The PlanetGSD dataset comprises 6527 soil samples from three planetary bodies – Earth, Moon, and Mars – integrated within a unified analytical framework based on the UGSD function. This compilation enables systematic comparison of granular characteristics across terrestrial and extraterrestrial environments with contrasting surface processes, climatic regimes, and geological histories. The database is organized with multi-level metadata, including geographic coordinates, geomorphic context, mission information, and original data sources, all standardized to facilitate cross-planetary analysis.

Earth component. The terrestrial samples represent the full spectrum of Earth's surface environments, spanning five continents and encompassing 20 distinct geomorphic settings. These include gravity-driven deposits (debris flows, landslides, moraines), fluvial systems (river valleys, floodplains), aeolian environments (deserts, loess), pedogenic landscapes (vegetated slopes, grasslands), and coastal zones. This diversity captures grain-size signatures associated with major process regimes, providing a comprehensive terrestrial reference for comparison with extraterrestrial materials. The Earth data integrate field-collected samples and literature-derived measurements, all processed using standardized laboratory protocols (sieving and laser diffraction).

Table 1 summarizes the key characteristics of the three planetary datasets, including sample size, geographic coverage, analytical methods, grain-size ranges, and data sources. Table 2 provides detailed information for each sampling site, including sample counts, dominant soil types, and associated surface processes.

The UGSD provide quantitative constraints on method-induced variability. Rather than erasing method differences, the UGSD framework makes them explicit and traceable by transferring the measured textural fractions to the parameter set (μ, Dc, n). Moreover, as discussed below, the probability distribution of the exponent μ provides a potential method to recover the “loss” of fine content in the sampling methods, e.g., as in the case of Mars.

The UGSD parameters (μ, Dc, n) exhibit spatial variability across sampling sites, reflecting the inherent heterogeneity of granular materials under different surface processes. Among these, the exponent μ – which correlates positively with fine-grain content – serves as the most sensitive indicator of local textural characteristics and is therefore the primary focus of our probabilistic analysis. The characteristic size Dc also varies spatially and its distribution can be analyzed similarly, while the shape exponent n primarily reflects broad-scale differences between planetary bodies or major process regimes (e.g., n≈ 1 for Earth and Moon, n>1 for Mars) and is not considered for site-scale probability distributions. Figure 4a displays the (μ, Dc) points of all the samples from the three bodies. Soils from the Earth, Moon, and Mars. The cluster of the Earth covers soils of almost all the types from continents, while clusters of the Moon and Mars are limited to samples in the landing sites. Future explorations of the Moon and Mars are expected to fill the data gaps.

The (μ, Dc) value can be assigned to each point in principle; while for a target area, no matter of the scale, they present a probability distribution. For example, at the site level, the population of μ values is found to follow the three-parameter Weibull distribution: 2fμ;λ,k,μ0=kλμ-μ0λk-1exp⁡-μ-μ0λk where μ0 is the location parameter (the minimum measured μ value), λ is the scale parameter (characteristic spread), and k is the shape parameter (controlling distribution tail behavior). The three-parameter form was adopted because it provides greater flexibility than the two-parameter version, accommodating potential shifts in the minimum μ value that may arise from site-specific depositional or weathering histories. Fitting was performed using maximum likelihood estimation (scipy.stats.weibull_min.fit in Python).

Table 6 summarizes the ranges of fitted Weibull parameters across the three planetary bodies, and Fig. 4b displays the site-level (k, λ) points for all 42 locations of the samples in the dataset. One sees distinct point groups for the three planetary bodies, supporting the utility of the k–λ parameter pair as a diagnostic descriptor for cross-planetary comparison. Complete site-level Weibull parameters, including geological context and dominant process annotations, are provided in Dataset S4 (Sheet 2).

The preceding sections established a two-step transformation of raw grain-size measurements into a unified probabilistic framework (Fig. 5). First, the UGSD function converts the original GSD curves into a compact set of interpretable parameters (μ, Dc, n), reducing high-dimensional textural data to physically meaningful indices (Sect. 2.3). Second, at each sampling site, the population of μ values is characterized by a three-parameter Weibull distribution (Sect. 2.4), capturing the local spatial variability of fine-grain content in a probabilistic form.

This transformation fundamentally changes the nature of the dataset: from a conventional collection of discrete observations to a probabilistic GSD parameter library. In this library, each site is represented not by individual measurements but by the probability distribution of its granular characteristics (μ0, k, λ), enabling stochastic simulation and extrapolation beyond the original sampling locations.

