Empirical Ground-motion Models Research Articles

Generative Adversarial Networks-Based Ground-Motion Model for Crustal Earthquakes in Japan Considering Detailed Site Conditions

ABSTRACT We develop a ground-motion model (GMM) for crustal earthquakes in Japan that can directly model the probability distribution of ground-motion acceleration time histories based on generative adversarial networks (GANs). The proposed model can generate ground motions conditioned on moment magnitude, rupture distance, and detailed site conditions defined by the average shear-wave velocity in the top 5, 10, and 20 m (VS5, VS10, and VS20) and the depth to shear-wave velocities of 1.0 km/s and 1.4 km/s (Z1.0 and Z1.4). We construct the neural networks based on styleGAN2 and introduce a novel neural network architecture to generate ground motions considering the effect of source, path, and such detailed site conditions. The resulting 5% damped spectral acceleration from the proposed GMM is consistent with empirical GMMs in terms of magnitude and distance scaling. The proposed GMM can also generate ground motions accounting for the shear-wave velocity profiles of surface soil with different magnitudes and distances and represent characteristics that are not explained solely byVS30.

Analysis of site-response residuals from empirical ground-motion models to account for observed sedimentary basin effects in Wellington, New Zealand

Analysis of prediction–observation residuals from the empirical ground-motion models (GMMs) used in the 2022 New Zealand National Seismic Hazard Model (NZ NSHM) update indicates a general underprediction of ground motions in the period range of [Formula: see text] s for soft sedimentary basin sites in Wellington. This study uses residual analysis to quantify this underprediction, understand the spatial distribution of these residuals and the specific conditions that cause them, and investigate options for the development of non-ergodic site-response adjustments to the GMMs. All 15 GMMs used in the NZ NSHM were evaluated, and the variability in site-response residuals between different models and different tectonic types of earthquake sources was quantified. Sites are regionalized based on different geomorphic features, such as individual basins and valleys. For example, average site terms are calculated for Te Aro, Thorndon, Miramar, Lower Hutt, Upper Hutt, and several smaller valleys. The period at which maximum underprediction occurs at these sedimentary basin and valley sites was found to correlate well with the fundamental site period of the soil profile [Formula: see text], suggesting improvements can be made to regionalized GMMs by incorporating site period into the site-response prediction for sedimentary basin sites.

High‐Frequency Ground Motions of Earthquakes Correlate With Fault Network Complexity

AbstractUnderstanding the generation of damaging, high‐frequency ground motions during earthquakes is essential both for fundamental science and for effective hazard preparation. Various theories exist regarding the origin of high‐frequency ground motions, including the standard paradigm linked to slip heterogeneity on the rupture plane, and alternative perspectives associated with fault complexity. To assess these competing hypotheses, we measure ground motion amplitudes in different frequency bands for 3 ≤ M ≤ 5.8 earthquakes in Southern California and compare them to empirical ground motion models. We utilize a Bayesian inference technique called the Integrated Nested Laplace Approximation (INLA) to identify earthquake source regions that produce higher or lower ground motions than expected. Our analysis reveals a strong correlation between fault complexity measurements and the high‐frequency ground motion event terms identified by INLA. These findings suggest that earthquakes on complex faults (or fault networks) lead to stronger‐than‐expected ground motions at high frequencies.

A machine learning estimator trained on synthetic data for real-time earthquake ground-shaking predictions in Southern California

After large-magnitude earthquakes, a crucial task for impact assessment is to rapidly and accurately estimate the ground shaking in the affected region. To satisfy real-time constraints, intensity measures are traditionally evaluated with empirical Ground Motion Models that can drastically limit the accuracy of the estimated values. As an alternative, here we present Machine Learning strategies trained on physics-based simulations that require similar evaluation times. We trained and validated the proposed Machine Learning-based Estimator for ground shaking maps with one of the largest existing datasets (<100M simulated seismograms) from CyberShake developed by the Southern California Earthquake Center covering the Los Angeles basin. For a well-tailored synthetic database, our predictions outperform empirical Ground Motion Models provided that the events considered are compatible with the training data. Using the proposed strategy we show significant error reductions not only for synthetic, but also for five real historical earthquakes, relative to empirical Ground Motion Models.

