the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 20 Apr 2026
Use of Spatial Embeddings in Genosoil Identification
Abstract. Genosoils are minimally disturbed reference states within pedogenons, that is, soil units shaped by similar pedogenic processes within the Soil Security framework. They are central to assessing human impacts on soil functions, services, and resistance to threats. At present, genosoil delineation relies on the Human Modification Index (HMI), yet in intensively managed landscapes HMI thresholds may exclude all local pixels, leaving no local reference state available. Because the same pedogenon may occur across geographically distant regions, non-local occurrences may provide an alternative source of reference information. Using the United Kingdom as a case study, we tested whether satellite-derived spatial embeddings can detect genosoil signatures at 10 m resolution and whether these signatures can be transferred to regions with limited or absent local low-human-modification examples. We evaluated two satellite foundation-model embedding products, AlphaEarth and Tessera, across three contrasting pedogenons selected from the Global Pedogenon Map. Within each pedogenon, pixels with lower HMI values were generally more similar to the genosoil reference, indicating that the embeddings capture a reproducible low-modification surface-state signal. At the global scale, similarity to the UK genosoil was largely confined to biogeographically coherent regions. Cross-border substitution of local UK genosoil delineation was mostly limited, with meaningful partial recovery observed primarily in the highly modified agricultural pedogenon. These results indicate that satellite foundation-model embeddings can support higher-resolution genosoil delineation than is currently possible from global human modification products alone, extending the operational framework from 90 m to 10 m. They also suggest a pathway towards future genosoil identification frameworks that rely less on coarse disturbance proxies and more on validated surface-state similarity.
- Preprint
(1429 KB) - Metadata XML
-
Supplement
(252 KB) - BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-1944', Marijn van der Meij, 08 May 2026
-
AC1: 'Reply on RC1', Julio Pachon, 09 Jul 2026
Pachón Maldonado et al. aim to identify genosoils (minimally disturbed soils in homogeneous soil geographic regions called pedogenons) using spatial embeddings (vectorized information summarizing land surface properties derived from remote sensing). The hypothesis is that undisturbed soils share similar land surface characteristics that can be represented, and transferred, using the spatial embeddings. The manuscript provides an extensive statistical analysis of different spatial datasets representing land surface properties, soil properties and human disturbances. The findings indicate limited geographical transferability of the spatial embeddings for genosoil identification.
I have three main comments regarding the manuscript: 1) I have some doubts about the quality of the datasets for their intended purpose, 2) the manuscript will be difficult to understand by the more general soil-scientific audience of SOIL due to the strong statistical focus and insufficient connection to existing pedogenic frameworks, and 3) the methods and results should be better structured for improved clarity.
I have detailed my main comments below, followed by a list of minor comments and technical corrections.
We thank Dr. van der Meij for a thorough and constructive review. The comments have substantially improved the manuscript. We address each point in turn below.
---
- Quality of datasets
1.1. The HMI dataset is used to identify undisturbed genosoils. However, especially for pedogenon 1564, I have doubts whether this dataset actually shows minimally disturbed soils. Their minimal presence and highly scattered occurrence (Table 1, Fig. 3) raise the question whether they are actually genosoils, or just wrongly classified phenosoils in the HMI dataset. An evaluation of Table S2 shows that the most similar countries to the UK also have a very small percentage of genosoils in their pedogenon, while countries with much higher percentage of genosoils show lower similarity. This makes me wonder if the HMI actually identifies genosoils for P1564, or whether wrongly classified agricultural soils or built-up areas are compared with each other, which could also be an explanation for their similarity in European context. A thorough evaluation of the accuracy of the HMI dataset is essential before spatial embeddings of derived genosoils can be reliably compared.
We agree with the reviewer that the reliability of the Human Modification Index (HMI) defined reference population is central to the interpretation of this study and to operational applications of the genosoil framework. Theobald et al. (2025) assessed HMI v3 against the independent Global Land Use Emergent Dataset, based on visual interpretation of high-resolution imagery in approximately 1,000 spatially balanced 1 km² plots, reporting RMSE values of 0.180 and 0.178 at 300 and 90 m, respectively; this was supplemented by expert review and comparisons with authoritative land-cover and ecoregional summaries.
