Pub Date : 2025-07-11DOI: 10.1016/j.cageo.2025.106011
Jundong He , Weirong Yang , Zhengbo Yu , Cheng Tan , Binbin Li
Geochemical anomalies are critical indicators for mineral exploration and resource evaluation. However, due to the diversity and complexity of geological processes, identifying geochemical anomalies remains challenging. This study proposes a dual-branch convolutional neural network based on adaptive feature fusion (1-2D AFFCNN) to simultaneously extract the spectral compositional relationships and spatial structural features of geochemical elements. The model incorporates an Adaptive Feature Fusion Module (AFFM) to effectively integrate features from different branches, significantly improving predictive performance and robustness. Experimental results demonstrate that the 1-2D AFFCNN outperforms traditional single models in terms of accuracy (92.3 %), recall (92.0 %), and AUC value (0.98). The three-stage training strategy effectively mitigates the vanishing gradient problem, enhancing training efficiency and stability. In the application to the Changba ore-concentrated area in Gansu Province, the high-probability anomaly zones generated by the model are highly consistent with the spatial distribution of known lead-zinc deposits, and several high-potential mineralization areas were identified. This study not only provides a novel approach for the comprehensive analysis of multidimensional geochemical data but also opens new avenues for mineral resource prediction and target area localization.
{"title":"Recognition of multiple geochemical anomalies by dual-branch convolutional neural network with adaptive feature fusion","authors":"Jundong He , Weirong Yang , Zhengbo Yu , Cheng Tan , Binbin Li","doi":"10.1016/j.cageo.2025.106011","DOIUrl":"10.1016/j.cageo.2025.106011","url":null,"abstract":"<div><div>Geochemical anomalies are critical indicators for mineral exploration and resource evaluation. However, due to the diversity and complexity of geological processes, identifying geochemical anomalies remains challenging. This study proposes a dual-branch convolutional neural network based on adaptive feature fusion (1-2D AFFCNN) to simultaneously extract the spectral compositional relationships and spatial structural features of geochemical elements. The model incorporates an Adaptive Feature Fusion Module (AFFM) to effectively integrate features from different branches, significantly improving predictive performance and robustness. Experimental results demonstrate that the 1-2D AFFCNN outperforms traditional single models in terms of accuracy (92.3 %), recall (92.0 %), and AUC value (0.98). The three-stage training strategy effectively mitigates the vanishing gradient problem, enhancing training efficiency and stability. In the application to the Changba ore-concentrated area in Gansu Province, the high-probability anomaly zones generated by the model are highly consistent with the spatial distribution of known lead-zinc deposits, and several high-potential mineralization areas were identified. This study not only provides a novel approach for the comprehensive analysis of multidimensional geochemical data but also opens new avenues for mineral resource prediction and target area localization.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"205 ","pages":"Article 106011"},"PeriodicalIF":4.2,"publicationDate":"2025-07-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144632361","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}
Accurate and timely correction of numerically forecasted sea surface wind fields is essential for sea ice drift prediction. However, current oceanic element prediction systems face two major challenges. The numerically forecasted sea surface wind fields are timely, but their accuracy is often limited. In contrast, reanalysis sea surface wind fields are more accurate but lack timeliness, limiting their applicability in urgent requirements. To address these challenges, a self-attention convolutional long and short-term memory network (SaCLN) has been developed for intelligently correcting the numerically forecasted sea surface wind fields. This approach combines the timeliness of the numerically forecasted wind fields with the accuracy of reanalysis wind fields to generate corrected wind fields that closely approximate the reanalysis wind fields. This network consists of a self-attention network and a convolutional long and short-term memory network (CLN). The self-attention network captures the global spatial correlations of a numerically forecasted sea surface wind field sequence. The CLN extracts the spatial and temporal characteristics of an attention weighted wind field sequence. The trained SaCLN model can effectively generate accurate and timely corrected wind fields, thereby enhancing the accuracy of sea ice drift prediction. The effectiveness of the SaCLN was validated through experiments predicting the drift of Arctic sea ice and Antarctic icebergs. Experimental results show that the drift results based on wind fields corrected by the SaCLN are more accurate than those based on numerically forecasted sea surface wind fields. This method has demonstrated its effectiveness in sea ice drift prediction, assisting researchers in better addressing the challenges posed by sea ice variability.
