Accurate determination of the Global Navigation Satellite System (GNSS) velocities plays a very important role in describing the tectonic process, crustal deformation, and many other geodetic applications. This study aims to compare the efficiency of linear regression (LR) models with machine learning (ML) algorithms, such as Decision Trees (DTs), Random Forest (RF), and Gaussian process regression (GPR), in estimating horizontal GNSS velocities. In this context, data from 20 Continuously Operating Reference Stations (CORSs) on the southern Scandinavian peninsula were used. The findings show that the RF and GPR ML techniques perform much better than the LR models, giving an R2, higher than 0.94 while having a lower root-mean-square error and mean absolute error. The velocities estimated with the help of DTs are comparable to LR results and show that DTs are applicable in GNSS time series velocity estimation. However, in this study, it is seen that DTs improve the results even further in capturing nonlinear patterns compared to the LR-based approach. This study improves the accuracy of horizontal velocity estimates based on GNSS time series data, considering the limitations of classical models, and highlights the potential success of ML techniques in GNSS velocity estimation. These findings can provide a critical framework for improving geophysical and geodetic analyses in infrastructure planning, crustal movement monitoring, and disaster risk reduction.
Regression methods are widely used to model, smooth, and forecast datasets, whether in time series, scattered data, or logistic variables. Besides, linear regression is often preferred due to its simplicity and efficiency in identifying trends. However, many real-world phenomena, such as climate change, economic fluctuations, gambling, and biological processes, exhibit nonlinear behaviours that require more advanced regression techniques. For this reason, some studies include polynomial, exponential, and dynamic approaches derived from linear models. This study focuses on ocean acidification and carbon dioxide (CO2) emissions, two interconnected factors influencing climate change and marine ecosystems. Both datasets are structured as time series, where ocean acidification is measured in seawater pH levels, and CO2 emissions are recorded in parts per million. To perform a deep data analysis, this paper sets up a particular seasonal autoregressive integrated moving average model for prediction in order to process volatile time series with complex trends, providing, thus, a scalable and effective longterm forecasting technique.
Formation evaluation requires the analysis of well-log data that are not available for all wells, especially old wells. In the case of an unstable borehole well, the formation bulk density (RHOB) log cannot reflect the true values of the formation density, and we cannot record it another time. However, the use of artificial intelligence approaches, such as artificial neural networks (ANNs), which are strong tools for real-time prediction without additional costs, can provide acceptable results with high accuracy. This work, by leveraging ANNs for continuous and reliable in-situ prediction from other wireline logs, enables the prediction of RHOBs without additional costly logging operations and overcomes the limitations of traditional methods in terms of availability, accuracy, cost, and real-time applicability. The dataset of two wells was used to train and test the model. Unseen data from another well within the same field were utilised for validation. The findings revealed that the predicted RHOB values significantly matched the actual values with a coefficient of determination of 0.98 and root-mean-square error of 0.12. This score confirms the generalisation capability of the model, overcoming, thus, limitations of traditional logging methods that require direct measurement and stable boreholes. This demonstrates the advantage of ANN in filling data gaps and enhancing well-log interpretation under challenging conditions.
Ambient noise cross-correlation has been successfully carried out in many cases by utilising seismic data with vertical and transverse components from broadband seismographs, while both the processing of data or discussion about correlation of different components has been less demonstrated. This study concentrated on the juncture area of the Anhui, Jiangxi, and Zhejiang provinces of China, from where seismic data, between October 2014 and August 2015, were acquired through 10 domestic seismographs. Empirical Green's functions were yielded from vertical correlations or the superposition of radial and vertical correlations for the Rayleigh waves and from transverse correlations the Love waves. Then, the group-velocity and phase-velocity dispersion curves for Rayleigh and Love waves were measured by applying a frequency-time analysis method. Finally, a onedimensional S-wave velocity structure was obtained through the joint inversion of the above dispersion measurements. The conclusions out of data processing were drawn as follows. i) The superposition of vertical and radial correlations presented the interference phenomenon: when the differences between the ratio of vertical-correlation signal-tonoise ratios (SNRs) to radial-correlation SNRs and value 1 are less than, or equal to, 0.35, the SNRs of the stacked vertical and radial correlations are higher. ii) The effective length of some dispersion measurements from the stacked vertical and radial correlations is longer by 2-8 s than those from the vertical correlations.
