) to the recorded predictor curves used in training (GR, REID). Artificial neural networks offer significant advantages over traditional deterministic methods. They do not require a precise mathematical model equation that describes the dependency between the predictor values and the target values and, unlike linear regression techniques, neural network methods do not overpredict mean values and thereby preserve original data variability. One of their most important advantages is that their predictions can be validated and confirmed through back-prediction of the input data. This procedure was applied to predict the presence of overpressured zones in the Anadarko Basin, Oklahoma. The results are promising and encouraging.     

深度学习在地震定位中的应用综述

近年来,深度学习的发展给科研人员开辟了地震定位研究的新思路,科研人员将深度学习技术应用于地震定位并取得了较好的效果。文章首先介绍根据神经网络的编码与解码对深度神经网络的分类,然后对深度学习的基本流程进行总结,最后对深度学习中广泛应用于地震定位的方法进行综述,总结不同方法的特点和实际应用情况。结果表明：深度学习方法能够实现地震事件的自动定位,且定位的精度较高,缩短了地震定位所需时间,在处理地震大数据方面也具有明显优势,能够克服目前传统地球物理方法在地震定位方面的一些不足之处。相信随着深度学习技术的进一步发展,必将更为广泛地应用于地震定位研究中。     

Model data selection using gamma test for daily solar radiation estimation

Hydrological modelling is a complicated procedure and there are many tough questions facing all modellers: what input data should be used? how much data is required? and what model should be used? In this paper, the gamma test (GT) has been used for the first time in modelling one of the key hydrological components: solar radiation. The study aimed to resolve the questions about the relative importance of input variables and to determine the optimum number of data points required to construct a reliable smooth model. The proposed methodology has been studied through the estimation of daily solar radiation in the Brue Catchment, the UK. The relationship between input and output in the meteorological data sets was achieved through error variance estimation before the modelling using the GT. This work has demonstrated how the GT helps model development in nonlinear modelling techniques such as local linear regression (LLR) and artificial neural networks (ANN). It was found that the GT provided very useful information for input data selection and subsequent model development. The study has wider implications for various hydrological modelling practices and suggests further exploration of this technique for improving informed data and model selection, which has been a difficult field in hydrology in past decades. Copyright © 2008 John Wiley & Sons, Ltd.     

Using recurrent neural networks to reconstruct rainfall‐runoff processes

Many novel techniques for reconstructing rainfall‐runoff processes require hydrometeorologic and geomorphologic information for modelling. However, certain information is not always measurable. In this paper, we employ a special recurrent neural network to reconstruct the rainfall‐runoff process by using collected rainfall data. In addition, we propose an indirect system identification to overcome the drawback of a traditional, time‐consuming trial‐and‐error search. The indirect system identification is an efficient method to recognize the structure of a recurrent neural network. The unit hydrograph can be derived directly from the weights of the network due to a state‐space form embedded in the recurrent neural network. This improves the link between the weights of the network and the physical concepts that most neural networks fail to connect. The case studies of 41 events from 1966 to 1997 have been implemented in Taiwan's Wu‐Tu watershed, where the runoff path‐lines are short and steep. Two recurrent neural networks and one state‐space model are utilized to simulate the rainfall‐runoff processes for comparison. The results are validated by four criteria: coefficient of efficiency; peak discharge error; time to peak arrival error; total discharge volume error. The resulting data from the recurrent neural network reveal that the neural network proposed herein is appropriate for hydrological systems. Copyright © 2005 John Wiley & Sons, Ltd.     

Deep neural network application for 4D seismic inversion to changes in pressure and saturation: Optimizing the use of synthetic training datasets

In this work, we tackle the challenge of quantitative estimation of reservoir dynamic property variations during a period of production, directly from four-dimensional seismic data in the amplitude domain. We employ a deep neural network to invert four-dimensional seismic amplitude maps to the simultaneous changes in pressure, water and gas saturations. The method is applied to a real field data case, where, as is common in such applications, the data measured at the wells are insufficient for properly training deep neural networks, thus, the network is trained on synthetic data. Training on synthetic data offers much freedom in designing a training dataset, therefore, it is important to understand the impact of the data distribution on the inversion results. To define the best way to construct a synthetic training dataset, we perform a study on four different approaches to populating the training set making remarks on data sizes, network generality and the impact of physics-based constraints. Using the results of a reservoir simulation model to populate our training datasets, we demonstrate the benefits of restricting training samples to fluid flow consistent combinations in the dynamic reservoir property domain. With this the network learns the physical correlations present in the training set, incorporating this information into the inference process, which allows it to make inferences on properties to which the seismic data are most uncertain. Additionally, we demonstrate the importance of applying regularization techniques such as adding noise to the synthetic data for training and show a possibility of estimating uncertainties in the inversion results by training multiple networks.     

南北构造带岩石圈结构与地震的研究

南北构造带是中国大陆东西部大地构造的主要分界,也是大陆内部强烈地震发生的主要地区之一.2008年汶川MS8.0地震发生后,在南北构造带及周边地区进行了大量的野外科学考察、深部地球物理探测和流动地震观测,在岩石圈结构与构造、强震发生的深部构造环境和动力学过程等方面获得了重要的进展.本文综述近年来发表的一批研究成果,包括岩石圈结构的深部地球物理探测和成像,地震层析成像,地震各向异性和壳幔变形,与近期发生的强烈地震相关研究,以及与大陆动力学有关的研究等.自2000年以来,我国建成了具有1000多个地震台站的国家和区域地震台网.它们在实时为地震监测服务的同时,其产出的海量数据还提供用于地球科学研究.一批作者在国内外发表了研究成果,大大提高了对南北构造带的认识.我们虽然取得许多共识性的重要成果,但是也存在一些问题,发现不同作者的结论是相互矛盾的.其原因之一可能是,现有台网的数据成像分辨率和精度仍不足以识别在地壳深处的细节,例如在孕震尺度概念下的地震危险区.加强流动地震观测,提高台站分布的密度,取得高可信度的目标模型是解决问题的重要方面.近年来"中国地震科学台阵观测"计划在南北构造带上实施的大型流动台阵观测,结合固定地震台网的资料,加上高分辨率深部地球物理探测,以获得高可信度的地壳上地幔三维精细结构及物性成像,是提高地震科学和大陆动力学研究水平的一个有效途径.     