The probabilistic parameterization naturally leads to the concept of granular fields – spatially continuous representations of soil heterogeneity derived from the site-level Weibull parameters. Following the workflow of Zhang et al. (2022), μ-fields are generated through a three-step Monte Carlo procedure:

UGSD fitting of measured GSD data to obtain μ for each sample (as described in Sect. 2.3).

Weibull parameterization and Monte Carlo simulation – For each sampling site, the population of μ values is fitted to the three-parameter Weibull distribution (Sect. 2.4). An ensemble of N random μ values is then generated by sampling from this fitted distribution, with N specified by the user (default N= 1000).

Spatial interpolation – The simulated μ values are assigned to a Digital Elevation Model (DEM)-registered grid via inverse-distance-weighted (IDW) interpolation (power = 2).

To ensure that the generated μ-fields respect realistic spatial continuity – an essential characteristic of geological and geomorphic units – the IDW interpolation is applied within spatial autocorrelation constraints defined by an exponential variogram model, with correlation length empirically determined from field observations (10 grid units at hillslope scale, 50 m at watershed scale). Users can specify DEM resolution, interpolation radius, and the number of Monte Carlo realizations.

The resulting μ-fields provide spatially explicit estimates of grain-size characteristics that preserve local heterogeneity while enabling upscaling to regional scales. Because soil properties such as porosity, permeability, and shear strength can be expressed as functions of μ (Jiang et al., 2024), these granular fields can be further transformed into property fields for hydro-geophysical modeling, rover mobility simulation, or in-situ resource utilization assessment. The complete workflow, from raw measurements to μ-fields, is illustrated in Fig. 5, and the implementation codes are provided in Codes S1–S3.

Importantly, the probabilistic framework presented here – specifically, the finding that site-level μ populations follow a three-parameter Weibull distribution – offers a principled way to partially overcome the spatial sampling bias inherent in mission-specific datasets. Because the Weibull model captures the statistical variability of GSD parameters, it enables stochastic simulation of grain-size characteristics for unsampled terrains (e.g., Martian highlands or dune fields not yet visited by rovers). Thus, while PlanetGSD 1.0 does not eliminate spatial bias, it provides a probabilistic pathway to propagate observed heterogeneity to unobserved locations – a point we return to in Sect. 6.3.

The PlanetGSD 1.0 database is organized into five core datasets (Tables S1–S5), each containing raw GSD measurements, fitted parameters, and associated metadata:

Dataset S1 (Earth GSD) contains 4419 terrestrial grain-size distributions with complete cumulative curves (17 size fractions from 0.0001 to 20 mm) and georeferenced metadata, integrating field-collected samples (2847) and literature-derived data (1572).

All datasets are provided in Microsoft Excel format (.xlsx) with multi-sheet structures to organize raw data, metadata, and quality flags within single files. Users are referred to the Figshare repository Zhang (2026a) for direct data access.

To facilitate correct interpretation and usage of the PlanetGSD 1.0 datasets, Table 8 provides complete field definitions for all variables appearing in Datasets S1–S5. The table includes field names, descriptions, data types, units, and example values, serving as a detailed reference to complement the database structure overview in Sect. 3.2.

All 6527 samples were fitted to the UGSD function (Eq. 3) using nonlinear least-squares optimization. Goodness-of-fit was quantified using the coefficient of determination (R2). Figure 3 shows the rescaled UGSD curves for Earth (a), Moon (b), and Mars (c), demonstrating the unified fitting performance across planetary bodies.

Median R2 values are 0.991 (Earth), 0.989 (Moon), and 0.983 (Mars), with 97.8 % of all samples yielding R2>0.95 (Table 9). The terrestrial and lunar populations exhibit strongly left-skewed R2 distributions, with > 85 % of samples achieving R2>0.98. The slightly lower Martian median is attributable to the truncation of the fine tail below ∼ 0.04–0.15 mm imposed by imaging resolution limits, which reduces the number of constraining data points available for curve fitting. Although a small number of lunar data points in Fig. 3b deviate slightly from the reference curve, these points correspond predominantly to coarse grains (D>1). Because soil physical properties are governed primarily by fine grains (Zhang et al., 2025), these minor deviations have negligible influence on the overall validity of our approach.

Following the outlier screening protocol (Sect. 2.3.2), 23 samples (0.4 % of total) with R2<0.94 were excluded after manual review due to data-entry errors or highly irregular curve shapes inconsistent with natural GSD. An additional 24 samples (0.4 %) with 0.90 <R2<0.94 were retained but flagged (quality_flag = 0) in Dataset S4 for user awareness. The remaining 6480 samples (99.3 %) carry quality_flag = 1 and are recommended for analysis without restriction.

