Pith. sign in

REVIEW 2 major objections 5 minor 9 references

CNN-Based Reanalysis of Optical Turbulence at the Canary Islands Observatories (OCAN)

T0 review · 2 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read This paper tries to establish that four turbulence regimes found by unsupervised clustering of CNN embeddings of Canary Islands SCIDAR heatmaps are physically meaningful atmospheric states, because each regime has a distinct surface wind si

desk verdict A promising but under-validated CNN reanalysis of Cn2 profiles; the four-cluster result is plausible, but the only independent check is confounded by season. read the letter →

arxiv 2607.25940 v1 pith:ODZLTE7G submitted 2026-07-28 astro-ph.IM

Airam Salas , Begoña García-Lorenzo , Julio A. Castro-Almazán , Gelys Trancho This is my paper · ORCID
classification astro-ph.IM
keywords opticalturbulenceGeneralizedSCIDARsitecharacterizationdeeplearningconvolutionalneuralnetworksfeatureextractionclusteringadaptiveoptics
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

By reprocessing the ~300-night Canary Islands Generalized SCIDAR database into 60-minute time–altitude turbulence heatmaps, feeding them through a frozen EfficientNetB1 backbone, and K-means clustering the resulting embeddings, this paper aims to show that the data contain recurrent, physically distinct turbulence regimes. The four clusters differ in the altitude and strength of turbulent layers, show sharp seasonal transitions, and — critically — have clearly different wind-rose patterns at both ORM and OT, even though wind information was excluded from the embeddings and clustering. The paper argues that these wind signatures independently validate the physical relevance of the classes, since only genuine atmospheric states would co-vary with circulation. A sympathetic reader would care because this offers an unsupervised, data-driven way to decompose a large turbulence archive into interpretable states, with potential uses in adaptive optics scheduling and site climatology.

What carries the argument

The central object is the turbulence heatmap: a time–altitude matrix of log C_n^2 values (60-minute window, 180-second bins) made into a three-channel tensor by appending sine and cosine of day-of-year as two extra channels. The tensor is passed through the frozen convolutional backbone of EfficientNetB1 (pre-trained on a large image corpus), which yields a compact embedding per heatmap. K-means clustering on these embeddings partitions the 238 high-quality heatmaps into four groups. The key simplifying mechanism is transfer learning: without any fine-tuning on atmospheric data, the pre-trained filters, as padding-preserving 'feature extractors,' are claimed to retain enough of the heatmaps'

What would settle it

Run the same pipeline on the same 238 heatmaps after randomly permuting the turbulence values within each heatmap (or blanking the first channel, leaving only the sine/cosine channels); if K-means still returns four clusters with the same seasonal distributions and the same wind-rose separations, the clustering is driven by the seasonal channels, not by turbulence structure. Conversely, if clusters collapse or lose wind signatures when the two seasonal channels are removed, that would confirm the turbulence structure is necessary.

Watch

Extended reading notes

Core claim

The central discovery claimed is that unsupervised clustering of deep embeddings extracted from turbulence heatmaps recovers four atmospheric turbulence regimes with distinct vertical-layer structures and distinct associated wind-direction and speed distributions. At ORM, clusters show different prevailing wind directions; at OT, the regimes also separate in wind space despite both sites sharing similar large-scale turbulence behaviour. Because wind data were not included in any stage of the CNN feature extraction or K-means clustering, the cluster-specific wind roses are presented as independent evidence that the embeddings preserve physically meaningful information about the atmospheric st

Load-bearing premise

The whole interpretation rests on the assumption that the CNN embeddings encode the physical vertical-turbulence structure rather than the injected day-of-year channels, because two of the three input channels are pure seasonality and no ablation or null model is provided to rule out that the clusters merely recapitulate the seasonal input.

Editorial extensions

If this is right

  • Each identified regime provides a compact label for a full 60-minute turbulence state, enabling night-by-night or hour-by-hour classification from C_n^2 profiles alone.
  • Regime membership can be correlated with integrated AO parameters (seeing, isoplanatic angle, coherence time), potentially allowing regime-aware scheduling of adaptive-optics instruments.
  • The same pipeline can be applied to other long-term SCIDAR or DIMM archives to check whether analogous turbulence regimes exist at other sites.
  • The link between clusters and wind roses suggests that regime transitions track synoptic circulation changes, offering a data-driven connection from turbulence profiles to meteorology.
  • Because clusters emerge without fixed calendar bins, the analysis defines 'natural' seasons of turbulence rather than imposing predefined ones.