{"title":"A self-attention convolutional long and short-term memory network for correcting sea surface wind field forecasts to facilitate sea ice drift prediction","authors":"Qing Xu , Qilin Jia , Yongqing Li , Hao Zhang , Peng Ren","doi":"10.1016/j.cageo.2025.105997","DOIUrl":"10.1016/j.cageo.2025.105997","url":null,"abstract":"<div><div>Accurate and timely correction of numerically forecasted sea surface wind fields is essential for sea ice drift prediction. However, current oceanic element prediction systems face two major challenges. The numerically forecasted sea surface wind fields are timely, but their accuracy is often limited. In contrast, reanalysis sea surface wind fields are more accurate but lack timeliness, limiting their applicability in urgent requirements. To address these challenges, a self-attention convolutional long and short-term memory network (SaCLN) has been developed for intelligently correcting the numerically forecasted sea surface wind fields. This approach combines the timeliness of the numerically forecasted wind fields with the accuracy of reanalysis wind fields to generate corrected wind fields that closely approximate the reanalysis wind fields. This network consists of a self-attention network and a convolutional long and short-term memory network (CLN). The self-attention network captures the global spatial correlations of a numerically forecasted sea surface wind field sequence. The CLN extracts the spatial and temporal characteristics of an attention weighted wind field sequence. The trained SaCLN model can effectively generate accurate and timely corrected wind fields, thereby enhancing the accuracy of sea ice drift prediction. The effectiveness of the SaCLN was validated through experiments predicting the drift of Arctic sea ice and Antarctic icebergs. Experimental results show that the drift results based on wind fields corrected by the SaCLN are more accurate than those based on numerically forecasted sea surface wind fields. This method has demonstrated its effectiveness in sea ice drift prediction, assisting researchers in better addressing the challenges posed by sea ice variability.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"205 ","pages":"Article 105997"},"PeriodicalIF":4.2,"publicationDate":"2025-07-08","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144588365","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}
Pub Date : 2025-07-07DOI: 10.1016/j.cageo.2025.105994
Ivan Javier Sánchez-Galvis , Herling Gonzalez-Alvarez , William Agudelo , Daniel O. Trad , Daniel A. Sierra
Simulating seismic wave propagation in complex geological structures is a challenging task in exploration geophysics, especially in foothill regions characterized by rough topography, irregular bedrock interfaces, low-velocity surface sediments, and significant heterogeneities. Although existing numerical methods can address such scenarios, they often require highly refined grids that lead to elevated computational costs. To address this, we introduce ElasWave3D, a three-dimensional solver based on the finite difference method for elastic wave propagation in the presence of irregular topography, specifically designed for GPU acceleration. The solver employs a novel Irregular Subdomain Index Array (ISIA) strategy to implement the parameter-modified (PM) formulation, thus enforcing the free-surface condition for arbitrary topographic variations. We validated ElasWave3D against the well-known SPECFEM3D solver in scenarios with rough topography and heterogeneous media, observing misfit errors below 1% and correlation values exceeding 99% in most cases. Additionally, our solver achieves more than an order-of-magnitude speedup (13×) over its CPU-OpenMP implementation on 24 cores. Consequently, ElasWave3D enables cost-effective, realistic, and detailed simulations of near-surface seismic scattering in heterogeneous Earth models with irregular topography.