Our method uses HMI-based delineation as an operational proxy following Román Dobarco et al. (2023) and Francos et al. (2025a); in the revised manuscript, we state explicitly that threshold-based delineation is a pragmatic means of identifying candidate minimally disturbed areas within pedogenons (Section 2.3). The submitted manuscript already used the term "candidate genosoil"; we have tightened that existing terminology so that all HMI-derived reference pixels, populations, centroids, footprints, and within-population statistics are consistently described as "candidate genosoils" or "HMI-defined candidate genosoils", while retaining "genosoil" for the conceptual class in the established genosoil/phenosoil framework. The manuscript also acknowledges that a low HMI value does not necessarily guarantee a genosoil-like embedding signature, and that a genosoil-like surface may be present in cells exceeding the conventional HMI threshold (Section 3.6). Figure 4 and Section 3.2 show that increasing human modification is generally associated with greater embedding distance from each country's low-HMI candidate-genosoil centroid, indicating a detectable HMI gradient in embedding space.
Our threshold of HMI ≤ 0.10 corresponds to the union of Theobald et al.'s (2025) "very low" (HMI ≤ 0.01) and "low" (0.01 < HMI ≤ 0.10) modification strata and its practicality lays in having a signal even in small pedogenons since neither P845 nor P1564 have HMI ≤ 0.05 or HMI ≤ 0.01. We have updated the Methods to reflect this. Román Dobarco et al. (2023) themselves disclaim pristineness for their HMI-derived genosoils: "by genosoils we do not refer to pristine soils without human influence" and frame HM as "a good proxy for anthropogenic pressures on soils" whose "resolution … is not high enough for locating genosoil profiles in the field".
We tested the reviewer’s specific hypothesis that HMI ≤ 0.10 pixels may be “wrongly classified agricultural soils or built-up areas” by comparing the HMI-defined candidate populations with ESA WorldCover v200 (2021) and CORINE Land Cover 2018. The result does not support that hypothesis. For P845, the HMI ≤ 0.10 candidate cells are 100% CORINE Natural grasslands. For P1564, the complete candidate footprint is 99.4% WorldCover tree cover, 0% built-up, and only 6.15% CORINE Non-irrigated arable. For P2932, the HMI ≤ 0.10 candidate population is 92.2% WorldCover grassland and 7.75% tree cover, with 0% cropland and 0% built-up; even the higher-HMI inverse fraction contains only 0.11% built-up and 0.53% cropland. We have added a supplementary table S20 and 2022 Sentinel-2 imagery for these diagnostics and modified the manuscript accordingly. These results support HMI ≤ 0.10 as an effective operational disturbance screen for identifying candidate genosoils in this study, while not replacing direct field validation of the soil profiles represented by those pixels.
Pedogenon
WorldCover Description (value)
CORINE leading classes (value)
845
83.3% Grassland (30)
16.7% Tree cover (10)
0% Cropland (40)
0% Built-up (50)100% Natural grasslands (321)
1564
99.4% Tree cover (10)
0.55% Grassland (30)
0.03% Cropland (40)
0% Built-up (50)67.5% Mixed forest (313) + 14.6% Coniferous (312) + 7.8% Broad-leaved (311) = 89.9% forest;
3.96% Transitional woodland-shrub (324);
6.15% Non-irrigated arable (211)2932
92.2% Grassland (30)
7.75% Tree cover (10)
0.00% Cropland (40)
0.00% Built-up (50)51.8% Moors and heathland (322)
27.3% Peat bogs (412)7.0% Natural grasslands (321)
6.3% Coniferous forest (312)
5.6% Sparsely vegetated (333)
1.1% Transitional woodland-shrub (324)1.2. Another concern comes from the uniformity of the pedogenons. Figure 5 shows occurrences of the same pedogenon in countries with a wide variety of climatic and topographic conditions, and the pedogenons sometimes contain soils that are formed under contrasting climatic conditions (Table S4). This heterogeneity will make it difficult to define one genosoil type with corresponding spatial embedding for each pedogenon. This point is addressed in Section 3.6, but in my opinion with insufficient detail. I would like to see a more extensive discussion of how the quality of the used datasets could influence the outcomes of this study, and also see this reflected in the conclusions.