Reading between the lines

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

  • The sine/cosine day-of-year channels are two of the three input channels, so the strong seasonal separation of clusters in Fig. 5 is at least partly a product of the input representation; an ablation that removes these channels would show how much genuinely turbulence-structural information the embeddings retain.
  • The 238 heatmaps are treated as independent samples, but if consecutive heatmaps overlap temporally, cluster assignments are autocorrelated and the wind-rose statistics may overstate separation; reporting the temporal stride would clarify effective sample size.
  • A direct null-model test — e.g., shuffling the turbulence pixels within each heatmap while keeping the seasonal channels intact — would determine whether the clusters and their wind signatures depend on vertical turbulence structure at all, or only on season.
  • The paper's own statement that embedding quality depends on matching heatmap resolution to the CNN's receptive scales suggests a tunable hyperparameter; this could be used to search for an optimal heatmap aspect ratio for other sites.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper presents a machine-learning reanalysis of the Generalized SCIDAR C_n^2(h) database at the two Canary Islands observatories (ORM and OT). The authors build an ETL pipeline, aggregate individual turbulence profiles into time-altitude heatmaps, and convert each accepted heatmap into a three-channel tensor containing normalized turbulence intensity plus sine and cosine encodings of the day of the year. These tensors are passed through a frozen pre-trained EfficientNet CNN to produce embeddings, which are clustered with K-means. A four-cluster solution is adopted and visualized in terms of mean turbulence maps, seasonal occurrence, and surface wind roses. The central claim is that the resulting clusters capture physically meaningful atmospheric circulation regimes, based mainly on the observation that wind-rose patterns differ across clusters even though wind information was not used in the clustering.

Significance. If validated, this approach would provide a novel data-driven way to extract recurrent turbulence regimes from large astronomical site-characterization archives, with potential utility for adaptive-optics scheduling and site monitoring. The paper builds an original end-to-end processing framework and applies modern representation-learning tools to a valuable public-domain dataset. The central claim, however, is not yet established. The analysis currently lacks controls for a strong seasonal confound, quantitative cluster validation, and an explicit treatment of sample non-independence. The methodology is promising, but the physical interpretation goes beyond what the presented evidence supports.

major comments (2)
  1. [§3.5 and §5, Figs. 5-7] The analysis uses 238 high-quality heatmaps, but the paper never reports the temporal stride or overlap between heatmaps, nor the quantitative thresholds for the selection criteria listed in §3.5 ('temporal coverage, minimum profile density, maximum interpolated interval'). If the heatmaps are generated from overlapping windows, the 238 samples are not independent. This would inflate the apparent significance of the cluster structure and of the wind-rose differences. Please report the full sampling schedule of the heatmaps, the effective number of independent samples, and repeat the clustering and wind-rose analysis with non-overlapping windows as a robustness check.
  2. [§3.6 and §4] The wind-rose differences in Figures 6 and 7 are presented visually, without any statistical test or statement of the number of wind measurements per cluster. Even if the seasonal confound were removed, the visual differences could be within sampling noise, especially if some clusters contain only a small number of heatmaps. A formal comparison of circular distributions, such as a permutation test on cluster labels or a two-sample Kuiper test, is needed to quantify the strength of the wind-rose separation.
minor comments (5)
  1. [§4] The text introducing Figure 5 says that seasonal patterns 'emerge naturally from the underlying turbulence structure'. This phrasing is misleading, because two of the three input channels explicitly encode the day of year. Consider rewording to acknowledge the role of the seasonal channels in the input representation.
  2. [Figures 6-7] The description of the wind stations notes that the JKT anemometer is on a rooftop and less than 5 m from the telescope dome on the prevailing leeward side, and the GONG station is also on a roof. These local obstructions may affect the measured wind directions. This limitation should be stated in the interpretation of the wind-rose results.
  3. [§3.4] The 'normalized arbitrary units' used for the turbulence intensity colour scale in Figure 4 are not defined. Please specify how the normalization was performed so that quantitative comparisons are meaningful.
  4. [General] No data or code availability statement is included. Since the paper introduces a complete processing framework, sharing the ETL and heatmap-generation code would significantly improve reproducibility and would also allow readers to test alternative clustering choices.
  5. [§1] The reference list correctly identifies previous statistical analyses of the OCAN database, but the reader would benefit from a more explicit statement of what those analyses found that the present clustering approach is meant to complement or improve.

Circularity Check

1 steps flagged · score 4.0 of 10

Seasonal channels are two of the three CNN input channels, so the seasonal separation of the clusters—and the wind-rose 'independent validation'—is partly forced by construction and not an emergent turbulence discovery.

