{"id":"daed0b37-dd82-4532-a2c1-bd1e337c17c1","arxiv_id":"2507.08096","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An object-based deep learning model estimates building heights from single COSMO-SkyMed SAR images plus building footprints, reaching about 2.20 m MAE in Munich under leave-one-city-out evaluation.","lead":"This paper trains a neural network to estimate building heights from single radar satellite images, using building footprints as guides. In European cities it reaches errors around one building story, but accuracy drops sharply in Asian megacities with many skyscrapers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (2) is internally inconsistent with the stated ground-range geometry: in ground range the offset is L = h*tan(theta), so the correct inversion is h = L/tan(theta), not h = L/cos(theta).","rationale":"The reader's weakest assumption focuses on the need for building footprints at inference time, which is a legitimate limitation that the title understates. However, the more load-bearing issue is internal: the method's stated geometry and its published height equation contradict each other. If the equation is wrong, the reported quantitative results cannot be explained by the described pipeline, and the central claim of accurate height estimation from single SAR images plus footprints is unsupported as written. Because the paper provides no code and only limited data, a reader cannot resolve whether the implementation uses a corrected formula or a slant-range measurement. This is a correctness risk, not merely an issue of presentation, and it should be settled before the paper can be accepted. If the authors can demonstrate that Eq. (2) is a typo and the implementation uses the correct ground-range relation, then a conditional acceptance with a corrected derivation would be appropriate; as currently written, the paper's central argument is not internally consistent.","tokens_in":11314,"tokens_out":7254,"duration_ms":84965,"concrete_test":"Run a controlled ground-range SAR simulation: place a single vertical building of known height h in a synthetic ground-range scene, measure L_BBB - L_FBB along the range direction, and compare the measured offset with h*tan(theta) and h*cos(theta). If the measured offset matches h*tan(theta), then Eq. (2) is wrong and the paper must be revised to h = L/tan(theta) or must explicitly document that L is a slant-range quantity despite the claimed ground-range processing.","verdict_should_be":"REJECT","load_bearing_attack":"The central height-inversion step, Eq. (2), is not consistent with the processing chain described in the Introduction and Section 3.1. The authors state that the method 'explicitly operates in the ground range domain, avoiding the need to reproject data into the SAR image (slant range) plane,' and that FBB/BBB are computed in ground coordinates by rotating the footprint-aligned box. In a ground-range (geocoded) SAR image, the layover displacement between the building bounding box and the footprint bounding box is L = h*tan(theta), not L = h*cos(theta). The correct inversion is therefore h = L/tan(theta), not h = L/cos(theta). For Munich, with incidence angle theta = 20.04 degrees (Table 1), 1/cos(theta) is about 1.064 while 1/tan(theta) is about 2.75, so the two formulas differ by roughly a factor of 2.6 for the same measured offset. If the implementation truly follows the stated ground-range geometry, the reported MAE of 2.20 m in Table 3 could not be produced by Eq. (2); if Eq. (2) was actually used, then the method must be operating in slant range or an undocumented conversion or learned correction is present. Either way, the paper as written does not establish the central claim, because the stated relationship between the regressed bounding boxes and physical height is geometrically wrong.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes an object-based method for estimating building heights from a single very high resolution (VHR) COSMO-SkyMed SAR image. Given a building footprint polygon and the corresponding SAR patch, a ResNet-101 regressor augmented with footprint-derived geometric features predicts bounding-box information, from which the height is recovered through the incidence-angle relation in Eq. (2). The method is evaluated with leave-one-city-out experiments over eight cities on three continents, reporting test MAEs of about 2.2-2.3 m for several European and North American cities and 8-11 m for dense or high-rise Asian cities, together with an additional 70/30 in-distribution experiment. The paper claims that this constitutes state-of-the-art out-of-distribution performance for single-image SAR height estimation.","tokens_in":11607,"tokens_out":7665,"duration_ms":86538,"significance":"If the reported