{"title":"ElasWave3D: A GPU-accelerated 3D finite-difference elastic wave solver for complex topography using irregular subdomain index arrays","authors":"Ivan Javier Sánchez-Galvis , Herling Gonzalez-Alvarez , William Agudelo , Daniel O. Trad , Daniel A. Sierra","doi":"10.1016/j.cageo.2025.105994","DOIUrl":"10.1016/j.cageo.2025.105994","url":null,"abstract":"<div><div>Simulating seismic wave propagation in complex geological structures is a challenging task in exploration geophysics, especially in foothill regions characterized by rough topography, irregular bedrock interfaces, low-velocity surface sediments, and significant heterogeneities. Although existing numerical methods can address such scenarios, they often require highly refined grids that lead to elevated computational costs. To address this, we introduce ElasWave3D, a three-dimensional solver based on the finite difference method for elastic wave propagation in the presence of irregular topography, specifically designed for GPU acceleration. The solver employs a novel Irregular Subdomain Index Array (ISIA) strategy to implement the parameter-modified (PM) formulation, thus enforcing the free-surface condition for arbitrary topographic variations. We validated ElasWave3D against the well-known SPECFEM3D solver in scenarios with rough topography and heterogeneous media, observing misfit errors below 1% and correlation values exceeding 99% in most cases. Additionally, our solver achieves more than an order-of-magnitude speedup (13×) over its CPU-OpenMP implementation on 24 cores. Consequently, ElasWave3D enables cost-effective, realistic, and detailed simulations of near-surface seismic scattering in heterogeneous Earth models with irregular topography.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"205 ","pages":"Article 105994"},"PeriodicalIF":4.2,"publicationDate":"2025-07-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144588363","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}
Pub Date : 2025-07-05DOI: 10.1016/j.cageo.2025.105989
Stella V. Paronuzzi-Ticco , Gladys Utrera , Mario C. Acosta
Driven by the need to improve computational efficiency, the technique of reducing variable precision in model calculations has recently attracted a lot of attention, particularly in the field of weather and climate simulations models, where computational gains are crucial to produce operational results faster and make better use of HPC resources.
However, the source of computational improvements resulting from working in reduced precision, an aspect that could help facilitate the transition in many applications, has never been thoroughly explained. In this paper, we make a step in this direction, shedding light on how to efficiently apply variable precision reduction in chaotic applications, and presenting a computational study methodology to make this possible.
For this purpose, we employ a tool for automatic porting of oceanographic code to mixed precision recently developed at the Barcelona Supercomputing Center and consider as case studies one of the most widely employed ocean models, NEMO, in one of the most ambitious initiatives to date, Destination Earth, because it aims at creating interactive digital replicas of the Earth with unprecedented precision, supporting real-time decision-making and long-term adaptation strategies, which also entails an unprecedented computational cost in terms of supercomputing. We analyze in depth the impact of mixed precision on the most representative functions of the model, providing a clear step forward in understanding where to focus efforts in precision reduction. These results can guide scientists in significantly speeding up weather and climate models using mixed precision by targeting computationally intensive functions and optimizing communications.
{"title":"Efficient variable precision reduction in chaotic climate models: Analysis of the NEMO case in the destination earth project","authors":"Stella V. Paronuzzi-Ticco , Gladys Utrera , Mario C. Acosta","doi":"10.1016/j.cageo.2025.105989","DOIUrl":"10.1016/j.cageo.2025.105989","url":null,"abstract":"<div><div>Driven by the need to improve computational efficiency, the technique of reducing variable precision in model calculations has recently attracted a lot of attention, particularly in the field of weather and climate simulations models, where computational gains are crucial to produce operational results faster and make better use of HPC resources.</div><div>However, the source of computational improvements resulting from working in reduced precision, an aspect that could help facilitate the transition in many applications, has never been thoroughly explained. In this paper, we make a step in this direction, shedding light on how to efficiently apply variable precision reduction in chaotic applications, and presenting a computational study methodology to make this possible.</div><div>For this purpose, we employ a tool for automatic porting of oceanographic code to mixed precision recently developed at the Barcelona Supercomputing Center and consider as case studies one of the most widely employed ocean models, NEMO, in one of the most ambitious initiatives to date, Destination Earth, because it aims at creating interactive digital replicas of the Earth with unprecedented precision, supporting real-time decision-making and long-term adaptation strategies, which also entails an unprecedented computational cost in terms of supercomputing. We analyze in depth the impact of mixed precision on the most representative functions of the model, providing a clear step forward in understanding where to focus efforts in precision reduction. These results can guide scientists in significantly speeding up weather and climate models using mixed precision by targeting computationally intensive functions and optimizing communications.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"205 ","pages":"Article 105989"},"PeriodicalIF":4.2,"publicationDate":"2025-07-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144580570","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}
Pub Date : 2025-07-05DOI: 10.1016/j.cageo.2025.106001
Michele Delchiaro , Valeria Ruscitto , Wolfgang Schwanghart , Eleonora Brignone , Daniela Piacentini , Francesco Troiani
Understanding river channel bankfull geometry is crucial for fluvial monitoring and flood prediction. The bankfull stage, typically reached every 1–2 years, marks when water spills onto the floodplain and strongly influences channel morphology. In our study, we present a novel approach for detecting river channel bankfull levels, utilizing a specialized MATLAB tool we developed, called BankfullMapper. The tool divides rivers into evenly spaced sections and computes a hydraulic depth function, plotting elevation above the thalweg against the area-to-width ratio. Bankfull levels are identified through (i) the lowest breakpoints from the thalweg or (ii) the most prominent breakpoints. Using Manning’s equation, the tool also estimates bankfull discharge.
