{"id":"35ca9146-d8b6-4e76-a52c-b21c71c1e105","arxiv_id":"2505.24638","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A U-Net with channel attention and angle-conditioned normalization retrieves cloud optical thickness from two-band simulated radiance with lower MSE than single-angle deep models and IPA.","lead":"A U-Net with attention and angle embeddings, CAAC, estimates cloud optical thickness from two-band satellite radiance images under varying sun and viewing angles. On simulated LES/SHDOM data it beats single-angle deep models and IPA by about nine-fold, but only about 1.5-fold against comparably trained multi-angle baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 9x MSE gain rests on an undisclosed train/test split; overlapping stride-10 windows from 102 cloud fields make random-window splits a severe leakage risk.","rationale":"The reader identified the same weakest assumption: the absence of a described train/test split despite overlapping, stride-10 windows from only 102 cloud fields. This is the most load-bearing concern because the headline empirical result, a factor-of-nine MSE reduction, cannot be interpreted without knowing whether test windows share cloud fields with training windows. If the split is random at the window level, the reported comparison is not an evaluation of generalization to new cloud scenes, and the quantitative claim collapses. If the split is at the field level, the concern is resolved and the result may stand, albeit with a much smaller advantage over multi-angle baselines. The concern is concrete, testable, and fixable rather than a fundamental flaw in the architecture. I agree with the reader's conditional verdict: the paper should not be accepted as-is without disclosing the split protocol, releasing code/data, or otherwise demonstrating field-disjoint evaluation. No independent formal verification or code artifact is provided, so the manuscript's reliability rests entirely on the described experiments. The main reason I do not move to REJECT is that the architecture and angle-conditioning idea are plausible, and the missing split information could be supplied by the authors; the concern is about the evidence for the central claim, not about internal mathematical inconsistency.","tokens_in":8192,"tokens_out":3021,"duration_ms":40267,"concrete_test":"Run a controlled split experiment with the released (or re-generated) dataset: randomly split the 102 cloud fields into training and test sets at the field level, e.g., 82 training fields and 20 test fields, and also perform a random window-level split into the same total number of training and test windows. Train CAAC, UNet, and CloudUNet under both regimes using identical seeds, learning rate search, and early stopping, then recompute the Section 5/Table 2 MSE ratios. If the field-level split reduces CAAC's MSE advantage over single-angle models from roughly 9x to below 2x, or makes the CAAC-versus-multi-angle-UNet gap statistically indistinguishable from noise, the headline conclusion is an artifact of leakage. Report test-set size and standard deviations over at least five seeds.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim in Section 5 and the abstract is that CAAC achieves roughly 9x, 7x, and 11x lower MSE than IPA and single-angle deep learning baselines. That claim is only meaningful if the test set contains cloud scenes, not just windows, that the model has not seen during training. The Implementation Details in Section 4 specify only that 64x64 windows with stride 10 are extracted; no train/test split is described anywhere in the paper. With only 102 LES cloud fields, each approximately 140x140 pixels at 100 m resolution, stride-10 extraction yields about 64 heavily overlapping windows per field. If windows are assigned to training and test sets randomly rather than at the cloud-field level, then adjacent and overlapping windows from the same cloud field appear in both sets, allowing the model to memorize local spatial structure rather than learn a general radiance-to-COT mapping. This leakage would differentially improve CAAC, whose attention and AdaIN modules can exploit the repeated context, and would inflate the reported ratios against IPA and single-angle baselines. The paper also does not state the number of test fields/windows or report repeated-split variability, so the factor-of-nine headline is not currently independently checkable. A secondary issue is that even under an honest field-level split, CAAC's advantage over multi-angle UNet and CloudUNet in Table 2 is about 1.5x, not 9x; the nine-fold claim is thus tied specifically to the weakest baselines and to the unresolved split question.