Abstract In high‐altitude regions characterized by cold climates and complex terrain, the engineering performance of geosynthetic‐reinforced soil slopes requires careful consideration. This paper discusses the design and construction of a reinforced soil slope based on an ultra‐high voltage converter station project in a high‐altitude region and verifies the reliability of the design scheme through cross‐validation using two numerical software programmes. A 40 m high reinforced soil slope was selected for monitoring and analysis of temperature, slope deformation, settlement, deep horizontal displacement, acceleration, earth pressure and geogrid strain. The results indicate that the safety factors of the slope under normal, earthquake and heavy rainfall conditions all satisfy relevant standard requirements. Both the lateral displacement and settlement of the slope surface increased rapidly during construction and then gradually stabilized. Twenty‐six months after completion, the maximum lateral displacement and settlement were 523.41 and 420.30 mm, respectively, both within the allowable range specified in the standards. Soil temperature exhibited a lag relative to ambient temperature, and temperature variations induced periodic daily fluctuations in deformation. Lateral displacement and settlement showed negative and positive correlations with soil temperature, respectively, whereas repeated freeze–thaw cycles led to the gradual accumulation of deformation over time. The vertical earth pressure within the slope decreased gradually with increasing slope height and showed notable deviations from theoretical values due to the shielding effect of the geogrids. Geogrid strain was higher near the slope surface and decreased with increasing slope height, with a maximum strain of 1.8% observed across all layers. The deep horizontal displacement gradually stabilized 1 year after the slope was completed. No seismic events were recorded at the site during the 2‐year period following slope completion.
Abstract This study aims to perform a comprehensive site characterization in the Balıkesir region of north‐western Turkey, with a particular focus on the Bandırma and Erdek districts. Microtremor measurements were carried out at systematically selected locations, and the horizontal‐to‐vertical spectral ratio (HVSR) technique was applied to determine the dominant period and HVSR peak amplitudes ( A 0 ). In addition, V s 30 values and soil classification types were obtained using surface wave analysis and evaluated in accordance with the Building Seismic Safety Council guidelines. Based on these results, seismic microzonation maps were generated using ArcGIS to illustrate the spatial distribution of dominant period, HVSR peak amplitudes ( A 0 ), V s 30 values and soil classes. These maps were then compared to assess the consistency among HVSR peak amplitudes, dominant period, V s 30 values and estimated soil depth. In the Bandırma district, HVSR peak amplitudes ( A 0 ) varied between 0.40 and 8.22, whereas dominant periods ranged from 0.065 s to 2.100 s. In the Erdek district, HVSR peak amplitudes ( A 0 ) ranged between 0.97 and 6.14, with dominant periods varying from 0.051 s to 3.100 s. The results emphasize the spatial variability of local soil conditions and highlight their potential influence on site response characteristics in urbanized areas. This study contributes to future seismic microzonation efforts and provides fundamental data for earthquake risk mitigation, land‐use planning and seismic hazard assessment in the region.
Abstract In studies of subsurface hydrology, monitoring and imaging techniques are used to estimate the moisture content, or to track moisture or injected tracer movement. One applied imaging technique is electrical resistance tomography (ERT), in which measurements collected using electrodes placed on the ground surface or in boreholes are used to reconstruct the subsurface electrical conductivity, or its temporal changes. ERT image reconstructions are, however, very sensitive to measurement noise and especially modelling errors. In geophysical ERT, especially when using borehole electrodes, a significant modelling error may result from uncertain electrode positions. In this work, we study the effects of such modelling errors and apply the non‐linear difference (NLD) imaging and Bayesian approximation error approach to compensate for them. The approach is tested with numerical simulations representing tracer injection experiments into the subsurface. The results show that when electrode positions are uncertain, adopting the NLD imaging approach can yield feasible estimates for the conductivity change from the initial state in cases where the conventional linearized difference and absolute reconstructions are biased. Moreover, the study demonstrates that while the absolute values of the conductivity obtained by NLD imaging can be biased due to electrode position errors, enhancing the NLD reconstruction with Bayesian approximation error modelling improves the reconstruction of the initial conductivity and the temporal change of conductivity, leading to improvement in the absolute conductivity estimates.
