TerraProbe: Multidomain Probing Framework

• TerraProbe is a research motif defined by integrating heterogeneous sensing, structured data encoding, and latent inference across domains like soil analytics, terramechanics, and geophysical surveys.
• It combines methodologies such as GIS-based predictive modeling, empirical simulations, transformer-driven subsurface reconstruction, and layered oracle frameworks to robustly handle operational uncertainty.
• Practical applications include terrain mobility estimation, soil property prediction, and detecting deceptive fixes in Terraform repairs, underscoring its broad impact in geosciences and software engineering.

TerraProbe denotes both a concrete evaluation framework and a broader class of probing systems that recur across recent research in geospatial soil analytics, terramechanics, planetary geophysics, probabilistic subsurface inversion, and terrain-aware robotic mapping. In the narrowest sense, TerraProbe is a five-layer oracle framework for evaluating LLM-assisted Terraform security repair (Alsaid et al., 25 Jun 2026). In a wider, inferential sense, several adjacent works describe “TerraProbe-like” platforms that collect heterogeneous observations, encode them in structured representations, and infer latent environmental properties such as soil characteristics, slip and sinkage, thermal inertia, subsurface conductivity, or terrain semantics (Piccoli et al., 2022, Kern et al., 8 Jan 2026, Grimm et al., 2021, Mazumder et al., 2 Sep 2025).

1. Scope and recurrent structure

The literature does not present a single canonical TerraProbe implementation across all domains. Instead, the name appears explicitly in empirical software engineering and implicitly in a set of systems whose function is probing under heterogeneity, sparse observations, and operational uncertainty. This suggests that TerraProbe is best understood as a research motif rather than a standardized platform.

Domain	Representative system	Probing target
Precision agriculture	Pignoletto	Soil characteristics from GIS, laboratory, airborne, and spaceborne data
Planetary mobility	Regression terramechanics and self-burrowing probes	Slip, sinkage, penetration, and terrain response
Planetary geophysics	EMS and VL2SP	Intrashell water, oceans, seismicity, and interior structure
Subsurface and terrain AI	Transparent Earth, NPE radar inversion, Terra 3DSG	Subsurface fields, terrain parameters, and outdoor semantic structure
Empirical software engineering	TerraProbe	Deceptive fixes in LLM-assisted Terraform repair

Across these works, the recurrent pipeline is structurally similar: acquisition of heterogeneous measurements, transformation into an intermediate representation, inference over latent properties, and operational use in planning, interpretation, or validation. The intermediate representation varies by domain—PostGIS layers, apparent conductivity curves, transformer latents, normalizing-flow posteriors, hierarchical scene graphs, or oracle-layer outcomes—but the architectural pattern is stable.

2. GIS-centered soil analytics and precision agriculture

A direct precursor to a soil-oriented TerraProbe is the unified GIS platform Pignoletto, which was designed for the collection, visualization, management and analysis of heterogeneous and multisource data for soil characteristics estimation and precision agriculture (Piccoli et al., 2022). Its architecture comprises a Map web client based on QGIS Server and Lizmap, a Flask-based Tabular client, a RESTful API, a PostgreSQL/PostGIS data layer, and a middle layer that dynamically creates QGIS projects, exposes CRUD operations, and manages predictive workflows. Raster data are stored as pyramidal views, vector data include laboratory and drone samples, and interoperability is provided through WMS and WFS.

The platform formalizes laboratory and drone observations as

$$\mathcal{L} = \{(h_i, p_i)\}_{i=1}^{N}, \qquad \mathcal{D} = \{h_j\}_{j=1}^{M},$$

with a predictive model

$M(h) = \hat{p}.$

Here, $h$ is a hyperspectral measurement vector and $p$ is a soil-property vector. This representation supports the integration of proximity, airborne, and spaceborne data, including hyperspectral, thermal multispectral, ionizing gamma radiation, PRISMA, and Copernicus products. Administrators can upload custom predictive systems, and model execution is exposed through Lizmap WPS tools. Usability was assessed through a Task-Driven Survey and the System Usability Scale, with tasks such as displaying “Argilla” acquisitions, filtering samples by Argilla ranges, and visualizing their map locations; the reported outcome was that tasks were generally perceived as easy and the overall usability was good (Piccoli et al., 2022).

