Computers & Geosciences最新文献_第5页

Enhanced global soil moisture prediction through a sampling-weighted sensitive learning strategy applied to various LSTM-based models 基于lstm模型的采样加权敏感学习策略增强了全球土壤湿度预测

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-10-17 DOI: 10.1016/j.cageo.2025.106068

Xiaoning Li , Zhichao Zhong , Qingliang Li , Cheng Zhang , Hongwei Zhao , Xiaofeng Li , Jinlong Zhu , Sen Yan

Soil moisture (SM) plays a critical role in land-atmosphere interactions, influencing both water and carbon cycles. Accurate SM predictions are essential for effective disaster response, optimized irrigation practices, and progress in environmental research. Deep learning (DL) models have become increasingly popular for predicting SM. However, many existing approaches overlook the imbalance in observed data—where moderate moisture levels are far more common than extreme dry or wet conditions. This skewed distribution limits the models' ability to accurately capture rare but critical extremes, ultimately reducing their overall effectiveness. To overcome this limitation, we propose a Sampling-Weighted Sensitive Learning Strategy that improves model generalization by assigning greater importance to rare samples during training. We evaluated this approach using three widely used DL architectures: Long Short-Term Memory Network (LSTM), Bidirectional Long Short-Term Memory Network (BiLSTM), and Gated Recurrent Unit (GRU). To ensure consistency across experiments, the same random seed was applied throughout. Our results demonstrate notable improvements in prediction accuracy when applying the proposed strategy. The BiLSTM model, in particular, showed the most significant gains: unbiased Root Mean Square Error (ubRMSE) decreased by 7.38 %, and Bias was reduced by 11.64 %. Its Kling-Gupta Efficiency (KGE) improved by 2.73 %—slightly below the 5.35 % gain observed for the unidirectional LSTM—but regional results were particularly strong. In data-scarce areas, especially North Africa and Western Asia, BiLSTM KGE improvements frequently exceeded 20 %. Models trained with the proposed strategy also produced narrower 95 % confidence intervals during high-variability periods (e.g., summer and dry seasons), indicating greater predictive robustness under challenging environmental. These findings underscore the importance of addressing sample imbalance in training data and demonstrate the effectiveness of our strategy in enhancing DL models for SM prediction.

土壤水分在陆地-大气相互作用中起着至关重要的作用，影响着水和碳的循环。准确的SM预测对于有效的灾害响应、优化灌溉实践和环境研究进展至关重要。深度学习（DL）模型在预测SM方面越来越受欢迎。然而，许多现有的方法忽略了观测数据的不平衡——适度的湿度水平远比极端干燥或潮湿的条件更常见。这种倾斜的分布限制了模型准确捕捉罕见但关键的极端情况的能力，最终降低了模型的总体有效性。为了克服这一限制，我们提出了一种抽样加权敏感学习策略，该策略通过在训练过程中赋予稀有样本更大的重要性来提高模型泛化。我们使用三种广泛使用的深度学习架构来评估这种方法：长短期记忆网络（LSTM）、双向长短期记忆网络（BiLSTM）和门控循环单元（GRU）。为了确保实验的一致性，在整个过程中使用相同的随机种子。我们的结果表明，当应用所提出的策略时，预测精度显着提高。特别是BiLSTM模型，显示出最显著的增益：无偏均方根误差（ubRMSE）降低了7.38%，偏差降低了11.64%。它的克林-古普塔效率（KGE）提高了2.73%，略低于单向lstm观察到的5.35%的增益，但区域结果特别强劲。在数据匮乏的地区，特别是北非和西亚，BiLSTM KGE的改进经常超过20%。采用该策略训练的模型在高变异性时期（如夏季和旱季）也产生了更窄的95%置信区间，表明在具有挑战性的环境下具有更强的预测稳健性。这些发现强调了解决训练数据中样本不平衡的重要性，并证明了我们的策略在增强深度学习模型用于SM预测方面的有效性。

{"title":"Enhanced global soil moisture prediction through a sampling-weighted sensitive learning strategy applied to various LSTM-based models","authors":"Xiaoning Li , Zhichao Zhong , Qingliang Li , Cheng Zhang , Hongwei Zhao , Xiaofeng Li , Jinlong Zhu , Sen Yan","doi":"10.1016/j.cageo.2025.106068","DOIUrl":"10.1016/j.cageo.2025.106068","url":null,"abstract":"<div><div>Soil moisture (SM) plays a critical role in land-atmosphere interactions, influencing both water and carbon cycles. Accurate SM predictions are essential for effective disaster response, optimized irrigation practices, and progress in environmental research. Deep learning (DL) models have become increasingly popular for predicting SM. However, many existing approaches overlook the imbalance in observed data—where moderate moisture levels are far more common than extreme dry or wet conditions. This skewed distribution limits the models' ability to accurately capture rare but critical extremes, ultimately reducing their overall effectiveness. To overcome this limitation, we propose a Sampling-Weighted Sensitive Learning Strategy that improves model generalization by assigning greater importance to rare samples during training. We evaluated this approach using three widely used DL architectures: Long Short-Term Memory Network (LSTM), Bidirectional Long Short-Term Memory Network (BiLSTM), and Gated Recurrent Unit (GRU). To ensure consistency across experiments, the same random seed was applied throughout. Our results demonstrate notable improvements in prediction accuracy when applying the proposed strategy. The BiLSTM model, in particular, showed the most significant gains: unbiased Root Mean Square Error (<em>ubRMSE</em>) decreased by 7.38 %, and <em>Bias</em> was reduced by 11.64 %. Its Kling-Gupta Efficiency (<em>KGE</em>) improved by 2.73 %—slightly below the 5.35 % gain observed for the unidirectional LSTM—but regional results were particularly strong. In data-scarce areas, especially North Africa and Western Asia, BiLSTM KGE improvements frequently exceeded 20 %. Models trained with the proposed strategy also produced narrower 95 % confidence intervals during high-variability periods (e.g., summer and dry seasons), indicating greater predictive robustness under challenging environmental. These findings underscore the importance of addressing sample imbalance in training data and demonstrate the effectiveness of our strategy in enhancing DL models for SM prediction.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106068"},"PeriodicalIF":4.4,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145321635","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Invertible neural network for real-time inversion and uncertainty quantification of ultra-deep resistivity measurements 超深电阻率测量实时反演与不确定度量化的可逆神经网络