Comparison with alternative parameterizations. To demonstrate the advantage of the UGSD function, we fitted all 6527 samples to three alternative models (lognormal, Rosin-Rammler, and Fredlund three-parameter) and found that UGSD achieved higher median R2 values than all three alternatives.

Kolmogorov-Smirnov test at significance level α= 0.05; pass rate indicates percentage of samples where the UGSD fit is not significantly different from the measured distribution (p>0.05).

The physical interpretation of UGSD parameters is supported by controlled experiments and field observations (Yong et al., 2017; Zhang et al., 2023): μ increases with fine-grain content (with μ>1 typical of aeolian settings, μ≈ 0.5 for fluvial-aeolian transitions, and μ<0.5 for lag or impact-mixed deposits), Dc marks the grain-size break point related to transport energy, and n reflects coarse-tail sorting efficiency. All four UGSD parameters are physically constrained (C>0, Dc>0, n>0); μ has no sign constraint because grain sizes below 1 µm yield negative ln(diameter) values, which is common in fine-grained samples. Statistically, the model is well-constrained: each GSD curve is fitted from > 10 independent size fractions (Earth/Moon) or > 150 grain measurements (Mars), and subsampling cross-validation confirms that the fitted UGSD parameters are stable, with standard deviations substantially smaller than the site-level variability across samples.

Three-parameter Weibull distributions were fitted to site-level μ populations for all 42 sites (20 Earth, 6 Moon, 16 Mars). Fit quality was assessed using two complementary approaches: the Kolmogorov–Smirnov (K–S) goodness-of-fit test at significance level α= 0.05, and visual inspection of Q–Q (quantile-quantile) plots comparing observed μ quantiles against theoretical Weibull quantiles.

Of the 42 sites, 39 (93 %) passed the K–S test, indicating that the three-parameter Weibull model adequately describes the μ population structure at most sampling locations. The three sites that did not pass are all terrestrial (MJ, DF, JZG) and exhibit bimodal μ distributions attributable to the mixing of geomorphically distinct lithological units within a single sampling area. Their Weibull parameters are retained in Dataset S4 (Sheet 2) with explanatory notes in the geological_description field; users may choose to exclude these sites or to fit separate Weibull distributions to the identified sub-populations.

To verify the accuracy of compiled and processed data, we cross-checked subsets of samples from each planetary body against original source publications. This procedure tests for systematic errors introduced during data transcription, digitization, and processing.

Lunar samples. Median grain sizes (D50) for 50 Apollo samples were compared with values reported in the Lunar Sourcebook (Heiken et al., 1991). The mean absolute deviation between our compiled values and the Lunar Sourcebook was 4.2 µm (range: 0.3–11.8 µm), within the reported measurement uncertainty of ±5 µm for the original sieving analyses (Graf, 1993). No systematic bias was observed (mean signed deviation: +0.8 µm), confirming that digitization from the NASA Lunar Soils Grain Size Catalog did not introduce directional errors.

These checks confirm that data transcription and processing introduced no systematic errors into PlanetGSD 1.0.

Because all Martian GSD data were derived from semi-automated image analysis, inter-operator variability represents an important component of measurement uncertainty. To quantify this variability, three independent analysts processed identical image sets (n= 30 images spanning all four landing sites) using the same grain-segmentation software (Shi et al., 2021; Zhao et al., 2023).

Results are summarized in Table 10. Inter-operator agreement was assessed using three metrics: (i) the coefficient of variation (CV) of D50 across the three analysts; (ii) the mean absolute percentage difference in cumulative grain-size fractions; and (iii) the resulting variation in fitted UGSD parameters.

The primary source of inter-operator variability is grain boundary delineation for overlapping or partially obscured grains. Images with > 30 % grain overlap showed systematically higher CV (mean: 14.2 %) compared to images with < 10 % overlap (mean: 3.8 %). This operator-dependent uncertainty is propagated into the UGSD parameter confidence intervals reported in Dataset S4.

These inter-operator variability estimates (CV ≈ 5 %–8 % for key parameters) are comparable to those reported in previous Mars grain-size studies using similar methods (Fedo et al., 2015; Weitz et al., 2018) and are substantially smaller than the inter-site variability observed across the four Martian landing sites, confirming that measurement uncertainty does not obscure geological signal in the dataset. However, caution is warranted in attributing the observed inter-site variability solely to geological signals. Alternative factors may also contribute to the relatively small variability observed among the four Martian landing sites.

First, the inherent resolution limits of image-based measurements (Table 3) truncate the fine tail of all Martian GSD curves below ∼ 0.04–0.15 mm, depending on the instrument. This right-censoring reduces the number of constraining size fractions available for UGSD fitting, which may artificially reduce apparent inter-site variability in the derived parameters – meaning that some of the observed similarity across sites could reflect what the instruments cannot resolve, rather than a true geological signal.