Abstract Traditional geotechnical site investigations often depend on invasive borings, which require significant time and resources. In contrast, surface wave methods gather data without disturbing the soil and are faster and more cost‐effective, allowing investigation of more extensive areas. However, these techniques involve an inversion process which necessitates significant computational effort and often leads to multiple solutions. The convergence of the inversion process and the accuracy of its output heavily rely on the initial assumptions about soil layering, including the number of layers, shear‐wave velocity and density. In this study, a machine learning (ML) approach was developed to refine the estimation of the most appropriate number of soil layers for a model. Nine hundred thousand synthetic shear‐wave velocity soil profiles were generated, each with a total thickness ranging from 20 m to 60 m. The number of layers ranges between two and seven, with each layer having a randomly determined thickness and shear‐wave velocity, then converted into Rayleigh and Love dispersion curves and horizontal‐to‐vertical spectral ratio (HVSR) curves for ML training. Three separate tail‐masked gated recurrent units with decay classifiers were then trained for each data type. To determine the best‐performing models, various network architectures with different optimizers were tested, and their confusion matrices and test accuracies were evaluated. The selected models were then used to produce a probability vector indicating the most likely number of layers for each individual Rayleigh dispersion curve, Love dispersion curve or HVSR curve. Based on these probability vectors, the final layer count prediction was made. These models demonstrate satisfactory accuracy when validated against field data, offering a practical way to refine initial assumptions in the inversion process for prediction of the number of model soil layers for the inversion process. All three classifiers are bundled in the soil layer estimator GUI, which visualizes input curves, enables adjustment of per‐model weights and exports Dinver‐compatible parameter files for the inversion setup. This approach provides a versatile tool for improving both the speed and precision of the inversion process for surface wave investigations.
This study concurrently applies multiple cutting‐edge survey techniques, including those based on airborne and shipborne geophysical methods, to investigate buried, potentially archaeological targets in the Taijiang Waters of Taiwan. The Taijiang Waters are archaeologically significant because they served as essential shipping channels in the seventeenth century, when many ships sank in their waters. However, conventional archaeological surveying methods are not applicable due to rapid morphological changes, river currents and poor water visibility. Through integrated analysis of various geophysical data, seven suspected underwater cultural heritage (UCH) targets were identified at burial depths ranging from 0.3 to 5.1 m below the riverbed, with approximate lengths ranging from 1 to 10 m. By comparing the geophysical results with historical records and the regional sedimentation rate of 0.2–0.4 cm/year, inferred from geological dating, our results suggest that the suspected UCH targets are possibly linked to the Dutch East India Company shipwrecks, including the vessels Maarssen , Koudekerke , Valk and Immenhorn , and therefore warrant further archaeological investigation.
Near-surface characterization of the lunar subsurface is essential for future exploration and infrastructure development, particularly for the construction of underground habitats that provide protection against radiation and micrometeorites. However, conventional seismic approaches for estimating subsurface properties typically rely on prior velocity models or multicomponent data, which are not available for legacy Apollo datasets. In this study, we propose a data-driven approach for estimating the critical angle (CA) and phase rotation of P-wave reflection events using active-source seismic data from the Apollo 16 mission. The CA is directly related to subsurface velocity contrasts, making it a key parameter for constraining near-surface structure. Our method is based on spectral recomposition, in which an inversion scheme is used to reconstruct the seismic spectrum of individual wavelets and extract their fundamental properties without requiring prior geological information. The proposed workflow is applied to near-surface lunar seismic data, allowing the detection of the CA and estimation of phase variations along reflection events. The results show that the estimated CA values are consistent with previously reported velocity structures, with differences on the order of a few per cent. These findings indicate that the method can reliably characterize near-surface lunar layers under limited data conditions. This work demonstrates the potential of data-driven spectral methods for extracting physically meaningful parameters from sparse and degraded seismic datasets, providing a practical tool for future lunar geophysical investigations and site characterization.
Abstract In June 2023, we measured ten near‐surface seismic profiles to image the critical zone of dolerite (diabase) dykes that cut the sandstone at Cap Fréhel in Brittany (France). Seismic ray tracing and inversion are used to determine the thickness of the regolith layer from first‐arrival P‐wave traveltimes. The 2D subsurface models show strong lateral velocity contrasts between weathered dolerite and sandstone. The weathered dolerite layer is between 6 m and 15 m thick at locations several hundred metres apart. The weathered sandstone is often only 1–2 m thick, sometimes up to 4 m. Seismic P‐waves propagate with in the weathered dolerite, depending on the degree of alteration, and with or more at the top of the unweathered dolerite. In weathered sandstone, P‐wave velocities increase from at the surface to at a depth of 1 m. The unweathered sandstone at a depth of 3–4 m has high P‐wave velocities of over , mostly exceeding , depending on the degree of fracturing. We have introduced smoothness constraints for the model parameters of the ray tracing and inversion software Rayinvr , so that the regularization parameter is replaced by the prior covariance of the constraints. The resolution matrix shows that the interface depth is always coupled with the velocity above the interface and sometimes also with the velocity of the head wave below the interface. The head wave velocities in the unweathered layer are not well resolved, which we attribute to a limited source–receiver distance and strong lateral velocity inhomogeneities when a seismic profile intersects both dolerite and sandstone. Seismic tomography based on wave propagation models without first‐order discontinuities does not allow to accurately determine the regolith thickness. The Akaike information criterion favours the Rayinvr models with fewer model parameters over seismic tomography models calculated by solving the eikonal equation or the shortest path method.