Adjusting an active shallow crustal ground motion model to regions with scarce data: application to France

The objective of this work is to test whether an empirical Ground Motion Model (GMM) developed for high-seismicity regions can be effectively adapted to a neighbouring region with lower seismic activity. We select the ITA18 suite of GMMs (Lanzano et al. in Bull Seismol Soc Am 109(2): 525-540, 2019a), developed for Italy, which is a region dominated by moderate-to-strong shallow crustal earthquakes, and assess their applicability to Continental France, where the seismic activity is less frequent and characterised by lower magnitudes. Based on a dataset of more than 2300 records of events with 3.0 ≤ MW ≤ 5.2, occurred in France in the time interval 1996–2019 (named FR20), we perform a residual analysis and calibrate an adjustment factor for both horizontal and vertical-to-horizontal (VH) components of Peak Ground Acceleration, Peak Ground Velocity, and 5% damped Spectral Acceleration (SA). Apart from the median correction, no modification of the scaling with magnitude, focal mechanism, and VS,30 is introduced, while the distance scaling is adjusted to capture the lower anelastic attenuation of the French data. In addition, to overcome the underestimation of the ITA18 model for the short period VH spectral amplitudes in the near-source region (Repi < 15 km), an additional empirical corrective factor is introduced. In spite of the good agreement of the adjusted model with respect to the median trends of the FR20 dataset, a regionalization of the source effects is introduced to reduce the relatively high between-event variability of the proposed model. The proposed model provides predictions similar to ITA18 in the most seismically active regions (Alps or Pyrenees), while, in the other zones, the predicted amplitudes are richer at high frequencies. Given the paucity of seismic records in these zones, this behavior should be confirmed on the basis of additional data (e.g. physics-based simulations, geologic and tectonic features). The use of the proposed model for hazard applications is recommended within the validity limits of the data (3.0 ≤ MW ≤ 5.2). However, the similarity of the ground motion in the Alps and Pyrenees with the predictions of ITA18 suggests that the adjusted model could be also employed for higher magnitudes, upon suitable checks.

Hybrid Empirical Ground-Motion Models for the Island of Hawaii Based on an Updated Strong Ground-Motion Database

ABSTRACT In this study, we develop ground-motion models (GMMs) for the Island of Hawaii. This area has been the site of several significant earthquake events with a growing database of strong ground-motion observations. Ground-motion modeling on the Island of Hawaii is challenging due to different anelastic attenuation characteristics, the volcanic origin of some of the events, and event depth distribution. Only a few GMMs have been developed for the Island of Hawaii. In this study, we apply a hybrid empirical method (HEM) to develop two separate GMMs for shallow (hypocentral depth ≤ 20 km) and deep (hypocentral depth &amp;gt; 20 km) earthquakes on the Island of Hawaii. We utilize the ratio of the stochastic point-source model in the target and host regions as an appropriate adjustment factor. We apply these adjustment factors to convert the GMMs from the host (western North America) to the target (Island of Hawaii) region. We considered five GMMs proposed in the Next Generation Attenuation Phase 2 project by the Pacific Earthquake Engineering Research Center to model ground motions in the host region. We developed GMMs to predict peak ground acceleration and 5%-damped pseudospectral acceleration at periods T = 0.01–10 s, for moment magnitudes (M) in the range of 3–7.5, and for Joyner–Boore distances in the RJB≤200 km range. The applicability of HEM to develop GMMs for the Island of Hawaii and the growing strong ground-motion data result in further improvements in the capability of GMMs.

Constraining Between-Event Variability of Kinematic Rupture Scenarios by Empirical Ground-Motion Model: A Case Study in Central Italy

ABSTRACT The region of central Italy is well known for its moderate-to-large earthquakes. Events such as 2016 Mw 6.2 Amatrice, generated in the shallow extensional tectonic regime, motivate numerical simulations to gain insights into source-related ground-motion complexities. We utilize a hybrid integral–composite kinematic rupture model by Gallovič and Brokešová (2007) to predict ground motions for other hypothetical Amatrice fault rupture scenarios (scenario events). The synthetic seismograms are computed in 1D crustal velocity models, including region-specific 1D profiles for selected stations up to 10 Hz. We create more than ten thousand rupture scenarios by varying source parameters. The resulting distributions of synthetic spectral accelerations at periods 0.2–2 s agree with the empirical nonergodic ground-motion model of Sgobba et al. (2021) for central Italy in terms of the mean and total variability. However, statistical mixed-effect analysis of the residuals indicates that the between-event variability of the scenarios exceeds the empirical one significantly. We quantify the role of source model parameters in the modeling and demonstrate the pivotal role of the so-called stress parameter that controls high-frequency radiation. We propose restricting the scenario variability to keep the between-event variability within the empirical value. The presented validation of the scenario variability can be generally utilized in scenario modeling for more realistic physics-based seismic hazard assessment.