We explored this concern do not treat the Global Pedogenon Map (Francos et al., 2025b) as a final soil taxonomy or as ground truth, but as a globally consistent, covariate-based stratification of soil-forming environments at 90 m, developed by the same approach under which the published genosoil/phenosoil research has been produced (Román Dobarco et al., 2023; Francos et al., 2025a).
Conventional soil maps such as national surveys, gSSURGO, and global products such as SoilGrids attempt to depict the soil as it is now, including its anthropogenic modification; none defines a globally consistent reference state against which that modification can be measured. The pedogenon/genosoil framework is, in this sense, a conceptual map rather than an inventory of observed profiles: it asks where comparable soil-forming environments recur, so that a minimally disturbed expression can be sought within each. Francos et al (2025b) purposefully avoid the top 30 cm of the soil maps data in their GPM development. That conceptual role is precisely why we begin from pedogenons rather than from WRB, which classify soils from profile morphology, answer a different question. SSURGOs development and their soil series are more in line but are limited to the US.
We are forthcoming on the intra-pedogenon heterogeneity in Section 3.6 and will have add to the conclusions that “Transferability of genosoil signal was strongest for internally coherent pedogenons and weakest where a single class spanned contrasting soil-forming environments. Improvements in future iterations of globally consistent pedogenon products will improve the genosoil signal”.
In the revised submission we will also deposit a per-pedogenon heterogeneity table, together with the code that generates it, in the data archive and reference it from Section 3.6. The analysis code and derived tables for the submitted manuscript are currently available to the reviewer here: https://zenodo.org/records/19424156?preview=1&token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6IjU4NzA5MTQwLTk1Y2EtNGZmNS05MmMwLTcyNmNjMTZhODk5ZCIsImRhdGEiOnt9LCJyYW5kb20iOiIzYTcwNzg3YTQwNWIyODQwZjJkYWVhNTRhNDY5ZmNmNSJ9.6Dwt8tRhJQox3evJN0T63nh-wan6--UBnE6Ut8VptmMqf_b8CGBWJ-2PsFL31t5ks2U6TYhwcICYf0gZ5hYZmQ
- Understandability for general soil-scientific audience
2.1. The authors remark that spatial embeddings “do not directly encode pedogenesis” (lines 83-84). However, many of the land surface properties of the spatial embeddings correspond to the soil forming factors, where Tessera seems to mainly focus on organisms through land cover from Sentinel, while AlphaEarth seems to represent topography and climate as well (lines 76-82). I think it would benefit the paper to frame the spatial embeddings in the context of soil forming factors, or a comparable model, to connect to more familiar frameworks in soil science. It would also be interesting to see a discussion on how the lack of representation of the other factors, especially time, could have an influence on the outcomes.
We thank the reviewer for this suggestion which we have adopted with additions to Section 2.2 and Section 3.6. We also clarify the architecturally distinct ways in which the two models relate to the individual soil-forming factors, as the mapping is more nuanced than a simple factor-by-factor assignment.
AlphaEarth is generated at inference from Sentinel-2, Sentinel-1, and Landsat 8/9 imagery only. However, its Space Time Precision (STP) training objective calibrates the embedding space against a broader set of reconstruction targets, including ERA5-Land climate variables (precipitation, air temperature, surface pressure), and the GLO-30 digital elevation model, as well as GEDI LiDAR canopy metrics, land cover, and biodiversity occurrence records (Brown et al., 2025, Table S1). These are training targets, not inference inputs: at inference time, only satellite imagery is required, but the learned embedding geometry is organised to be predictive of those environmental covariates. Rahman et al. (2026) confirm this empirically: 12 of 26 environmental variables, including air temperature (R² ≈ 0.97) and elevation (R² ≈ 0.97), are recoverable from AlphaEarth embeddings.
Tessera uses no climate or topographic targets at any stage of training. Its Barlow Twins temporal invariance objective trains exclusively on Sentinel-1 and Sentinel-2 time series (Feng et al., 2025). Two independently sampled subsets of the annual cloud-free observation sequence at the same pixel are used as the two views: the model learns to produce similar embeddings from both, enforcing invariance to which particular observation dates are available. The result is a representation of what is spectrally and temporally stable across the growing season. As such Tessera captures the stable phenological signature that all soil-forming factors jointly produce. We have clarified this in the revised manuscript.
