{"id":"8cae2dc9-bf9e-4f3a-a9e7-001b53cfe06d","arxiv_id":"2507.18099","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"On Cartosat-3 imagery of Hyderabad, a Cross-Pseudo-Supervision model with dynamic weighting achieves the best MIoU (42.14) while plain CPS attains the best recall (79.59); the analysis documents tree loss and building growth in the IT corridor.","lead":"This paper compares deep learning models for mapping land use from Indian satellite images, after correcting for atmospheric haze. It reports which model scores best and uses the best results to show how Hyderabad changed between 2020 and 2023.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth noise and a single un-swept threshold make the headline rankings (CPS best Recall, Dist-Diff best MIoU) unverified; the paper itself concedes frequent incorrect labels in Section 3.","rationale":"Good-faith reading: the paper contributes an end-to-end LUT atmospheric correction and segmentation pipeline with a real case study; the 60,848 training patches across multiple cities and the explicit equations are genuine evidence. The central claim, however, is an empirical ranking of models, and for that ranking to be true the evaluation labels must be accurate enough that Recall and MIoU differences (e.g., Recall 79.59 vs 52.52; MIoU 42.14 vs 34.96) reflect model quality rather than vector completeness. The paper itself undermines this precondition by conceding sparse and incorrectly annotated ground truth. This is an internal correctness risk, not a disagreement with consensus. The proposed test, a threshold sweep combined with ground-truth perturbation, would settle whether the ordering is an artifact of label noise or a stable property. If the ordering is stable, the concern is resolved; if not, the conditional verdict should be downgraded. I agree with the reader's weakest assumption and therefore recommend no change to the CONDITIONAL verdict.","tokens_in":10825,"tokens_out":3925,"duration_ms":45334,"concrete_test":"Rerun the Table 1 evaluation over a threshold sweep (0.2, 0.3, 0.4, 0.5, 0.6, 0.7) for every model, and for at least one threshold recompute scores after eroding or dilating the ground-truth masks by 1-3 pixels to simulate vector misalignment. If CPS (highest Recall) and CPS-Dist-Diff (highest MIoU) are not the argmax at the original 0.4 threshold and under mask perturbation, the headline ranking is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central comparison is not robust because both metrics are computed against an evaluation mask that the paper admits is unreliable. Section 2.1.2 says OSM/Microsoft/Google vectors 'can suffer from sparse labeling, misclassification, and outdated information,' and Section 3 states 'there are many instances where the ground truth labels are incorrectly annotated.' On such a mask, Recall only divides by labeled positives, so unlabeled true objects are invisible to the numerator, while MIoU counts them as false positives in the union. Rankings therefore depend on where the vectors happen to be complete. The water row illustrates the risk: CPS has water Recall 86.96 but water MIoU 10.70, while CPS-Dist-Diff has water MIoU 49.65 and Recall 73.03; such a dramatic inversion is exactly what sparse or misaligned labels produce. In addition, the 0.4 threshold in Eq. 15 was chosen after the fact, and no sensitivity analysis is reported. Because threshold moves the Recall/MIoU trade-off, the abstract's pairing of 'CPS highest Recall' with 'Dist-Diff highest MIoU' could simply be the point on the threshold curve where these two methods cross, rather than a stable property of the models. Coupled with single-run scores on one 24.6 km2 patch and no released code or data, this makes the central claim conditional rather than established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an atmospheric-correction pipeline based on 6S Look-Up Tables applied to Cartosat MX multispectral imagery, followed by supervised (DeepLabV3+) and semi-supervised (Cross-Pseudo Supervision, CPS, and variants with distribution/difficulty-aware dynamic weighting) semantic segmentation for Land Use/Land Cover mapping. The central empirical claim is that, on a single 24.6 km² evaluation patch over Hyderabad, CPS achieves the highest mean Recall (79.59) while CPS with Dist-Diff dynamic weighting achieves the best MIoU (42.14), and that both outperform the supervised DeepLabV3+ baseline (Recall 52.52, MIoU 34.96). The paper also presents a Hyderabad case study over 2020–2023 quantifying urban expansion and vegetation loss. The contribution is framed as an application-oriented comparison of methods on sparse, vector-derived labels rather than a new algorithmic proposal.","tokens_in":11158,"tokens_out":2217,"duration_ms":25137,"significance":"If the reported rankings were robust, the paper would provide a useful practical reference for LULC practitioners using high-resolution Indian satellite data with