Development of an empirical ground-motion model for post-mining induced seismicity near Gardanne, France

Development of an empirical ground-motion model for post-mining induced seismicity near Gardanne, France

Ground-Motion Variability for Ruptures on Rough Faults

ABSTRACT Fault roughness influences earthquake rupture dynamics, seismic energy radiation, and, hence, resulting ground motion and its variability. Using 3D dynamic rupture simulations considering a range of rough-fault realizations, we investigate the effects of rupture complexity caused by fault roughness on ground-motion variability, that is, the variability of peak ground acceleration (PGA) and velocity (PGV) as a function of distance. In our analysis, we vary hypocenter locations (leading to unilateral and bilateral ruptures) and fault roughness amplitude to generate a set of magnitude M ≈ 7 strike-slip dynamic rupture simulations. Synthetic seismic waveforms computed on a dense set of surface sites (maximum resolved frequency 5.75 Hz) form our database for detailed statistical analyses. For unilateral ruptures, our simulations reveal that ground-shaking variability (in terms of PGA and PGV) remains nearly constant with increasing distance from the fault. In contrast, bilateral ruptures lead to slowly decreasing ground-motion variability with increasing distance in the near field (less than 20 km). The variability becomes almost constant at large fault distances. We also find that low-amplitude fault roughness leads to ruptures that are likely to generate higher PGA variability than events on faults with high-amplitude roughness. Increasing fault roughness distorts the radiation pattern, thereby reducing directivity effects and, hence, potentially lowering ground-motion variability. The average PGV variability from our rough-fault rupture models is consistent with estimates from empirical ground-motion models (GMMs). However, the average PGA variability exceeds the variability encoded in empirical GMMs by nearly 20%. Hence, our findings have implications for near-source ground-motion prediction in seismic hazard studies, because ground-motion variability depends on details of the earthquake rupture process and is larger than GMM estimates.

Combining observed linear basin amplification factors with 1D nonlinear site-response analyses to predict site response for strong ground motions: Application to Wellington, New Zealand

This study develops a method for estimating site amplification that combines instrumentally observed site-specific amplification factors with adjustment factors from nonlinear site-response analyses. This approach provides estimates of site response for large-strain motions based on observations and sophisticated nonlinear modeling. A database of weak-to-moderate intensity ground motions recorded in three basins of Wellington, New Zealand is used to study the observed site amplification. A subset of nine strong-motion stations was selected to perform nonlinear site-response analyses with scaled strong ground motions to assess the influence of nonlinearity on site amplification factors and demonstrate the approach. Different shear-wave velocity ( VS) profiles, constitutive models, and modeling approaches (e.g. one-dimensional (1D) site-response analyses vs empirical [Formula: see text]-based approaches) are used to quantify the sensitivity and modeling uncertainty in the nonlinear site-response analyses. It was found that for soft sites subjected to strong ground motions, there may be a decrease in spectral acceleration amplification factors for periods up to approximately 2 s, relative to the expected linear site response. For longer periods, there is little to no amplification from the effects of soil nonlinearity. However, at stiffer sites, which generally experience less basin amplification in observations, there may be moderate amplification at longer periods when nonlinearity is considered due to softening of the soil profile. Empirical ground-motion models were found to under-represent the observed amplification between basin sites and the nearby reference site, especially at intermediate to long periods, corresponding to resonant frequencies of these basin sites. In addition, the empirical nonlinear site amplification models ([Formula: see text]-based) were found to deviate from nonlinear analyses at large strains, where such models are poorly constrained due to such a limited number of observations.