As the predominant binder in collapsible loess reinforcement, cement is a major source of greenhouse gas emissions, accounting for 8% of global anthropogenic CO2; this necessitates an urgent transition toward sustainable alternatives. This study quantitatively demonstrates that the industrial by-products fly ash (FA) and cement-slag blend can serve as effective cement replacements while enhancing both mechanical and hydraulic performance. Through standardized column tests on collapsible loess specimens (binder contents: 6%-12%, curing period: 1-28 days), cement-slag blend consistently outperformed both cement and FA across several critical metrics: (1) Cement-slag blend-treated samples achieved a 28-day unconfined compressive strength (UCS) up to 750 kPa, surpassing FA counterparts by 15% and California bearing ratio (CBR) values exceeding 90. (2) Cement-slag blend uniquely exhibited superior strength without the typical brittleness, as evidenced by stress-strain curves maintaining post-peak ductility. (3) Cement-slag blend reduced hydraulic conductivity to 2.8 & times; 10(-7) m/s, approximately 38% lower than that of FA (4.54 & times; 10(-7) m/s), due to the formation of a denser calcium silicate hydrate (C-S-H) microstructure. Mechanistically, the 7-day performance inflection point observed exclusively in cement-slag blend specimens is attributed to their high CaO (41.6 wt%) and Al2O3 (12.4 wt%) content, which facilitate faster pozzolanic kinetics. In contrast, FA exhibited delayed reactivity stemming from a reliance on externally sourced Ca-2(+). These findings establish a threshold-driven activation model, advancing fundamental understanding of soil stabilization beyond prior linear hydration models. Practically, our data provides validated mix design guidelines: Specimens with 6% cement-slag blend content met subgrade construction standards (CBR > 90, UCS > 750 kPa) while reducing carbon footprint by 48% compared to conventional cement stabilization, offering actionable pathways for infrastructure decarbonization through scientific waste valorization.
Abstract Long‐period magnetotelluric (LMT) data have long recording durations and low frequencies, with exploration depths reaching up to 1000 km. However, the LMT method is highly susceptible to anthropogenic noise, and the long acquisition time of the instruments makes rework costly; therefore, denoising of the data is particularly critical. At present, relatively few methods are available for LMT data processing, and most of them suffer from a low degree of automation, large bias caused by manual processing, and excessive parameter dependence. To address these issues, this paper proposes a denoising method for LMT signals. The proposed method employs an improved task‐specific network, batch normalization residual temporal convolutional network, to automatically identify noise‐contaminated segments in the signal and retain high‐quality segments and then combines Hampel filtering, interval‐point difference detection, step compensation and other processing methods to remove different types of noise in a targeted manner. The experimental results show that the improved deep learning model achieves a recognition accuracy of 0.9994. Compared with the denoising results without signal‐noise identification, after introducing the model, the mean squared errors of the BX and EX channels are reduced by 97.94% and 96.38%, respectively, whereas the signal‐to‐noise ratios are improved by 16.84 and 7.62 dB, respectively. In addition, the proposed method significantly reduces the dispersion of the apparent resistivity and phase curves. The method can reliably improve the quality of LMT data and help further enhance the effectiveness of deep LMT exploration.
