Deep learning selection of analogues for Mars landing sites in the Qaidam Basin, Qinghai-Tibet Plateau

Andr\'e Antunes; Biqiong Wu; Fanwei Meng; Guoliang Zhou; Jie Zhao; Tianqi Hao; Xiaopeng Wang

REVIEW 4 major objections 6 minor 3 references

Deep learning selection of analogues for Mars landing sites in the Qaidam Basin, Qinghai-Tibet Plateau

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A Siamese CNN trained on Mars and Qaidam Basin images learns a quantitative geomorphological similarity metric, recovers the four known analogue pairings, and uses the resulting distances to propose priority landing sites.

desk verdict A useful idea wrapped in a circular experiment: the reported matches reproduce the same manual correspondences used to build the training labels, so the landing-site claim is unsupported as presented. read the letter →

arxiv 2501.08584 v1 pith:2L77BOMX submitted 2024-12-31 physics.geo-ph physics.space-ph

Fanwei Meng , Xiaopeng Wang , André Antunes , Jie Zhao , Guoliang Zhou , Biqiong Wu , Tianqi Hao This is my paper · ORCID

classification physics.geo-phphysics.space-ph

keywords MarslandingsitesQaidamBasinevaporitesSiamesenetworkcontrastivelossgeomorphologicalsimilaritydeeplearningplanetaryanalogues

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes using a Siamese convolutional neural network trained with a contrastive loss to turn visual similarity between Mars and Qaidam Basin surface images into a quantitative number. The network embeds each 100×100 image patch into a five-dimensional feature vector, and the Euclidean distance between two vectors is the terrain-similarity score. On four geomorphic classes, the closest matches recovered are exactly the pairings a geologist would choose: yardangs, linear dunes, barchan chains, and polygonal salt crusts, with distances 0.11, 0.26, 0.38, and 0.05. The paper's aim is to make Mars landing-site selection an automatic, repeatable ranking procedure, with priority going to evaporite polygonal terrains where halite fluid inclusions could preserve signs of life.

What carries the argument

The mechanism is a Siamese network: two identical convolutional networks that share the same weights and map an input image to a five-dimensional feature vector, with similarity defined as the Euclidean distance $D_W(X_1,X_2)=\|G_W(X_1)-G_W(X_2)\|_2$. Training uses a contrastive loss (the DrLIM formulation from Hadsell et al., 2006, applied to pairs labeled similar or dissimilar by the researchers): the loss shrinks the distance for same-location pairs and, for different-location pairs, increases it up to a margin $m$. This learned embedding is what carries the argument, because after training, matching a new Mars image to a Qaidam image is just feature extraction plus a distance lookup, with no retraining.

What would settle it

Take the trained network and feed it Martian terrain images from regions that were not among the four training targets. If its closest Qaidam Basin matches do not match an independent classification (for example, expert geomorphological labels or hyperspectral evidence of evaporites), the landing-site-selection claim fails. A faster check is label permutation: retrain with the Mars–Qaidam pairings randomly shuffled; if accuracy stays near 99%, the metric is not learning the intended analogue structure.

Watch

Extended reading notes

Core claim

The central claim is that a weight-sharing Siamese CNN with a contrastive loss can learn a globally coherent nonlinear mapping in which images from the same landform type are pulled close together and images from different types are pushed apart, so that a new terrain image can be matched to its best Earth analogue by nearest-neighbour distance alone. The paper reports that the trained model reaches a test-set accuracy of 99.82% and that the four minimum distances between the Qaidam sub-regions and the four Martian target regions are all below 0.5, with the polygonal-structure pairing D-D' the closest at 0.05. In the authors' words, 'By identifying different landforms, priority landing sites on Mars can be selected.' The application-level corollary is that evaporite-rich polygonal terrain in the Qaidam Basin is the best Earth model for deciding where to land a Mars rover in search of preserved biosignatures.

Load-bearing premise

The load-bearing premise is that the four manually chosen Mars–Qaidam pairings used to label the training data are the right ground truth, so the network is only learning to reproduce those pairings rather than to recognise geomorphological similarity that would transfer to unlabelled Mars regions.