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CAAC, a UNet-style deep learning model with channel attention and AdaIN-based angle conditioning, for retrieving cloud optical thickness (COT) from two-wavelength radiance images. Experiments use 102 LES cloud fields with SHDOM-simulated radiances at multiple solar and view zenith angles, evaluated against IPA, UNet, and CloudUNet baselines. The central reported result is that multi-angle CAAC achieves 0.0459 MSE on the all-angle partition, which the authors describe as a nine-fold improvement over single-angle methods and IPA; the abstract and conclusion generalize this to 'at least a factor of nine' reduction in retrieval errors.","tokens_in":8420,"tokens_out":4486,"duration_ms":54622,"significance":"If the claims hold, the work is a useful step in angle-robust COT retrieval: the use of 3D radiative transfer simulation, explicit angle conditioning, and multi-angle training is sensible, and the ablations (CAAC with/without attention and angle coding) help isolate the contribution of each component. The paper's strengths include the physically grounded simulated dataset, the clear architectural motivation, and the attempt to evaluate on unseen angles. However, the headline factor-of-nine claim is overstated relative to the multi-angle baselines in Table 2, and the paper omits a crucial train/test split description, which currently prevents the main quantitative result from being independently verified.","major_comments":[{"comment":"The paper does not state whether the train/validation/test split is performed at the cloud-field level or at the level of individual windows. With 102 LES cloud fields and 64x64 windows extracted at stride 10, each field yields roughly 64 overlapping windows; a random window split would place spatially overlapping windows from the same field in both training and test sets, allowing the model to memorize local spatial structure and inflating the reported MSE improvements. Please specify the split granularity, report the number of test fields and test windows, and provide results under a field-level split or repeated cross-validation.","section":"Section 4, Implementation details"},{"comment":"The 'at least a factor of nine' claim is not supported by the full comparison in Table 2. In the SZA:all, VZA:all partition, CAAC (0.0459 MSE) is only 1.48x lower than multi-angle UNet (0.0678) and 1.69x lower than multi-angle CloudUNet (0.0778); the 9x figure arises only when comparing against IPA (0.4212) and single-angle UNet/CloudUNet (0.4097). The abstract, Section 5, and conclusion should state the baseline-specific ratios rather than a blanket 'at least nine-fold' improvement.","section":"Section 5, Table 2 and abstract"},{"comment":"The angle-invariance claim rests primarily on Figure 7, which is presented only as a heatmap with no numerical MSE values or comparison table for the 165 unseen angle combinations. It is also not stated whether the 'different set of LES cloud profiles' used in this evaluation is disjoint from the profiles used during training. Please provide quantitative results and confirm field-level disjointness.","section":"Section 5, Evaluation on unseen SZA and VZA"},{"comment":"The angle-coding mechanism is not fully specified. AdaIN is defined in Eq. (1), but the paper does not describe how the scalar SZA and VZA values are embedded, what network maps the angle code to the AdaIN mean and standard deviation, or at which decoder resolutions the AdaIN blocks are applied. Without this information, CAAC is not reproducible from the manuscript.","section":"Section 3 and Figure 2"}],"minor_comments":[{"comment":"There are typographical errors: the title contains 'RETRIEV ALS' instead of 'RETRIEVALS', and Table 1 has the header 'Parititon' instead of 'Partition'.","section":"Title and Table 1"},{"comment":"The shifted log transformation is written as log(COT+1); please specify whether this is the natural logarithm and clarify whether the reported MSE is computed on the transformed scale or after inverse transformation.","section":"Section 4, Data Pre-processing"},{"comment":"Figure 3 lacks axis labels and a precise explanation of what '20% of multi-angle training data' means, which makes the reported 3.4x and 7.5x improvements difficult to interpret or reproduce.","section":"Figure 3"},{"comment":"Reference [7] lists 'Rintaro Okamura et al.' without the full author list; please complete the citation.","section":"References"},{"comment":"The paper would benefit from a data/code availability statement and from explicit reporting of the number of training, validation, and test windows used in Table 2.