Second, the four landing sites – Gale, Jezero, Gusev, and Meridiani – are all located in low-latitude regions with broadly similar aeolian regimes. This geographic concentration does not capture the full range of Martian surface textural diversity (e.g., high-latitude mantling terrains, dust-dominated regions, or polar layered deposits).

We note that while Martian dust storms are known to mobilize and transport fine grains globally (Kahre et al., 2017; Senel et al., 2021), the grain-size fractions most affected by this process (silt and clay, < 62 µm) are below the detection limits of the rover imagers used in this study (Table 3). Therefore, the potential homogenizing effect of dust storms cannot be directly evaluated with our data and is not invoked as an explanation for the observed inter-site variability.

The inter-operator variability reported above (CV ≈ 5 %–8 %) inherently captures the combined effects of operator-dependent grain boundary delineation and the spherical shape approximation, as both contribute to the observed dispersion in fitted parameters.

The UGSD parameterization enables not only direct comparison of GSD characteristics but also continuous representations of spatial heterogeneity. Using the site-level Weibull statistics derived in Sect. 2.4 and the Monte Carlo workflow described in Sect. 2.5, we generated μ-fields – spatially explicit maps of the UGSD parameter μ – for three representative sites, one from each planetary body (Fig. 6).

The Laowa Gully (Earth) simulation produces a spatially homogeneous μ-field (range = 0.422), representing the null hypothesis of uniform grain-size distribution under uniform pedogenic processes (Zhang et al., 2022). This pattern can serve as a baseline for detecting anthropogenically-induced or naturally-occurring spatial heterogeneity in regions where limited field data suggest uniformity. The Apollo 17 Taurus-Littrow valley (Moon) simulation exhibits fine-scale random heterogeneity (∼ 100 m correlation length) without persistent spatial gradients (range = 0.125, Moran's I = 0.08). This pattern represents the stochastic mixing expected from impact gardening – dominant regolith – forming process on airless bodies – where meteoroid bombardment produces spatially uncorrelated grain-size distributions (McKay et al., 1991; Szalay et al., 2019). The Vera Rubin Ridge–Greenheugh Pediment area in Gale Crater (Mars) simulation shows a pronounced spatial gradient (range = 1.612, Moran's I = 0.76). This pattern demonstrates how directional geomorphic features (e.g., valley networks, crater rims, or lithological contacts extracted from DEMs) can induce systematic spatial organization in modeled grain-size distributions.

These contrasting patterns illustrate the flexibility of our spatial parameterization approach (Sect. 2.4) in generating diverse heterogeneity patterns under varying geomorphic constraints. While direct field validation of these simulations remains challenging for planetary surfaces due to the scarcity of in-situ grain-size transects at comparable scales, the demonstrated patterns provide testable hypotheses for future mission targets. For terrestrial applications (e.g., Laowa Gully), the same methodology can generate spatially explicit heterogeneity fields when user-provided DEM data are available, with field validation deferred to future work. The provided code (Code S1) enables users to generate analogous μ-fields for any study site using their own DEM data.

The overlap between the lunar field and a subset of terrestrial sites (e.g., Jiangjia Gully, Kunlun Mountains) in Fig. 4 warrants explanation. Lunar regolith is dominated by impact comminution (Devine et al., 1982; King, 1977), producing extremely fine grained, poorly sorted distributions with low μ values (Table 5). The overlapping terrestrial samples are dominated by physical fragmentation under cold or arid conditions (e.g., glacial till, periglacial colluvium, debris flow deposits), where chemical weathering is minimal (Attal et al., 2015; Weinman et al., 2011). Their Weibull parameters (low k, low μ) quantitatively mirror the textural signature of impact dominated fragmentation. This overlap thus highlights a fundamental insight: different geological processes that share a dominance of physical comminution over chemical alteration can produce similar grain size distribution fingerprints (Ovalle and Dano, 2011; Toneva and Peukert, 2007). The parameter space therefore discriminates process regimes (physical vs. chemical weathering) rather than planetary provenance per se. We have elaborated this interpretation in the revised Fig. 4 caption and Sect. 6.3.

For a complementary view of the parameter space itself, the three planetary populations occupy largely distinct domains: Earth spans the μ range from -0.012 to +0.41, Moon occupies a narrow near-zero range (0.001 to 0.126), and Mars displays the broadest μ range (from -1.531 to +1.081). The non-overlapping Weibull shape parameter k ranges (Table 6) further confirm systematic differences among the three bodies, supporting the utility of the k–λ parameter pair as a diagnostic descriptor for cross-planetary discrimination.