sparse open-source labels. The work explicitly addresses an important gap: adapting semi-supervised segmentation methods to geographically heterogeneous, sparsely labeled remote-sensing imagery. Strengths include the use of a physically based LUT atmospheric correction, a clearly described data-preparation pipeline, and an end-to-end workflow that is reproducible in principle from the textual description. However, the significance is currently conditional because the evaluation rests on a single patch, a post-hoc threshold choice, and a ground-truth mask that the authors themselves describe as frequently mislabeled. The comparative claim about model ordering is therefore not yet established to the standard expected for a journal publication.","major_comments":[{"comment":"The headline rankings are based on a single evaluation patch (24.6 km²) with no error bars, no multiple runs, and no statistical significance testing. Given that training is stochastic, the 2–3 point differences between CPS (Recall 79.59, MIoU 33.95), CPS with Dist-Diff (Recall 76.89, MIoU 42.14), and the next-best methods could easily be within run-to-run variance. The paper should report means and variances over at least several seeds, and ideally evaluate over additional patches or cities.","section":"§3, Table 1, Eq. (15)"},{"comment":"The binary threshold is set to 0.4 instead of the standard 0.5, and the paper itself (Figure 4) shows that this choice changes the prediction masks. No sensitivity analysis is provided, so the central pairing of 'CPS highest Recall' with 'CPS-Dist-Diff highest MIoU' may simply reflect the point on the precision/recall trade-off curve where the two methods cross. The authors should report metrics across a range of thresholds (e.g., 0.3–0.7) or justify 0.4 independently of the test set.","section":"§3, Eq. (15)"},{"comment":"The evaluation ground truth is derived from OpenStreetMap, Microsoft, and Google vector data plus an NDVI threshold chosen by manual inspection. The paper explicitly concedes that these sources 'can suffer from sparse labeling, misclassification, and outdated information' (Section 2.1.2) and that 'there are many instances where the ground truth labels are incorrectly annotated' (Section 3). On such a mask, Recall only counts true positives among labeled positives, while MIoU penalizes unlabeled true objects as false positives. The dramatic class-level inversion for water (CPS Recall 86.96 but MIoU 10.70; CPS-Dist-Diff Recall 73.03 but MIoU 49.65) is exactly the signature of label incompleteness or misalignment. The paper needs a quantitative assessment of label noise, e.g., manual verification on a stratified sample or comparison against a higher-quality reference, before method rankings can be trusted.","section":"§2.1.2, §3"},{"comment":"Several hand-set hyperparameters and filtering choices are load-bearing for the comparison but are neither swept nor justified: the 65% 'Other' patch-filtering threshold, the 3-pixel road buffer, the NDVI vegetation threshold, the ramp-up weight 0.1 in Eq. (7), and the Dist-Diff exponents/momentum. The abstract's claim that dynamic weighting 'enhanc[es] pseudo-label reliability during training' is not supported by any direct measurement of pseudo-label quality; in fact, Table 1 shows that adding Dist-Diff lowers Recall (76.89 vs. 79.59). The authors should either provide a sensitivity analysis showing the rankings are stable, or temper the claim about reliability improvement.","section":"§2.2, §2.4.3, §3"}],"minor_comments":[{"comment":"The abstract states that CPS 'is further refined with dynamic weighting, enhancing pseudo-label reliability during training,' but this is not directly measured; please rephrase to describe what is actually observed (higher MIoU, lower recall).","section":"Abstract and throughout"},{"comment":"The sentence 'These sources can suffer from sparse labeling, misclassification, and outdated information due to ongoing construction and deconstruction' is an important limitation but is buried in the introduction; it should be echoed in the results/discussion where the evaluation is interpreted.","section":"§1, last paragraph"},{"comment":"Typographical errors: 'atmoshperic' should be 'atmospheric'; 'Xarrary' should be 'Xarray'; the parenthetical 'Zarr format [7] instead of Xarrary [6]' is confusingly phrased.","section":"§2.1.1"},{"comment":"The sentence 'patches with more than 50% missing values on the multi-spectral image, or more than 65% pixels belonging to the Other class were discarded' should specify whether the 50% threshold applies to the raster data values or to NaN/cloud masks.","section":"§2.2"},{"comment":"Equation (3) contains an unexplained '5X' notation; the summation limits and the indices (i, w) are not defined completely. Also, 'lhd' and 'lce' are not formally introduced before use.","section":"§2.4.2, Eqs. (2)–(5)"},{"comment":"In the 'Distribution Aware (No Post Processing)' row, the