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-10-17 DOI: 10.1016/j.cageo.2025.106067

George Bittar , Sihong Wu , Yawei Su , Shubin Zeng , Jiajia Sun , Xuqing Wu , Yueqin Huang , Jiefu Chen

Real-time geosteering, formation evaluation, and wellbore placement decisions hinge on the ability to invert electromagnetic (EM) well logging measurements in a fast manner while understanding the associated uncertainties. Conventional deterministic inversion methods, such as the Levenberg–Marquardt algorithm (LMA) and Occam’s inversion, often get trapped in local minima and yield a single optimal solution, neglecting the impact of the non-uniqueness of solutions. Bayesian approaches like Markov Chain Monte Carlo (MCMC) can provide the posterior distribution but are computationally expensive, making them impractical for real-time inversion. In this study, we develop a deep learning-based invertible neural network (INN) that performs rapid approximate Bayesian inversion under a specific likelihood and provides uncertainty quantification (UQ) for ultra-deep resistivity measurements. Synthetic tests demonstrate that the INN recovers the posterior distribution and generates an ensemble of predictions to quantify uncertainty within seconds. We compare its performance with conventional inversion algorithms, including LMA and Occam’s inversion, evaluating accuracy and inference efficiency. The results show that the INN delivers reliable resistivity inversion with uncertainty information at a fraction of the computational cost, highlighting its potential for real-time geosteering and other drilling-related decision-making tasks.

实时地质导向、地层评价和井筒布置决策取决于能否快速反演电磁测井数据，同时了解相关的不确定性。传统的确定性反演方法，如Levenberg-Marquardt算法（LMA）和Occam反演，往往陷入局部最小值，只得到一个最优解，而忽略了解的非唯一性的影响。马尔可夫链蒙特卡罗（MCMC）等贝叶斯方法可以提供后验分布，但计算成本高，无法实现实时反演。在本研究中，我们开发了一种基于深度学习的可逆神经网络（INN），该网络在特定似然下执行快速近似贝叶斯反演，并为超深电阻率测量提供不确定性量化（UQ）。综合测试表明，INN可以在数秒内恢复后验分布并生成预测集合以量化不确定性。我们将其性能与传统的反演算法，包括LMA和Occam的反演进行了比较，评估了精度和推理效率。结果表明，INN能够以很小的计算成本提供可靠的不确定信息电阻率反演，突出了其在实时地质导向和其他钻井相关决策任务方面的潜力。

{"title":"Invertible neural network for real-time inversion and uncertainty quantification of ultra-deep resistivity measurements","authors":"George Bittar , Sihong Wu , Yawei Su , Shubin Zeng , Jiajia Sun , Xuqing Wu , Yueqin Huang , Jiefu Chen","doi":"10.1016/j.cageo.2025.106067","DOIUrl":"10.1016/j.cageo.2025.106067","url":null,"abstract":"<div><div>Real-time geosteering, formation evaluation, and wellbore placement decisions hinge on the ability to invert electromagnetic (EM) well logging measurements in a fast manner while understanding the associated uncertainties. Conventional deterministic inversion methods, such as the Levenberg–Marquardt algorithm (LMA) and Occam’s inversion, often get trapped in local minima and yield a single optimal solution, neglecting the impact of the non-uniqueness of solutions. Bayesian approaches like Markov Chain Monte Carlo (MCMC) can provide the posterior distribution but are computationally expensive, making them impractical for real-time inversion. In this study, we develop a deep learning-based invertible neural network (INN) that performs rapid approximate Bayesian inversion under a specific likelihood and provides uncertainty quantification (UQ) for ultra-deep resistivity measurements. Synthetic tests demonstrate that the INN recovers the posterior distribution and generates an ensemble of predictions to quantify uncertainty within seconds. We compare its performance with conventional inversion algorithms, including LMA and Occam’s inversion, evaluating accuracy and inference efficiency. The results show that the INN delivers reliable resistivity inversion with uncertainty information at a fraction of the computational cost, highlighting its potential for real-time geosteering and other drilling-related decision-making tasks.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106067"},"PeriodicalIF":4.4,"publicationDate":"2025-10-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145364037","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

A new elastic wave equation for decoupling P-wave and S-waves and its application 纵波与横波解耦的弹性波动方程及其应用

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-10-14 DOI: 10.1016/j.cageo.2025.106065

Meng Guo , Bingshou He , Qianqian Ci

The imaging of P-wave and S-wave in reverse time migration (RTM) of elastic waves is often achieved by cross-correlating P-waves or S-waves with different propagation directions. This requires us to obtain the Poynting vector or optical flow vector of each imaging point at different times during the wavefield extrapolation process and use it to indicate the direction of wave propagation. But we can only obtain the Poynting vector of the mixed wavefield of P-wave and S- waves, and we cannot obtain the Poynting vector of pure P-wave or pure S-wave when using the existing velocity-stress elastic wave equations for the wavefield extrapolation process. Therefore, the propagation direction obtained is also a mixed wavefield rather than pure P-wave or pure S-wave, and this does not meet the requirements for elastic wave RTM and will cause errors. The existing first-order velocity-dilation-rotation elastic wave equation, although it overcomes the aforementioned issues, cannot accurately describe the law of wave propagation at the wave impedance interface due to the assumption of a homogeneous medium. Especially when the interface of P-wave and S-wave velocities is not consistent, it will lead to errors in the reflection, transmission, and conversion wavefields when using this equation for elastic wavefield extrapolation. In addition, severe energy leakage effects will occur at the interface of S-wave velocity when using this equation, which will lead to inaccurate S-wave imaging. In this paper, we propose a new elastic wave equation for decoupling P-wave and S-waves based on the assumption of an inhomogeneous medium, which not only gives the propagation direction of pure P-wave and pure S-wave, but also completely overcomes the above problems. Using the new equation of the Poynting vector in the elastic wave field to perform cross-correlation imaging, the model calculations show that the imaging results eliminate the noise generated by RTM, demonstrating the accuracy and applicability of the equation.