We applied the method to two Italian rivers with contrasting hydrological settings: the single-channel Potenza River and the braided-to-wandering Marecchia River. Potenza was used for checking the tool's spatial analysis capability, while Marecchia served for spatio-temporal testing (2009 vs. 2022). Modelled bankfull extents were validated against expert-mapped active channel polygons using accuracy, precision, sensitivity, and specificity metrics.
For Potenza, bankfull discharges (33.9–52 m3 s⁻1) closely matched gauge data (2010–2023) using Gumbel distribution. The method showed high accuracy (0.90–0.92), sensitivity (0.94–0.95), and specificity (0.89–0.92), with moderate precision (0.53–0.61). For Marecchia, sensitivity ranged from 0.63 to 0.92, specificity from 0.73 to 0.89, accuracy from 0.80 to 0.83, and precision from 0.56 to 0.65.
Overall, the semi-automated approach reliably captures spatial and temporal changes in bankfull geometry and discharge across diverse river systems. It performs best using the lowest morphological breakpoints and offers a robust, detailed tool for hydrological research and river management.
{"title":"BankfullMapper: a semi-automated MATLAB tool on high-resolution digital terrain models for spatio-temporal monitoring of bankfull geometry and discharge","authors":"Michele Delchiaro , Valeria Ruscitto , Wolfgang Schwanghart , Eleonora Brignone , Daniela Piacentini , Francesco Troiani","doi":"10.1016/j.cageo.2025.106001","DOIUrl":"10.1016/j.cageo.2025.106001","url":null,"abstract":"<div><div>Understanding river channel bankfull geometry is crucial for fluvial monitoring and flood prediction. The bankfull stage, typically reached every 1–2 years, marks when water spills onto the floodplain and strongly influences channel morphology. In our study, we present a novel approach for detecting river channel bankfull levels, utilizing a specialized MATLAB tool we developed, called BankfullMapper. The tool divides rivers into evenly spaced sections and computes a hydraulic depth function, plotting elevation above the thalweg against the area-to-width ratio. Bankfull levels are identified through (i) the lowest breakpoints from the thalweg or (ii) the most prominent breakpoints. Using Manning’s equation, the tool also estimates bankfull discharge.</div><div>We applied the method to two Italian rivers with contrasting hydrological settings: the single-channel Potenza River and the braided-to-wandering Marecchia River. Potenza was used for checking the tool's spatial analysis capability, while Marecchia served for spatio-temporal testing (2009 vs. 2022). Modelled bankfull extents were validated against expert-mapped active channel polygons using accuracy, precision, sensitivity, and specificity metrics.</div><div>For Potenza, bankfull discharges (33.9–52 m<sup>3</sup> s⁻<sup>1</sup>) closely matched gauge data (2010–2023) using Gumbel distribution. The method showed high accuracy (0.90–0.92), sensitivity (0.94–0.95), and specificity (0.89–0.92), with moderate precision (0.53–0.61). For Marecchia, sensitivity ranged from 0.63 to 0.92, specificity from 0.73 to 0.89, accuracy from 0.80 to 0.83, and precision from 0.56 to 0.65.</div><div>Overall, the semi-automated approach reliably captures spatial and temporal changes in bankfull geometry and discharge across diverse river systems. It performs best using the lowest morphological breakpoints and offers a robust, detailed tool for hydrological research and river management.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"205 ","pages":"Article 106001"},"PeriodicalIF":4.2,"publicationDate":"2025-07-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144632302","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}
Pub Date : 2025-07-01DOI: 10.1016/j.cageo.2025.106000
Ruiyou Li, Yong Zhang, Jiayi Ju, Rongqiang Liu
The transient electromagnetic method (TEM) is a widely used geophysical technique for investigating complex geological conditions. Deep learning (DL) provides a novel approach for solving the complex, nonlinear TEM inversion problem. However, most current DL inversion methods for TEM survey data depend heavily on labeled data (real resistivity models), which are difficult to acquire from field surveys. In this study, we propose an unsupervised DL inversion method for TEM survey data based on the physical laws that govern electric field propagation. First, we integrate forward modeling into the training process, allowing the predicted resistivity model to be converted into simulated data. This simulated data is then compared with observed data to calculate a data misfit. Then, unsupervised training (label-independent) is achieved using the data misfit as the loss function, with dynamic smoothing constraints employed to alleviate the ill-posed inversion problem. Furthermore, the DL network incorporates an Attention mechanism to extract crucial feature information for TEM inversion. Finally, the multivariate variational mode decomposition (MVMD) technique optimized by the whale optimization algorithm (WOA) is adopted to reduce noise in the survey data and enhance TEM inversion precision. Both synthetic examples and field surveys show that our proposed approach accurately delineates subsurface model structures, offering an innovative solution for TEM inversion.