Evaluation of Empirical Ground-Motion Models for the 2022 New Zealand National Seismic Hazard Model Revision

ABSTRACT This article presents an evaluation of empirical ground-motion models (GMMs) for active shallow crustal, subduction interface, and subduction slab earthquakes using a recently developed New Zealand (NZ) ground-motion database for the 2022 New Zealand National Seismic Hazard Model revision. This study considers both NZ-specific and global models, which require evaluation to inform of their applicability in an NZ context. A quantitative comparison between the models is conducted based on intensity measure residuals and a mixed-effects regression framework. The results are subsequently investigated to assess how the models are performing in terms of overall accuracy and precision, as well as to identify the presence of any biases in the model predictions when applied to NZ data. Many models showed reasonable performance and could be considered appropriate for inclusion within suites of models to properly represent ground-motion predictions and epistemic uncertainty. In general, the recent models that are NZ-specific or developed on large international databases performed the best. This evaluation of models helped inform suitable GMMs for the ground-motion characterization model logic tree. In addition, spatial trends in systematic site-to-site residuals to the west of the Taupō Volcanic Zone demonstrated the need for backarc attenuation modifications for slab earthquakes.

Development of the Site Characterization Database for the 2022 New Zealand National Seismic Hazard Model

Abstract This article presents the development of the site characterization database for the 2022 New Zealand National Seismic Hazard Model update. This database summarizes the site characterization parameters at past and present GeoNet seismic monitoring network instrument locations, including strong-motion, short-period, and broadband seismometer stations. Site characterization parameters required to assess and improve empirical ground-motion models and those used in codified seismic design frameworks internationally have been included in the database. Measurement uncertainty was assigned, and the quality of the data used to assign each parameter was classified. The site period (T0) was the most well constrained of all the site parameters, with almost half of the database classified based on high-quality measurements, with these dominated by microtremor-based horizontal-to-vertical spectral ratio. Although there was an improvement in the quality of the parameters representing the time-averaged shear-wave velocity in the uppermost 30 m of the profile (VS30), little site-specific data were available, with almost no information for rock sites. Most of these classifications were based on national maps or geologic interpretation. Depth-based parameters (Z1.0 and Z2.5) had the lowest quality overall, with very few direct measurements available to constrain these values. Despite these limitations, the quality of parameters assigned to instrument locations has improved and greatly expanded previous databases through the assignment of parameter values to the entire GeoNet seismic network.

Variability of response spectra of single-path ground-motion record pairs

To improve the accuracy of seismic hazard analysis, understanding the uncertainty of the prediction of the ground-motion amplitude for a specific site owing to a specific source is important. In this study, the variability of the acceleration response spectra of single-path ground motion was evaluated using record pairs obtained from two earthquakes at the same site with the same magnitude, location, and focal mechanism, recorded by high-density strong-motion observation networks in Japan. The natural log standard deviations of the single-path ground motions varied with the period and were distributed in the range of approximately 0.3–0.5. Even if an empirical ground-motion prediction model that accurately models the characteristics of single-path ground motions could be developed, avoiding this variability in the prediction is difficult. The standard deviation of the within-event residuals was smaller at larger distances, and the period of the maximum standard deviation tended to increase as the earthquake magnitude increased. These features may be attributed to differences in the source characteristics of each earthquake. These findings will be useful for the development of advanced empirical ground-motion models and for improving the accuracy of seismic hazard analysis.

3D Wave Propagation Simulations of Mw 6.5+ Earthquakes on the Tacoma Fault, Washington State, Considering the Effects of Topography, a Geotechnical Gradient, and a Fault Damage Zone

ABSTRACT We simulate shaking in Tacoma, Washington, and surrounding areas from Mw 6.5 and 7.0 earthquakes on the Tacoma fault. Ground motions are directly modeled up to 2.5 Hz using kinematic, finite-fault sources; a 3D seismic velocity model considering regional geology; and a model mesh with 30 m sampling at the ground surface. In addition, we explore how adjustments to the seismic velocity model affect predicted shaking over a range of periods. These adjustments include the addition of a region-specific geotechnical gradient, surface topography, and a fault damage zone. We find that the simulated shaking tends to be near estimates from empirical ground-motion models (GMMs). However, long-period (T = 5.0 s) shaking within the Tacoma basin is typically underpredicted by the GMMs. The fit between simulated and GMM-derived short-period (T = 0.5 s) shaking is significantly improved with the addition of the geotechnical gradient. From comparing different Mw 6.5 earthquake scenarios, we also find that the response of the Tacoma basin is sensitive to the azimuth of incoming seismic waves. In adding surface topography to the simulation, we find that average ground motion is similar to that produced from the nontopography model. However, shaking is often amplified at topographic highs and deamplified at topographic lows, and the wavefield undergoes extensive scattering. Adding a fault damage zone has the effect of amplifying short-period shaking adjacent to the fault, while reducing far-field shaking. Intermediate-period shaking is amplified within the Tacoma basin, likely due to enhanced surface-wave generation attributable to the fault damage zone waveguide. When applied in the same model, the topography and fault damage zone adjustments often enhance or reduce the effects of one another, adding further complexity to the wavefield. These results emphasize the importance of improving near-surface velocity model resolution as waveform simulations progress toward higher frequencies.