弹性波逆时偏移（RTM）中的纵波和横波成像通常是通过不同传播方向的纵波或横波相互关联来实现的。这就要求我们在波场外推过程中，获取每个成像点在不同时刻的坡印亭矢量或光流矢量，并用它来指示波的传播方向。但我们只能得到纵波和横波混合波场的Poynting矢量，而用现有的速度-应力弹性波方程进行波场外推时，无法得到纯纵波或纯横波的Poynting矢量。因此，得到的传播方向也是混合波场，而不是纯p波或纯s波，这不符合弹性波RTM的要求，会产生误差。现有的一阶速度-膨胀-旋转弹性波动方程虽然克服了上述问题，但由于假设介质均质，无法准确描述波在波阻抗界面处的传播规律。特别是当纵波和横波速度界面不一致时，用该方程进行弹性波场外推时，会导致反射、透射和转换波场出现误差。此外，使用该方程时，在横波速度界面处会产生严重的能量泄漏效应，导致横波成像不准确。本文基于非均匀介质的假设，提出了一种新的纵波与横波解耦的弹性波动方程，不仅给出了纯纵波和纯横波的传播方向，而且完全克服了上述问题。利用弹性波场中新的Poynting矢量方程进行互相关成像，模型计算表明，成像结果消除了RTM产生的噪声，证明了该方程的准确性和适用性。

{"title":"A new elastic wave equation for decoupling P-wave and S-waves and its application","authors":"Meng Guo , Bingshou He , Qianqian Ci","doi":"10.1016/j.cageo.2025.106065","DOIUrl":"10.1016/j.cageo.2025.106065","url":null,"abstract":"<div><div>The imaging of P-wave and S-wave in reverse time migration (RTM) of elastic waves is often achieved by cross-correlating P-waves or S-waves with different propagation directions. This requires us to obtain the Poynting vector or optical flow vector of each imaging point at different times during the wavefield extrapolation process and use it to indicate the direction of wave propagation. But we can only obtain the Poynting vector of the mixed wavefield of P-wave and S- waves, and we cannot obtain the Poynting vector of pure P-wave or pure S-wave when using the existing velocity-stress elastic wave equations for the wavefield extrapolation process. Therefore, the propagation direction obtained is also a mixed wavefield rather than pure P-wave or pure S-wave, and this does not meet the requirements for elastic wave RTM and will cause errors. The existing first-order velocity-dilation-rotation elastic wave equation, although it overcomes the aforementioned issues, cannot accurately describe the law of wave propagation at the wave impedance interface due to the assumption of a homogeneous medium. Especially when the interface of P-wave and S-wave velocities is not consistent, it will lead to errors in the reflection, transmission, and conversion wavefields when using this equation for elastic wavefield extrapolation. In addition, severe energy leakage effects will occur at the interface of S-wave velocity when using this equation, which will lead to inaccurate S-wave imaging. In this paper, we propose a new elastic wave equation for decoupling P-wave and S-waves based on the assumption of an inhomogeneous medium, which not only gives the propagation direction of pure P-wave and pure S-wave, but also completely overcomes the above problems. Using the new equation of the Poynting vector in the elastic wave field to perform cross-correlation imaging, the model calculations show that the imaging results eliminate the noise generated by RTM, demonstrating the accuracy and applicability of the equation.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106065"},"PeriodicalIF":4.4,"publicationDate":"2025-10-14","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145364036","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Parallel finite element forward modeling of 3-D magnetotelluric conductivity and permeability anisotropy with coupled PML boundary conditions 耦合PML边界条件下三维大地电磁导电性和渗透率各向异性的平行有限元正演模拟

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-10-10 DOI: 10.1016/j.cageo.2025.106064

Shuaiying Qiao , Tiaojie Xiao , Junjun Zhou , Chunye Gong , Bo Yang , Jie Liu , Yun Wang , Qinglin Wang

Magnetotelluric sounding (MT) is a crucial geophysical exploration method, with its response primarily influenced by two physical parameters: conductivity and magnetic permeability. MT forward modeling typically presents as a large-scale, open-domain problem, necessitating boundary truncation and computational acceleration for the simulation area. Compared to traditional boundary conditions, the Perfectly Matched Layer (PML) offers a more efficient and accurate truncation method. However, the current application of the PML is confined to scenarios involving variations in conductivity alone, and is unable to accommodate simultaneous variations in both conductivity and permeability, as well as complex anisotropic models. Therefore, this paper proposes a PML that accounts for both conductivity and permeability parameters, as well as anisotropy, making it suitable for complex anisotropic models. Furthermore, by integrating the Multi-Processing Interface (MPI) to design a multi-level parallel processing scheme, we have achieved parallel vector finite element forward modeling of three-dimensional magnetotelluric conductivity and permeability anisotropy with coupled PML boundary conditions. In comparison with previous results, the PML boundary conditions have been validated to possess the advantages of high efficiency, high precision, and stable performance. Numerical experimental results indicate that, compared with traditional boundary conditions, the PML reduces the degrees of freedom (DOFs) by over 85%, and decreases both computation time and memory usage by more than 90%. Compared with the conventional method with 888,822 DOFs, the proposed method, which integrates the PML and a multi-level parallelization strategy, achieves a speedup of approximately 85.24 for a single frequency using 32 processes and approximately 649.63 for 8 frequencies using 512 processes. The PML boasts a wider range of applicability, better performance, and thus holds broader prospects for application.

