{"id":"36d3612e-20a7-40fc-9cbd-6f0b5470db7d","arxiv_id":"2506.17944","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"SegChange-R1 combines a Swin encoder with a Phi-1.5 text encoder and a linear-attention BEV module, reporting gains on three change-detection benchmarks and a new drone-view dataset.","lead":"This paper proposes SegChange-R1, a change detection model that adds language-model text embeddings and a linear-attention 'BEV' module to remote sensing image pairs. It also introduces a drone-view building change detection dataset and reports state-of-the-art results on three of four benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 contradicts the central SOTA claim: on WHU-CD, SegChange-R1's F1 (0.968) is below ChangeCLIP's 0.982, so the paper's own results do not support 'better performance in all benchmark tests' (Section 4) or 'state-of-the-art in F1' (Section 6).","rationale":"The reader's strongest claim already identifies the WHU-CD contradiction, and I agree that this invalidates the paper's universal SOTA claim. I do not make the BEV-naming issue the primary attack because even a misnamed linear-attention module would not refute the empirical results; the empirical contradiction does. The same table shows SegChange-R1 wins DSIFN-CD, CDD, and DVCD, so the work may have incremental value, but the advertised central claim is unsupported. Given the absence of released code or data and the inconsistent DVCD size statements, the contradiction cannot currently be resolved by re-analysis; a direct reproduction is needed. My verdict is unchanged from the reader's REJECT: the central claim is contradicted by the paper's own evidence, and the remaining contribution is not sufficiently supported to accept.","tokens_in":10269,"tokens_out":5286,"duration_ms":52792,"concrete_test":"Recompute the WHU-CD row using the released SegChange-R1 checkpoint and the stated test protocol (threshold 0.5, batch size 1), and compare F1/IoU/OA against the ChangeCLIP numbers in Table 2, ideally by also running ChangeCLIP under identical evaluation. If SegChange-R1's F1 remains 0.968 < 0.982, the 'all benchmark tests' and 'state-of-the-art in F1' statements must be revised to exclude WHU-CD and to specify metrics where the claim holds. If the reproduced numbers differ materially from Table 2, the table itself needs correction before any comparison is meaningful.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that SegChange-R1 'demonstrated better performance in all benchmark tests' (Section 4, Results) and 'achieves state-of-the-art performance in F1 score, IoU, and overall accuracy' (Section 6). Table 2 contradicts this. On WHU-CD, SegChange-R1 obtains F1 = 0.968, IoU = 0.926, OA = 99.60, while ChangeCLIP obtains F1 = 0.982, IoU = 0.915, OA = 99.52. The text itself concedes ChangeCLIP 'achieves good results on WHU-CD (F1: 0.982).' Since F1 is the headline metric and the claimed advantage is about all benchmarks, the unqualified SOTA statement is false on the paper's own data. This is load-bearing: the method's contribution is defined by outperforming prior work; a three-of-four-datasets, metric-dependent win is a different, much weaker claim. A secondary concern is that the BEV module (Eqs. 1-2) is a learned linear projection plus linear attention with no camera model, depth estimate, or coordinate transform, so calling it a Bird's Eye View transformation is unjustified. However, the immediate blocker to acceptance is the empirical contradiction. The abstract promises code and models, but no code or data are released in the manuscript, so the table cannot be independently checked. The DVCD dataset size is also stated as 13,800 pairs in the abstract and conclusion, 12,833 pairs in Section 4, and 13,833 pairs by summing the Table 1 splits, so the dataset claims are not self-consistent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SegChange-R1, a remote sensing change detection method that augments a vision model with an LLM text encoder (Microsoft/Phi-1.5) and a so-called BEV Space Converter built from a linear projection and linear attention, and introduces DVCD, a new drone-view building change detection dataset. The authors report experiments on WHU-CD, DSIFN-CD, CDD, and DVCD, plus ablations on backbone, prompt configuration, and the BEV module, and claim state-of-the-art F1, IoU, and overall accuracy on all benchmarks.","tokens_in":10661,"tokens_out":6362,"duration_ms":62712,"significance":"If established, the direction of LLM-guided text conditioning for change detection and a new drone-view