Vagueness and non-uniqueness of geophysical data inversion can be reduced through the integration of different geophysical methods. Joint interpretation is one of the most common ways to combine datasets of two or more geophysical methods. In this paper, magnetometry inversion, induced polarisation (IP), and direct current (DC) resistivity data are utilised as a joint interpretation for the investigation of sulphide mineralisation in the Sar Balla village, near Eshghabad city, in Iran. Six profiles with a 100-metre separation are used to collect data. The IP and DC resistivity data are measured using pole-dipole configuration with unit electrode spacings of 20 and 40 m. The damped weighted minimum length solution algorithm is used for inverting both magnetic and DC resistivity data, whereas smoothness constrained inversion is implemented for IP data. Primary investigation using inversion of magnetic data is illustrative of a huge anomaly extending to the north of the profile with a depth range from 100 to 800 m. IP and DC resistivity data enable a more detailed examination of the subsurface down to 300 m, where inverse sections indicate a considerable amount of sulphide mineral deposits with high chargeability and low resistivity. According to these results, 29 regions have high potential for sulphide mineralisation, among which seven anomalies appear to continue to greater depths and can be suggested for deep drilling.
Amplitude-versus-angle (AVA) inversion is a widely used geophysical technique for estimating subsurface elastic properties from seismic reflection data. However, its illposed nature necessitates robustand efficient methods for uncertainty quantification. This work presents a novel Bayesian AVA inversion framework that integrates annealed Stein variational gradient descent (A-SVGD) with discrete cosine transform (DCT) compression of the model space. A-SVGD introduces an annealing schedule to enhance posterior exploration, while DCT significantly reduces problem dimensionality, improving stability and convergence. The inversion employs full Zoeppritz equations as the forward model and leverages automatic differentiation for efficient and accurate gradient computation. Synthetic tests demonstrate that A-SVGD outperforms standard SVGD in convergence speed and uncertainty quantification. The DCT-based model-space compression achieves performance comparable to full-domain inversion, but at a fraction of the computational cost. Further tests under realistic conditions (e.g. where errors are introduced in both the estimated source wavelet and noise statistics) highlight the robustness of the proposed approach. Additionally, replacing the Jacobian with a DCT-projected analytical approximation shows promise for accelerating computation, though its applicability is model-dependent. Overall, this study demonstrates that A-SVGD in a DCT-compressed domain provides a powerful and efficient solution for probabilistic AVA inversion and achieves final predictions comparable with those obtained with a gradient-based Markov chain Monte Carlo method.
Besides the Istituto di Geofisica e Vulcanologia's [the Italian National Institute of Geophysics and Volcanology] traditional web pages, since 2010 the INGV Earthquake Deparment (INGVterremoti) platform has been making information about Italy's ongoing seismicity and earthquake and tsunami research activities available to everyone, also through a series of web and social communication channels, such as Blog Wordpress, X (Twitter), Facebook, YouTube, apps for iPhone and Android, and StoryMaps. Lately, the new INGVterremoti Instagram channel has landed on the platform: it was launched on 27 September 2024 during the European Researchers' Night. The main task of this new social profile is to show the world of earthquakes and tsunamis through visual-based formats and contents so as to reach out and involve everyone, especially the new generations. The launch of the new Instagram channel is the result of several months of groundwork that brought together collaborators with a specific communication background and experts on natural hazard communication. The long planning and drafting work of a new editorial line led INGVterremoti to take on a new challenge: the communication of risk on a platform that is totally visual-based and mostly populated by a younger audience. In order to do so, it was necessary to develop a specific editorial plan that could identify and organise new content to share on the new account as well as new ways and means of publication. This paper will describe all the stages involved in the development of the INGVterremoti Instagram channel, as well as the results achieved during the first eight months of its activity.
The shift to renewable energy has heightened the demand for efficient, cost-effective subsurface imaging, especially for offshore wind farms and underground storage of carbon or hydrogen. These projects frequently depend on short-offset or single-channel seismic data due to logistical and financial constraints, which restrict the resolution of traditional velocity analysis. This work presents a novel method to improve seismic imaging from post-stack data by integrating image focusing into full-waveform inversion (FWI). We include minimum entropy velocity analysis as a cost function within FWI, evaluating subsurface models based on the focusing quality of reverse time migration images using the minimum entropy (ME) norm. Validation on a synthetic dataset shows that, although the ME-norm itself can be ambiguous, its derivative reliably indicates the correct velocity. This insight leads to a new cost function whose gradient, computed via the adjoint-state method, effectively guides model updates through the steepestdescent method. Application to a marine dataset from the Viking Graben (North Sea) demonstrates enhanced image quality, with better reflector alignment and contrast. These results highlight the potential of ME-norm derivatives to drive FWI, thus enabling advanced velocity model building from short-offset or single-fold data with lower computational costs.