大地电磁测深是一种重要的地球物理勘探方法，其响应主要受电导率和磁导率两个物理参数的影响。MT正演建模通常是一个大规模的开放域问题，需要对模拟区域进行边界截断和计算加速。与传统的边界条件相比，完美匹配层（PML）提供了一种更有效、更准确的截断方法。然而，目前PML的应用仅限于仅涉及电导率变化的场景，无法适应电导率和渗透率的同时变化，以及复杂的各向异性模型。因此，本文提出了一种同时考虑电导率和渗透率参数以及各向异性的PML，使其适用于复杂的各向异性模型。此外，通过集成多处理接口（MPI）设计多层次并行处理方案，实现了具有耦合PML边界条件的三维大地电磁导电性和渗透率各向异性的并行矢量有限元正演模拟。通过与已有结果的比较，验证了PML边界条件具有效率高、精度高、性能稳定等优点。数值实验结果表明，与传统边界条件相比，该边界条件的自由度降低了85%以上，计算时间和内存占用均降低了90%以上。与传统方法的888,822 dof相比，该方法集成了PML和多级并行化策略，使用32个进程实现了单个频率约85.24的加速，使用512个进程实现了8个频率约649.63的加速。PML的适用范围更广，性能更好，具有更广阔的应用前景。

{"title":"Parallel finite element forward modeling of 3-D magnetotelluric conductivity and permeability anisotropy with coupled PML boundary conditions","authors":"Shuaiying Qiao , Tiaojie Xiao , Junjun Zhou , Chunye Gong , Bo Yang , Jie Liu , Yun Wang , Qinglin Wang","doi":"10.1016/j.cageo.2025.106064","DOIUrl":"10.1016/j.cageo.2025.106064","url":null,"abstract":"<div><div>Magnetotelluric sounding (MT) is a crucial geophysical exploration method, with its response primarily influenced by two physical parameters: conductivity and magnetic permeability. MT forward modeling typically presents as a large-scale, open-domain problem, necessitating boundary truncation and computational acceleration for the simulation area. Compared to traditional boundary conditions, the Perfectly Matched Layer (PML) offers a more efficient and accurate truncation method. However, the current application of the PML is confined to scenarios involving variations in conductivity alone, and is unable to accommodate simultaneous variations in both conductivity and permeability, as well as complex anisotropic models. Therefore, this paper proposes a PML that accounts for both conductivity and permeability parameters, as well as anisotropy, making it suitable for complex anisotropic models. Furthermore, by integrating the Multi-Processing Interface (MPI) to design a multi-level parallel processing scheme, we have achieved parallel vector finite element forward modeling of three-dimensional magnetotelluric conductivity and permeability anisotropy with coupled PML boundary conditions. In comparison with previous results, the PML boundary conditions have been validated to possess the advantages of high efficiency, high precision, and stable performance. Numerical experimental results indicate that, compared with traditional boundary conditions, the PML reduces the degrees of freedom (DOFs) by over 85%, and decreases both computation time and memory usage by more than 90%. Compared with the conventional method with 888,822 DOFs, the proposed method, which integrates the PML and a multi-level parallelization strategy, achieves a speedup of approximately 85.24 for a single frequency using 32 processes and approximately 649.63 for 8 frequencies using 512 processes. The PML boasts a wider range of applicability, better performance, and thus holds broader prospects for application.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106064"},"PeriodicalIF":4.4,"publicationDate":"2025-10-10","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145321623","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Eigenvector decomposition for joint analysis of spatial characteristics in the North Atlantic from 1979 to 2024 1979 - 2024年北大西洋空间特征联合分析的特征向量分解

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-10-09 DOI: 10.1016/j.cageo.2025.106062

Andrey K. Gorshenin , Anastasiia A. Osipova , Konstantin P. Belyaev

The extension of the use of Itô stochastic differential equations (SDEs) for joint analysis of spatio-temporal characteristics in the North Atlantic region, such as sea surface temperature (SST), the sum of sensible and latent heat fluxes, and surface atmospheric pressure for the period between 1979 and 2024 is introduced. Previously, this model was used only for the fluxes. The joint point estimates for the random coefficients of SDEs as multidimensional matrices (the drift vector and the diffusion matrix) are obtained for the entire considered period. The numerical estimations of these values were carried out using high-performance computing equipment with software implementation in Python language using the reanalysis data from the ERA5 database. Developed methods and tools are used for the statistical analysis of the temporal evolution of the coefficients of the Itô equation, analysis of joint and marginal diffusion matrices, their finite-dimensional Karhunen–Loéve’s decomposition into eigenvalues and eigenvectors, determination of their interrelations, temporal trends, as well as dynamic visualization on geographical maps of the region under study. The spatial structure of the eigenvectors of the diffusion matrix, their time evolution and the relationship to jet streams and large-scale heat waves that determine latitudinal heat transfer in the North Atlantic are shown. It is also demonstrated that there is a positive trend in the interannual variability in drift and diffusion coefficients. This indicates a quantitative and qualitative increase in the air–sea interaction and the relationship between heat fluxes and ocean surface temperature. It also makes it possible to quantify the energy exchange between the ocean and atmosphere on an interannual scale. The way of using quantities from a stochastic model to improve the neural network forecasts is also discussed.