{"title":"Physics-embedded deep learning inversion for transient electromagnetic method survey data","authors":"Ruiyou Li, Yong Zhang, Jiayi Ju, Rongqiang Liu","doi":"10.1016/j.cageo.2025.106000","DOIUrl":"10.1016/j.cageo.2025.106000","url":null,"abstract":"<div><div>The transient electromagnetic method (TEM) is a widely used geophysical technique for investigating complex geological conditions. Deep learning (DL) provides a novel approach for solving the complex, nonlinear TEM inversion problem. However, most current DL inversion methods for TEM survey data depend heavily on labeled data (real resistivity models), which are difficult to acquire from field surveys. In this study, we propose an unsupervised DL inversion method for TEM survey data based on the physical laws that govern electric field propagation. First, we integrate forward modeling into the training process, allowing the predicted resistivity model to be converted into simulated data. This simulated data is then compared with observed data to calculate a data misfit. Then, unsupervised training (label-independent) is achieved using the data misfit as the loss function, with dynamic smoothing constraints employed to alleviate the ill-posed inversion problem. Furthermore, the DL network incorporates an Attention mechanism to extract crucial feature information for TEM inversion. Finally, the multivariate variational mode decomposition (MVMD) technique optimized by the whale optimization algorithm (WOA) is adopted to reduce noise in the survey data and enhance TEM inversion precision. Both synthetic examples and field surveys show that our proposed approach accurately delineates subsurface model structures, offering an innovative solution for TEM inversion.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"204 ","pages":"Article 106000"},"PeriodicalIF":4.2,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144549666","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}
Pub Date : 2025-06-25DOI: 10.1016/j.cageo.2025.105996
Wenyao Fan , Leonardo Azevedo , Gang Liu , Qiyu Chen , Xuechao Wu , Yang Li
The spatiotemporal evolution of sedimentary bodies is difficult to model with traditional geological modeling tools due to its non-stationarity nature. Deep learning algorithms, based on Convolutional Long-Short Term Memory (ConvLSTM) networks, allow to alleviate these limitations as the spatial and temporal dynamics of the sedimentary environment can be explicitly modeled, with structural and attribute information being constructed layer-by-layer. However, due to memory flow limitations and hierarchical visual representations of ConvLSTM, both low-level and high-level semantic features cannot be simultaneously captured. Consequently, small-scale geological features are often overlooked. In addition, long-term modeling and predicting capabilities of ConvLSTM are insufficient during geological sections encoding and forecasting processes. All these challenges might impact the application of ConvLSTM for geo-modeling. To overcome these limitations, we propose herein a geological modeling Recurrent Neural Network (GM-RNN) framework. Specifically, we use zigzag transition path of spatiotemporal memory flow, which allow spatial dynamics at different recurrent layers to interact with each other. Besides, Spatiotemporal LSTM (ST-LSTM) units with memory decoupling are introduced, in which long-term and short-term modeling capabilities for complex spatiotemporal variations can be improved. Finally, Reverse Schedule Sampling (RSS) strategies are used to improve the long-term prediction performances of GM-RNN. Two kinds of Training Images (TIs) are used to assess the simulation performance of GM-RNN. Numerical experiments show that diverse simulations match the corresponding TI in terms of spatial variability, channel connectivity, facies type proportion and spatial distribution patterns. Additionally, we show that 2D geological sections with different scales can be the input of a trained GM-RNN and geobodies are predicted at these scales without compromising the quality of the models.