We will add in section 3.6 that time for pedogenesis is not uniform geographically. Soils on recently deglaciated terrain or on fresh mass-movement deposits will be less developed that soils from long-stable surfaces. No clear covariate can account for this and thus time is indeed missing from pedogenon maps. Research in incorporating time into pedogenons maps will be very valuable and a way to reduce the within-class heterogeneity.
2.2. Next to that, despite the remark that spatial embeddings “do not directly encode pedogenesis” (lines 83-84), the authors actually suggest to use the spatial embeddings to “flag pedogenons whose global extent may conflate distinct soil-forming environments” (lines 409-410). I think these two statements contradict each other and need revising.
We thank the reviewer for pointing this out. We intended lines 83–84 to be a mechanistic statement while lines 409–410 to be a diagnostic one. Precisely because embeddings capture the integrated surface expression of the soil-forming environment, systematic divergence between the expected and observed embedding patterns within a named pedogenon class reveals that the class may span multiple distinct soil-forming environments. In the latter, the embedding is used as an independent biogeographic test of whether pedogenon boundaries are internally coherent and not as a direct measure of pedogenesis. Embedding divergence does not measure pedogenesis, but embedding incoherence within a nominally uniform pedogenon can flag a class that conflates distinct pedogenic contexts. Nonetheless to avoid confusion, lines 409-410 have been removed.
2.3. Statistical terminology. The manuscript introduces various statistical concepts and metrics that are not consistently named and referenced throughout the manuscript. For example, lines 164 – 166 uses the terms “internal coherence”, “cohesion” and “cosine similarity” to describe the representativeness of a reference embedding to its population. Throughout the rest of the manuscript, these terms are used interchangeably. Other important metrics are not defined by equation, but as a quick mention in the main text (e.g. cosine distance, lines 176-177). I think that the paper can benefit from more strict use of statistical terminology and a clear overview of their definitions and descriptions, for example in a table.
We agree and the revised manuscript will create a table to be referenced and aid consistency.
- Structure of the manuscript
Most Tables and Figures present results for three pedogenons and two spatial embeddings. Their order of presentation is however not consistent, where some Figures present the pedogenons as rows (Fig. 3) and others as columns (Fig. 4). The order of presentation of the pedogenons also varies between Figures, Tables and their description in the text. I think the manuscript will be much clearer when the pedogenons and spatial embeddings are consequently presented and described in the same order throughout the Methods and Results.
Thank you for this suggestion we will adopt a single consistent presentation order throughout the revised manuscript.
---
Minor comments and technical corrections
The abstract contains a lot of technical terms which will not be understandable without reading the manuscript first (e.g. spatial embeddings , line 12; reproducible low-modification surface-state signal, line 16-17). Please make sure that the abstract is in itself understandable to the audience of SOIL.
We have revised the abstract to define key terms on first use. "Spatial embeddings" is now followed by a parenthetical: "(vectorised representations of land surface properties derived from satellite imagery)." "Reproducible low-modification surface-state signal" has been rewritten as "a consistent reference signature in minimally disturbed pixels."
Figure 1. The Caption should better describe what is visible in the two panes. Pane B would fit better in the Methodology than in the Introduction.
Caption updated to better describe what is visible in describe both panels and we have adopted the recommendation to move B to the Methods section.
Section 2.1. Another argument for selecting these pedogenons is that they have different levels of occurrence and modification (lines 224-225), which could be mentioned here as well.
Agreed and adopted
I was wondering whether the cohesions derived from AlphaEarth and Tessera are comparable, as the spatial embeddings have different dimensions and are based on different datasets. I can imagine that Tessera has less internal variation as it is based on less variable data, which could lead to a higher cohesion. Could you add a remark about this in the manuscript?
AlphaEarth and Tessera are not directly comparable in magnitude because the two embeddings have different dimensionalities (64-D vs. 128-D) and are derived from different training objectives. We also think that an embedding based on less variable input data can show systematically higher cohesion for reasons unrelated to genosoil signal strength. We have added a caveat to §2.3 and to the caption of Table 2.