介绍了Itô随机微分方程（SDEs）在1979 - 2024年北大西洋地区海温、感热通量和潜热通量和大气压力等时空特征联合分析中的推广应用。以前，该模型仅用于通量。在整个考虑周期内，得到了SDEs随机系数作为多维矩阵（漂移向量和扩散矩阵）的结合点估计。利用ERA5数据库的再分析数据，利用高性能计算设备和Python语言软件实现对这些值的数值估计。已开发的方法和工具用于统计分析Itô方程系数的时间演变，分析联合和边际扩散矩阵，分析它们的有限维karhunen - lo分解为特征值和特征向量，确定它们的相互关系，时间趋势，以及在研究区域的地理地图上动态可视化。给出了扩散矩阵特征向量的空间结构、时间演化及其与决定北大西洋纬向传热的急流和大尺度热浪的关系。漂移系数和扩散系数的年际变化也呈正趋势。这表明海气相互作用以及热通量与海洋表面温度之间的关系在数量和质量上都有所增加。这也使得在年际尺度上量化海洋和大气之间的能量交换成为可能。本文还讨论了利用随机模型中的量来改进神经网络预测的方法。

{"title":"Eigenvector decomposition for joint analysis of spatial characteristics in the North Atlantic from 1979 to 2024","authors":"Andrey K. Gorshenin , Anastasiia A. Osipova , Konstantin P. Belyaev","doi":"10.1016/j.cageo.2025.106062","DOIUrl":"10.1016/j.cageo.2025.106062","url":null,"abstract":"<div><div>The extension of the use of Itô stochastic differential equations (SDEs) for joint analysis of spatio-temporal characteristics in the North Atlantic region, such as sea surface temperature (SST), the sum of sensible and latent heat fluxes, and surface atmospheric pressure for the period between 1979 and 2024 is introduced. Previously, this model was used only for the fluxes. The joint point estimates for the random coefficients of SDEs as multidimensional matrices (the drift vector and the diffusion matrix) are obtained for the entire considered period. The numerical estimations of these values were carried out using high-performance computing equipment with software implementation in Python language using the reanalysis data from the ERA5 database. Developed methods and tools are used for the statistical analysis of the temporal evolution of the coefficients of the Itô equation, analysis of joint and marginal diffusion matrices, their finite-dimensional Karhunen–Loéve’s decomposition into eigenvalues and eigenvectors, determination of their interrelations, temporal trends, as well as dynamic visualization on geographical maps of the region under study. The spatial structure of the eigenvectors of the diffusion matrix, their time evolution and the relationship to jet streams and large-scale heat waves that determine latitudinal heat transfer in the North Atlantic are shown. It is also demonstrated that there is a positive trend in the interannual variability in drift and diffusion coefficients. This indicates a quantitative and qualitative increase in the air–sea interaction and the relationship between heat fluxes and ocean surface temperature. It also makes it possible to quantify the energy exchange between the ocean and atmosphere on an interannual scale. The way of using quantities from a stochastic model to improve the neural network forecasts is also discussed.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106062"},"PeriodicalIF":4.4,"publicationDate":"2025-10-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145269659","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

TorchTEM3D: PyTorch-Driven forward modeling platform for fast 3D transient electromagnetic modeling and efficient sensitivity matrix calculation TorchTEM3D: pytorch驱动的正演建模平台，用于快速3D瞬变电磁建模和高效灵敏度矩阵计算

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-10-04 DOI: 10.1016/j.cageo.2025.106063

Ziteng Li , Hai Li , Keying Li , Ahmed M. Beshr

The three-dimensional (3D) forward modeling of transient electromagnetic (TEM) data is often computationally demanding due to its high complexity and limited hardware acceleration, which also affects the efficiency of sensitivity matrix calculation. In recent years, deep learning frameworks, particularly PyTorch, have been widely used in various fields due to their high flexibility, parallel computing capabilities, and powerful automatic differentiation function. In this paper, we develop a time-domain finite-difference forward modeling platform for 3D TEM, named TorchTEM3D, based on the powerful parallel computing and GPU acceleration capabilities of PyTorch. By fully utilizing the automatic differentiation function of PyTorch, we achieve efficient and fast calculation of sensitivity matrix (the gradient of the electromagnetic response to the geoelectric model). Compared with existing open-source Python computing platforms such as SimPEG and custEM, our method improves computing speed by 15–60 times. Furthermore, high-precision sensitivity matrices can be obtained with a single forward modeling run.

瞬变电磁（TEM）数据的三维正演建模由于其高复杂性和有限的硬件加速，往往需要大量的计算量，这也影响了灵敏度矩阵计算的效率。近年来，深度学习框架，特别是PyTorch，由于其高灵活性、并行计算能力和强大的自动微分功能，在各个领域得到了广泛的应用。本文基于PyTorch强大的并行计算和GPU加速能力，开发了三维瞬变电磁法时域有限差分正演建模平台TorchTEM3D。充分利用PyTorch的自动微分功能，实现了灵敏度矩阵（电磁响应对地电模型的梯度）的高效快速计算。与现有的开源Python计算平台SimPEG和custEM相比，我们的方法将计算速度提高了15-60倍。此外，单次正演模拟可以获得高精度的灵敏度矩阵。

{"title":"TorchTEM3D: PyTorch-Driven forward modeling platform for fast 3D transient electromagnetic modeling and efficient sensitivity matrix calculation","authors":"Ziteng Li , Hai Li , Keying Li , Ahmed M. Beshr","doi":"10.1016/j.cageo.2025.106063","DOIUrl":"10.1016/j.cageo.2025.106063","url":null,"abstract":"<div><div>The three-dimensional (3D) forward modeling of transient electromagnetic (TEM) data is often computationally demanding due to its high complexity and limited hardware acceleration, which also affects the efficiency of sensitivity matrix calculation. In recent years, deep learning frameworks, particularly PyTorch, have been widely used in various fields due to their high flexibility, parallel computing capabilities, and powerful automatic differentiation function. In this paper, we develop a time-domain finite-difference forward modeling platform for 3D TEM, named TorchTEM3D, based on the powerful parallel computing and GPU acceleration capabilities of PyTorch. By fully utilizing the automatic differentiation function of PyTorch, we achieve efficient and fast calculation of sensitivity matrix (the gradient of the electromagnetic response to the geoelectric model). Compared with existing open-source Python computing platforms such as SimPEG and custEM, our method improves computing speed by 15–60 times. Furthermore, high-precision sensitivity matrices can be obtained with a single forward modeling run.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106063"},"PeriodicalIF":4.4,"publicationDate":"2025-10-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145269658","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Distance Transform Loss: Boundary-aware segmentation of seismic data 距离变换损失：地震数据的边界感知分割