Trees MIoU entry is '58,07' with a comma instead of a decimal point. Also, the table caption does not state that the threshold is 0.4.","section":"Table 1"},{"comment":"The case study reports area changes over 2020–2023, but the changes are computed from model predictions on two images; the text should state whether these predictions were validated against any ground truth or visually inspected, and should acknowledge that the observed changes may contain false positives from the segmentation errors.","section":"§3.1 and Table 2"},{"comment":"Reference [8] and [20] are self-citations to arXiv preprints; please provide DOIs or published venue information if available. Also, the in-text citation for the DHC framework [31] appears only in the introduction, not in the experimental section where the method is implemented.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful practical engineering contribution, but the evaluation protocol is not yet at the level required for a journal. The central comparison lacks statistical robustness and the ground-truth quality is acknowledged to be poor. If the authors add multi-run results, threshold sensitivity, and a quantitative label-noise assessment, the paper could become publishable. I would not recommend rejection because the issues are fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is an honest, incremental application paper that is transparent about its noisy ground truth, but the headline ranking (CPS beats DeepLabV3+, Dist-Diff beats CPS on MIoU) is not established by the evidence as presented. The evaluation is a single 24.6 km2 patch, no error bars, and a decision threshold (0.4) chosen after the fact without a sensitivity analysis.\n\nWhat is actually new: adapting Cross-Pseudo Supervision with DHC-style dynamic weighting to sparsely labeled Cartosat MX imagery over Indian cities, plus a LUT-based atmospheric correction pipeline and a Hyderabad case study. That combination is not in the cited literature. The paper does a decent job describing the data preparation, and it explicitly acknowledges that OSM/Microsoft/Google vectors are sparse, misclassified, and outdated, and that many ground truth labels are wrong. That is more honest than most remote sensing papers.\n\nWhere the soft spots are: the central comparison is underdetermined. Section 2.1.2 and Section 3 both admit the evaluation mask is unreliable. On a mask where only some objects are labeled, Recall is inflated when true positives happen to fall on labeled pixels, and MIoU counts unlabeled objects as false positives in the union. The water row in Table 1 is exactly what you would expect from this: CPS water Recall 86.96 with MIoU 10.70, while Dist-Diff water Recall 73.03 with MIoU 49.65 — a dramatic inversion that likely reflects where the vectors are complete, not model quality. The choice of 0.4 as the binarization threshold is post-hoc, and the abstract's pairing of CPS best Recall with Dist-Diff best MIoU could be a threshold artifact. Also there are no error bars, no multiple runs, and no code or data release. The claim in the abstract that dynamic weighting 'enhanc[es] pseudo-label reliability' sits awkwardly with the lower Recall numbers, and the paper does not discuss that tension.\n\nBottom line: this is a useful case study for practitioners working with sparse OSM labels on Indian cities, but as a comparative evaluation it needs error bars, a threshold sweep, and ideally a second evaluation site before the rankings can be trusted. I would send it to a competent referee, but I would expect heavy revision. It is not something I would cite for the comparative numbers.","headline":"Honest incremental application paper whose headline model rankings are not yet supported by the single-patch, noisy-label, post-hoc-threshold evaluation.","tokens_in":11643,"tokens_out":2058,"would_cite":false,"duration_ms":20756,"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":"Atmospheric correction plus Cross-Pseudo Supervision beats DeeplabV3+ in land use mapping from Cartosat MX imagery.","keywords":["Land Use Land Cover","semantic segmentation","atmospheric correction","Cross-Pseudo Supervision","remote sensing","Cartosat MX","semi-supervised learning","Hyderabad"],"falsifier":"A concrete test would be to train the same DeeplabV3+ and CPS models on the identical Hyderabad patches but evaluate against a densely, manually annotated ground truth for the same area, and then compare the Recall and MIoU rankings. If the rankings change, the reported advantage of CPS over DeeplabV3+ would be shown to be partly an artifact of label noise. Likewise, removing the atmospheric-correction step in an ablation would directly test the paper's claim that AC is the key enabler of the recall improvements.","tokens_in":10626,"feed_emoji":"🛰️","tokens_out":1772,"duration_ms":17730,"temperature":0.7,"pith_summary":"This paper claims that applying LUT-based atmospheric correction to