A complementary non-contact TerraProbe logic appears in thermal-vision soil assessment under Martian conditions (Castilla-Arquillo et al., 2023). In that framework, a multipurpose environmental chamber reproduces Earth-like and Mars-like pressure conditions, specifically 1000 mbar and 8 mbar with 95% $\mathrm{CO_2}$, while a Long-Wave Infrared camera measures representative diurnal thermal cycles. Thermal inertia is defined as

$I = \sqrt{k\,\rho\,c},$

and a sinusoidal estimator is used: $$I_{\text{sin}} = \frac{\Delta G_s}{\Delta T_s}\sqrt{\frac{2\pi}{P}}.$$ Four samples—Bedrock, Soil A, Soil B, and Soil C—were studied, and the resulting paired datasets comprise 9225 radiometric thermal images together with heater, air, and subsurface temperatures. At Mars-like pressure, discriminability across granularities increases markedly: bedrock exhibits the highest $I_{\text{sin}}$, Soil C the lowest, and the reported trend is consistent with Mars rover measurements (Castilla-Arquillo et al., 2023). In practical terms, this supplies a TerraProbe-style remote pathway from thermal behavior to terrain class and, plausibly, to trafficability-relevant soil state.

3. Mobility, burrowing, and in-situ geotechnical probing

A second major TerraProbe lineage concerns the direct estimation of terrain mechanical behavior. In a real-time lunar-rover simulator, terramechanics effects are represented by empirical regressions rather than by DEM or SCM. Slip is modeled as

$s(v_w) = 0.0265\, v_w + 0.0256,$

with slope dependence

$$s(v_w,\alpha) = (0.00522\, v_w + 0.00105)\,\alpha^2 + s(v_w),$$

and sinkage as

$M(h) = \hat{p}.$0

These models are embedded in NVIDIA IsaacSim/OmniLRS at a 30 Hz physics update rate, using full-rover field tests, single-wheel experiments, and DEM simulations as training and validation sources. The simulator reproduces steady-state and dynamic slip and sinkage on flat terrain and slopes up to $M(h) = \hat{p}.$1, with implementation-verification errors reported as slip ratio error $M(h) = \hat{p}.$2 on flat terrain, mean absolute slip error $M(h) = \hat{p}.$3 on slopes, and sinkage error $M(h) = \hat{p}.$4 mm within the tested load range (Kern et al., 8 Jan 2026). In effect, this is a mobility-probing TerraProbe: it does not merely render terrain, but estimates how terrain will alter commanded motion.

The same probing logic appears in a razor-clam–inspired self-burrowing probe for granular soils under variable gravity (Wang et al., 2024). The probe uses a dual-anchor cycle—shaft expansion, tip penetration with oscillation, tip anchor expansion, shaft contraction, shaft retraction, and tip anchor contraction—and was studied in 3D DEM at $M(h) = \hat{p}.$5, $M(h) = \hat{p}.$6, $M(h) = \hat{p}.$7, $M(h) = \hat{p}.$8, $M(h) = \hat{p}.$9, and $h$0. Total work is decomposed as

$h$1

and penetration efficiency as

$h$2

The reported behavior is sharply structured: penetration per cycle is about 2 cm at low gravity and about 4 cm at high gravity, total work scales approximately linearly with gravity, and energy efficiency is higher in low gravity despite the smaller absolute penetration distance (Wang et al., 2024). This gives TerraProbe a mechanical interpretation in which probing is itself a locomotion strategy.

For direct geotechnical characterization of lunar regolith analogs, a third strand evaluates small field instruments and finds that most are inadequate for low-cohesion, frictional soils. The modified shear vane tester is the exception: the study concludes that a modified version allows extraction of several important soil parameters and may be useful on the lunar surface by astronauts or a robotic lander (Rahmatian et al., 2023). The analysis further reports that JSC-1A does not behave mechanically like the higher-fidelity simulants NU-LHT-2M and CHENOBI, probably because its particle shapes are more rounded, whereas BP-1 behaves very similarly to NU-LHT-2M and CHENOBI and has a natural particle size distribution similar to that of lunar soil (Rahmatian et al., 2023). In TerraProbe terms, this is a minimalistic in-situ geotechnical laboratory.

4. Planetary electromagnetic and long-life geophysical packages

A planetary geophysical TerraProbe is exemplified by the Europa Magnetotelluric Sounder, a single-station magnetotelluric instrument intended to detect water within 30 km as well as characterize Europa’s subsurface ocean (Grimm et al., 2021). The method relies on the large conductivity contrast between slightly salty water and ice and uses the MT relation between horizontal electric and magnetic fields. Depth sensitivity is governed by skin depth,

$h$3

and the target frequency range for intrashell water is approximately $h$4–$h$5 Hz. EMS comprises central electronics, a fluxgate magnetometer on a mast, and three ballistically deployed electrodes that measure differences in surface electric potential. The prototype was tested in relevant thermal, vacuum, and radiation environments, and the instrument is reported to be capable of uniquely determining the occurrence of intrashell water on Europa (Grimm et al., 2021). Kinematic source periods of roughly 3–700 hr constrain the ocean, while higher-frequency magnetospheric waves probe shallower depths.