Induced polarization (IP) and electrical resistivity surveys were conducted in the northern region of the Peddavura Schist Belt, part of the Eastern Dharwar Craton, India, to evaluate the potential for sulphide mineralization within altered Banded Magnetite Quartzite (BMQ), metabasalt and metarhyolite. The primary lithological units include metabasalt, BMQ, granitic gneiss, quartz reefs, quartz veins, metarhyolite and dykes. These geological settings alter the physical properties and structural configurations of the rock units, making them suitable targets for geophysical methods in mineral exploration. The findings suggest that copper and sulphide mineralization are associated with sheared lithologies, particularly quartz veins and zones of alteration. A total number of 11 NE-SW gradient array survey traverses, oriented perpendicular to the local geological strike, were laid down at approximately 100 m intervals with 10 m station spacing and a length of 1000 m. Three-dimensional (3D) inversion of two-dimensional (2D) parallel profiles was performed using a Python-based Boundless Electrical Resistivity Tomography code. The analysis of 11 2D resistivity and IP profiles, along with 3D models, showed that potential mineralized zones have medium resistivity range (1000-5000 ohm m) and moderate-to-high IP values (phase) (>5 mrad). These zones are prominent near the surface and extend to depths of up to 100 m, although their signatures diminish at greater depths. Surrounding the moderate resistivity zones are areas of high resistivity, suggesting concentrated sulphide minerals along altered BMQ and weathered or fractured zones. The highly chargeable zones with moderate resistivity generally occur at shallow depths, up to approximately 100 m below the surface. The subsurface resistivity and IP data have effectively delineated potential copper-gold exploration zones, marked by extensive high-chargeability regions and moderate resistivity anomalies.
Electrical resistivity tomography (ERT) is one of the most widely used geophysical techniques for hydrogeological questions, such as investigating the groundwater transition zone between land and sea. The interpretation of 2D coast-parallel ERT profiles is challenging because most of the electrical current flows through the highly conductive seawater and not the ground beneath the profile. We study this effect using synthetic data and a field data example and propose an approach for improved evaluation of coast-parallel geoelectric profiles. The synthetic study shows that the impact of the seawater on apparent resistivities is negligible (errors below 1.5%) only if the distance between current injection electrodes (a) is equal to or smaller than the distance (d) between ERT profile and shoreline (). Ignoring this criterion, standard 2D geoelectric inversion results may diverge by 50% from the actual resistivity distribution at depths of . To improve the inversion also for profiles with we tested an approach to consider the seawater effect through a correction factor applied to the measured apparent resistivity values prior to the inversion. The correction factor depends on , the seafloor slope and the resistivity ratio of seawater and average underground. We present an analytical function approximating the correction factor well for seafloor slopes up to 10 degrees and seawater-to-average-land resistivity ratios of 0.040-0.002. We found that this approach is effective at removing errors in apparent resistivities for . Inversion results then show more reliable resistivities up to relative depths of . We applied the correction approach to field data from a coast-parallel profile at Konyaalt & imath; Beach (Antalya, T & uuml;rkiye). The field study demonstrates the importance of considering the seawater effect to interpret possible locations of fresh submarine groundwater discharges and saltwater intrusions. It also shows the practical applicability of the new correction approach.
To address the challenges of low recognition accuracy and high rates of missed and false detections of small targets in ground-penetrating radar images of road internal defects, this paper proposes a lightweight detection algorithm named YOLOv11 (You Only Look Once)-PME, based on an improved YOLOv11 architecture. The method introduces innovations at both the network structure and loss function levels: a lightweight CPU convolutional neural network (PP-LCNet) is adopted as the backbone to enhance feature extraction capability while reducing parameter count; a multi-scale cross-reference attention module (MSCRSimAM) is designed to improve perception and fusion of multi-scale defect features; and the enhanced MPDIoU (EnMPDIoU) loss function is proposed, which incorporates bounding box diagonal constraints and optimizes centre distance weighting to increase sensitivity to the location and shape of defects. Experimental results demonstrate that YOLOv11-PME achieves a mean average precision (mAP@0.5) of 96.8% on a self-established road defect dataset, outperforming the original YOLOv11 by 5.3%, while reducing both parameter count and computational cost by 11.5% and 4.8%, respectively. Field detection results show strong agreement with core validation, indicating that the proposed algorithm exhibits high reliability, generalization capability and strong potential for engineering applications.
This paper presents an f-x domain Transformer network for multi-frequency ground-penetrating radar data fusion, which processes single-sided amplitude spectra to preserve complementary spectral features while avoiding distortions common in time-domain approaches. The method introduces full-range two-dimensional sinc interpolation for calibration, an energy-driven frequency-window selection strategy and an encoder-only Transformer adapted for spectral inputs. Experiments on synthetic and field data show that the proposed fusion yields higher spectral entropy and reduces training time by 70%-90% on CPU and similar to 40% on GPU, offering a physically consistent and computationally efficient solution for enhanced subsurface interpretation in engineering and geophysical surveys.