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central unresolved issue is the undisclosed train/test split; if the authors can show field-level splitting and re-report the comparisons, the contribution is defensible, though the actual gain over multi-angle UNet/CloudUNet is modest (1.5x) rather than nine-fold. The paper is likely from a workshop/conference context, and the overclaim in the abstract should be corrected before considering publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's real contribution is a multi-angle deep learning framework for cloud optical thickness retrieval with explicit angle coding via AdaIN, evaluated on LES-based simulated radiance. That's a sensible and useful increment. The headline claim that CAAC cuts MSE by at least 9x is not supported by the paper's own Table 2: against multi-angle trained UNet it is about 1.5x, against multi-angle CloudUNet about 1.7x. The 9x figure comes from comparing to IPA and single-angle DL baselines, which is not the same as outperforming the current best DL baselines.\n\nThe paper does several things right. The data generation is careful: 102 LES cloud fields, SHDOM 3D radiative transfer at multiple SZAs and VZAs, realistic cloud fields. The ablation (CAAC-alpha, CAAC-beta) separates attention and angle coding, and the unseen-angle generalization test is a good idea. The AdaIN angle embedding is a clean way to inject viewing geometry without inflating the model. The direction is worth pursuing.\n\nThe soft spots are real. Most important: the paper never describes how the train/test split was done. With 102 cloud fields and heavily overlapping 64x64 windows extracted at stride 10, a random window split would leak shared spatial structure into training, inflating the reported numbers. The attention model, with more capacity to memorize local context, could plausibly benefit more from that leakage. The stress-test note is right to flag this; it is not a manufactured concern. The paper needs to specify a field-level split, report the number of test fields/windows, and ideally show repeated-split variability.\n\nSecond, the abstract overstates the result. The 'at least a factor of nine' language should be tied to the actual comparison set. Against the strongest baselines the gain is modest but still positive. That is fine, but it is not ninefold.\n\nThird, no code or data are released. For a paper whose main contribution is empirical, that is a real limitation.\n\nThe paper is worth sending to peer review. The methodology is sound in outline, the data generation is credible, and the angle-coding idea is useful to the remote sensing community. But the split question must be answered before the quantitative claim is believable, and the headline should be rewritten to match Table 2. A serious referee would be doing the field a favor.","headline":"A sensible angle-conditioned U-Net for cloud optical thickness retrieval, but the 9x headline overstates Table 2 and the missing train/test split description makes the main result hard to trust.","tokens_in":9010,"tokens_out":2186,"would_cite":false,"duration_ms":25698,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CAAC, an attention-based neural network with angle embeddings, is claimed to retrieve cloud optical thickness from two-wavelength radiance with at least ninefold lower mean squared error than single-angle baselines.","keywords":["cloud optical thickness","deep learning retrieval","attention mechanism","angle invariance","multi-angle training","remote sensing","radiative transfer","UNet"],"falsifier":"Train CAAC and all baselines on 64-by-64 windows from some LES cloud fields, hold out complete cloud fields for testing, and recompute the MSE ratios. If the ninefold advantage over single-angle methods shrinks substantially, or if CAAC no longer beats multi-angle trained UNet, the central claim is contradicted.","tokens_in":7949,"feed_emoji":"☁️","tokens_out":9891,"duration_ms":115371,"temperature":0.7,"pith_summary":"The paper sets out to show that cloud optical thickness (COT), the quantity that controls how much sunlight a cloud reflects, can be retrieved from two-wavelength satellite radiance without losing accuracy as the sun's angle or the satellite's viewing angle changes. The proposed model, Cloud-Attention-Net with Angle Coding (CAAC), is a compact UNet-style network whose attention gating sharpens cloud features and whose angle embeddings are injected through adaptive instance normalization. On a simulated dataset built from large-eddy-simulation cloud fields and a 3D radiative transfer model, the paper reports that multi-angle trained CAAC has the lowest