Line 160, 167. The terms L2-mean and L2-normalization need explanation.
We will add that L2 normalization refers to scaling each vector to a length of exactly 1, creating a hypersphere where only the cosine distances matter. L2 mean being the mean direction of the vector on the hypersphere.
Eq. 3. Can you explain more extensively what the Jaccard Index calculates and represents?
We will expand on the explanation of Jaccard index which is a statistic used to measure the similarity and diversity of sample sets. As shown in eq. 3, it is defined as the size of the intersection of two sample sets divided by the size of the union of the sample sets.
Line 192. Silhouette confidence intervals are mentioned here but not provided in the results. Can you use uniform confidence-interval widths for silhouettes and bootstrapping?
In the revised results we will report bootstrap confidence intervals for all silhouette scores, obtained by resampling the pixel populations underlying each silhouette, and we will apply a single uniform method to every bootstrapped metric (95% percentile intervals from 2000 resamples) so that the silhouette, cohesion, and Table S3 uncertainties are all reported on the same basis.
Lines 216-219: Should be moved to methods, and Fig 1b should be referenced here.
Agreed and done.
Figure 3. This Figure needs some modification:
Axis descriptions and labels are not readable due to their size and because they sometimes overlap with the Figures.
Axis labels enlarged and rotated to prevent overlap.
Can you add the pedogenon code as a row label instead of as an inset in one of the panes?
Pedogenon code added as a row label on the left margin.
Can you indicate where each area is located within the UK?
UK inset map added showing the approximate geographic location of each pedogenon's reference pixels.
The details on the maps of P1564 are barely visible. Is there a way to improve the readability of these maps?
P1564's sparse pixel coverage in the UK is the a result and although we have increased the contrast and enlarged the map slightly, we are unable to further improve it.
Table 2. Can you add values for coherence / cohesion, as these are mentioned in the text as well (e.g. lines 247-248, 250)?
Cohesion column added for all countries and both models.
Figure 4.
Shouldn’t the Y-axis label be “cosine distance from each country’s genosoil centroid” instead of “distance from UK genosoil centroid”?
Y-axis label corrected
Could you add a legend for the line widths instead of mentioning the scaling in the caption? You could, for example, group lines based on certain thresholds of pedogenon area. Also make the line width of the UK pedogenon consistent with the other line widths.
Line width legend added; lines grouped by pedogenon area quartile. UK line width made consistent with the other widths.
Figure 5. Can you use colors that are distinguishable for colorblind people and for black-and-white prints?
These colors were used and tested for colorblind using https://www.color-blindness.com/coblis-color-blindness-simulator/. We have added texture to aid when black and white printing.
Line 337. You state that AlphaEarth shows a greater degree of local environmental context, which would be relevant for understanding soil formation. Yet, this spatial embedding scores systematically lower than Tessera. Could you discuss the reasons for this in the context of soil formation, data quality and used statistics? See also comment 7.
We appreciate this observation as an opportunity to clarify. We assume the reviewer refers to Figure 5 and how AlphaEarth has lower cosine similarity to the UK for a given country. We think AlphaEarth's lower values arise because it implicitly encodes climate and relief (Sect. 2.2), and therefore separates countries whose climatic and topographic settings differ from the UK - whereas Tessera, trained only on Sentinel-1/2 phenology with no climate or relief targets, cannot make that separation and compresses distant countries toward a high similarity floor. The two embeddings agree on the genuine near-analogues (for the closest matches, cosine > 0.85, Tessera is never higher than AlphaEarth) and diverge only for distant occurrences: there AlphaEarth returns low or near-zero similarity (overall range ≈ 1.06, floor ≈ −0.06), whereas Tessera compresses the scale to a high floor (range ≈ 0.66, floor ≈ +0.34), so even climatically and geomorphically contrasting countries retain apparently high similarity to the UK. Consistent with this, AlphaEarth scores higher on the inter-country silhouette (e.g. P1564: −0.35 vs −0.54 for Tessera). We will add greater discussion on Sections 3.4 and 3.6: the lower cosine values under AlphaEarth indicate stronger discrimination due to climate and relief.