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-09-30 DOI: 10.1016/j.cageo.2025.106061

Rafael Henrique Vareto , Ricardo Szczerbacki , Luiz A. Lima , Pedro O.S. Vaz-de-Melo , William Robson Schwartz

The segmentation of seismic data is a challenging exercise given the complexity and high variability of subsurface sources. This arduous task is effective in the identification of geological features, including facies classification, fault detection, and horizon interpretation. As a result, this work introduces a new cost function entitled Distance Transform Loss (DTL) that punishes deep networks when class boundaries are misclassified in exchange for more accurate contour delineations, an important aspect in the geological field. DTL consists of four key steps: contour detection, distance transform mapping, pixel-wise multiplication, and the summation of all grid elements. We conduct a comprehensive evaluation of deep convolutional architectures using publicly available seismic datasets, demonstrating that the proposed approach consistently enhances semantic segmentation performance. The results highlight DTL as a robust and architecture-agnostic loss function, capable of addressing class imbalance and boundary delineation challenges that commonly arise in seismic interpretation tasks.

考虑到地下震源的复杂性和高变异性，地震数据的分割是一项具有挑战性的工作。这项艰巨的任务有效地识别了地质特征，包括相分类、断层检测和层位解释。因此，这项工作引入了一个名为距离变换损失（DTL）的新成本函数，当类边界被错误分类时，它会惩罚深度网络，以换取更准确的等高线描绘，这是地质领域的一个重要方面。DTL包括四个关键步骤：轮廓检测、距离变换映射、逐像素乘法和所有网格元素的求和。我们使用公开可用的地震数据集对深度卷积架构进行了全面评估，证明了所提出的方法始终提高了语义分割性能。结果表明，DTL是一种鲁棒的、与体系结构无关的损失函数，能够解决地震解释任务中常见的类不平衡和边界描绘挑战。

{"title":"Distance Transform Loss: Boundary-aware segmentation of seismic data","authors":"Rafael Henrique Vareto , Ricardo Szczerbacki , Luiz A. Lima , Pedro O.S. Vaz-de-Melo , William Robson Schwartz","doi":"10.1016/j.cageo.2025.106061","DOIUrl":"10.1016/j.cageo.2025.106061","url":null,"abstract":"<div><div>The segmentation of seismic data is a challenging exercise given the complexity and high variability of subsurface sources. This arduous task is effective in the identification of geological features, including facies classification, fault detection, and horizon interpretation. As a result, this work introduces a new cost function entitled Distance Transform Loss (DTL) that punishes deep networks when class boundaries are misclassified in exchange for more accurate contour delineations, an important aspect in the geological field. DTL consists of four key steps: contour detection, distance transform mapping, pixel-wise multiplication, and the summation of all grid elements. We conduct a comprehensive evaluation of deep convolutional architectures using publicly available seismic datasets, demonstrating that the proposed approach consistently enhances semantic segmentation performance. The results highlight DTL as a robust and architecture-agnostic loss function, capable of addressing class imbalance and boundary delineation challenges that commonly arise in seismic interpretation tasks.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106061"},"PeriodicalIF":4.4,"publicationDate":"2025-09-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145269660","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Weakly supervised semantic segmentation of microscopic carbonates on marginal devices 微碳酸盐在边缘装置上的弱监督语义分割

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-09-27 DOI: 10.1016/j.cageo.2025.106059

Keran Li , Yujie Gao , Yingjie Ma , Chengkun Li , Junjie Ye , Hao Yu , Yiming Xu , Dongyu Zheng , Ardiansyah Koeshidayatullah

Microscopic analysis is the cornerstone to uncover petrological and mineralogical characteristics of carbonate rocks. In addition, such information is critical for precise identification of carbonate microfacies and diagenetic evolution. This type of information is important, but relies too much on manual experience, which is time-consuming and laborious. Recently, several successful deep learning models showed great potential in the identification process. However, current deep learning models have typically complex model architectures greatly hinder the deployment-inference in practical and lightweight environments. To overcome the difficulty of deep learning models in reasoning in actual edge scenes, a three-stage segmentation method by weakly supervised learning was proposed. The approach embeds class activation mapping (CAM), grey level co-occurrence matrix (GLCM), and knowledge distillation (KD) modules to achieve attention transfer to the lightweight network (CamNet). Furthermore, based on the performance of the model algorithm and application requirements, a lightweight carbonate thin section image-assistant recognition system has been developed. Through ingenious control flow design, this system achieves an effective balance between runtime latency and resource consumption, demonstrating superior performance metrics. Experimental results indicate that CamNet’s total parameter count is only 800k. When deployed in embedded systems, CamNet achieves an inference speed of 6.87 fps. Our successful development verifies the efficiency and practicality in marginal devices.

微观分析是揭示碳酸盐岩岩石矿物学特征的基石。此外，这些信息对于精确识别碳酸盐岩微相和成岩演化具有重要意义。这种类型的信息很重要，但过于依赖于人工经验，这既耗时又费力。最近，一些成功的深度学习模型在识别过程中显示出巨大的潜力。然而，当前的深度学习模型通常具有复杂的模型架构，这极大地阻碍了在实际和轻量级环境中的部署推理。为了克服深度学习模型在实际边缘场景中的推理困难，提出了一种基于弱监督学习的三阶段分割方法。该方法通过嵌入类激活映射（CAM）、灰度共生矩阵（GLCM）和知识蒸馏（KD）模块来实现对轻量级网络（CamNet）的注意力转移。在此基础上，根据模型算法的性能和应用需求，开发了轻质碳酸盐薄壁图像辅助识别系统。通过巧妙的控制流设计，该系统实现了运行时延迟和资源消耗之间的有效平衡，展示了卓越的性能指标。实验结果表明，CamNet的总参数数仅为800k。在嵌入式系统中部署时，CamNet的推理速度为6.87 fps。我们的成功开发验证了边际装置的效率和实用性。