  1. self definitional [Section 3.5 (High-Quality Tensor Generation), Section 4 (Fig. 5), Section 5]
    "Each accepted heatmap was converted into a three-channel tensor. The first channel contains the normalized turbulence intensity, while the remaining two channels encode the seasonal phase through sine and cosine transformations of the day of the year. ... Because wind information was not included in the CNN embeddings or clustering procedure, the observed wind-regime signatures suggest that the identified turbulence classes capture physically meaningful differences in the atmospheric conditions."

    Two of the three input channels directly encode day-of-year, so cluster membership is a function of season by construction. The seasonal organization shown in Fig. 5 is therefore in part a direct product of the input representation, not solely emergent from the turbulence channel. The wind-rose comparison is offered as independent because wind was not an input, but wind roses at these sites have a strong seasonal climatology, and season was an input; the same wind-rose separation could arise from the seasonal channels even if the turbulence channel carried no physical information. No ablation removing the seasonal channels or season-only null model is provided, and the paper itself concedes 'additional validation is required.'

full rationale

The central derivation chain is not a fitted-parameter-then-prediction loop: the CNN is a frozen pretrained feature extractor, clustering is unsupervised, and no parameter is fitted to the OCAN data in order to predict a related quantity. The self-citations ([2]–[4]) support data provenance and prior statistical context, not a load-bearing uniqueness theorem or ansatz, and the main clustering results are computed from real Cn2 profiles. However, one genuinely circular element is present: Section 3.5 injects sine and cosine of day-of-year into every tensor, so the seasonal dominance of the clusters in Fig. 5 is partly built into the inputs. The paper frames this as patterns that 'emerge naturally from the underlying turbulence structure,' and then uses wind-rose differences as independent validation on the grounds that wind was not used in clustering. Because wind direction at the Canary observatories has a seasonal climatology and season is a direct input, wind-rose separation across clusters is not independent of the input representation. The missing ablation/null-model control leaves the physical-relevance claim less supported than the paper suggests, though the turbulence-channel differences and external wind data give the work some independent content. Thus partial circularity, not full reduction.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The work inherits all SCIDAR calibration assumptions, adds domain-transfer assumptions for ImageNet features, and chooses several free parameters by hand. No new physical entities are proposed; the four regimes are labels for clusters over existing measurements.

free parameters (6)
  • heatmap timeframe = 60 min (final display)
    Total temporal extent of each heatmap; chosen by exploratory testing among temporal configurations, affecting the scale seen by the CNN.
  • bin duration = 180 s (final display)
    Temporal averaging interval for profiles; selected during experimentation; affects smoothness and temporal resolution of the heatmaps.
  • number of clusters k = 4
    Selected from inertia and silhouette analyses; all regime descriptions depend on this model choice.
  • maximum altitude and vertical grid = 20 km a.s.l.
    Profiles are truncated to 20 km; grid resolution and normalization choices influence the embeddings but are not fully specified.
  • heatmap quality thresholds = unspecified
    Criteria for temporal coverage, minimum profile density, and maximum interpolation gap determine the 238 accepted heatmaps; values are not reported.
  • CNN backbone architecture = EfficientNetB1
    One of several architectures tested; chosen based on matching to heatmap resolution, not derived from theory.
assumptions (6)
  • domain assumption Generalized SCIDAR Cn2(h) profiles are accurate, calibrated, and representative of the two observatories.
    The entire analysis ingests these measurements without re-validation (Section 2).
  • domain assumption ImageNet-pretrained CNN features transfer to turbulence heatmaps.
    Frozen EfficientNet features trained on natural images are assumed to preserve distinctions in Cn2 heatmaps (Section 3.6).
  • domain assumption K-means Euclidean distance on CNN embeddings captures atmospheric regime similarity.
    The clustering step assumes spherical, Euclidean-separable structure in embedding space (Section 3.6).
  • ad hoc to paper The seasonal sine/cosine channels act as conditioning, not as the dominant driver of cluster separation.
    Two of three input channels encode day-of-year phase (Section 3.5); the paper does not test whether clusters form without them.
  • domain assumption The 238 high-quality heatmaps are treated as independent samples.
    Heatmaps may be generated from overlapping 60-minute windows, but no stride or temporal-autocorrelation analysis is provided (Section 4).
  • domain assumption Surface wind measurements from rooftop stations represent the circulation associated with each cluster.
    Used as independent validation in Section 4 (Figures 6-7); the JKT anemometer is described as less than 5 m from the dome on the leeward side, so contamination is possible.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CNN-Based Reanalysis of Optical Turbulence at the Canary Islands Observatories (OCAN)." pith.science (2026). https://pith.science/paper/ODZLTE7G

@misc{pith2026260725940,
  author       = {Pith},
  title        = {Pith review of: CNN-Based Reanalysis of Optical Turbulence at the Canary Islands Observatories (OCAN)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ODZLTE7G}},
  note         = {Machine review of arXiv:2607.25940}
}
read the original abstract