{"title":"Automatic reconstruction of 3D geological models based on recurrent neural network and predictive learning","authors":"Wenyao Fan , Leonardo Azevedo , Gang Liu , Qiyu Chen , Xuechao Wu , Yang Li","doi":"10.1016/j.cageo.2025.105996","DOIUrl":"10.1016/j.cageo.2025.105996","url":null,"abstract":"<div><div>The spatiotemporal evolution of sedimentary bodies is difficult to model with traditional geological modeling tools due to its non-stationarity nature. Deep learning algorithms, based on Convolutional Long-Short Term Memory (ConvLSTM) networks, allow to alleviate these limitations as the spatial and temporal dynamics of the sedimentary environment can be explicitly modeled, with structural and attribute information being constructed layer-by-layer. However, due to memory flow limitations and hierarchical visual representations of ConvLSTM, both low-level and high-level semantic features cannot be simultaneously captured. Consequently, small-scale geological features are often overlooked. In addition, long-term modeling and predicting capabilities of ConvLSTM are insufficient during geological sections encoding and forecasting processes. All these challenges might impact the application of ConvLSTM for geo-modeling. To overcome these limitations, we propose herein a geological modeling Recurrent Neural Network (GM-RNN) framework. Specifically, we use zigzag transition path of spatiotemporal memory flow, which allow spatial dynamics at different recurrent layers to interact with each other. Besides, Spatiotemporal LSTM (ST-LSTM) units with memory decoupling are introduced, in which long-term and short-term modeling capabilities for complex spatiotemporal variations can be improved. Finally, Reverse Schedule Sampling (RSS) strategies are used to improve the long-term prediction performances of GM-RNN. Two kinds of Training Images (TIs) are used to assess the simulation performance of GM-RNN. Numerical experiments show that diverse simulations match the corresponding TI in terms of spatial variability, channel connectivity, facies type proportion and spatial distribution patterns. Additionally, we show that 2D geological sections with different scales can be the input of a trained GM-RNN and geobodies are predicted at these scales without compromising the quality of the models.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"204 ","pages":"Article 105996"},"PeriodicalIF":4.2,"publicationDate":"2025-06-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144534725","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}
Pub Date : 2025-06-22DOI: 10.1016/j.cageo.2025.105995
Jing Chen , Masaru Nagaso , Mijian Xu , Ping Tong
TomoATT is an open-source software package, aiming at determining seismic velocity and azimuthal anisotropy based on adjoint-state traveltime tomography methods. Key features of TomoATT include Eikonal equation modeling, adjoint-state method, sensitivity kernel regularization, and multi-level parallelization. Through several toy experiments, we demonstrate TomoATT's capability in accurate forward modeling, handling multipathing phenomenon, delivering reliable tomographic results, and achieving high-performance parallelization. Additionally, TomoATT is benchmarked with a synthetic experiment and two real-data applications in central California near Parkfield and Thailand. The successful recovery of the synthetic model, along with the imaging results that are consistent with previous studies and regional tectonics, verifies the effectiveness of TomoATT. Each inversion starts with only three simple input files (about model, data, and parameters) and completes within 2 h using 64 processors. Overall, TomoATT offers an efficient and user-friendly tool for regional and teleseismic traveltime tomography, empowering researchers to image subsurface structures and deepen our understanding of the Earth's interior.