Table 4. Could you use consistent rounding of the decimals?
Checked throughout and fixed.
Code availability. The referenced Zenodo page does not exist.
A preview link was added in the submission and can be found here:
https://zenodo.org/records/19424156?preview=1&token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6IjU4NzA5MTQwLTk1Y2EtNGZmNS05MmMwLTcyNmNjMTZhODk5ZCIsImRhdGEiOnt9LCJyYW5kb20iOiIzYTcwNzg3YTQwNWIyODQwZjJkYWVhNTRhNDY5ZmNmNSJ9.6Dwt8tRhJQox3evJN0T63nh-wan6--UBnE6Ut8VptmMqf_b8CGBWJ-2PsFL31t5ks2U6TYhwcICYf0gZ5hYZmQ
Final version and DOI will be added once paper is fully accepted.
Table S2. Why are the surface areas of pedogenons and genosoils different for each embedding model? Aren’t these based on the independent HMI?
The surface areas differed slightly because the pedognon map derived from GPM and the genosoil derived from HMI masks, are resampled by nearest-neighbour onto each embedding's native pixel grid and these do not perfectly align. We verified this is a resampling artifact. We appreciate the reviewer pointing this out, and have simplified the table by reporting a single pedogenon and genosoil area taken directly from the GPM and HMI maps respectively, independent of the embedding model.
Table S6 is not referenced in the text.
Table S6 will be referenced in section 3.5
Citation: https://doi.org/10.5194/egusphere-2026-1944-AC1
-
AC1: 'Reply on RC1', Julio Pachon, 09 Jul 2026
-
RC2: 'Comment on egusphere-2026-1944', Anonymous Referee #2, 08 May 2026
This is a very good study on testing the behavior if either of two 10-m-resolution RS-based foundation model (AlphaEarth and Tessera) would be introduced to representatively identify three types of UK genosoil across different nations at the global scale. The hypothesis testing experiments were well designed from different aspects for discussing different behaviors of AlphaEarth and Tessera, as well as providing reasonable explanations on such behaviors. Limitations within current study had been presented in the end of manuscript.
Among the three pedogenons tested in current study, there are very unique characteristics. It’s good; meanwhile it raises my further question. How about two pedogenons with similar characteristics such as much closer distance on geography or taxonomy? How would be the differentiation behavior from either AlphaEarth or Tessera?
Another minor comment is on Figure 3: the bottom subfigures are difficult to read clearly.
Citation: https://doi.org/10.5194/egusphere-2026-1944-RC2
Data sets
Use of Spatial Embeddings in Genosoil Identification: code and tables Julio Pachon https://zenodo.org/records/19424156?preview=1&token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6IjU4NzA5MTQwLTk1Y2EtNGZmNS05MmMwLTcyNmNjMTZhODk5ZCIsImRhdGEiOnt9LCJyYW5kb20iOiIzYTcwNzg3YTQwNWIyODQwZjJkYWVhNTRhNDY5ZmNmNSJ9.6Dwt8tRhJQox3evJN0T63nh-wan6--UBnE6Ut8VptmMqf_b8CGBWJ-2PsFL31t5ks2U6TYhwcICYf0gZ5hYZmQ
Model code and software
Use of Spatial Embeddings in Genosoil Identification: code and tables Julio Pachon https://zenodo.org/records/19424156?preview=1&token=eyJhbGciOiJIUzUxMiJ9.eyJpZCI6IjU4NzA5MTQwLTk1Y2EtNGZmNS05MmMwLTcyNmNjMTZhODk5ZCIsImRhdGEiOnt9LCJyYW5kb20iOiIzYTcwNzg3YTQwNWIyODQwZjJkYWVhNTRhNDY5ZmNmNSJ9.6Dwt8tRhJQox3evJN0T63nh-wan6--UBnE6Ut8VptmMqf_b8CGBWJ-2PsFL31t5ks2U6TYhwcICYf0gZ5hYZmQ
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 190 | 106 | 22 | 318 | 28 | 12 | 17 |
- HTML: 190
- PDF: 106
- XML: 22
- Total: 318
- Supplement: 28
- BibTeX: 12
- EndNote: 17
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Pachón Maldonado et al. aim to identify genosoils (minimally disturbed soils in homogeneous soil geographic regions called pedogenons) using spatial embeddings (vectorized information summarizing land surface properties derived from remote sensing). The hypothesis is that undisturbed soils share similar land surface characteristics that can be represented, and transferred, using the spatial embeddings. The manuscript provides an extensive statistical analysis of different spatial datasets representing land surface properties, soil properties and human disturbances. The findings indicate limited geographical transferability of the spatial embeddings for genosoil identification.