results hold, the paper would represent a useful advance: story-level accuracy from single-date VHR SAR on unseen cities of similar morphology is a nontrivial and practically relevant result, and the assembled multi-city dataset (about 53,000 patches and 3.3 million buildings) is a valuable benchmarking resource. The leave-one-city-out protocol is an appropriate and demanding test of generalization. However, the central geometric derivation needs to be corrected or clarified before the numerical results can be interpreted, and the comparisons against prior work are not direct. The paper does not provide code, and the SAR data are restricted by an ASI research agreement, so independent replication is not immediately possible.","major_comments":[{"comment":"Eq. (2), h = L/cos(theta), is inconsistent with the paper's explicit statement that the method operates in the ground-range domain. In a geocoded ground-range SAR image, the layover offset between the footprint and the building top is L = h * tan(theta), so the correct inversion would be h = L/tan(theta), not h = L/cos(theta). For the Munich image, with theta = 20.04 degrees (Table 1), the two formulas differ by about a factor of 2.6 for the same offset L. The paper must state unambiguously whether L is measured in slant range or ground range; if it is measured in ground range, Eq. (2) is wrong, and if it is measured in slant range, the claims in Section 3.1 about avoiding the slant-range plane are contradicted. Please provide a derivation and, if Eq. (2) is used as written, explain how the network can produce the reported accuracies despite the factor-of-2.6 discrepancy.","section":"Section 3, Eq. (2) and Section 3.1"},{"comment":"The description of what the network actually regresses is internally inconsistent. Section 2.4 states that the model regresses a single height value per building and uses an MSE loss on height, while Section 3.2 states that the output of the fully connected layer is used to estimate bounding boxes from which the height is then computed. This distinction is material: if the loss is on height, Eq. (2) is not part of the training objective and the bounding-box interpretation is only a post-hoc explanation; if the loss is on box offsets, the generation of ground-truth bounding-box labels from footprint and height must be specified exactly, including which geometric formula is used to create those labels. Please clarify the training loss and the label-generation procedure.","section":"Sections 2.4 and 3.2"},{"comment":"The comparisons in Table 5 do not support the abstract's claim of outperforming state-of-the-art methods in similar out-of-distribution scenarios. The cited methods use different sensors and data sources (Sentinel-1/2 time series, ICESat-2/GEDI LiDAR, GaoFen-7 multi-view photogrammetry) and different evaluation protocols. In particular, there is no direct comparison with Sun et al. [14], which is the direct predecessor and the baseline against which the two claimed innovations (ground-range processing and removal of center coordinates) are defined. Please add a direct comparison with [14] on the same COSMO-SkyMed data and the same leave-one-city-out protocol, or substantially soften the state-of-the-art claim.","section":"Table 5 and Related Work"},{"comment":"All reported results come from a single leave-one-city-out run with no error bars, no multiple seeds, and no significance tests. Small differences between cities, such as Munich 2.20 m, Los Angeles 2.24 m, and Milan 2.26 m, are likely within seed-to-seed or bootstrap variability. Please add repeated runs (at least a few seeds) or bootstrap confidence intervals for the headline MAE and RMSE values, especially for the claim of state-of-the-art accuracy.","section":"Table 3"}],"minor_comments":[{"comment":"The manuscript contains numerous typographical and grammatical errors that should be corrected, including 'ibjecxt' in Section 4, 'compatisons' in the text before Table 5, 'specificaly' in Section 2.3, 'rigourous' and 'othere' in Section 2.2, and 'amd' in the Shenzhen comparison paragraph.","section":"Throughout"},{"comment":"The caption says 'Scatter plots of absolute errors per city relative to reference heights,' but the surrounding text refers to 'orange regions' and 'yellow and red zones,' which suggests a spatial map or image. Please make the figure type and caption consistent.","section":"Figure 7"},{"comment":"The 'inclination angle' is defined as the angle between the orbital plane and the equator, but its role in rotating the footprint-aligned bounding box is not derived or explained. Please clarify how this angle relates to the SAR acquisition geometry (incidence angle and heading) and how the FBB rotation is computed.","section":"Section 3.1"},{"comment":"The