Sea surface clorophyll-a (SSC) represents a fundamental indicator of oceanic primary production and serves as a widely used proxy for phytoplankton biomass. Although upwelling is recognised as the dominant driver of SSC variability, precipitation and river discharge can substantially modulate its spatial and temporal distribution. This study examines the spatiotemporal dynamics of SSC in the Arafura Sea, utilising satellitederived oceanographic data from 1998 to 2022, with a focus on the influences of upwelling, precipitation, and oceanographic conditions. Seasonal analysis reveals that SSC peaks during the SE monsoon (June-August), coinciding with intensified upwelling, as indicated by positive Ekman pumping velocity (EPV), reduced sea surface height, and lowered sea surface temperature (SST). Coastal regions consistently exhibit higher SSC, driven by riverine discharge and precipitation-enhanced nutrient input. Regional correlation analysis confirms that offshore EPV and increased precipitation are the dominant mechanisms driving SSC enhancement during the SE monsoon. Climatological correlation analysis further identifies a strong positive relationship between SSC and EPV in key regions, with a maximum correlation coefficient of 0.85. The spatial distribution of SSC exhibits a pronounced inshore-offshore gradient across both monsoon seasons, reflecting the interplay between coastal upwelling, stable river discharge, and nutrient runoff. The empirical orthogonal function analysis of interannual variability suggests that SSC fluctuations are influenced by additional drivers beyond the El-Nino Southern Oscillation, including SST co-variability and anomalous precipitation patterns over central and north-western Papua. These findings show that phytoplankton variability in the Arafura Sea is chiefly controlled by local ocean, atmosphere, and land interactions, underscoring their importance for ecosystem forecasting.
The borehole nuclear magnetic resonance technique is a technique widely used in exploration geophysics for subsurface structure imaging. It is primarily utilised in various projects including reservoir characterisation, hydrocarbon exploration, groundwater studies, and fracture characterisation. To interpret porosity, nuclear magnetic resonance (NMR) measurements are typically conducted alongside other logging methods such as caliper, resistivity, and gamma ray. However, due to cost and operational constraints, NMR data may not be acquired in all wells. Consequently, to predict this parameter in wells with missing data, the use of logs from other wells becomes necessary. Machine learning (ML) algorithms, increasingly prevalent, are well-suited for regression problems due to their capacity to model complex and latent relationships within data. Given the inherent difficulty in comprehending the intricate relationships within multi-dimensional spaces involving these measurement parameters, we implemented ML regression algorithms to map predictor parameters to response parameters. Thus, this study evaluates various ML regressors, including their ensemble learning counterparts, to compare their effectiveness in predicting NMR data, both individually and in combined configurations.
Large earthquakes are often followed by aftershocks, which can cause further damage and cost lives. A pattern recognition approach called Next Strong Related Earthquake (NESTORE) has been developed to forecast whether one of these subsequent large events is to be expected in an occurring seismicity cluster. This method, already successfully applied in Italy, Slovenia, California, Greece, and Japan, has been optimised in the NESTOREv1.0 software written in MATLAB. Using machine learning, NESTOREv1.0 provides a probabilistic forecast of earthquake clusters where a mainshock is followed by a significant aftershock. It classifies clusters as type A (mainshock and strongest aftershock differ by <_ 1 magnitude unit) or type B (larger difference). NESTOREv1.0 adapts to specific regions through supervised training. It trains one-node decision trees on individual features at increasing time intervals, selects the best classifiers, and combines them using a Bayesian method to forecast type A clusters. Recent improvements to the algorithm added a new approach for identifying clusters based on Epidemic-Type Aftershock Sequence (ETAS) and an innovative method for detecting outliers before training. This study compares results from Greece, Italy, western Slovenia, California, and Japan, highlighting the performance on independent test sets and seismicity features in different regions and interpreting the differences between the regions.
YOLO (You Only Look Once) is one of the most popular computer vision algorithms. Computer vision has revolutionised the field of moving object detection in real time with its ability to analyse and understand visual content much like a human. This paper presents a comprehensive framework for fish detection, bounding box-based tracking, and counting in underwater environments using the YOLOv8 deep learning architecture. Accurately and efficiently identifying, tracking, and counting fish plays an important role in aquatic research, conservation efforts, and fishery management. The proposed system uses a pre-trained YOLOv8 model which is fine tuned using a large annotated dataset of underwater fish images. The model is improved using transfer learning to learn features specific to fish detection in water. Real-time underwater fish detection is performed on underwater video streams using a fine-tuned YOLOv8 model. The high speed and accuracy of YOLOv8 enables efficient localisation of fish instances at every frame. The analysis of such data enables accurate fish counts and facilitates effective monitoring and assessment of fish populations in water bodies. The true positive rate of 0.91 and accuracy of 92% indicated that the system successfully identified a significant proportion of fish instances present in the images.