benchmark would be useful contributions to the remote sensing community. The paper does provide a structured comparison against many baselines and attempts to ablate its design choices. However, the manuscript's central empirical claim is contradicted by its own Table 2: on WHU-CD, ChangeCLIP's F1 of 0.982 exceeds SegChange-R1's 0.968, and on CDD, ChangeCLIP's IoU of 0.975 exceeds SegChange-R1's 0.969. In addition, the BEV module as defined by Eqs. (1)-(2) has no geometric content, the dataset size is stated inconsistently (13,800, 12,833, and 13,833), and the LLM-text integration is not described at a level that permits reproduction. These are load-bearing issues rather than presentation matters, so the paper does not currently justify its headline claims.","major_comments":[{"comment":"The claim that SegChange-R1 \"demonstrated better performance in all benchmark tests\" and \"achieves state-of-the-art performance in F1 score, IoU, and overall accuracy\" is directly contradicted by the numbers in Table 2. On WHU-CD, ChangeCLIP achieves F1=0.982 while SegChange-R1 achieves F1=0.968; on CDD, ChangeCLIP achieves IoU=0.975 while SegChange-R1 achieves IoU=0.969. Since F1 is the headline metric, the unqualified claim of universal superiority is false on the paper's own data. The authors should either correct the claim to a narrower, metric-specific statement or provide evidence that the reported comparisons are not directly comparable.","section":"Section 4 Results and Section 6 Conclusion vs. Table 2"},{"comment":"The BEV Space Converter is defined by a learned linear projection z_t = W_z x_t + b_z and a linear attention score A_ij = w_a^T ReLU(W_a1 z_i + W_a2 z_j). These equations contain no geometric operation: there is no camera model, depth estimate, homography, or coordinate grid that would map features into a bird's-eye view. As written, the module is a generic linear-attention feature transformation, and the claims about \"unifying features into a BEV space\" and correcting \"perspective-induced geometric inconsistencies\" are unsupported. The ablation in Table 5 compares \"No BEV\", \"Transformer\", and \"Linear Attention\", which only varies the attention mechanism and does not test a geometric BEV transformation; either provide a concrete geometric formulation or rename the module and revise the corresponding claims.","section":"Section 3.2, Eqs. (1)-(2)"},{"comment":"The DVCD dataset size is inconsistent. The abstract and the introduction state 13,800 image pairs; Section 4 states that the dataset comprises 12,833 pairs; and Table 1's train/validation/test split (11,066 + 1,383 + 1,384) sums to 13,833. Since DVCD is presented as a new contribution, the correct total is load-bearing. In addition, the meaning of the \"Instruct\" column in Table 1 is not defined, and the abstract's promise that \"code and pre-trained models are available\" is not backed by any code, configuration files, or checkpoints in the manuscript, which prevents independent verification of the experiments.","section":"Section 4 and Table 1 vs. Abstract and Section 1"},{"comment":"The integration of the LLM text encoder is underspecified. The text states that Microsoft/Phi-1.5 is used as a text encoder, but it does not describe how token-level embeddings are pooled, how the text embeddings are projected into the visual feature space, how the \"dynamic sequence length control\" works, or what training objective and loss are used for the fusion. Table 4 reports that \"LLM-Enhanced\" improves F1 over \"Dynamic Descriptions\" by 1.63 points, but without specifying the exact prompt templates, the text encoder's architecture and weights, or the training protocol, the improvement cannot be attributed to LLM reasoning rather than to a different text encoder or increased model capacity.","section":"Section 3.1 and Table 4"}],"minor_comments":[{"comment":"The abstract ends with an incomplete sentence: \"Experiments on four widely-used datasets demonstrate significant improvements over existing method\" lacks a period and should say \"methods\".","section":"Abstract"},{"comment":"The third contribution bullet introduces the dataset as \"DVSC\" while the rest of the paper consistently calls it \"DVCD\"; this must be unified.","section":"Section 1"},{"comment":"The captions of Tables 4 and 5 contain unresolved placeholder strings, \"[Task/Model]\" and \"[Task/Metric]\", indicating an incomplete compilation; these should be replaced with the actual task and metric names.","section":"Tables 4 and 5"},{"comment":"Several reference entries are malformed, for example \"[2] et al. Bai, T.