Rare earth element (REE) continues to be a global concern for exploration and utilize as a raw material of modern industry. This research focuses on tracing of REE sources that determine their potential, predicting REE volumes and the scale of mining priorities in an area. The samples of this study were taken from Bromo Tengger Semeru (BTS) complex that was dominated by material volcanic according to the geological map, that is, lava andesite, tephra, pyroclastic fall, laharic, pumice, volcanic deposits and volcanic soils. The determination of REE potential is measured by Inductively Coupled Plasma-Optical Emission Spectrometry (ICP-OES). The dominant REE potential in BTS, with average values, respectively, is Ce (38.82 ppm), La (29.91 ppm), Nd (22.91 ppm), Gd (17.67 ppm), Pr (10.20), Tb (8.71 ppm), Sm (4.28 ppm) and Y (18.73 ppm). The REE volume prediction was carried out by integrating the geochemical data of the REE with the geological map of the research site by multiplying the volume of total REE from measurement and the total volume of the area by mapping, dividing by one million (ppm). The predicted BTS volume is limited to the selected area above sea level. Furthermore, the scale of mining priority in the BTS area shows that the Bromo area becomes first priority and then the Semeru area as the second priority with the same pattern of Ce, Nd and Gd, although opposite pattern to La.
The rapid and accurate detection of geological hazards on urban roads is an urgent matter requiring prompt action. Current continuous towed seismic detection methods provide an effective and economical way for repeated detection along the same survey line. However, the conventional weighted stacking method used for multiple detection data encounters two key challenges. First, the presence of a low signal-to-noise ratio (SNR) in the detection data can adversely affect the final stack result. Second, the presence of strong energy noise makes it difficult to identify weak energy reflected and diffracted waves. To address these constraints, this paper proposes a novel multi-scale stacking method using a deep belief network (DBN) to eliminate the dependence on the real labels and improve the resolution of the continuous towed seismic stack section. We have also redesigned the algorithm for training the weight matrix by utilizing a limited amount of data and extracting the reflected and diffracted waves separately. The results of the numerical simulation demonstrate that the proposed multi-scale stacking method offers distinct advantages compared to the weighted stacking method: (1) It effectively addresses the effects of low SNR on the final stack results; (2) it allows for the extraction of wavefields with different characteristics and enables the analysis of near-surface anomalous structures from multiple angles; (3) it suppresses noise and extracts weak energy reflected and diffracted waves. The proposed multi-scale stacking method is useful for interpreting continuous towed seismic data and has real-world engineering applications.
Electrical resistivity tomography (ERT) surveys produce multilevel apparent resistivity measurements whose internal redundancy is rarely quantified. Using 5 years of rolling Wenner-beta acquisition in Ireland, we construct a national-scale database. Building on it, we use an information-theoretic framework to estimate the proportion of measurable values that should, in principle, be inferable from neighbouring levels. To test these predictions, we implement a unified machine-learning pipeline, PCA-based preprocessing followed by supervised regression, and, in parallel, attempt a Monte Carlo data-augmentation scheme. The latter is obtained by fitting a three-component Gaussian mixture model using expectation-maximization to model the joint distribution of inputs and log-transformed targets, from which synthetic samples are generated to stabilize learning in sparsely sampled resistivity ranges. Seven algorithms (Neural Network, Random Forest, Extra Trees, XGBoost, CatBoost, SVM and Decision Tree) consistently achieved a relative MAE below 2%. Across these models, 70%-80% of apparent resistivity values on an independent survey were inferred within 5% error, closely matching the information-theory expectations on redundancy. These results demonstrate that multilevel ERT contains substantial predictable structure governed by regional geology, opening a pathway towards targeted acquisition-time reduction and more efficient near-surface characterization.
This study presents a method that integrates spectral recomposition (SR) with a neural network to improve near-surface seismic analysis. The approach incorporates SR-derived wavelet-timing attributes into a fully convolutional network (FCN) to enhance the characterization of shallow subsurface structures. Field evaluation was conducted using S-wave reflection data acquired near Rotterdam (the Netherlands). The results show good agreement with prior geological interpretations of key stratigraphic interfaces, including the water table (similar to 1 m), alternating Holocene clay-sand layers (2-12 m) and the stiff Pleistocene sand boundary (22-25 m). Compared with conventional approaches, the FCN trained with SR-derived features reduced depth-prediction errors and improved the continuity of interpreted interfaces, particularly in complex near-surface conditions. These results indicate that incorporating SR-derived signal parameters can enhance the robustness and reliability of neural-network-based near-surface velocity model building, although further validation across different geological settings is required.