Cartosat MX satellite images, followed by a supervised or semi-supervised segmentation model, yields accurate Land Use Land Cover (LULC) maps. In a patch over Hyderabad, the Cross-Pseudo Supervision (CPS) model achieved the highest average Recall (79.59), while adding a distribution-difficulty dynamic weighting (Dist-Diff) pushed MIoU to 42.14, both outperforming the DeeplabV3+ baseline (Recall 52.52, MIoU 34.96). The work also demonstrates a practical use case: comparing 2020 and 2023 images to quantify urban expansion, tree loss, and road growth in Hyderabad's IT corridor. A sympathetic reader would care because the method addresses a common real-world constraint—training on sparsely labeled, noisy open-source vector data—rather than on clean curated datasets.","feed_headline":"Semi-supervised models beat DeeplabV3+ in LULC mapping from Cartosat MX","feed_subtitle":"Atmospheric-corrected Cartosat MX patches show CPS reaching 79.59 Recall and 42.14 MIoU, with a case study quantifying Hyderabad's urban…","key_machinery":"The core mechanism is a Look-Up Table (LUT)-based atmospheric correction that converts raw Digital Number values to Bottom-of-Atmosphere reflectance using coefficients from the 6S radiative transfer model, making the multispectral input 'analysis-ready'. On top of this, the paper applies Cross-Pseudo Supervision, where two separately initialized DeeplabV3+ models with EfficientNet backbones train on the same sparsely labeled patches, each using the other's pseudo-labels as supervision, with a sigmoid ramp-up on the pseudo-loss weight. For the dynamic weighting variant, it borrows the DHC 'Dist-Diff' strategy: distribution-aware weights (based on class pixel ratios) and difficulty-aware weights (based on per-class Dice score trends) are combined and applied to both supervised and pseudo-supervised losses. These components together are intended to handle class imbalance and noisy sparse labels, which are the main obstacles in real-world LULC mapping from open data.","core_discovery":"The central discovery is that a semi-supervised training scheme, adaptive weighting, and a specific atmospheric-correction preprocessing step together outperform a standard supervised segmentation baseline on a sparse-label, high-resolution remote sensing task. The study reports that CPS with dynamic weighting (Dist-Diff) achieves the best Mean Intersection-over-Union (42.14 average), while vanilla CPS achieves the best Recall (79.59 average), and both beat DeeplabV3+ (34.96 MIoU and 52.52 Recall) on the same Hyderabad patch. The authors attribute the gains to the LUT-based correction converting Top-of-Atmosphere to Bottom-of-Atmosphere reflectance, and to class-imbalance handling via dynamic class weights. They also present a change-detection case study showing, in their evaluation area, a 31.47% increase in road area, a 10.30% increase in building area, and a 30.79% decrease in tree area between December 2020 and March 2023.","pith_inferences":["An implication the authors do not fully draw is that the noisy OSM/Microsoft/Google-derived ground truth likely depresses the MIoU of all models, and the ranking could shift if a curated, densely annotated evaluation set were used for the same Hyderabad patch.","The paper's choice of threshold 0.4 (instead of the standard 0.5) to binarize softmax outputs is a testable lever: a sensitivity analysis at several thresholds would show whether the reported margins are robust, especially for roads and water where per-class scores fluctuate widely.","The CPS improvement over DeeplabV3+ is claimed to be amplified by the atmospheric correction, but the paper does not include an ablation without AC on the same models; that ablation would isolate how much of the gain comes from preprocessing rather than from the semi-supervised architecture.","Extending the same pipeline to other Indian cities (e.g., merging training data from Bangalore, Mumbai, Pune, Varanasi, and Delhi) could test whether the model generalizes across the diverse terrains the introduction mentions, since only Hyderabad is used for evaluation."],"forward_implications":["If the reported rankings hold, practitioners on sparse-label remote sensing tasks with similar data should prefer CPS or CPS+Dist-Diff over a plain DeeplabV3+ baseline, since both net Recall and MIoU gains.","The paper's atmospheric-correction step should be treated as a prerequisite for training on Cartosat MX imagery, since uncorrected TOA reflectance is claimed to be hazy and less reliable for surface characterization.","The dynamic weighting (Dist-Diff) improves MIoU at the cost of some Recall, so the choice between CPS variants should depend on whether a user cares more about average completeness or about IoU on minority classes like water.","The change-detection case study suggests that even with imperfect open-source ground truth, the pipeline can produce plausible LULC change statistics that could inform urban planning