Broadband ground motion simulation using a hybrid approach of the May 21, 2021 M7.4 earthquake in Maduo, Qinghai, China

In this study, the broadband ground motions of the 2021 M7.4 Maduo earthquake were simulated to overcome the scarcity of ground motion recordings and the low resolution of macroseismic intensity map in sparsely populated high-altitude regions. The simulation was conducted with a hybrid methodology, combining a stochastic high-frequency simulation with a low-frequency ground motion simulation, from the regional 1-D velocity structure model and the Wang WM et al. (2022) source rupture model, respectively. We found that the three-component waveforms simulated for specific stations matched the waveforms recorded at those stations, in terms of amplitude, duration, and frequency content. The validation results demonstrate the ability of the hybrid simulation method to reproduce the main characteristics of the observed ground motions for the 2021 Maduo earthquake over a broad frequency range. Our simulations suggest that the official map of macroseismic intensity tends to overestimate shaking by one intensity unit. Comparisons of simulations with empirical ground motion models indicate generally good consistency between the simulated and empirically predicted intensity measures. The high-frequency components of ground motions were found to be more prominent, while the low-frequency components were not, which is unexpected for large earthquakes. Our simulations provide valuable insight into the effects of source complexity on the level and variability of the resulting ground motions. The acceleration and velocity time histories and corresponding response spectra were provided for selected representative sites where no records were available. The simulated results have important implications for evaluating the performance of engineering structures in the epicentral regions of this earthquake and for estimating seismic hazards in the Tibetan regions where no strong ground motion records are available for large earthquakes.

Dynamic Rupture Models, Fault Interaction and Ground Motion Simulations for the Segmented Húsavík‐Flatey Fault Zone, Northern Iceland

AbstractThe Húsavík‐Flatey Fault Zone (HFFZ) is the largest strike‐slip fault in Iceland and poses a high seismic risk to coastal communities. To investigate physics‐based constraints on earthquake hazards, we construct three fault system models of varying geometric complexity and model 79 3‐D multi‐fault dynamic rupture scenarios in the HFFZ. By assuming a simple regional prestress and varying hypocenter locations, we analyze the rupture dynamics, fault interactions, and the associated ground motions up to 2.5 Hz. All models account for regional seismotectonics, topo‐bathymetry, 3‐D subsurface velocity, viscoelastic attenuation, and off‐fault plasticity, and we explore the effect of fault roughness. The rupture scenarios obey earthquake scaling relations and predict magnitudes comparable to those of historical events. We show how fault system geometry and segmentation, hypocenter location, and prestress can affect the potential for rupture cascading, leading to varying slip distributions across different portions of the fault system. Our earthquake scenarios yield spatially heterogeneous near‐field ground motions modulated by geometric complexities, topography, and rupture directivity, particularly in the near‐field. The average ground motion attenuation characteristics of dynamic rupture scenarios of comparable magnitudes and mean stress drop are independent of variations in source complexity, magnitude‐consistent and in good agreement with the latest regional empirical ground motion models. However, physics‐based ground motion variability changes considerably with fault‐distance and increases for unilateral compared to bilateral ruptures. Systematic variations in physics‐based near‐fault ground motions provide important insights into the mechanics and potential earthquake hazard of large strike‐slip fault systems, such as the HFFZ.