I have three main comments regarding the manuscript: 1) I have some doubts about the quality of the datasets for their intended purpose, 2) the manuscript will be difficult to understand by the more general soil-scientific audience of SOIL due to the strong statistical focus and insufficient connection to existing pedogenic frameworks, and 3) the methods and results should be better structured for improved clarity.
I have detailed my main comments below, followed by a list of minor comments and technical corrections.
---
1. Quality of datasets
1.1. The HMI dataset is used to identify undisturbed genosoils. However, especially for pedogenon 1564, I have doubts whether this dataset actually shows minimally disturbed soils. Their minimal presence and highly scattered occurrence (Table 1, Fig. 3) raise the question whether they are actually genosoils, or just wrongly classified phenosoils in the HMI dataset. An evaluation of Table S2 shows that the most similar countries to the UK also have a very small percentage of genosoils in their pedogenon, while countries with much higher percentage of genosoils show lower similarity. This makes me wonder if the HMI actually identifies genosoils for P1564, or whether wrongly classified agricultural soils or built-up areas are compared with each other, which could also be an explanation for their similarity in European context. A thorough evaluation of the accuracy of the HMI dataset is essential before spatial embeddings of derived genosoils can be reliably compared.
1.2. Another concern comes from the uniformity of the pedogenons. Figure 5 shows occurrences of the same pedogenon in countries with a wide variety of climatic and topographic conditions, and the pedogenons sometimes contain soils that are formed under contrasting climatic conditions (Table S4). This heterogeneity will make it difficult to define one genosoil type with corresponding spatial embedding for each pedogenon. This point is addressed in Section 3.6, but in my opinion with insufficient detail. I would like to see a more extensive discussion of how the quality of the used datasets could influence the outcomes of this study, and also see this reflected in the conclusions.
2. Understandability for general soil-scientific audience
2.1. The authors remark that spatial embeddings “do not directly encode pedogenesis” (lines 83-84). However, many of the land surface properties of the spatial embeddings correspond to the soil forming factors, where Tessera seems to mainly focus on organisms through land cover from Sentinel, while AlphaEarth seems to represent topography and climate as well (lines 76-82). I think it would benefit the paper to frame the spatial embeddings in the context of soil forming factors, or a comparable model, to connect to more familiar frameworks in soil science. It would also be interesting to see a discussion on how the lack of representation of the other factors, especially time, could have an influence on the outcomes.
2.2. Next to that, despite the remark that spatial embeddings “do not directly encode pedogenesis” (lines 83-84), the authors actually suggest to use the spatial embeddings to “flag pedogenons whose global extent may conflate distinct soil-forming environments” (lines 409-410). I think these two statements contradict each other and need revising.
2.3. Statistical terminology. The manuscript introduces various statistical concepts and metrics that are not consistently named and referenced throughout the manuscript. For example, lines 164 – 166 uses the terms “internal coherence”, “cohesion” and “cosine similarity” to describe the representativeness of a reference embedding to its population. Throughout the rest of the manuscript, these terms are used interchangeably. Other important metrics are not defined by equation, but as a quick mention in the main text (e.g. cosine distance, lines 176-177). I think that the paper can benefit from more strict use of statistical terminology and a clear overview of their definitions and descriptions, for example in a table.
3. Structure of the manuscript
Most Tables and Figures present results for three pedogenons and two spatial embeddings. Their order of presentation is however not consistent, where some Figures present the pedogenons as rows (Fig. 3) and others as columns (Fig. 4). The order of presentation of the pedogenons also varies between Figures, Tables and their description in the text. I think the manuscript will be much clearer when the pedogenons and spatial embeddings are consequently presented and described in the same order throughout the Methods and Results.
---
Minor comments and technical corrections