{"id":"a9d8ee3b-6b08-453e-8aea-8d28abd6fdcb","arxiv_id":"2606.07374","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A Residual U-Net learns to predict InSAR coherence from detected SAR backscatter magnitudes, trained on Sentinel-1 12-day pairs and generalizing to unseen temporal baselines and diverse locations with better accuracy than intensity-based methods.","lead":"The paper presents a deep learning framework using a Residual U-Net to regress InSAR coherence directly from detected SAR backscatter images, trained on coregistered Sentinel-1 SLC pairs. Smart generalists might read it because it could simplify large-scale coherence mapping for change detection using publicly available analysis-ready data without complex coregistration.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Training on coregistered magnitude pairs does not guarantee the mapping works on misaligned detected-image inputs","rationale":"The concern matches the reader's weakest assumption exactly and is the single point that must hold for the headline claim (coherence regression from detected images without coregistration) to be true. No stronger internal inconsistency is visible from the given material.","tokens_in":1677,"tokens_out":318,"duration_ms":18804,"concrete_test":"Take a held-out coregistered SLC test pair, apply a controlled 5-pixel range and 5-pixel azimuth shift to one magnitude image, run the trained model, and compare the output coherence map against the original ground-truth coherence; if mean absolute error rises by more than 0.15 or correlation falls below 0.65, the model relies on alignment and the no-coregistration claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Residual U-Net is trained on backscatter magnitude pairs taken from precisely coregistered 12-day SLC data, with supervision from the corresponding InSAR coherence. The central claim requires that the same network, when presented with a pair of detected SAR images that have not been coregistered, still produces a usable coherence map. Because the training inputs are spatially registered, any learned features that exploit pixel-wise correspondence between the two magnitude channels will be invalid once the inputs are geometrically offset. The abstract provides no indication that misalignment was deliberately introduced during training or that robustness to typical GRD/SLC geometric differences was measured.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes a Residual U-Net for regressing InSAR coherence directly from pairs of detected SAR backscatter magnitude images. The network is trained on coherence targets derived from precisely coregistered 12-day Sentinel-1 SLC pairs and is evaluated on multiple datasets spanning coregistered SLC products and analysis-ready data (including GRD) with varying radiometric properties, geometries, and locations. The central claims are that the approach achieves high-resolution coherence estimation without requiring accurate coregistration at inference time, outperforms existing intensity-based methods, and generalizes across unseen geographical locations and temporal baselines.","tokens_in":1818,"tokens_out":536,"duration_ms":13389,"significance":"If the generalization to misaligned detected-image inputs holds, the method would enable coherence-based applications at scale using globally distributed analysis-ready SAR products (e.g., via Google Earth Engine) without SLC processing or coregistration steps, which would be a practical advance for change monitoring and mapping tasks.","major_comments":[{"comment":"The training inputs are magnitude pairs taken from precisely coregistered SLC data, yet the central claim requires the same weights to produce usable coherence on detected images that have not been coregistered. No description is given of deliberate misalignment augmentation during training or of controlled experiments that measure performance degradation as a function of geometric offset; this assumption is load-bearing for the claim of operating “without the need for accurate coregistration.”","section":"Methods / Training procedure"},{"comment":"Evaluation on “open access analysis-ready data” is reported to demonstrate generalization, but the manuscript does not state whether these test pairs were coregistered, what the typical residual misalignment is, or provide separate metrics for aligned versus unaligned inputs. Without such quantification, the transfer from the training distribution to the claimed inference distribution cannot be verified.","section":"Experiments / Evaluation datasets"},{"comment":"The abstract asserts “improved accuracy compared to existing intensity-based approaches,” yet the provided summary contains no numerical error metrics, baseline tables, or train/validation split details. If these are absent from the full results section as well, the quantitative support for the accuracy claim is insufficient.","section":"Abstract / Results"}],"minor_comments":[{"comment":"The abstract states that the network “generalizes well … across different temporal baselines that were never seen at training time,” but does not name the specific temporal baselines used in training versus testing.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive and detailed review. We address each major comment point by point below, indicating where revisions will be made to strengthen the manuscript.","responses":[{"response":"We agree that the manuscript does not describe misalignment augmentation or controlled offset experiments. Training used coregistered SLC pairs to obtain reliable coherence targets, with the model intended to generalize to detected products at inference. To support the claim, the revised manuscript will include a new subsection with controlled misalignment experiments reporting accuracy degradation versus offset.","revision_made":"yes","referee_comment":"[Methods / Training procedure] The training inputs are magnitude pairs taken from precisely coregistered SLC data, yet the central claim requires the same weights to produce usable coherence on detected images that have not been coregistered. No description is given of deliberate misalignment augmentation during training or of controlled experiments that measure performance degradation as a function of geometric offset; this assumption is load-bearing for the claim of operating “without the need for accurate coregistration.”"},{"response":"The analysis-ready datasets (e.g., GRD) were used as distributed without additional coregistration. We will add explicit statements on their processing status and residual misalignment characteristics in the revised manuscript. Separate aligned/unaligned metrics will be provided where the data permits; otherwise we will note the limitation.","revision_made":"partial","referee_comment":"[Experiments / Evaluation datasets] Evaluation on “open access analysis-ready data” is reported to demonstrate generalization, but the manuscript does not state whether these test pairs were coregistered, what the typical residual misalignment is, or provide separate metrics for aligned versus unaligned inputs. Without such quantification, the transfer from the training distribution to the claimed inference distribution cannot be verified."