arxiv: 2604.14756 · v1 · submitted 2026-04-16 · ⚛️ physics.geo-ph

Recognition: unknown

Subsurface Property Mapping using Google AlphaEarth Foundations

Nori Nakata , Jingxiao Liu , Guodong Chen , Rie Nakata , Charuleka Varadharajan

Authors on Pith no claims yet

Pith reviewed 2026-05-10 08:42 UTC · model grok-4.3

classification ⚛️ physics.geo-ph

keywords AlphaEarth embeddingssubsurface mappingVS30seismic site characterizationsubsurface temperaturefoundation modelsregressionsurface-to-subsurface inference

0 comments

The pith

AlphaEarth surface embeddings support estimation of subsurface properties from indirect surface observations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines whether embeddings from the AlphaEarth foundation model can estimate subsurface properties despite only indirect and non-unique links to surface features. It tests the approach in two applications across the conterminous United States: predicting shallow shear-wave velocity for seismic site characterization using embeddings alone or with topographic and tectonic covariates, and reconstructing subsurface temperatures via nonlinear regression on the embeddings. In both cases the models produce spatially coherent and physically plausible maps that outperform simpler baselines. This approach matters because direct subsurface measurements remain sparse and uneven, limiting hazard assessment, energy planning, and infrastructure work. The results indicate that foundation-model surface representations can enable regional surface-to-subsurface inference while underscoring the need for strong spatial validation under heterogeneous data conditions.

Core claim

We test this idea in two conterminous U.S. applications: shallow seismic site characterization using VS30 with embedding features alone and with conventional covariates (topographic slope and a tectonic-status indicator), and subsurface temperature reconstruction using embedding-based nonlinear regression. Across both applications, embedding-informed models recover spatially coherent, physically plausible patterns and outperform simpler baselines. The comparison also highlights a key difference: domain covariates materially stabilize VS30 regression, whereas temperature mapping relies primarily on embedding features. Overall, the results support the feasibility of foundation-model surface

What carries the argument

AlphaEarth embeddings used as input features for regression models that predict subsurface properties such as VS30 and temperature.

If this is right

Embedding features recover coherent patterns in both VS30 and temperature mapping.
Conventional covariates improve stability for VS30 estimates but are not required for temperature reconstruction.
Foundation-model surface representations enable regional inference from surface data to subsurface properties.
Robust spatial validation remains essential given uneven label distribution and data coverage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The approach may extend to additional subsurface variables that share indirect surface signals.
Global application becomes feasible if similar embeddings exist for other regions.
Hybrid models that blend embeddings with physical constraints could further reduce errors in data-sparse zones.

Load-bearing premise

The indirect and non-unique physical links between AlphaEarth surface embeddings and subsurface properties are sufficient to produce reliable estimates despite sparse, uneven labels and heterogeneous data coverage.

What would settle it

A region with dense independent subsurface observations where embedding-based models produce incoherent patterns or show no improvement over baselines would challenge the central claim.

Figures

Figures reproduced from arXiv: 2604.14756 by Charuleka Varadharajan, Guodong Chen, Jingxiao Liu, Nori Nakata, Rie Nakata.

**Figure 1.** Figure 1: Overview of the VS30 dataset, model performance, national prediction, and model interpretation. (a) Spatial distribution of the measured VS30 stations used in this study across the conterminous United States, colored by observed VS30, with the west-east tectonic boundary at 105◦W. (b) Comparison between predicted and measured VS30 for the held-out test set, shown on log-log axes with the 1:1 line. (c) Pred… view at source ↗

**Figure 2.** Figure 2: (a) Depth spatial distribution of boreholes from the database. (b) Observed bottom-hole temperature distribution of boreholes from the database. (c) Observed versus predicted bottom-hole temperature distribution for the machine learning method without AlphaEarth embeddings. (d) Observed versus predicted bottom-hole temperature distribution for the machine learning method with AlphaEarth embeddings. Applyin… view at source ↗