Editorial extensions

If this is right

Landing-site preselection becomes a quantitative ranking: a future mission can tile a candidate region into patches, embed them, and rank them by distance to terrestrial analogue images.
The metric can serve as a validation tool for geologist-chosen analogues, since it recovered the A-A', B-B', C-C', and D-D' pairings with distances below 0.5.
Polygonal salt-crust terrains are singled out as the highest-priority targets, because halite and gypsum can carry fluid inclusions that may preserve biological signatures over geological time.
The same Siamese pipeline transfers to other analogue deserts (for example, Atacama or Antarctic dry valleys) and to other planetary bodies, provided paired training images exist.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The reported 99.82% accuracy is measured on random splits of pairs built from the same four labelled Mars–Qaidam correspondences, so I read it as evidence that the network reproduces the manual labels rather than proof that it can discover new analogues; a held-out-region test would separate the two.
Because the embedding is five-dimensional and the inputs are 100×100 optical patches, the similarity is likely dominated by texture and pattern statistics; folding in mineralogical or topographic channels would test whether the same metric tracks geological process, not just visual resemblance.
Tiling all of Mars through the same embedding would yield a global distance-to-Qaidam map, turning the four point matches into a planet-wide similarity surface that could be inspected for unanticipated analogues.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Short version: this is a useful idea wrapped in a circular experiment. The new bit is applying a Siamese CNN with contrastive loss to Mars–Earth landform matching, and the geological background on Qaidam Basin evaporites as a Mars analogue is competent and well referenced. But the headline matching distances are not evidence of predictive ability, because the evaluation is built from the same four manual correspondences used to construct the training labels.

What it does well: the paper makes a genuine case for why quantitative, repeatable analogue selection would help landing-site screening, and the descriptions of halite, fluid inclusions, and the Qaidam Basin literature are solid. It is also transparent about the dataset construction, which makes the flaw easy to trace.

Where it falls down: in §5.1 the training pairs are built by pairing one Qaidam image from each region A–D with all images from the corresponding Mars region A'–D' as similar. Section 5.3 then reports that exactly A–A', B–B', C–C', D–D' come out as closest. That is the network reproducing its own labels. The 9:1 train/test split is applied after 10× augmentation, so augmented copies of the same source image can appear in both sets; 99.82% accuracy is consistent with near-duplicate recognition rather than geomorphological understanding. There is no held-out Mars region, no baseline, no error bars, and no code or data to assess how much of the result is image-level memorization.

This is a fixable design problem, not a broken concept. The proper test would train on some areas and evaluate on held-out pairs that never contributed to labels, with a spatial split that keeps augmented versions of a source image in one fold, plus simple baselines like pixel-level distance or a generic pretrained CNN embedding. Until that is done, the landing-site selection claim is unsupported.

The paper deserves a serious referee, because the application is timely and the flaw is instructive, but my own verdict is that the experimental core needs to be redesigned before the claims can stand. It is a good reading-group example of circular validation in learned similarity.

Referee Report

4 major / 6 minor

Summary. The paper proposes a Siamese convolutional neural network trained with a contrastive loss to learn a similarity metric between terrestrial and Martian surface images, and applies it to four hand-selected geomorphic regions in the Qaidam Basin and four corresponding regions on Mars. The authors report a 99.82% test-set accuracy and minimum Euclidean distances for the pairs A-A', B-B', C-C', and D-D', concluding that the model identifies the same analogue correspondences that geologists selected and can therefore be used to select priority Mars landing sites. The geological background on evaporites, halite, and fluid inclusions as biosignature repositories is appropriate for an astrobiology-motivated analogue study, but the supporting experiment is severely compromised by the way training labels are constructed and by the absence of any held-out evaluation on new terrain.

Significance. If the method were shown to transfer to unseen Martian terrains, it would offer a useful quantitative complement to expert geological judgement for planetary analogue site selection. The paper is, to the best of my knowledge, novel in applying deep metric learning to Mars-Earth geomorphological matching, and the underlying geological motivation (evaporites and halite fluid inclusions as preservation windows) is well referenced. The authors also make the standard choice of a Siamese architecture and contrastive loss, following Hadsell et al. (2006). However, the experimental validation is not independent of the label construction: the four positive correspondences used to build training pairs are the same four correspondences later reported as the minimally distant matches, and the train/test splitting procedure allows augmented copies of identical source images to appear in both sets. As a result, the reported accuracy and distance values do not provide evidence for the paper's central claim that the model can select landing sites by discovering analogues in unlabelled Mars imagery.

major comments (4)