A second geophysical embodiment is the Venus Long-Life Surface Package, a compact RTG-powered, high-temperature surface package designed to survive directly in the Venus surface environment of approximately $h$6C and 92 bar for months to years (Wilson et al., 2016). Its baseline includes an RTG providing approximately 26 W electrical and approximately 500 W thermal power, a 3-axis MEMS capacitive accelerometer–based seismometer, meteorological sensors, high-temperature SiC-based electronics, and a low-profile mechanical configuration optimized for ground coupling and reduced wind-induced noise. The primary science objective is long-duration seismometry for at least 100 days and preferably at least one Venus year; secondary objectives include atmospheric boundary-layer characterization, potential heat-flow measurements, possible gas monitoring, and radio science (Wilson et al., 2016). This extends TerraProbe from episodic sensing to persistent, ambient-environment operation.

These planetary systems also clarify an important distinction. In Europa MT, TerraProbe is fundamentally an inversion instrument that exploits naturally occurring electromagnetic forcing; in the Venus package, it is a long-life geophysical station designed to live in the ambient environment rather than being shielded from it. The commonality is not the sensor modality but the operational doctrine: infer inaccessible subsurface or interior state from sparse, physically coupled measurements.

5. Learned subsurface and terrain representations

A data-driven TerraProbe emerges most clearly in the transformer-based foundation model Transparent Earth (Mazumder et al., 2 Sep 2025). The model reconstructs subsurface properties from heterogeneous datasets whose modalities include stress angle, strain angle, sediment thickness, mantle temperature, tectonic plates, fault type, basin type, and basin age. Positional encoding is defined over latitude, longitude, and depth as

$h$7

while modality encodings are derived from a text embedding model applied to a description of each modality. The architecture supports in-context learning with zero inputs or an arbitrary subset of modalities, and on validation data it reduces errors in predicting stress angle by more than a factor of three (Mazumder et al., 2 Sep 2025). This is a TerraProbe in the literal sense of a queryable latent geophysical model.

A probabilistic inversion counterpart is provided by neural posterior estimation for radar sounder data (Corso et al., 5 May 2026). The terrain parameter vector is

$h$8

where $h$9 is dielectric constant, $p$0 is RMS height, and $p$1 is RMS slope. Synthetic observations are generated by a GPU-based simulator, the peak returned power is extracted as

$p$2

and a normalized observation

$p$3

is passed to a normalizing-flow density estimator for $p$4. The model is reported to be well calibrated on simulated data and transferable to SHARAD profiles over Cerberus Palus and Zephyria Planum, while also allowing systematic evaluation of posterior robustness to reference surface variability (Corso et al., 5 May 2026). Relative to point-estimate inversion, this TerraProbe explicitly preserves parameter correlations and calibration dependence.

A related surface-level representation appears in Terra, a hierarchical terrain-aware 3D scene graph for outdoor mapping (Samuelson et al., 23 Sep 2025). Although it does not use the TerraProbe name, it provides terrain-aware place nodes built from per-terrain generalized Voronoi diagrams, hierarchical region nodes obtained by agglomerative or spectral clustering, and a task-agnostic metric-semantic sparse map that supports object retrieval, region monitoring, and terrain-aware navigation. The method performs on par with state-of-the-art camera-based 3DSG methods in object retrieval, surpasses them in region classification, and remains memory efficient (Samuelson et al., 23 Sep 2025). This suggests a semantic-mapping interpretation of TerraProbe: probing not only physical substrate but also navigable environmental structure.

6. TerraProbe in empirical software engineering

In empirical software engineering, TerraProbe is an explicit framework rather than a conceptual extrapolation. It is a five-layer oracle framework for detecting deceptive fixes in LLM-assisted Terraform security repair (Alsaid et al., 25 Jun 2026). The underlying problem is that prior evaluations often count a repair as successful when the targeted static-analysis finding disappears, without checking planning validity, behavioral change, or security intent. TerraProbe formalizes a stricter funnel.