{"title":"TomoATT: An open-source package for Eikonal equation-based adjoint-state traveltime tomography for seismic velocity and azimuthal anisotropy","authors":"Jing Chen , Masaru Nagaso , Mijian Xu , Ping Tong","doi":"10.1016/j.cageo.2025.105995","DOIUrl":"10.1016/j.cageo.2025.105995","url":null,"abstract":"<div><div>TomoATT is an open-source software package, aiming at determining seismic velocity and azimuthal anisotropy based on adjoint-state traveltime tomography methods. Key features of TomoATT include Eikonal equation modeling, adjoint-state method, sensitivity kernel regularization, and multi-level parallelization. Through several toy experiments, we demonstrate TomoATT's capability in accurate forward modeling, handling multipathing phenomenon, delivering reliable tomographic results, and achieving high-performance parallelization. Additionally, TomoATT is benchmarked with a synthetic experiment and two real-data applications in central California near Parkfield and Thailand. The successful recovery of the synthetic model, along with the imaging results that are consistent with previous studies and regional tectonics, verifies the effectiveness of TomoATT. Each inversion starts with only three simple input files (about model, data, and parameters) and completes within 2 h using 64 processors. Overall, TomoATT offers an efficient and user-friendly tool for regional and teleseismic traveltime tomography, empowering researchers to image subsurface structures and deepen our understanding of the Earth's interior.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"204 ","pages":"Article 105995"},"PeriodicalIF":4.2,"publicationDate":"2025-06-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144481586","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}
Pub Date : 2025-06-18DOI: 10.1016/j.cageo.2025.105965
Adrian Baddeley , Warick Brown , Gopalan Nair , Robin Milne , Suman Rakshit , Shih Ching Fu
In mineral prospectivity mapping, the spatial coordinates of mineral deposits and other geological features are often recorded originally in vector form, and converted to a grid of cells (a raster of pixels) for analysis. Although the results of the analysis clearly depend on the choice of pixel size, it is widely believed that, if pixel size is progressively reduced, results should converge to a stable value. However, we show that this is not true. Using a database of gold deposits in the Murchison region of Western Australia, the Weights of Evidence (WofE) contrast statistic was calculated for raster conversions with pixel widths varying from 5 km to 100 m, using the vector-to-raster conversion algorithms common in mainstream GIS packages. In response to even the slightest changes in pixel width, the calculated value of fluctuated by 1.5 units, and the calculated probability of a deposit fluctuated by a factor of 4.5. As pixel size was progressively reduced, the results did not converge. We investigate this instability phenomenon experimentally and theoretically, and establish that it could be widespread. It could arise in any form of prospectivity analysis (including logistic regression, machine learning and deep learning) where the explanatory variables are discontinuous. We have confirmed that it also occurs with logistic regression. Instability is primarily associated with deposit points which lie close to a discontinuity such as a feature boundary, and could be characterised as a failure to respect “ground truth” at the deposit location. Accordingly, instability can persist even with very small pixel sizes (as small as 3 m in the Murchison example). We propose a new algorithm for vector-to-raster conversion which respects ground truth, and produces results which converge rapidly as pixel size decreases. In the Murchison example, this algorithm provides stable results for pixel widths of 500 m or less. Our theoretical results predict the maximum error as a function of pixel width, and allow the geologist to select an appropriate pixel size for the data available. Potential fields of application include species distribution modelling and geospatial risk analysis.
{"title":"Mineral prospectivity analysis is unstable to changes in pixel size","authors":"Adrian Baddeley , Warick Brown , Gopalan Nair , Robin Milne , Suman Rakshit , Shih Ching Fu","doi":"10.1016/j.cageo.2025.105965","DOIUrl":"10.1016/j.cageo.2025.105965","url":null,"abstract":"<div><div>In mineral prospectivity mapping, the spatial coordinates of mineral deposits and other geological features are often recorded originally in vector form, and converted to a grid of cells (a raster of pixels) for analysis. Although the results of the analysis clearly depend on the choice of pixel size, it is widely believed that, if pixel size is progressively reduced, results should converge to a stable value. However, we show that this is not true. Using a database of gold deposits in the Murchison region of Western Australia, the Weights of Evidence (WofE) contrast statistic <span><math><mi>C</mi></math></span> was calculated for raster conversions with pixel widths varying from 5 km to 100 m, using the vector-to-raster conversion algorithms common in mainstream GIS packages. In response to even the slightest changes in pixel width, the calculated value of <span><math><mi>C</mi></math></span> fluctuated by 1.5 units, and the calculated probability of a deposit fluctuated by a factor of 4.5. As pixel size was progressively reduced, the results did not converge. We investigate this instability phenomenon experimentally and