[§5.1 and §5.3] The experimental evaluation is circular. In §5.1, similar pairs are defined by taking one image from each Qaidam region and pairing it with 'all images from the corresponding research area on Mars' as similar, and with images from the other three Mars regions as dissimilar. The 'corresponding' Mars areas are exactly the four regions later reported in §5.3 to have the smallest distances (A-A', B-B', C-C', D-D'). The network is therefore trained to minimize distances for these four manual correspondences and then evaluated on those same correspondences. The reported minimum distances 0.11, 0.26, 0.38, and 0.05 are not a test of whether the model has discovered the same analogue pairings independently; they are a measure of how well the model reproduces the training labels. This undermines the landing-site selection claim.
[§5.1] The 9:1 random split of labeled pairs does not prevent data leakage. The images are augmented by a factor of 10 through rotation, blurring, scaling, and grayscaling, and the pairs are built from these augmented images. A random split of pairs, rather than a split by source image identity or by geographic sub-region, means that augmented copies of the same original image can appear in both training and test sets. The reported 99.82% test accuracy can therefore reflect near-duplicate recognition rather than generalization to new terrain. To validate the method, the authors would need to evaluate on geographically disjoint areas or on Mars regions not used during training.
[Abstract and §5.3] The paper's central claim that the model can select priority landing sites on Mars is not supported by the experiment. All four Mars regions (A', B', C', D') were pre-selected by the authors as known or suspected analogues based on prior geological literature, and the experiment only asks whether the model reproduces those four correspondences. No test is conducted on unlabelled Mars terrains, no candidate set is ranked beyond the four pre-chosen regions, and no external validation against expert judgement on a held-out regional set is provided. The conclusion 'By identifying different landforms, priority landing sites on Mars can be selected' therefore extends beyond what the experiments demonstrate.
[§4.2 and §4.3] The equations defining the distance function and the contrastive loss are missing. In §4.2 the text states that the Euclidean distance 'can be defined as:' followed by an empty reference '(1)', and in §4.3 the loss function is referenced as equations (2), (3), and (4) without any of the formulas being displayed. Since these definitions are the core of the method and are needed to assess the correctness of the training procedure, the manuscript must show the explicit equations for D_W, the contrastive loss, and its partial losses L_S and L_D.

minor comments (6)

[Throughout] There are numerous typographical and grammatical errors, including 'Qadam' for 'Qaidam', 'Accrodingly', 'geomorpholigcal', 'analaogue', 'cystals', and inconsistent punctuation around citations. These should be corrected.
[§5.3] The coordinates for region D' are given as '(341.61°E, 50.70°N)', which reverses the latitude/longitude order used for the other regions (e.g., A'=10.461°S, 176.446°E). Please standardize the coordinate format.
[§5.3] The phrase 'Their differences are all less than 0.5, indicating a very high similarity' introduces an arbitrary threshold of 0.5 with no justification or calibration. Please justify the threshold or present the full distributions of distances rather than only minima.
[§5.2] The paper reports a single final loss value of 0.0088 and a test accuracy of 99.82%, but provides no training curves, validation loss, or error analysis. Adding convergence diagnostics and a confusion matrix or per-class accuracy would help the reader assess the training process.
[§4.1] The dataset description is incomplete: the number of distinct images per region, the number of pairs generated, and the exact division of images among the four regions are not reported. This information is essential for reproducibility.
[Experimental Methods] The manuscript does not state whether code or trained model weights will be made available. Given the reproducibility standard in deep learning research, a data/code availability statement should be added.

Circularity Check

2 steps flagged · score 8.0 of 10

The reported A-A', B-B', C-C', D-D' minimum-distance matches reproduce the manual correspondences used to build training labels; the landing-site 'prediction' is forced by construction.

fitted input called prediction [Section 5.1 (Dataset) and Section 5.3 (Selection of Mars analog sites in the Qaidam Basin)]
"One image sample was selected from each of the four research areas in the Qaidam Basin and paired with all images from the corresponding research area on Mars, marked as similar. It was also paired with all images from different research areas on Mars, marked as dissimilar. The labeled dataset was randomly divided into training and testing sets at a ratio of 9:1. [...] We found that the matches A-A', B-B', C-C', and D-D' have the smallest differences of: 0.11, 0.26, 0.38, and 0.05 respectively."