Introduction states that center coordinates contribute marginally to height prediction, but no ablation or experiment is presented to support this claim. Please either add an ablation study or remove the causal claim and present the choice as a design decision.","section":"Introduction"},{"comment":"The h >= 40 m columns are based on very small samples in several cities, for example 80 buildings in Munich and 72 in Milan (Table 2), making the reported MAE and RMSE values for tall buildings unstable. Please report the number of buildings in each height category for each city and consider suppressing or clearly caveating these tall-building metrics.","section":"Table 3"},{"comment":"The manuscript header contains placeholder submission metadata ('Received:', 'Revised:', 'Accepted:', 'Published:', 'Citation: ... Remote Sens. 2025, 1, 0') that should be removed or completed before publication. The keyword list also repeats 'SAR'.","section":"Header"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of Remote Sensing and the empirical effort is substantial. The main load-bearing issue is the geometric inconsistency of Eq. (2) with the stated ground-range processing; this is fixable in principle by clarifying the coordinate system or correcting the formula and re-running the experiments, so I do not recommend rejection. I would also ask the editor to require a direct comparison with the direct predecessor [14] before considering the paper acceptable. I saw no evidence of misconduct; the issues are technical and presentational."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis paper reports an object-based building height estimator from single COSMO-SkyMed SAR images, trained and tested on an eight-city, three-continent dataset with leave-one-city-out evaluation. The dataset itself is a genuine asset, and the protocol is appropriate for testing out-of-distribution generalization. On European cities the reported MAE around 2.2 m is impressive, and the authors are honest enough to include Shenzhen, where a stronger competitor beats them. That is the good part.\n\nThe soft spot is serious. The method is stated to operate in the ground range domain, with bounding boxes computed in geocoded coordinates. The height is then recovered from the difference between building bounding box and footprint bounding box via h = L/cos(theta). But in a ground-range image, the layover displacement of a building is L = h*tan(theta), not h*cos(theta). The correct inversion is h = L/tan(theta). For Munich's incidence angle of 20 degrees, the two formulas differ by about a factor of 2.6. If the authors genuinely measured L from ground-range, geocoded imagery, Eq. (2) cannot yield the reported MAE; the network would have to be learning an implicit rescaling, which is not described. If they actually used Eq. (2), then the image must be in slant range, which contradicts the paper's explicit claim. The ground-truth bounding box generation is not specified in enough detail to see how the labels were created, so this is not just a typo—it is a load-bearing ambiguity in the central inversion step.\n\nOther issues are milder: no direct comparison with the immediate predecessor (Sun et al.) on this dataset, no error bars or significance tests, and no code or data release (though the ASI data restriction is understandable). The title also overstates things: the method needs building footprints at inference, so it is not from SAR alone.\n\nVerdict: the dataset and evaluation design deserve referee time, and this is a paper worth discussing. But the geometry issue has to be resolved before the headline numbers can be trusted. I would not cite it as-is.","headline":"A useful new multi-city dataset and an honest OOD evaluation, but the height-inversion equation is inconsistent with the claimed ground-range processing and needs resolution before the numbers can be trusted.","tokens_in":12137,"tokens_out":13156,"would_cite":false,"duration_ms":135337,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Building heights can be estimated from a single SAR image plus building footprints, with about one-storey error on European-style cities.","keywords":["building height estimation","single SAR image","COSMO-SkyMed","object-based regression","bounding box regression","out-of-distribution generalization","urban remote sensing","deep learning"],"falsifier":"Take a held-out city with LiDAR ground-truth heights and feed the network footprint polygons that are deliberately shifted by one ground-range pixel (2.5 m) while keeping the SAR image unchanged. The height formula $h = (L_{\\mathrm{BBB}}-L_{\\mathrm{FBB}})/\\cos\\theta$ predicts a systematic bias near $2.5/\\cos\\theta$ m, so a much smaller or inconsistent bias would mean the bounding-box geometry