mean squared error (MSE) of all methods tested, with a ninefold or larger reduction in MSE reported against the Independent Pixel Approximation and the single-angle deep learning baselines. The paper also reports that CAAC remains accurate on combinations of sun and viewing angles it never saw in training, which is the condition a real satellite retrieval would face. A reader should care because cloud optical thickness retrieval errors feed directly into biases in climate and weather model evaluations, and current deep learning baselines either ignore viewing geometry or must be retrained for each angle.","feed_headline":"Angle-aware attention model cuts cloud-thickness error ninefold","feed_subtitle":"Angle-aware deep network keeps error low as sun and viewing angles shift in simulated cloud scenes.","key_machinery":"The central object is the CAAC architecture: a UNet whose skip connections pass through a Channel Attention Module (CAM) that reweights feature maps channel by channel, and whose decoder layers receive the viewing geometry through Adaptive Instance Normalization (AdaIN). The AdaIN operation is the identity $$\\mathrm{AdaIN}(x,y)=\\$\\sigma$(y)\\,((x-\\mu(x))/\\$\\sigma$(x))+\\mu(y)$$, where $x$ is a feature map and $y$ is the angle information: the feature map is normalized by its own mean and standard deviation, then rescaled and shifted by statistics learned from the angle. This one operation is what lets a single shared network adapt its features to different solar and viewing angles instead of needing one model per angle. Multi-angle training supplies the angle-conditioned statistics, while the channel attention is what the paper credits with preserving informative cloud-structure features along the encoder-decoder path.","core_discovery":"On the paper's own terms, the discovery is that the same cloud field viewed under different solar zenith and view zenith angles produces radiance fields with changed intensity, shadowing, and apparent horizontal shifts, and that a retrieval model can be made invariant to those changes by conditioning its internal features on the angles themselves. The paper claims that CAAC, trained jointly on four solar zenith angles and three view zenith angles, outperforms the Independent Pixel Approximation, single-angle trained UNet and CloudUNet, and multi-angle trained versions of those networks on every data partition reported. It gives the headline margins as ninefold, sevenfold, and elevenfold lower MSE than the single-angle methods, and attributes the gain to channel attention in place of plain skip connections plus AdaIN-based angle coding. In a separate experiment on cloud profiles and angles not used in training, CAAC is reported to give the best and most stable performance across the tested grid of sun and viewing angles. The evidence is synthetic but physically motivated: large-eddy-simulation cloud fields with radiance computed by a 3D radiative transfer solver at 0.66 and 2.13 microns.","pith_inferences":["The ninefold headline is anchored to single-angle baselines; the paper's own table shows a smaller margin over multi-angle trained UNet and CloudUNet, so the practical size of the gain depends on which baseline is the point of comparison.","Because the paper does not specify whether the 64-by-64 windows were split by cloud field or randomly, and the windows overlap with stride 10, the reported factors should be treated as upper bounds until a whole-cloud-field held-out test confirms them.","The angle-conditioning mechanism is not specific to cloud optics: the same AdaIN-plus-attention pattern could be tested on other remote-sensing inversions whose signal is distorted by viewing geometry, such as aerosol optical depth or surface reflectance.","Demonstrating the model on real satellite radiance with independent cloud references, rather than simulated LES radiance, is the natural next test of whether the angle invariance survives the gap between simulated and observed scenes."],"forward_implications":["One model trained on multiple angles could replace angle-specific lookup tables or per-angle trained networks, saving memory and compute in operational satellite pipelines.","Angle-conditioned retrieval should reduce COT bias at large solar and viewing angles, exactly where the Independent Pixel Approximation shows the most over- and under-estimation due to shadows and cloud distortion.","The same architecture can be extended to jointly retrieve cloud optical thickness and effective radius in multi-angle settings, which is listed as the immediate next step.","If the reported generalization to unseen angles holds, satellite instruments with varying