{"title":"Weakly supervised semantic segmentation of microscopic carbonates on marginal devices","authors":"Keran Li , Yujie Gao , Yingjie Ma , Chengkun Li , Junjie Ye , Hao Yu , Yiming Xu , Dongyu Zheng , Ardiansyah Koeshidayatullah","doi":"10.1016/j.cageo.2025.106059","DOIUrl":"10.1016/j.cageo.2025.106059","url":null,"abstract":"<div><div>Microscopic analysis is the cornerstone to uncover petrological and mineralogical characteristics of carbonate rocks. In addition, such information is critical for precise identification of carbonate microfacies and diagenetic evolution. This type of information is important, but relies too much on manual experience, which is time-consuming and laborious. Recently, several successful deep learning models showed great potential in the identification process. However, current deep learning models have typically complex model architectures greatly hinder the deployment-inference in practical and lightweight environments. To overcome the difficulty of deep learning models in reasoning in actual edge scenes, a three-stage segmentation method by weakly supervised learning was proposed. The approach embeds class activation mapping (CAM), grey level co-occurrence matrix (GLCM), and knowledge distillation (KD) modules to achieve attention transfer to the lightweight network (CamNet). Furthermore, based on the performance of the model algorithm and application requirements, a lightweight carbonate thin section image-assistant recognition system has been developed. Through ingenious control flow design, this system achieves an effective balance between runtime latency and resource consumption, demonstrating superior performance metrics. Experimental results indicate that CamNet’s total parameter count is only 800k. When deployed in embedded systems, CamNet achieves an inference speed of 6.87 fps. Our successful development verifies the efficiency and practicality in marginal devices.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106059"},"PeriodicalIF":4.4,"publicationDate":"2025-09-27","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145269657","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

GeoFedNet: Federated learning for privacy-aware, robust, and generalizable seismic interpretation GeoFedNet：用于隐私感知、鲁棒和通用地震解释的联邦学习

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-09-26 DOI: 10.1016/j.cageo.2025.106060

Muhammad Saif ul Islam, Aamir Wali

Seismic structural interpretation is crucial for understanding subsurface geology, particularly in hydrocarbon exploration, as it aids in identifying reservoir formations, assessing drilling risks, and optimizing resource extraction. However, developing a widely generalizable model for seismic interpretation remains challenging due to the limited availability of large-scale public datasets, variations in seismic surveys, and privacy constraints that hinder data sharing. These factors lead to inconsistencies in model performance across diverse datasets, limiting the applicability of existing approaches. To address this gap, we propose a federated learning-based framework for seismic interpretation, enabling distributed model training without requiring direct data sharing. In this approach, local models are trained independently across different clients, and a global model is aggregated to improve generalization across heterogeneous datasets. This method not only preserves data confidentiality but also mitigates challenges related to labeled data scarcity and class imbalance, allowing clients with limited data to benefit from collaborative learning. We evaluate GeoFedNet on three key seismic interpretation tasks: seismic structure classification, salt detection, and facies segmentation. Across all tasks, GeoFedNet achieves performance within 1%–3% of centralized models while significantly outperforming isolated local models by up to 15% in accuracy and generalization. These results demonstrate that our framework can effectively learn from non-IID and imbalanced data without compromising performance. GeoFedNet also shows improved robustness to client variability and better minority class recognition, which are critical in real-world subsurface interpretation scenarios. These findings highlight the potential of federated learning in enabling hydrocarbon companies to collaboratively train robust seismic interpretation models while maintaining data privacy, ultimately improving exploration efficiency and informed decision-making.

地震构造解释对于了解地下地质至关重要，特别是在油气勘探中，因为它有助于识别储层、评估钻井风险和优化资源开采。然而，由于大规模公共数据集的可用性有限，地震调查的变化，以及阻碍数据共享的隐私限制，开发一个广泛推广的地震解释模型仍然具有挑战性。这些因素导致不同数据集的模型性能不一致，限制了现有方法的适用性。为了解决这一差距，我们提出了一个基于联邦学习的地震解释框架，实现分布式模型训练，而不需要直接共享数据。在这种方法中，局部模型在不同的客户端上独立训练，全局模型被聚合以提高跨异构数据集的泛化。这种方法不仅保护了数据的机密性，还减轻了与标记数据稀缺和类别不平衡相关的挑战，使数据有限的客户能够从协作学习中受益。我们在三个关键的地震解释任务上对GeoFedNet进行了评估：地震构造分类、盐检测和相分割。在所有任务中，GeoFedNet在集中式模型的1%-3%内实现了性能，而在精度和泛化方面明显优于孤立的局部模型，最高可达15%。这些结果表明，我们的框架可以在不影响性能的情况下有效地从非iid和不平衡数据中学习。GeoFedNet还显示了对客户端可变性的更好的鲁棒性和更好的少数类识别，这在真实的地下解释场景中至关重要。这些发现突出了联合学习的潜力，使油气公司能够在保持数据隐私的同时协同训练强大的地震解释模型，最终提高勘探效率和明智的决策。