This study presents a comparative assessment of equivalent linear (EL) and nonlinear (NL) site response analyses for the Kathmandu Valley, Nepal, using DEEPSOILTM based on borehole data from multiple locations and seismic input motions focusing on two significant events: the Gorkha (Mw 7.8, 25 April 2015) and Kobe, Japan (Mw 6.9, 17 January 1995) earthquakes. A statistical analysis, including Pearson correlation, normality tests, t-tests, and Mann-Whitney U tests, was conducted using MATLAB to examine the relationship between peak ground acceleration (PGA) and displacement under EL and NL conditions. Results show that the NL analysis better captures soil nonlinearity. The Gorkha NL case showed a strong inverse correlation (slope = -20.195, p = 0), indicating significant displacement reduction with increasing PGA. EL cases, in contrast, exhibited weaker trends, often overestimating. Boxplots further confirmed significant differences in PGA and displacement between events, especially under EL conditions. These findings highlight the critical importance of incorporating NL site response analysis in seismic hazard assessments. The Gorkha NL results reveal that NL models can reduce predicted displacement by more than 50% compared to EL models. This has direct implications for infrastructure resilience in earthquake-prone regions like the Kathmandu Valley, where site-specific NL analysis is essential for accurate seismic design and safety evaluation.
The Italian area is affected by on-going deformations and/or mass transfers, both of geophysical and human origin, which act on very differenttemporal scales and significantly modifythe gravity field overtime. The current Italian gravity database contains gravimetric measurements taken on land and sea by various research institutions and services. Data were acquired at very different epochs, with different instruments, some now obsolete and mostly lacking inter-comparison, and also with non-uniform operational procedures of data acquisition and analysis. Thus, these data are not homogeneous and do not represent an updated image of the Italian gravity field. This paper shows the main steps followed towards the realisation of a new reference network for absolute gravity in the Italian area. The choices are in line with the resolutions approved by the International Association of Geodesy during its 2015 general assembly. The goal is to update the existing absolute gravity network by adding gravity stations according to the new standards so as to align this infrastructure to the highest level of precision and accuracy. Based on this, an overall revision of the Italian gravity database will be possible, thus, leadingto a better estimate of the gravity potential.
Shale hydrocarbons are sources of oil and gas trapped in shale rock formations which are rich in organic material. Oil and gas reservoirs can be detected by knowing the characteristics of the rocks. This study attempts to determine whether permeability, mobility, transmissibility, and the brittleness index (BI) can be used for identifying unconventional reservoirs. The research area includes the geological formations in the OD Field of the NE Java basin. We perform acoustic impedance seismic inversion on 48 two-dimensional seismic sections and we analyse well logs from three wells (namely M-1, M-2, and M-3). The results obtained show that wells M-1, M-2, and M-3 have small permeability, mobility, and transmissibility parameters. The distribution of the BI values is 0.46-0.54 falling within a less brittle category. The Kujung formation has a BI value of 0.46-0.76 falling within a low brittle category. The Ngimbang formation has a high BI value of 0.52-0.82 with a brittle category.
Accurate determination of shear slowness (DTS) is essential for well placement optimisation, mechanical rock property estimation, and reservoir characterisation. However, direct DTS measurement is expensive, limited in availability, and lacking repeatability. Traditional empirical methods require extensive calibration and are only valid for specific rock types. This study presents the voting regressor (VR), an ensemble machine learning (ML) technique combining Extra Trees, Random Forest, Gradient Boosting, LightGBM, support vector regressor, and multi-layer perceptron to enhance DTS prediction. The model uses well logs including gamma ray (GR), bulk density (RHOB), porosity index (PHIX), and compressional slowness (DT), applied tothe Tensleep Formation in the Teapot Dome field, Wyoming, USA. Each model was evaluated using the coefficient of determination (R2), mean absolute error (MAE), mean-squared error (MSE), and root mean-squared error (RMSE). A weighted averaging approach based on R2 performance was used to build the VR model, achieving an R2 of 0.96, RMSE of 0.19, MAE of 0.12, and MSE of 0.037. DT, PHIX, and RHOB were the most important features, while GR showed minimal impact. Validation on two unseen wells confirmed a strong generalisation of the VR (R2 = 0.91-0.92). This work highlights the potential of ML for accurate DTS prediction and improved subsurface characterisation in data-limited settings.