geometry could be processed without retraining for every possible solar and view configuration."],"supporting_citations":[{"why":"Defines the Independent Pixel Approximation baseline whose biases the paper aims to overcome.","marker":"[3]"},{"why":"Early multi-pixel deep learning retrieval that motivates learning COT from neighborhoods rather than single pixels.","marker":"[7]"},{"why":"A convolutional segmentation-style COT retrieval baseline included in the model comparisons.","marker":"[8]"},{"why":"Supplies the CloudUNet architecture and the window-based preprocessing framework that CAAC builds on and compares with.","marker":"[9]"},{"why":"Provides the LASSO large-eddy-simulation cloud fields used as the basis for the synthetic radiance data.","marker":"[15]"},{"why":"Provides the SHDOM 3D radiative transfer solver used to compute the two-wavelength radiance observations.","marker":"[16]"},{"why":"Introduces the convolutional block attention module that CAAC uses for channel attention.","marker":"[17]"},{"why":"Introduces adaptive instance normalization, which CAAC adapts to inject angle information into its feature maps.","marker":"[18]"}],"fun_headline_variants":["Angle-coded attention cuts cloud-thickness error ninefold","CAAC: angle-aware deep net slashes cloud retrieval error","Attention net with angle coding beats IPA by 9x in COT","Angle-invariant attention model cuts COT error ninefold"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reported error reductions are not inflated by the unspecified train/test split of the overlapping 64-by-64 image windows, since the paper does not say whether entire cloud fields were held out from training.","fun_headline_variants_meta":{"raw":{"variants":["Angle-coded attention cuts cloud-thickness error ninefold","CAAC: angle-aware deep net slashes cloud retrieval error","Attention net with angle coding beats IPA by 9x in COT","Angle-invariant attention model cuts COT error ninefold"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000542,"raw_usage":{"total_tokens":2618,"prompt_tokens":990,"completion_tokens":1628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":606,"completion_tokens_details":{"reasoning_tokens":1557}},"tokens_in":606,"tokens_out":1628,"duration_ms":13635,"temperature":1.0,"reasoning_tokens":1557,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:16:32.830314+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CAAC and all baselines on 64-by-64 windows from some LES cloud fields, hold out complete cloud fields for testing, and recompute the MSE ratios. If the ninefold advantage over single-angle methods shrinks substantially, or if CAAC no longer beats multi-angle trained UNet, the central claim is contradicted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Independent Pixel Approximation baseline whose biases the paper aims to overcome."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Early multi-pixel deep learning retrieval that motivates learning COT from neighborhoods rather than single pixels."},{"cited_title":"Influence of cloud retrieval errors due to three dimensional radiative effects on calculations of broadband cloud radiative effect,","cited_arxiv_id":null,"evidence_quote":"A convolutional segmentation-style COT retrieval baseline included in the model comparisons."},{"cited_title":"Determination of the optical thickness and effective particle radius of clouds from reflected solar radiation measurements. part i: Theory,","cited_arxiv_id":null,"evidence_quote":"Supplies the CloudUNet architecture and the window-based preprocessing framework that CAAC builds on and compares with."},{"cited_title":"Cloudunet: Adapt- ing unet for retrieving cloud properties,","cited_arxiv_id":null,"evidence_quote":"Provides the LASSO large-eddy-simulation cloud fields used as the basis for the synthetic radiance data."},{"cited_title":"Transfer-learning-based approach to retrieve the cloud proper- ties using diverse remote sensing datasets,","cited_arxiv_id":null,"evidence_quote":"Provides the SHDOM 3D radiative transfer solver used to compute the two-wavelength radiance observations."},{"cited_title":"Cloud identification and properties retrieval of the fengyun-4a satellite using a resunet model,","cited_arxiv_id":null,"evidence_quote":"Introduces the convolutional block attention module that CAAC uses for channel attention."},{"cited_title":"Retrieval of cloud properties from thermal infrared radiometry using convolutional neural network,","cited_arxiv_id":null,"evidence_quote":"Introduces adaptive instance normalization, which CAAC adapts to inject angle information into its feature maps."}],"review_version":1}