theoretically, and establish that it could be widespread. It could arise in any form of prospectivity analysis (including logistic regression, machine learning and deep learning) where the explanatory variables are discontinuous. We have confirmed that it also occurs with logistic regression. Instability is primarily associated with deposit points which lie close to a discontinuity such as a feature boundary, and could be characterised as a failure to respect “ground truth” at the deposit location. Accordingly, instability can persist even with very small pixel sizes (as small as 3 m in the Murchison example). We propose a new algorithm for vector-to-raster conversion which respects ground truth, and produces results which converge rapidly as pixel size decreases. In the Murchison example, this algorithm provides stable results for pixel widths of 500 m or less. Our theoretical results predict the maximum error as a function of pixel width, and allow the geologist to select an appropriate pixel size for the data available. Potential fields of application include species distribution modelling and geospatial risk analysis.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"204 ","pages":"Article 105965"},"PeriodicalIF":4.2,"publicationDate":"2025-06-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144320944","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}
Pub Date : 2025-06-15DOI: 10.1016/j.cageo.2025.105991
Ling Peng , Lei Li , S. Mostafa Mousavi , Xiaobao Zeng , Gregory C. Beroza
Seismic event detection, phase picking, and phase association are the most fundamental and critical steps in seismic network data processing. We propose a two-stream neural network that integrates the time domain and time-frequency domain representations for microseismic phase detection and picking. This model builds on the EQTransformer (EQT) by incorporating an additional time-frequency stream using a Short-Time Fourier Transform as input. This preserves the original time-domain network structure, while enabling the fusion of features from both domains through lateral interactions. We explore two feature-fusion strategies: fixed weighting addition and a cross-attention mechanism, resulting in two two-stream EQT (TS-EQT) models: AddTwoStream-EQT (ATS-EQT) and CrossTwoStream-EQT (CTS-EQT). We enhance the data through a multi-model average picking strategy to reduce the labeling errors. We train the models with the STEAD dataset and test them on the STEAD, DiTing and Geysers datasets. We find that the TS-EQT models are superior to the original EQT model in both learning ability and generalization performance. The cross-attention mechanism feature fusion strategy is superior to the fixed weighting addition strategy. Specifically, ATS-EQT detects 45 % more events than EQT on the Geysers microseismic dataset, the number of P-wave and S-wave picks increases by about 44 % and 48 %, respectively. CTS-EQT detects 48 % more events, and the number of P-wave and S-wave picks increases by about 52 % and 56 %, respectively. This study verifies that the frequency domain features improve the training of the original model and suggests the potential of two-stream approaches for other geophysical tasks.
{"title":"TwoStream-EQT: A microseismic phase picking model combining time and frequency domain inputs","authors":"Ling Peng , Lei Li , S. Mostafa Mousavi , Xiaobao Zeng , Gregory C. Beroza","doi":"10.1016/j.cageo.2025.105991","DOIUrl":"10.1016/j.cageo.2025.105991","url":null,"abstract":"<div><div>Seismic event detection, phase picking, and phase association are the most fundamental and critical steps in seismic network data processing. We propose a two-stream neural network that integrates the time domain and time-frequency domain representations for microseismic phase detection and picking. This model builds on the EQTransformer (EQT) by incorporating an additional time-frequency stream using a Short-Time Fourier Transform as input. This preserves the original time-domain network structure, while enabling the fusion of features from both domains through lateral interactions. We explore two feature-fusion strategies: fixed weighting addition and a cross-attention mechanism, resulting in two two-stream EQT (TS-EQT) models: AddTwoStream-EQT (ATS-EQT) and CrossTwoStream-EQT (CTS-EQT). We enhance the data through a multi-model average picking strategy to reduce the labeling errors. We train the models with the STEAD dataset and test them on the STEAD, DiTing and Geysers datasets. We find that the TS-EQT models are superior to the original EQT model in both learning ability and generalization performance. The cross-attention mechanism feature fusion strategy is superior to the fixed weighting addition strategy. Specifically, ATS-EQT detects 45 % more events than EQT on the Geysers microseismic dataset, the number of P-wave and S-wave picks increases by about 44 % and 48 %, respectively. CTS-EQT detects 48 % more events, and the number of P-wave and S-wave picks increases by about 52 % and 56 %, respectively. This study verifies that the frequency domain features improve the training of the original model and suggests the potential of two-stream approaches for other geophysical tasks.</div></div>","PeriodicalId":55221,"journal":{"name":"Computers & Geosciences","volume":"204 ","pages":"Article 105991"},"PeriodicalIF":4.2,"publicationDate":"2025-06-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144307172","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}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号