Layer	Check	Function
L1	Targeted finding removal	Targeted Checkov removal
L2	Full scanner rerun	Full-scanner cleanliness
L3	terraform validate	Structural validation
L4	terraform plan	Planning success
L5	terraform show -json diff	Plan comparison

The study evaluates 288 first-pass repairs generated by gemini-2.5-flash-lite, GPT-4o, and Claude 3.5 Sonnet across 68 real-world TerraDS modules and 28 controlled injected-defect modules. For the primary model, targeted removal reaches 83.3 percent, but full-scanner cleanliness drops to 10.4 percent, Terraform planning succeeds for 39.6 percent, and plan comparison is reachable for 38.5 percent (Alsaid et al., 25 Jun 2026). Human adjudication on plan-compared real-world cases shows that 71.4 percent are deceptive fixes: they pass automated checks while leaving the underlying vulnerability in place. Across the three models, deceptive-fix rates range from 57.1 percent to 71.4 percent, with pairwise Fisher exact $p$5-values above 0.10, indicating no statistically significant model-specific difference (Alsaid et al., 25 Jun 2026).

The framework defines a repair $p$6, an oracle layer $p$7, and a deceptive fix as a check-passing repair that still violates the security intent of the original finding. Its four-dimensional taxonomy covers mechanism, intent alignment, security impact, and detection difficulty, and is validated with Cohen’s $p$8 and Krippendorff’s $p$9 (Alsaid et al., 25 Jun 2026). A particularly consequential result is the IAM analysis for CKV2_AWS_11: wildcard Resource: "*" grants persist in all nine deceptive-fix cases, despite the targeted scanner finding being removed (Alsaid et al., 25 Jun 2026). Here TerraProbe functions as a layered evaluation methodology that distinguishes intent-aligned repair from scanner-passing false success.

7. Methodological themes, limitations, and future directions

Taken together, these works suggest a common TerraProbe pattern: heterogeneous sensing or artifact collection, explicit representation learning or structured storage, and inference that remains useful under sparse or uncertain evidence. Pignoletto uses GIS layers and predictive models; lunar terramechanics uses regression surrogates over wheel–soil interaction; EMS uses MT impedance; Transparent Earth uses multimodal transformer latents; radar inversion uses calibrated posteriors; Terra uses terrain-aware scene graphs; the Terraform framework uses layered oracles (Piccoli et al., 2022, Kern et al., 8 Jan 2026, Grimm et al., 2021, Mazumder et al., 2 Sep 2025, Corso et al., 5 May 2026, Alsaid et al., 25 Jun 2026).

The principal limitations are domain-specific but recurrent. Generalization is a central issue: Pignoletto’s models are trained on regional data and must be adapted to other locations; the lunar rover terramechanics model is specific to dry silica sand and does not model side slip; the burrowing probe was studied only in dry cohesionless sand with fixed geometry; EMS modeling is largely 1D and depends on source-field characterization; Transparent Earth does not yet provide explicit uncertainty quantification or physics constraints; radar NPE currently models surface scattering only and uses a single scalar observation; TerraProbe for Terraform remains tied to Checkov, Terraform planning harnesses, and the TerraDS corpus (Piccoli et al., 2022, Wang et al., 2024, Grimm et al., 2021, Mazumder et al., 2 Sep 2025, Corso et al., 5 May 2026, Alsaid et al., 25 Jun 2026).

A related misconception is that a successful probe or repair can be validated by a single shallow metric. Several of these systems were designed precisely to reject that assumption. Thermal imagery at Earth pressure can under-discriminate soils that separate cleanly at 8 mbar; targeted scanner removal can overstate repair success by nearly an order of magnitude relative to full-scanner cleanliness; deep-water detectability on Europa depends on broadband MT rather than on low-frequency induction alone (Castilla-Arquillo et al., 2023, Alsaid et al., 25 Jun 2026, Grimm et al., 2021). The broader implication is that TerraProbe, across its domains, is defined less by a sensor type than by layered evidential reasoning.

Future directions are correspondingly convergent. The geospatial and robotics systems point toward richer multimodal fusion, spatial uncertainty handling, and online ingestion. The planetary geophysics systems point toward long-duration operation, radiation-tolerant electronics, and joint inversion of complementary modalities. The learned subsurface models point toward more modalities, physics-informed constraints, and scalable query interfaces. The software-engineering framework points toward stronger semantic oracles and benchmarks that reward intent-aligned outcomes rather than superficial pass conditions (Wilson et al., 2016, Mazumder et al., 2 Sep 2025, Alsaid et al., 25 Jun 2026). This suggests that TerraProbe is evolving into a general research paradigm for converting heterogeneous observations into operationally meaningful latent-state estimates under strict validation.