The training labels define 'similar' exclusively as the four correspondences between a Qaidam sample from region A/B/C/D and all Mars images of the corresponding region A'/B'/C'/D'; all other Mars pairings are 'dissimilar'. The contrastive loss is then minimized for exactly those four pairings. Section 5.3 reports that the smallest final Euclidean distances occur for exactly A-A', B-B', C-C', and D-D'. This is not an independent prediction of similarity: the network was optimized to make these four correspondences closer than all others, so the reported minimum distances are the training objective read back as a result. No held-out Mars region or geologically unseen pairing is tested.
other [Section 5.1 (Dataset) and Section 5.2 (Establishment and training of the Siamese network model)]
"The images in the dataset were augmented by a factor of 10 through rotation, blurring, scaling, and grayscaling. [...] The labeled dataset was randomly divided into training and testing sets at a ratio of 9:1. [...] After iterating over the entire dataset, the final loss value is 0.0088, and the accuracy of the model is verified on the test set to be 99.82%."

The random 9:1 split is applied to pairs after 10x augmentation and after all positive pairs are restricted to one Qaidam sample from each of the four regions paired with all images of the corresponding Mars region. Consequently, augmented copies of the same source images can appear in both training and test sets, and the test pairs are drawn from the same four manual correspondences used to fit the embedding. The 99.82% accuracy therefore measures near-duplicate recognition within the labelled correspondences rather than generalization to new, unlabelled Martian terrain, so it cannot validate the landing-site selection claim.

full rationale

The central experimental claim is the set of minimum distances in Section 5.3 and the 99.82% test accuracy in Section 5.2, offered as evidence that the Siamese CNN can select Martian landing sites by geomorphological similarity to Qaidam Basin analogues. The construction in Section 5.1 makes this claim circular: the 'similar' pairs are defined as Qaidam region A with Mars region A', B with B', C with C', and D with D', and these are exactly the pairs later reported as having the smallest distances. The contrastive loss explicitly minimizes the Euclidean distance for those labels, so the output distances are a read-out of the input labels rather than a discovered correspondence. Additionally, the train/test evaluation is performed on random splits of pairs built from the same four labelled correspondences after augmentation, so the high accuracy can be obtained by recognizing augmented near-duplicates. Because no Mars region, landform, or geographical area outside the four hand-picked correspondences is held out, the paper does not demonstrate predictive transfer of the learned similarity metric to new terrain. The geological and astrobiological discussion is not circular, but it does not rescue the load-bearing matching result, which is forced by the label construction. No self-citation chain is involved; the circularity is internal to the dataset and evaluation design.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central result depends on hand-chosen hyperparameters, a post-hoc similarity threshold, and domain assumptions about Qaidam as a Mars analogue. No new physical entity is introduced; the only constructed object is the learned five-dimensional feature representation.

free parameters (5)

contrastive loss margin m = not reported
The margin defining the radius around G_W(X) for dissimilar pairs in the loss function is not given in the paper, but it controls the learned embedding geometry.
similarity threshold 0.5 = 0.5
Section 5.3 declares differences below 0.5 as indicating very high similarity, but this threshold is chosen after seeing the result and is not calibrated on independent data.
feature vector dimension = 5
The output dimensionality of the embedding is set to 5 by hand with no ablation or justification in Section 5.3.
network hyperparameters = lr=0.001, weight decay=0.0005, batch size=64, epochs=300
These training choices in Section 5.2 are stated without sensitivity analysis, and the final results depend on them.
tiling and augmentation factors = 10 sub-images, 10x augmentation, 100x100 images
Section 5.1 fixes these data-construction choices, but their effect on generalization is never tested.

assumptions (4)

domain assumption Qaidam Basin is a valid Mars analogue for evaporite and aeolian landforms.
The paper relies on earlier geological studies (Zheng et al. 2013; Xiao et al. 2017) rather than demonstrating this equivalence within the deep-learning experiment.
domain assumption Visual similarity of surface images implies landing-site suitability for astrobiological targets.
Section 5.3 equates low Euclidean distance between CNN features with priority for life-detection missions, but no mineralogical, spectral, or habitability ground truth is linked to the similarity scores.
ad hoc to paper The hand-picked four Earth-Mars correspondences are the correct classes to learn and to evaluate.
Section 5.1 constructs the training labels from exactly these correspondences, so the experiment assumes the truth of the analogy it later reports as its main finding.
domain assumption Google Earth imagery and published Mars orbital images are directly comparable for cross-planet matching.
The paper does not correct or match image resolution, illumination, projection, or sensor characteristics before treating Earth and Mars images as comparable inputs.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Deep learning selection of analogues for Mars landing sites in the Qaidam Basin, Qinghai-Tibet Plateau." pith.science (2026). https://pith.science/paper/2L77BOMX