discussions.","The reported computational advantage of the Generic SSL framework (shorter training time and less space) makes it a practical fallback when compute is limited, even if its scores are lower than CPS."],"supporting_citations":[{"why":"Supplies the 6S radiative transfer model that underlies the LUT-based atmospheric correction coefficients.","marker":"[30]"},{"why":"Introduces the Cross-Pseudo Supervision method, which the paper adapts to sparsely labeled remote sensing data.","marker":"[4]"},{"why":"Provides the DHC distribution-difficulty weighting strategies (DistDW and DiffDW) that the paper integrates into CPS.","marker":"[31]"},{"why":"Defines the generic SSL framework with encoder and three decoders (Dist, Diff, Ens) that the paper re-implements.","marker":"[32]"},{"why":"OpenStreetMap vector data are the primary source for training ground-truth masks, and the paper's known limitation is their sparsity and mislabeling.","marker":"[22]"},{"why":"A comparison baseline: the paper contrasts its results with this earlier building-segmentation work on very high-resolution imagery.","marker":"[23]"},{"why":"The Segmentation Models PyTorch library provides the DeeplabV3+ architecture with EfficientNet backbone used as the baseline and in CPS.","marker":"[11]"}],"fun_headline_variants":["CPS semi-supervised model tops DeeplabV3+ in LULC segmentation","Adaptive class weights boost CPS for land cover mapping","Dynamic weighting lifts CPS MIoU over DeeplabV3+ baseline","Hyderabad LULC: CPS with Dist-Diff wins segmentation race","Best MIoU from CPS with adaptive weighting in remote sensing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation depends on ground truth masks that are built from open-source vector data (OpenStreetMap, Microsoft, Google) and an NDVI threshold chosen by manual inspection, and the paper itself admits these sources are often sparse, mislabeled, or outdated.","fun_headline_variants_meta":{"raw":{"variants":["CPS semi-supervised model tops DeeplabV3+ in LULC segmentation","Adaptive class weights boost CPS for land cover mapping","Dynamic weighting lifts CPS MIoU over DeeplabV3+ baseline","Hyderabad LULC: CPS with Dist-Diff wins segmentation race","Best MIoU from CPS with adaptive weighting in remote sensing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000799,"raw_usage":{"total_tokens":3511,"prompt_tokens":937,"completion_tokens":2574,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":2479}},"tokens_in":553,"tokens_out":2574,"duration_ms":16898,"temperature":1.0,"reasoning_tokens":2479,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:37:58.139721+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test would be to train the same DeeplabV3+ and CPS models on the identical Hyderabad patches but evaluate against a densely, manually annotated ground truth for the same area, and then compare the Recall and MIoU rankings. If the rankings change, the reported advantage of CPS over DeeplabV3+ would be shown to be partly an artifact of label noise. Likewise, removing the atmospheric-correction step in an ablation would directly test the paper's claim that AC is the key enabler of the recall improvements.","supporting_citations":[{"cited_title":"Vermote, D","cited_arxiv_id":null,"evidence_quote":"Supplies the 6S radiative transfer model that underlies the LUT-based atmospheric correction coefficients."},{"cited_title":"Semi-supervised semantic segmentation with cross pseudo supervision","cited_arxiv_id":null,"evidence_quote":"Introduces the Cross-Pseudo Supervision method, which the paper adapts to sparsely labeled remote sensing data."},{"cited_title":"Dhc: Dual- debiased heterogeneous co-training framework for class-imbalanced semi-supervised medical image seg- mentation","cited_arxiv_id":null,"evidence_quote":"Provides the DHC distribution-difficulty weighting strategies (DistDW and DiffDW) that the paper integrates into CPS."},{"cited_title":"Towards generic semi-supervised framework for volumetric medical image segmentation","cited_arxiv_id":null,"evidence_quote":"Defines the generic SSL framework with encoder and three decoders (Dist, Diff, Ens) that the paper re-implements."},{"cited_title":"Planet dump re- trieved from https://planet.osm.org","cited_arxiv_id":null,"evidence_quote":"OpenStreetMap vector data are the primary source for training ground-truth masks, and the paper's known limitation is their sparsity and mislabeling."},{"cited_title":"Fast building segmentation from satellite imagery and few local labels","cited_arxiv_id":null,"evidence_quote":"A comparison baseline: the paper contrasts its results with this earlier building-segmentation work on very high-resolution imagery."},{"cited_title":"Segmentation models py- torch","cited_arxiv_id":null,"evidence_quote":"The Segmentation Models PyTorch library provides the DeeplabV3+ architecture with EfficientNet backbone used as the baseline and in CPS."}],"review_version":1}