is not the mechanism driving the estimates.","tokens_in":11120,"feed_emoji":"🛰️","tokens_out":8540,"duration_ms":91070,"temperature":0.7,"pith_summary":"This paper claims that building heights can be estimated from a single very-high-resolution synthetic aperture radar (SAR) image, provided the building footprints are known. The proposed object-based network turns each footprint into an aligned bounding box and learns to predict the larger box that the building's height creates in radar geometry, with the length difference between the two boxes divided by the cosine of the incidence angle giving the height. Under leave-one-city-out tests on eight COSMO-SkyMed cities, the method reports mean absolute errors of about 2.20 m in Munich, 2.24 m in Los Angeles, and 2.26 m in Milan, roughly one building story. It also reports much larger errors, above 10 m, on Asian high-rise cities and on buildings taller than 40 m, so the claim is conditional on low- and mid-rise urban morphologies. If accepted, the result would make single-pass, single-sensor height mapping practical in cities that already have footprint databases.","feed_headline":"Building heights from one radar pass land within 2.2 m in Munich","feed_subtitle":"Why it matters: a city height map can come from one satellite pass, with no time series or optical data needed.","key_machinery":"The load-bearing identity is the SAR bounding-box relation $h = (L_{\\mathrm{BBB}} - L_{\\mathrm{FBB}})/\\cos\\theta$, where the FBB is the smallest rectangle enclosing the footprint, rotated to match the orbit inclination, and the BBB is the corresponding box that includes the height-induced displacement in the range direction. The network's job is to regress the BBB from the SAR image and footprint-derived features, using only the box dimensions as extra features rather than the center coordinates to reduce overfitting. Operating directly in ground-range coordinates rather than reprojecting into the radar slant-range plane, the method uses the acquisition incidence angle to convert the measured displacement into a height estimate.","core_discovery":"The central claim is that a footprint-guided object-based regressor can extract building heights from a single VHR SAR image with accuracy comparable to one storey for European-style urban fabrics. The network processes a SAR patch together with the binary footprint mask, computes a footprint-aligned bounding box (FBB) rotated according to the orbit inclination, and regresses a building bounding box (BBB) whose range-direction length exceeds the FBB length by $L$; the height is $h = L / \\cos\\theta$. Trained with a mean squared error loss on 160,000 patches from seven cities and tested one city at a time, it achieves a leave-one-city-out MAE of 2.20 m in Munich, 2.24 m in Los Angeles, and 2.26 m in Milan, while rising above 10 m for Shanghai and Shenzhen, and an overall in-distribution MAE of 4.95 m. The same experiments show systematic failure on buildings above 40 m and on cities dominated by tall, narrow-footprint structures, which the paper attributes to tall-building scarcity in the training data and to slant-range ambiguity. The paper's conclusion is therefore conditional: single-VHR-SAR height estimation generalizes across continents mainly for the low- and mid-rise building stock that dominates the training distribution.","pith_inferences":["Our inference: adding synthetic tall buildings or incidence-angle augmentation to the training set is a direct test of the paper's tall-building data-scarcity explanation; if Asian-city MAE drops sharply, that explanation is confirmed.","Our inference: running the same network on VHR images from other sensors with the height formula unchanged would show whether the learned features are geometric or radiometrically sensor-specific.","Our inference: in cities without footprint databases, the method would need an automated footprint extractor upstream, and footprint boundary errors would then enter the height estimate directly through the $L_{\\mathrm{FBB}}$ term."],"forward_implications":["If the central claim holds, a city with up-to-date footprints needs only one VHR SAR acquisition, not multi-date or multi-sensor stacks, to produce a building height map at roughly one-storey accuracy in low- to mid-rise areas.","The same trained model can be applied to an unseen European-style city without local training data, with expected MAE in the 2–3 m range for buildings under 40 m.","For cities with many buildings above 40 m or with dense high-rise cores, expected errors are much larger, above 10 m MAE in the reported tests, so height products from this method should be flagged as low-confidence there.","The 53,079 paired SAR and reference patches assembled for the eight cities become a reusable