The joint inversion of gravity and magnetic data presents a powerful solution to mitigate the inherent non-uniqueness of potential field methods. This paper introduces a novel structural coupling approach for joint inversion based on a modified cross-entropy function. The proposed stabilizer minimizes the divergence between the model distributions, thereby enforcing structural correlation, whereas a minimum-entropy condition promotes sparsity to recover focused subsurface models. The non-quadratic functional is transformed into a pseudo-quadratic form, enabling an efficient solution via the reweighted regularized conjugate gradient method. A key advantage of this formulation is its independence from the results of standalone inversions or explicit petrophysical relationships. The method's efficacy is first demonstrated on two different synthetic models, where it yields more compact and structurally coherent models compared to separate inversions, accurately resolving the geometry and physical properties. The technique was subsequently applied to airborne gravity gradiometer and magnetic data from the Nikka volcanogenic massive sulphide deposit in Ontario, Canada. The joint inversion successfully delineated the No. 3 Lens, resolving its lower boundary at approximately 300 m depth with greater precision than individual inversions, which produced smoother and more diffuse anomalies. The results confirm that the modified cross-entropy approach significantly enhances structural resolution and provides a robust tool for integrated geophysical interpretation, particularly in complex geological settings with limited prior information.
How to effectively reconstruct impaired ground penetrating radar (GPR) data has emerged as a crucial research topic for ensuring the detection accuracy. This paper presents a method for reconstructing impaired GPR data based on U-Net++. This method mainly consists of two stages: multi-scale structural similarity index (SSIM) missing positions detection and determination of U-Net++ reconstruction results by the maximum evaluation index method. This method is applicable regardless of whether the location and size of the missing area are known. U-Net++ is built upon the U-Net network structure, and the introduction of dense skip connections and nested subnetworks enhances the feature fusion and fine-grained detail recovery capabilities, thereby improving the reconstruction performance in complex environments. This paper systematically compiles a diverse dataset that includes simulation data and scanned data to conduct the experiment. The experimental results show that, compared with the original structure and some other variants of U-Net networks, U-Net++ achieves better reconstruction performance with approximately the same amount of time consumption. Compared with the interpolation algorithm reconstruction method, using U-Net++ for reconstruction has the advantages of shorter processing time and better results. This advantage is particularly evident in the reconstruction of large-size consecutive GPR trace missing. The method used in this article to determine the location and size of the missing area is currently only applicable to relatively simple scenarios. Further optimization is needed in future research.
Accurate estimation of inverted electrical resistivity values is essential for reliable geophysical exploration, particularly in identifying mineral deposits such as iron ore. Traditional geoelectrical methods, including combined resistivity profiling, gradient, pole-dipole and Schlumberger arrays, are valuable but face challenges such as data interpretation difficulties at unmeasured locations, noise sensitivity and high survey costs. This study applies advanced machine learning (ML) techniques to simultaneously improve the precision and spatial continuity of inverted resistivity estimation and modelling of iron ore deposits. The improved resistivity maps enable more reliable delineation of magnetite-rich zones and clearer discrimination of lithological units, thus supporting 3D ore-body visualization and more targeted drilling and resource evaluation. Consequently, the results have direct implications for exploration efficiency and mine planning by reducing uncertainty in subsurface interpretations. ML models such as random forests (RFs), categorical boost and decision trees are employed to analyse geoelectrical data, with a focus on overcoming issues related to outliers. The generalized extreme studentized deviate (GESD) test is used to identify and remove outliers, and hyperparameters are optimized through grid search and cross-validation to improve model performance. A dataset of 9562 data points from multiple array configurations is used, and model efficacy is evaluated using metrics including root mean square error (RMSE), standard deviation, Nash-Sutcliffe efficiency, mean absolute percentage error and r. Results indicate that the RF model, particularly when trained on GESD-processed data and optimized through grid search and cross-validation, delivers the most accurate inverted resistivity estimates. The RF model achieved an RMSE of 7.23 Omega m and a Pearson correlation coefficient (R) of 0.94, highlighting its robustness. The study demonstrates that integrating ML with geoelectrical data provides a robust framework for extending inverted resistivity coverage in unsampled zones, offering a more efficient and cost-effective approach compared to traditional methods. This research highlights the potential of combining conventional geophysical techniques with advanced data science to enhance subsurface exploration and resource evaluation.