@misc{pith2026250108584,
  author       = {Pith},
  title        = {Pith review of: Deep learning selection of analogues for Mars landing sites in the Qaidam Basin, Qinghai-Tibet Plateau},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2L77BOMX}},
  note         = {Machine review of arXiv:2501.08584}
}

read the original abstract

Remote sensing observations and Mars rover missions have recorded the presence of beaches, salt lakes, and wind erosion landforms in Martian sediments. All these observations indicate that Mars was hydrated in its early history. There used to be oceans on Mars, but they have now dried up. Therefore, signs of previous life on Mars could be preserved in the evaporites formed during this process. The study of evaporite regions has thus become a priority area for Mars' life exploration. This study proposes a method for training similarity metrics from surface land image data of Earth and Mars, which can be used for recognition or validation applications. The method will be applied in simulating tasks to select Mars landing sites using a selecting small-scale area of the Mars analaogue the evaporite region of Qaidam Basin, Qinghai-Tibet Plateau. This learning process minimizes discriminative loss function, which makes the similarity measure smaller for images from the same location and larger for images from different locations. This study selected a Convolutional Neural Networks (CNN) based model, which has been trained to explain various changes in image appearance and identify different landforms in Mars. By identifying different landforms, priority landing sites on Mars can be selected.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages

[1]

Artificial Intelligence

Introduction Based on its geological time scale, Mars is divided into the Noachian era (4 billion to 3.7 billion years ago), Hesperian era (3.7 -3 billion years ago), the Amazonian era (3 billion years ago to present). A large amount of geological data supports that the re was a rich hydrosphere on Mars during the Noachian period, with a scale equivalent ...

work page 2010
[2]

The climate of the Paleocene was dry and resulted in the input of red coarse debris deposits

Geological Settings The western part of the Qaidam Basin sank on the basis of the Mesozoic fault basin, transitioning from a small freshwater lake to a large depression lake. The climate of the Paleocene was dry and resulted in the input of red coarse debris deposits. In the early Oligocene, due to the dry climate and the uplifting of the Altun and Kunlun...

work page 2017
[126]

Origin of the salina salt of Michigan

doi: 10.1128/microbe.1.120.1 Dellwig, I.F., 1955. Origin of the salina salt of Michigan. J, Sediment. Petrol. 25, 83-110. Di Achille, G., Hynek, B. M., 2010. Ancient ocean on Mars supported by global distribution of deltas and valleys. Nat. Geosci. 3, 459-463. Edwards, C.S., Bandﬁeld, J.L., Christensen, P.R., Fergason, R.L., 2009. Global distribution of b...

arXiv 2015

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Artificial Intelligence

Introduction Based on its geological time scale, Mars is divided into the Noachian era (4 billion to 3.7 billion years ago), Hesperian era (3.7 -3 billion years ago), the Amazonian era (3 billion years ago to present). A large amount of geological data supports that the re was a rich hydrosphere on Mars during the Noachian period, with a scale equivalent ...

work page 2010

[2] [2]

The climate of the Paleocene was dry and resulted in the input of red coarse debris deposits

Geological Settings The western part of the Qaidam Basin sank on the basis of the Mesozoic fault basin, transitioning from a small freshwater lake to a large depression lake. The climate of the Paleocene was dry and resulted in the input of red coarse debris deposits. In the early Oligocene, due to the dry climate and the uplifting of the Altun and Kunlun...

work page 2017

[3] [126]

Origin of the salina salt of Michigan

doi: 10.1128/microbe.1.120.1 Dellwig, I.F., 1955. Origin of the salina salt of Michigan. J, Sediment. Petrol. 25, 83-110. Di Achille, G., Hynek, B. M., 2010. Ancient ocean on Mars supported by global distribution of deltas and valleys. Nat. Geosci. 3, 459-463. Edwards, C.S., Bandﬁeld, J.L., Christensen, P.R., Fergason, R.L., 2009. Global distribution of b...

arXiv 2015