Atmospheric optical turbulence, caused by refractive-index fluctuations driven by wind and temperature inhomogeneities, is the primary source of image degradation in ground-based telescopes. Its vertical distribution, described by the refractive-index structure constant (Cn2(h)), determines essential parameters such as seeing, isoplanatic angle, and coherence time, which characterize astronomical sites and are also crucial for the design and performance of high-resolution instruments. The Observatorios de Canarias (OCAN), comprising the Observatorio del Roque de los Muchachos on La Palma and the Observatorio del Teide on Tenerife, have been extensively characterized through long-term Generalized SCIDAR campaigns. In total, the database comprises nearly 300 observing nights and hundreds of thousands of individual Cn2(h) profiles. We present a reanalysis of the OCAN database using deep-learning techniques. An end-to-end software framework was developed to ingest, preprocess, and transform the original turbulence measurements into standardized numerical representations suitable for machine-learning analysis. Temporal-vertical turbulence heatmaps are generated under multiple temporal configurations and enriched with seasonal information. These representations are processed by pre-trained convolutional neural networks (CNN) used as frozen feature extractors to obtain compact deep embeddings. Unsupervised clustering is subsequently applied to identify latent atmospheric regimes within the dataset. The resulting clusters reveal coherent turbulence structures and offer a complementary perspective on atmospheric conditions at the Canary Islands observatories. This methodology provides new insights into turbulence variability at the Canary Islands sites and represents a novel approach to large-scale turbulence characterization using modern deep-learning techniques.

Figures

Figures reproduced from arXiv: 2607.25940 by the authors.

Figure 1
Figure 1. Monthly statistical C2 n profiles for the Roque de los Muchachos Observatory (ORM, purple profiles) and the Teide Observatory (OT, brown profiles) derived from Generalized SCIDAR measurements. Profiles were obtained by combining all available turbulence measurements acquired during each month over the full observing campaigns. The horizontal axis shows the turbulence strength, C2 n, in logarithmic scale (m−2/3 ), wh… view at source ↗
Figure 2
Figure 2. Workflow of the CNN-based reanalysis pipeline. Generalized SCIDAR turbulence profiles are transformed [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Cluster visualization for the full C2 n(h) database for OCAN (k=4) using EfficientNetB1 and 180 seconds bins within 60 minutes timeframe. 4. RESULTS The results presented below constitute a preliminary assessment of the proposed CNN-based analysis framework. The primary objective is to evaluate whether the methodology can identify coherent patterns within the tur￾bulence database; therefore, the physical interpretat… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Mean temporal–vertical turbulence maps associated with the four turbulence regimes identified from the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Seasonal distribution of the four turbulence regimes identified through CNN-based clustering of the complete [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Wind-rose distributions associated with the four turbulence regimes identified at the ORM. Each panel shows the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Same as Fig [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 2 canonical work pages

  1. [2]

    MNRAS , keywords =

    Adaptive optics parameters connection to wind speed at the Teide Observatory: corrigendum. MNRAS , keywords =. doi:10.1111/j.1365-2966.2011.17787.x , adsurl =

    arXiv 2011

  2. [3]

    Optical Turbulance: Astronomy Meets Meteorology , year = 2009, editor =

    Vertical Turbulence Profiles at Canary Islands Astronomical Sites. Optical Turbulance: Astronomy Meets Meteorology , year = 2009, editor =. doi:10.1142/9781848164864_0009 , adsurl =

    work page doi:10.1142/9781848164864_0009 2009

  3. [4]

    International Conference on Machine Learning , year=

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks , author=. International Conference on Machine Learning , year=

  4. [5]

    Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability , year=

    Some Methods for Classification and Analysis of Multivariate Observations , author=. Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability , year=

  5. [6]

    Journal of Machine Learning Research , year=

    Visualizing Data using t-SNE , author=. Journal of Machine Learning Research , year=

  6. [7]

    arXiv:1802.03426 , year=

    UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction , author=. arXiv:1802.03426 , year=

    arXiv

  7. [8]

    IEEE Transactions on Knowledge and Data Engineering , year=

    A Survey on Transfer Learning , author=. IEEE Transactions on Knowledge and Data Engineering , year=

  8. [9]

    Philosophical Magazine , year=

    On Lines and Planes of Closest Fit to Systems of Points in Space , author=. Philosophical Magazine , year=

Show all 9 references
  1. [10]

    Applied Optics , year = 1997, month = oct, volume =

    Whole atmospheric-turbulence profiling with generalized scidar. Applied Optics , year = 1997, month = oct, volume =. doi:10.1364/AO.36.007898 , adsurl =

    1997 doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.