**Figure 3.** Figure 3: Subsurface temperature reconstruction using AlphaEarth foundation model enabled neural network with [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗

read the original abstract

Subsurface properties are essential for hazard assessment, energy and environmental management, and infrastructure resilience, but direct observations are sparse and uneven, motivating the use of surface observations as indirect constraints. Here we explore whether AlphaEarth embeddings can be applied to subsurface estimation despite indirect and non-unique physical links between surface and depth. We test this idea in two conterminous U.S. applications: shallow seismic site characterization using $V_S 30$ with embedding features alone and with conventional covariates (topographic slope and a tectonic-status indicator), and subsurface temperature reconstruction using embedding-based nonlinear regression. Across both applications, embedding-informed models recover spatially coherent, physically plausible patterns and outperform simpler baselines. The comparison also highlights a key difference: domain covariates materially stabilize $V_S 30$ regression, whereas temperature mapping relies primarily on embedding features. Overall, the results support the feasibility of foundation-model surface representations for regional surface-to-subsurface inference, while emphasizing the need for robust spatial validation under heterogeneous labels and uneven data coverage.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

AlphaEarth embeddings produce plausible subsurface patterns in two US applications but the validation is too light to confirm real gains over baselines.

read the letter

The paper's main point is that AlphaEarth foundation-model embeddings can serve as features for estimating Vs30 and subsurface temperatures across the conterminous US, recovering coherent maps that beat simpler baselines in both cases. They also note a useful split: conventional covariates like slope and tectonics help stabilize the Vs30 results, while temperature mapping leans mostly on the embeddings themselves. That distinction comes through clearly and feels like the most concrete takeaway. The work is new in applying this specific model to these subsurface tasks with the reported comparisons, and it does a reasonable job framing the indirect surface-to-depth link as a starting point rather than a solved problem. The patterns look physically sensible on the surface, which is a low bar but still worth showing for an exploratory study. The soft spots are mostly around validation and reporting. The abstract claims outperformance without any numbers, error bars, or details on how the data were split or held out. In geophysics, observations are often clustered, so ordinary random or stratified splits can leak signal from nearby points and inflate apparent improvements. The heterogeneous coverage they mention makes this issue worse, and nothing indicates they ran spatial block or distance-based cross-validation to check it. That leaves the central claim resting on patterns that might not hold up under stricter tests. This is for geophysicists or hazard researchers who want to see whether general foundation models can transfer to depth inference without heavy domain engineering. A reader already working on ML for earth data might find the applications interesting but will want the missing quantitative checks before treating it as a method. It deserves a serious referee because the idea is timely and the applications matter for real problems, even if the current evidence is preliminary. Send it to review with instructions to add proper spatial validation, full metrics, and clearer data-handling rules.

Referee Report

2 major / 1 minor

Summary. The paper explores the use of Google AlphaEarth foundation-model surface embeddings as features for subsurface property estimation in two conterminous-U.S. applications: Vs30-based seismic site characterization (with and without topographic slope and tectonic covariates) and nonlinear regression for subsurface temperature. It reports that embedding-informed models recover spatially coherent, physically plausible patterns and outperform simpler baselines, while noting that domain covariates stabilize Vs30 regression but temperature mapping relies primarily on embeddings. The work concludes that the results support feasibility of foundation-model representations for regional surface-to-subsurface inference, subject to the need for robust spatial validation under heterogeneous labels.

Significance. If the outperformance and spatial coherence hold under proper spatial cross-validation, the approach could offer a scalable way to leverage abundant surface embeddings for sparse subsurface mapping, with potential value for seismic hazard, geothermal, and infrastructure applications. The explicit baseline comparisons and acknowledgment of indirect physical links are strengths, but the absence of quantitative metrics, error bars, and validation details currently limits the strength of the central claim.

major comments (2)