},{"response":"The full results section contains quantitative error metrics, baseline comparisons, and train/validation details supporting the abstract claim. We will add explicit cross-references from the abstract to the relevant tables and figures in the revised version to make this support clearer.","revision_made":"no","referee_comment":"[Abstract / Results] The abstract asserts “improved accuracy compared to existing intensity-based approaches,” yet the provided summary contains no numerical error metrics, baseline tables, or train/validation split details. If these are absent from the full results section as well, the quantitative support for the accuracy claim is insufficient."}],"tokens_in":1413,"tokens_out":517,"duration_ms":34701,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's main move is training a Residual U-Net to output InSAR coherence directly from pairs of detected SAR backscatter images. Targets come from standard coherence computed on coregistered 12-day Sentinel-1 SLC pairs. They then run the model on both coregistered SLC and on open analysis-ready detected products that cover different locations, radiometry, and geometries.\n\nWhat is actually new is the direct regression setup that skips the usual coregistration step at inference time. If it works, it would let people pull coherence maps from the large volumes of GRD data already distributed through places like Google Earth Engine. The claim that performance holds for temporal baselines never seen in training is the part worth watching.\n\nThe practical framing is reasonable. Many applications in change monitoring would benefit from coherence without full SLC processing pipelines.\n\nThe soft spot is the alignment issue. Training inputs are precisely registered, so any features that depend on exact pixel correspondence between the two magnitude channels will break once the inputs are offset, as they typically are in detected products. The abstract says they evaluate on analysis-ready data but supplies no numbers, no error maps, no description of how geometric differences were managed, and no baseline comparisons. That leaves the central generalization claim unverified from what is shown.\n\nThis is for SAR remote-sensing groups that need scalable coherence products and are willing to test a DL shortcut. A reader who already works with intensity-based coherence estimators would see the most direct value.\n\nIt deserves peer review so the methods and quantitative results can be examined in detail.","headline":"The U-Net regresses coherence from detected SAR magnitudes after training on aligned SLC pairs, but the transfer to misaligned inputs is the part that needs checking.","tokens_in":2306,"tokens_out":392,"would_cite":false,"duration_ms":11277,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A Residual U-Net regresses InSAR coherence directly from detected SAR backscatter images without coregistration.","keywords":["InSAR coherence","deep learning","Residual U-Net","Sentinel-1","detected SAR images","backscatter","coherence regression","SAR interferometry"],"falsifier":"The trained model producing lower accuracy than conventional intensity-based coherence estimators when tested on a new collection of detected SAR images from an unseen sensor or geographic region.","tokens_in":2576,"feed_emoji":"📡","tokens_out":615,"duration_ms":15492,"temperature":0.7,"pith_summary":"The paper develops a deep learning approach that estimates coherence maps using only detected SAR images instead of requiring precisely coregistered single-look complex pairs. A Residual U-Net is trained on coherence derived from Sentinel-1 12-day SLC data to learn the relationship between backscatter magnitudes and coherence values. This yields higher accuracy than prior intensity-based methods and generalizes to new geographic sites and to temporal baselines absent from training. The framework also runs on open analysis-ready detected products, such as ground-range detected data, supporting broader operational use.","feed_headline":"Neural net predicts InSAR coherence from detected SAR images","feed_subtitle":"Trained on coregistered pairs, the model works on analysis-ready data and generalizes to new locations and time gaps.","key_machinery":"Residual U-Net that maps pairs of backscatter magnitude images to coherence values.","core_discovery":"The authors show that a Residual U-Net trained on coherence maps from coregistered Sentinel-1 SLC pairs can predict coherence from backscatter magnitudes alone in detected SAR images. The model delivers high-resolution estimates with improved accuracy over existing intensity-based techniques. It maintains performance across diverse locations and on temporal baselines different from the 12-day training pairs, and functions on globally distributed analysis-ready detected data.","pith_inferences":["Operational pipelines could shift away from SLC data requirements toward simpler detected-image workflows.","The approach might combine with multi-temporal stacks to refine coherence estimates in time-series applications.","Transfer tests on data from non-Sentinel-1 sensors would clarify how far the learned magnitude-to-coherence relation extends."],"forward_implications":["Coherence estimation becomes possible on widely distributed analysis-ready detected products without coregistration processing.","Accuracy exceeds that of existing backscatter-only methods for high-resolution outputs.","The same model applies to temporal baselines not present in the training set.","Large-scale mapping and change-monitoring tasks can use standard open data sources such as Google Earth Engine products."],"fun_headline_variants":["U-Net regresses InSAR coherence from detected SAR backscatter","Coherence predicted from SAR intensity without coregistration","Residual network learns coherence from analysis-ready SAR data","Deep model estimates InSAR coherence across new baselines","Coherence from detected Sentinel-1 images via U-Net"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The statistical mapping from backscatter magnitudes to coherence learned on coregistered 12-day Sentinel-1 SLC pairs transfers to detected SAR images that lack coregistration and may have different radiometric or geometric properties.","fun_headline_variants_meta":{"raw":{"variants":["U-Net regresses InSAR coherence from detected SAR backscatter","Coherence predicted from SAR intensity without coregistration","Residual network learns coherence from analysis-ready SAR data","Deep model estimates InSAR coherence across new baselines","Coherence from detected Sentinel-1 images via U-Net"]},"model":"grok-4.3","cost_usd":0.002861,"raw_usage":{"total_tokens":1560,"prompt_tokens":616,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":28612000,"prompt_tokens_details":{"text_tokens":616,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":869,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":616,"tokens_out":75,"duration_ms":6407,"temperature":1.0,"reasoning_tokens":869,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T21:02:47.811920+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"The trained model producing lower accuracy than conventional intensity-based coherence estimators when tested on a new collection of detected SAR images from an unseen sensor or geographic region.","supporting_citations":[],"review_version":1}