PlanetGSD 1.0 provides reference GSD parameters against which lunar and Martian simulant materials can be quantitatively evaluated. Users can compute the UGSD and Weibull parameters of candidate simulants using the provided analysis codes (Codes S1–S3) and compare them with the site-level parameter ranges documented in Dataset S4. For example, the Weibull parameters for Apollo mare sites (k= 0.053–0.077, λ= 1.83–4.48) define a target envelope for lunar simulants intended to replicate mare regolith; simulants falling outside this envelope can be identified as texturally non-representative. Table 11 provides example analog matches linking target terrains to PlanetGSD reference sites to facilitate such comparisons.

It is important to note that Fig. 6 presents an interpolated μ field generated from the Monte Carlo workflow described in Sect. 2.5, which already accounts for spatial autocorrelation via an exponential variogram model to ensure realistic spatial continuity.

However, even with spatial autocorrelation, the current interpolation does not explicitly incorporate mapped geologic or geomorphic boundaries (e.g., lithological contacts, fault lines, or geomorphic units). Purely statistical spatial correlation without such contextual constraints is of limited value for predictive mapping, especially with sparse data points.

Therefore, the μ field in Fig. 6 is presented as a continuous visual summary of sparse point measurements, not as a final predictive soil map. It serves two modest but necessary purposes: (i) it provides an intuitive, spatially continuous visualization of the general trend of grain size, helping to generate hypotheses or guide future sampling; and (ii) it offers a baseline continuous covariate for future geostatistical models (e.g., kriging with external drift) or machine-learning approaches that will explicitly integrate geologic units, geomorphic boundaries, or remote-sensing data (planned for PlanetGSD 2.0).

Future work (PlanetGSD 2.0) will explicitly incorporate such spatial constraints to generate more robust, context-aware predictions.

PlanetGSD 1.0 opens several avenues for future research, but it also has limitations that should guide its use and development:

Gaps in geologic and environmental representation. The terrestrial component is dominated by gravity-driven and fluvial deposits (debris flows, landslides, alluvial fans), while other geomorphic settings remain underrepresented. These include playa lakes, aeolian dune fields, glacial outwash plains, coastal sediments, tropical weathering profiles, and cold-desert soils (e.g., Antarctic dry valleys). The current geographic bias toward Asia (73 % of samples) is less critical than the underrepresention of these process-based endmembers. Lunar samples are predominantly from nearside mare regions, with limited representation of highland, farside, and polar materials. Martian samples are limited to four low-latitude landing sites, lacking high-latitude and dust-dominated terrains. Priority targets for future data collection include underrepresented geomorphic settings on Earth (as listed above), lunar farside and polar regions, and new Martian landing sites (e.g., Zhurong at Utopia Planitia, and high-latitudes sites such as Arcadia Planitia). A future version of PlanetGSD (e.g., PlanetGSD 2.0) should further expand farside and polar lunar sampling beyond the current CE-6 sample and include Martian terrains that have not yet been visited by rovers (e.g., highlands, polar regions, dune fields).

PlanetGSD 1.0 is designed to support the following research applications:

Comparative planetology. The unified UGSD parameterization enables direct comparison of grain-size signatures across Earth, Moon, and Mars. The k–λ parameter domains (Fig. 4) provide a quantitative framework for classifying samples by planetary provenance.

To ensure robust application of PlanetGSD 1.0, users should consider the following guidance:

Parameter interpretation. While the UGSD parameter μ correlates positively with fine-grain content and the Weibull shape parameter k reflects the width of the μ distribution, these are empirical descriptors derived from curve fitting. Direct translation of μ to physical quantities such as porosity, permeability, or shear strength should be undertaken with caution. Such applications are most robust when μ is coupled with complementary site-specific data or established geotechnical transfer functions (e.g., Zhang et al., 2022; Jiang et al., 2024). The parameters are primarily intended for comparative purposes – e.g., ranking sites by relative fineness or sorting – rather than for absolute grain-size inference.

Users should be aware of three principal limitations when using PlanetGSD 1.0, organized around data completeness, data integrity, and appropriate application scale.

PlanetGSD is conceived as a living database that will evolve with new missions and community input. Future versions (e.g., PlanetGSD 2.0) will incorporate:

Additional terrestrial environments (aeolian dunes, glacial outwash, coastal sediments, loess sequences);

All contributed datasets will undergo the same validation and quality control checks applied to the original PlanetGSD compilation (see Technical Validation) before inclusion. Version history and a detailed changelog will be maintained in the repository README file.