{"title":"GeoFedNet: Federated learning for privacy-aware, robust, and generalizable seismic interpretation","authors":"Muhammad Saif ul Islam, Aamir Wali","doi":"10.1016/j.cageo.2025.106060","DOIUrl":"10.1016/j.cageo.2025.106060","url":null,"abstract":"<div><div>Seismic structural interpretation is crucial for understanding subsurface geology, particularly in hydrocarbon exploration, as it aids in identifying reservoir formations, assessing drilling risks, and optimizing resource extraction. However, developing a widely generalizable model for seismic interpretation remains challenging due to the limited availability of large-scale public datasets, variations in seismic surveys, and privacy constraints that hinder data sharing. These factors lead to inconsistencies in model performance across diverse datasets, limiting the applicability of existing approaches. To address this gap, we propose a federated learning-based framework for seismic interpretation, enabling distributed model training without requiring direct data sharing. In this approach, local models are trained independently across different clients, and a global model is aggregated to improve generalization across heterogeneous datasets. This method not only preserves data confidentiality but also mitigates challenges related to labeled data scarcity and class imbalance, allowing clients with limited data to benefit from collaborative learning. We evaluate GeoFedNet on three key seismic interpretation tasks: seismic structure classification, salt detection, and facies segmentation. Across all tasks, GeoFedNet achieves performance within 1%–3% of centralized models while significantly outperforming isolated local models by up to 15% in accuracy and generalization. These results demonstrate that our framework can effectively learn from non-IID and imbalanced data without compromising performance. GeoFedNet also shows improved robustness to client variability and better minority class recognition, which are critical in real-world subsurface interpretation scenarios. These findings highlight the potential of federated learning in enabling hydrocarbon companies to collaboratively train robust seismic interpretation models while maintaining data privacy, ultimately improving exploration efficiency and informed decision-making.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106060"},"PeriodicalIF":4.4,"publicationDate":"2025-09-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145222456","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Neural network-based framework for signal separation in spatio-temporal gravity data 基于神经网络的时空重力数据信号分离框架

IF 4.4 2区地球科学 Q1 COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS

Computers & Geosciences

Pub Date : 2025-09-23 DOI: 10.1016/j.cageo.2025.106057

Betty Heller-Kaikov, Roland Pail, Martin Werner

Global, temporal gravity data such as those provided by the Gravity Recovery and Climate Experiment (GRACE) and GRACE-Follow on (GRACE-FO) satellite missions contain signals from many mass redistribution processes on Earth. These include hydrological, atmospheric, oceanic, cryospheric and solid Earth-related processes. As the measured gravity changes represent the sum of all signals, an optimal exploitation of these data for scientific applications requires strategies for separating the individual contained signals. We provide a neural network algorithm using a multi-channel U-Net architecture that translates the sum of several signals to the individual contained components based on their typical space–time patterns. The software contains strategies for transforming spatio-temporal gravity data depending on latitude, longitude, and time to 2-D “image” training samples. The software also includes implementations of strategies for introducing additional knowledge about the physical behavior of the individual signals as constraints to the training. In a closed-loop simulation example, simulated gravity signals induced by processes in the atmosphere and oceans, hydrosphere, cryosphere and solid Earth are successfully separated at relative RMS prediction errors between 19 and 67%. This shows that neural network-based methods can help solving geodetic tasks if the considered data is transformed into a suitable data format. To apply the framework to real observational data, we suggest training the network on representative, physical forward-modeled signals and subsequently applying the trained network to real data. The latter will additionally require external validation strategies. The software is freely available on GitHub under https://github.com/Betty-Heller/neural-gravity and is, in general, also applicable for signal separation in any other dataset depending on three variables.

由重力恢复和气候实验（GRACE）和GRACE- follow on （GRACE- fo）卫星任务提供的全球时间重力数据包含了地球上许多质量再分配过程的信号。这些过程包括水文、大气、海洋、冰冻圈和与固体地球有关的过程。由于测量到的重力变化代表了所有信号的总和，为了科学应用的最佳利用这些数据，需要分离单个包含信号的策略。我们提供了一种使用多通道U-Net架构的神经网络算法，该算法将几个信号的总和转换为基于其典型时空模式的单个包含分量。该软件包含根据纬度、经度和时间将时空重力数据转换为二维“图像”训练样本的策略。该软件还包括用于引入关于单个信号的物理行为的附加知识作为训练约束的策略的实现。在闭环模拟实例中，大气和海洋、水圈、冰冻圈和固体地球过程诱导的模拟重力信号成功分离，相对RMS预测误差在19% ~ 67%之间。这表明，如果将考虑的数据转换为合适的数据格式，基于神经网络的方法可以帮助解决大地测量任务。为了将该框架应用于实际观测数据，我们建议在具有代表性的物理正演模拟信号上训练网络，然后将训练好的网络应用于实际数据。后者将额外需要外部验证策略。该软件在GitHub （https://github.com/Betty-Heller/neural-gravity）上免费提供，通常也适用于任何其他数据集的信号分离，具体取决于三个变量。

{"title":"Neural network-based framework for signal separation in spatio-temporal gravity data","authors":"Betty Heller-Kaikov, Roland Pail, Martin Werner","doi":"10.1016/j.cageo.2025.106057","DOIUrl":"10.1016/j.cageo.2025.106057","url":null,"abstract":"<div><div>Global, temporal gravity data such as those provided by the Gravity Recovery and Climate Experiment (GRACE) and GRACE-Follow on (GRACE-FO) satellite missions contain signals from many mass redistribution processes on Earth. These include hydrological, atmospheric, oceanic, cryospheric and solid Earth-related processes. As the measured gravity changes represent the sum of all signals, an optimal exploitation of these data for scientific applications requires strategies for separating the individual contained signals. We provide a neural network algorithm using a multi-channel U-Net architecture that translates the sum of several signals to the individual contained components based on their typical space–time patterns. The software contains strategies for transforming spatio-temporal gravity data depending on latitude, longitude, and time to 2-D “image” training samples. The software also includes implementations of strategies for introducing additional knowledge about the physical behavior of the individual signals as constraints to the training. In a closed-loop simulation example, simulated gravity signals induced by processes in the atmosphere and oceans, hydrosphere, cryosphere and solid Earth are successfully separated at relative RMS prediction errors between 19 and 67%. This shows that neural network-based methods can help solving geodetic tasks if the considered data is transformed into a suitable data format. To apply the framework to real observational data, we suggest training the network on representative, physical forward-modeled signals and subsequently applying the trained network to real data. The latter will additionally require external validation strategies. The software is freely available on GitHub under <span><span>https://github.com/Betty-Heller/neural-gravity</span><svg><path></path></svg></span> and is, in general, also applicable for signal separation in any other dataset depending on three variables.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"207 ","pages":"Article 106057"},"PeriodicalIF":4.4,"publicationDate":"2025-09-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"145160178","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"地球科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0