[Abstract] Abstract: the claim of outperformance and spatially coherent patterns is presented without any quantitative metrics, error bars, cross-validation details, or data-exclusion rules. Given the heterogeneous coverage and clustered nature of Vs30 and temperature observations, this omission leaves open whether reported gains reflect genuine embedding utility or spatial leakage from non-spatial splits.
[Abstract and Results] The central feasibility claim requires that embedding features produce reliable subsurface estimates despite indirect physical links and sparse labels. This hinges on the outperformance being robust to spatial structure; without block or distance-based cross-validation (standard for geographically clustered geophysical data), the spatially coherent patterns could be artifacts rather than evidence of useful mapping.

minor comments (1)

[Abstract] The abstract mentions 'nonlinear regression' for temperature but does not specify the model family or regularization; adding this would clarify how embedding features are used.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight important aspects of validation for geophysical mapping tasks. We address the concerns regarding quantitative metrics and spatial cross-validation below, and have made revisions to strengthen the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the claim of outperformance and spatially coherent patterns is presented without any quantitative metrics, error bars, cross-validation details, or data-exclusion rules. Given the heterogeneous coverage and clustered nature of Vs30 and temperature observations, this omission leaves open whether reported gains reflect genuine embedding utility or spatial leakage from non-spatial splits.

Authors: We agree that the abstract would benefit from including quantitative metrics to support the claims. In the revised manuscript, we have updated the abstract to report key performance indicators, such as the improvement in R² and reduction in RMSE when using AlphaEarth embeddings compared to baselines for both Vs30 and temperature predictions. We have also added a brief mention of the cross-validation strategy and data selection criteria. Full details, including error bars from multiple runs, are provided in the results section and supplementary information. revision: yes
Referee: [Abstract and Results] The central feasibility claim requires that embedding features produce reliable subsurface estimates despite indirect physical links and sparse labels. This hinges on the outperformance being robust to spatial structure; without block or distance-based cross-validation (standard for geographically clustered geophysical data), the spatially coherent patterns could be artifacts rather than evidence of useful mapping.

Authors: This is a valid concern for spatially autocorrelated data. While the original manuscript used random k-fold cross-validation, we acknowledge that this may not fully account for spatial clustering. In the revision, we have implemented spatial block cross-validation by dividing the conterminous U.S. into geographic blocks and ensuring training and test sets are spatially separated. The outperformance of embedding-based models persists under this more rigorous validation, supporting the feasibility claim. We have added these results, along with visualizations of the block partitions, to the revised Results section. revision: yes

Circularity Check

0 steps flagged

No circularity: embeddings treated as fixed external features in regression

full rationale

The paper treats AlphaEarth embeddings as pre-trained, fixed inputs to standard regression models for Vs30 and subsurface temperature. It explicitly compares embedding-informed models against simpler baselines (slope, tectonic indicator) and evaluates against independent observations. No derivation step reduces by construction to its own inputs, no fitted parameter is relabeled as a prediction, and no load-bearing claim rests on self-citation chains or ansatzes imported from prior author work. The central claim is therefore an empirical performance comparison rather than a self-referential derivation.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that foundation-model embeddings encode transferable information about subsurface properties despite indirect physical links. No free parameters or invented entities are introduced in the abstract.

axioms (1)

domain assumption Surface embeddings from AlphaEarth capture relevant information for subsurface properties despite indirect and non-unique physical links
Invoked to justify testing the embeddings for Vs30 and temperature estimation.

pith-pipeline@v0.9.0 · 5480 in / 1207 out tokens · 24513 ms · 2026-05-10T08:42:23.797614+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 2 canonical work pages

[1]

Accessed 2026-03-15

URL https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_SATELLITE_ EMBEDDING_V1_ANNUAL. Accessed 2026-03-15. C. Jin, X. Jiang, L. Wen, C. Wu, X. Xu, and J. Jiao. Assessing the utility of satellite embedding features for biomass prediction in subtropical forests with machine learning.Remote Sensing, 18(3):436, 2026. doi: 10.3390/rs18030436. ...

work page doi:10.3390/rs18030436 2026
[2]

doi: 10.1785/0120100262. 12

work page doi:10.1785/0120100262