Traffic-induced vibrations represent a problem in urban areas where repeated dynamic loading from heavy vehicles can affect both the structural integrity of buildings and occupant comfort. This study addresses the dynamic response of a historic masonry building in the Borgo Teresiano district of Trieste, a site with Quaternary sediments known to amplify ground motions due to their low stiffness and heterogeneity. The aim was to investigate the interaction between traffic-induced ground motion and the structure's response by using a rapid, non-invasive monitoring approach. During a shortterm campaign, low-cost seismic stations with triaxial velocimeters were installed at the base and top of the structure. Ambient vibration data were analysed using: i) Fourier amplitude spectra combined with the rotation of horizontal components, ii) spectral ratios between sensor pairs within the structure to identify main structural modes and amplification directions, iii) single station horizontal-to-vertical spectral ratio analysis to identify fundamental soil resonance frequency. The results showed a resonance peak at 11 Hz in the vertical component at the top floor, interpreted as responsible for the amplification of vibrations caused by repeated traffic combined with soil-structure interaction, with levels exceeding discomfort thresholds and impacting wellbeing. This highlights how cost-effective sensors and ambient noise analysis support mitigation strategies.
It is known that geoelectric fields, induced at ground level, during geomagnetic storms may represent a threat for infrastructures such as power lines, pipelines, and railway lines. Particularly, when one of these structures is located inside a region characterised by large lateral ground conductivity contrast (typically a coastal region where the seawater conductivity is significantly higher than that of the land), the commonly named geoelectric coast effect manifests and consists in an increased level of the induced geoelectric field near the coastline with penetration inside the land region for several kilometres. A convenient calculation method, capable of assessing the geoelectric coast effect, is the generalised thin sheet model. In literature, this method is generally applied by considering the idealised case of a constant-depth sea. On the contrary, in this paper, we extend its use to the more realistic case of a stepped-depth sea model. In such a way, we show that, by considering a shelf region, characterised by different sea depth levels and located between the coastline and the maximum sea depth region, the application of such model results in significant differences in the predicted geoelectric field with respect to the simpler constant-depth sea model. Moreover, the influence of certain parameters characterising the model (shelf region length and slope, and maximum sea depth) is put into evidence. These are the main novelties introduced by this paper.
Accurate monitoring of sea surface temperature (SST) is vital for understanding regional climate variability, marine ecosystem dynamics, and long-term climate change. In this study, the consistency between satellite-derived SST data from the Copernicus Marine Environment Monitoring Service (CMEMS) and in-situ observations from 21 coastal stations operated by the Turkish State Meteorological Service was evaluated across the Turkish coastline. Initial assessments were based on classical statistical comparisons using the root-mean-square deviation and Pearson correlation. Subsequently, four machine learning (ML) regression models, linear regression, support vector regression, gradient boosting, and artificial neural networks, were applied to assess the predictive capability of CMEMS data for estimating in-situ SST. Among the models, GB achieved the best overall performance (coefficient of determination = 0.97, root-mean-square error = 0.84 degrees C), owing to its ability to effectively capture complex nonlinear relationships between datasets. Based on these results, a spatial gap analysis was conducted, and eight statistically optimised proxy observation points (termed virtual SST stations) were proposed to enhance SST coverage in underserved coastal segments. This study demonstrates a scalable (regionally adaptable) and objective methodology for optimising SST monitoring networks by integrating ML with geospatial analysis. The proposed approach offers practical benefits in enhancing climate resilience, improving SST anomaly forecasting, and supporting evidence-based marine resource management, such as fishery zoning or coastal ecosystem protection.
This paper employs various combinations of geophysical data, including sequential inversion, joint interpretation, and joint inversion, to analyse two case studies in Iran and South Africa. The selection of integration types for geophysical datasets is dependent on the specific requirements of each case study. The first real case, which involves apparent resistivities and gravity gradient data to detect three tunnels, is particularly well suited for joint interpretation. The joint interpretation of reconstructed resistivity and density models serves as a reliable indicator of the accurate recovery of the tunnels. The second real case, comprising direct current (DC) resistivity, magnetometry, and electromagnetism at low induction number (EM-LIN) data, enables 1) the joint inversion of DC resistivity and magnetometry and 2) the sequential inversion of DC resistivity and EM-LIN data. The inversion of DC resistivity data reveals a two-layered medium, where the upper layer exhibits high conductivity and the lower layer exhibits high resistivity. In contrast, the inversions of magnetometry and EM-LIN data indicate the presence of a resistive and magnetised dolerite dyke, with depth ranges spanning from 3 to 15 m. The resistivity models generated through joint and sequential inversions provide conclusive evidence of a dyke situated within a two-layered medium.