benchmark for future single-image SAR height estimation methods."],"supporting_citations":[{"why":"defines the object-based bounding-box regression geometry that this method adapts, replacing slant-range boxes with ground-range footprint-aligned boxes.","marker":"[14]"},{"why":"shows single-VHR-SAR height reconstruction is feasible and provides the single-image approach this method extends.","marker":"[13]"},{"why":"is the multimodal, multitemporal benchmark compared in Milan.","marker":"[9]"},{"why":"is the LiDAR-altimetry and footprint benchmark compared in Los Angeles and New York.","marker":"[11]"},{"why":"is the photogrammetry-plus-deep-learning benchmark compared in Shenzhen, where it reports lower error than this method.","marker":"[20]"},{"why":"supplies the reference building footprints and heights for the European cities.","marker":"[16]"},{"why":"supplies the Chinese building footprint and height reference data used for Shanghai and Shenzhen.","marker":"[17]"},{"why":"provides the deep residual convolutional architecture used as the feature extractor.","marker":"[19]"}],"fun_headline_variants":["One radar pass, one story accuracy in Munich","Radar-only building heights: 2.2 m error in Munich","Deep learning reads building heights from one radar image","Cross-city radar heights work in Europe, falter in Asia"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that an accurate footprint polygon is available for every building at prediction time; if footprints are missing, misaligned, or stale, the footprint-aligned box that the height formula is anchored to is wrong.","fun_headline_variants_meta":{"raw":{"variants":["One radar pass, one story accuracy in Munich","Radar-only building heights: 2.2 m error in Munich","Deep learning reads building heights from one radar image","Cross-city radar heights work in Europe, falter in Asia"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000703,"raw_usage":{"total_tokens":3209,"prompt_tokens":1017,"completion_tokens":2192,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":2125}},"tokens_in":633,"tokens_out":2192,"duration_ms":16204,"temperature":1.0,"reasoning_tokens":2125,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:27:18.770991+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a held-out city with LiDAR ground-truth heights and feed the network footprint polygons that are deliberately shifted by one ground-range pixel (2.5 m) while keeping the SAR image unchanged. The height formula $h = (L_{\\mathrm{BBB}}-L_{\\mathrm{FBB}})/\\cos\\theta$ predicts a systematic bias near $2.5/\\cos\\theta$ m, so a much smaller or inconsistent bias would mean the bounding-box geometry is not the mechanism driving the estimates.","supporting_citations":[{"cited_title":"Large-scale building height retrieval from single SAR imagery based on bounding box regression networks","cited_arxiv_id":null,"evidence_quote":"defines the object-based bounding-box regression geometry that this method adapts, replacing slant-range boxes with ground-range footprint-aligned boxes."},{"cited_title":"The SAR2Height framework for urban height map reconstruction from single SAR intensity images","cited_arxiv_id":null,"evidence_quote":"shows single-VHR-SAR height reconstruction is feasible and provides the single-image approach this method extends."},{"cited_title":"How high are we? Large-scale building height estimation at 10 m using Sentinel-1 SAR and Sentinel-2 MSI time series","cited_arxiv_id":null,"evidence_quote":"is the multimodal, multitemporal benchmark compared in Milan."},{"cited_title":"Automated Estimation of Building Heights with ICESat-2 and GEDI LiDAR Altimeter and Building Footprints: The Case of New York City and Los Angeles","cited_arxiv_id":null,"evidence_quote":"is the LiDAR-altimetry and footprint benchmark compared in Los Angeles and New York."},{"cited_title":"Leveraging Chinese GaoFen-7 imagery for high-resolution building height estimation in multiple cities","cited_arxiv_id":null,"evidence_quote":"is the photogrammetry-plus-deep-learning benchmark compared in Shenzhen, where it reports lower error than this method."},{"cited_title":"EUBUCCO v0.1: European building stock characteristics in a common and open database for 200+ million individual buildings","cited_arxiv_id":null,"evidence_quote":"supplies the reference building footprints and heights for the European cities."},{"cited_title":"Building Floorspace in China: A Dataset and Learning Pipeline","cited_arxiv_id":"2303.02230","evidence_quote":"supplies the Chinese building footprint and height reference data used for Shanghai and Shenzhen."},{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"provides the deep residual convolutional architecture used as the feature extractor."}],"review_version":1}