Empirical ground-motion models (GMMs) and associated correlations for cumulative absolute velocity, Arias intensity, and significant durations calibrated on Iranian strong motion database

Empirical ground-motion models (GMMs) and associated correlations for cumulative absolute velocity, Arias intensity, and significant durations calibrated on Iranian strong motion database

A non-ergodic spectral acceleration ground motion model for California developed with random vibration theory

A new approach for creating a non-ergodic pseudo-spectral acceleration (PSA) ground-motion model (GMM) is presented, which accounts for the magnitude dependence of the non-ergodic effects. In this approach, the average PSA scaling is controlled by an ergodic PSA GMM, and the non-ergodic effects are captured with non-ergodic PSA factors, which are the adjustment that needs to be applied to an ergodic PSA GMM to incorporate the non-ergodic effects. The non-ergodic PSA factors are based on the effective amplitude spectrum (EAS) non-ergodic effects and are converted to PSA through Random Vibration Theory (RVT). The advantage of this approach is that it better captures the non-ergodic source, path, and site effects through small-magnitude earthquakes. Due to the linear properties of the Fourier Transform, the EAS non-ergodic effects of the small events can be applied directly to the large magnitude events. This is not the case for PSA, as response spectra are controlled by a range of frequencies, making PSA non-ergodic effects dependent on the spectral shape, which in turn is magnitude-dependent. Two PSA non-ergodic GMMs are derived using the ASK14 (Abrahamson et al. in Earthq Spectra 30:1025–1055, 2014) and CY14 (Chiou and Youngs in Earthq Spectra 30:1117–1153, 2014) GMMs as backbone models, respectively. The non-ergodic EAS effects are estimated with the LAK21 (Lavrentiadis et al. in Bull Earthq Eng ) GMM. The RVT calculations are performed with the V75 (Vanmarcke in ASCE Mech Eng Mech Division 98:425–446, 1972) peak factor model, the D_{a0.05-0.85} estimate of AS96 (Abrahamson and Silva in Apendix A: empirical ground motion models, description and validation of the stochastic ground motion model. Tech. rep.,. Brookhaven National Laboratory, New York) for the ground-motion duration, and BT15 (Boore and Thompson in Bull Seismol Soc Am 105:1029–1041, 2015) oscillator-duration model. The California subset of the NGAWest2 database (Ancheta et al. in Earthq Spectra 30:989–1005, 2014) is used to fit both models. The total aleatory standard deviation of each of the two non-ergodic PSA GMMs is approximately 25% smaller than the total aleatory standard deviation of the corresponding ergodic PSA GMMs. This reduction has a significant impact on hazard calculations at large return periods. In remote areas, far from stations and past events, the reduction of aleatory variability is accompanied by an increase in epistemic uncertainty.

Data-driven zonations for modelling the regional source and propagation effects into a Ground Motion Models in Italy

The main objective of this work is to provide two zonations of Italy useful to incorporate regional source and path effects in empirical Ground Motion Models (GMMs). To this end, we revise existing zonations developed for seismic and tsunami hazard studies, based on the results of a residual analysis between the observations of the ITACAext dataset and the predictions of ITA18, a GMM recently developed for Italy. The analysis consists in the decomposition of the residuals into repeatable terms, according to the well-established approach used for the calibration of non-ergodic models. Based on the spatial trend of the interpolated residuals, and geological and seismological considerations, the zonations are proposed and discussed, in the perspective of supporting regionalization of a future generation of GMMs for Italy.

U.S. National VS30 Models and Maps Informed by Remote Sensing and Machine Learning

Abstract The shear-wave velocity time averaged over the upper 30 m (VS30) is widely used as a proxy for site effects, forms the basis of seismic site class, and underpins site-amplification factors in empirical ground-motion models. Many earthquake simulations, therefore, require VS30. This presents a challenge at regional scale, given the infeasibility of subsurface testing over vast areas. Although various models for predicting VS30 have thus been proposed, the most popular U.S. national, or “background,” model is a regression equation based on just one variable. Given the growth of community data sets, satellite remote sensing, and algorithmic learning, more advanced and accurate solutions may be possible. Toward that end, we develop national VS30 models and maps using field data from over 7000 sites and machine learning (ML), wherein up to 17 geospatial parameters are used to predict subsurface conditions (i.e., VS30). Of the two models developed, that using geologic data performs marginally better, yet such data are not always available. Both models significantly outperform existing solutions in unbiased testing and are used to create new VS30 maps at ∼220 m resolution. These maps are updated in the vicinity of field measurements using regression kriging and cover the 50 U.S. states and Puerto Rico. Ultimately, and like any model, performance cannot be known where data is sparse. In this regard, alternative maps that use other models are proposed for steep slopes. More broadly, this study demonstrates the utility of ML for inferring below-ground conditions from geospatial data, a technique that could be applied to other data and objectives.