Long‐lead seasonal rainfall forecasting using time‐delay recurrent neural networks: a case study

A temporal artificial neural network‐based model is developed and applied for long‐lead rainfall forecasting. Tapped delay lines and recurrent connections are two different components that are used along with a static multilayer perceptron network to design a time‐delay recurrent neural network. The proposed model is, in fact, a combination of time‐delay and recurrent neural networks. The model is applied in three case studies of the Northwest, West, and Southwest basins of Iran. In addition, an autoregressive moving average with exogenous inputs (ARMAX) model is used as a baseline in order to be compared with the time‐delay recurrent neural networks developed in this study. Large‐scale climate signals, such as sea‐level pressure, that affect the rainfall of the study area are used as the predictors in the models, as well as the persistence between rainfall data. The results of winter‐spring rainfall forecasts are discussed thoroughly. It is demonstrated that in all cases the proposed neural network results in better forecasts in comparison with the statistical ARMAX model. Moreover, it is found that in two of three case studies the time‐delay recurrent neural networks perform better than either recurrent or time‐delay neural networks. The results demonstrate that the proposed method can significantly improve the long‐lead forecast by utilizing a non‐linear relationship between climatic predictors and rainfall in a region. Copyright © 2007 John Wiley & Sons, Ltd.     

利用卷积神经网络检测地震的方法与优化

本文以西昌台阵观测的8 321次近震数据为例,详细介绍了利用深度卷积神经网络检测地震的数据处理流程,包括数据预处理、模型训练、波形长度、网络层数、学习率和概率阈值等关键参数对检测结果的影响,并将训练得到的最优模型,应用于事件波形和连续波形的检测。研究表明,数据预处理和数据增强可以提升模型的检测精度和抗干扰能力。用于模型训练的波形窗口长度可近似于S-P到时差的最大值。不同网络层数(5—8层)的检测结果差别不大。对于地震检测,学习率设为10-4—10-3较为合适。卷积神经网络检测出的地震数量与选择的概率阈值有关,通过绘制精确率-召回率变化曲线,可以为选择合适的概率阈值提供参考。本文为进一步利用深度学习算法提高地震检测效果提供了参考。     

A multi-data training method for a deep neural network to improve the separation effect of simultaneous-source data

Within the field of seismic data acquisition with active sources, the technique of acquiring simultaneous data, also known as blended data, offers operational advantages. The preferred processing of blended data starts with a step of deblending, that is separation of the data acquired by the different sources, to produce data that mimic data from a conventional seismic acquisition and can be effectively processed by standard methods. Recently, deep learning methods based on the deep neural network have been applied to the deblending task with promising results, in particular using an iterative approach. We propose an enhancement to deblending with an iterative deep neural network, whereby we modify the training stage of the deep neural network in order to achieve better performance through the iterations. We refer to the method that only uses the blended data as the input data as the general training method. Our new multi-data training method allows the deep neural network to be trained by the data set with the input patches composed of blended data, noisy data with low amplitude crosstalk noise, and unblended data, which can improve the ability of the deep neural network to remove crosstalk noise and protect weak signal. Based on such an extended training data set, the multi-data training method embedded in the iterative separation framework can result in different outputs at different iterations and converge to the best result in a shorter iteration number. Transfer learning can further improve the generalization and separation efficacy of our proposed method to deblend the simultaneous-source data. Our proposed method is tested on two synthetic data and two field data to prove the effectiveness and superiority in the deblending of the simultaneous-source data compared with the general training method, generic noise attenuation network and low-rank matrix factorization methods.     

Evaluation of coalbed methane reservoirs from geophysical log data using an improved fuzzy comprehensive decision method and a homologous neural network

The evaluation of coalbed methane reservoirs using log data is an important approach in the exploration and development of coalbed methane reservoirs. Most commonly, regression techniques, fuzzy recognition and neural networks have been used to evaluate coalbed methane reservoirs. It is known that a coalbed is an unusual reservoir. There are many difficulties in using regression methods and empirical qualitative recognition to evaluate a coalbed, but fuzzy recognition, such as the fuzzy comprehensive decision method, and neural networks, such as the back-propagation (BP) network, are widely used. However, there are no effective methods for computing weights for the fuzzy comprehensive decision method, and the BP algorithm is a local optimization algorithm, easily trapped in local minima, which significantly affect the results. In this paper, the recognition method for coal formations is developed; the improved fuzzy comprehensive decision method, which uses an optimization approach for computing weighted coefficients, is developed for the qualitative recognition of coalbed methane reservoirs. The homologous neural network, using a homologous learning algorithm, which is a global search optimization, is presented for the quantitative analysis of parameters for coalbed methane reservoirs. The applied procedures for these methods and some problems related to their application are also discussed. Verification of the above methods is made using log data from the coalbed methane testing area in North China. The effectiveness of the methods is demonstrated by the analysis of results for real log data.