\" and \"[8] et al. Chen, P.\"; the author lists and formatting need to be corrected.","section":"References"},{"comment":"Table 2 has many missing values and appears to have misaligned columns (for instance, the FCCDN row and the SNUNet row do not align with the dataset blocks), which makes the comparison difficult to read; the table should be reformatted.","section":"Table 2"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early draft rather than a finished submission. The internal contradictions between Section 4's claims and Table 2 are not cosmetics: they invalidate the headline result. The BEV module, as defined, is not a geometric transformation, so validating that contribution would require a new method, not a local fix. I recommend rejection rather than major revision because the core claims are not established by the manuscript's own evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nShort version: the paper is not ready for review in its current form. The text-guided change detection recipe is a sensible combination of known parts, and the drone-view dataset is the most defensible contribution, but the central SOTA claim fails on the paper's own numbers and the dataset is not available to check.\n\nWhat is genuinely new: the DVCD dataset concept, and the specific integration of Phi-1.5 as a text encoder feeding a language-guided mask decoder. The ablations on text configuration and the BEV module are directionally consistent, which suggests the authors did run experiments. That is credit where it is due.\n\nThe soft spots are load-bearing. Table 2 shows SegChange-R1 at F1 0.968 on WHU-CD, below ChangeCLIP's 0.982, while the text claims better performance on all benchmark tests. A three-of-four-datasets, metric-dependent win is not state of the art in F1, and the conclusion's SOTA statement is false as written. The dataset size is stated as 13,800 in the abstract and conclusion, 12,833 in Section 4, and 13,833 if you add the Table 1 splits. That kind of inconsistency matters because the whole empirical contribution rests on that dataset. The BEV module is Eq. (1)-(2): a learned linear projection and linear attention, with no camera model, depth estimate, homography, or coordinate grid. Calling it BEV is not justified by the paper; it is a generic attention layer unless the authors show otherwise. And the DVCD data is not released, no code, no error bars, and the ablation tables still contain [Task/Model] placeholders. None of these are nitpicks.\n\nWho gets value: someone working on vision-language remote sensing might skim the architecture and dataset idea, but nobody should quote the numbers. As it stands, I would not send this to a serious referee; I would desk reject and invite resubmission after the authors correct the overclaim, release the dataset and code, fix the dataset statistics, and either provide real geometric grounding for the BEV name or drop it.","headline":"Plausible text-guided change-detection idea and a useful dataset concept, but the paper's own Table 2 sinks the SOTA claim and the rest is not verifiable.","tokens_in":11154,"tokens_out":5508,"would_cite":false,"duration_ms":57261,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"SegChange-R1 claims that adding LLM-generated text descriptions and a linear-attention BEV converter to a change-detection pipeline yields top accuracy on three of four building-change benchmarks and adds a drone-view dataset.","keywords":["change detection","remote sensing","large language model","text-guided segmentation","bird's eye view","linear attention","drone imagery","building change detection"],"falsifier":"Retrain the model on DVCD with the BEV Space Converter replaced by an identical linear-attention layer that makes no claim of spatial transformation; if F1 and IoU do not drop, the BEV claim is empty, and one can also inspect the learned matrix $W_z$ to see whether it encodes any position-dependent geometric mapping.","tokens_in":10058,"feed_emoji":"🛰️","tokens_out":6386,"duration_ms":62850,"temperature":0.7,"pith_summary":"SegChange-R1 is a remote-sensing change detection method that injects natural-language descriptions of what changed, generated by an LLM text encoder, into a visual segmentation pipeline so the model focuses on relevant change regions instead of all pixel differences. The paper's central claim is that this text guidance, combined with a linear-attention 'BEV' module that supposedly unifies features from two time phases into a shared bird's-eye-view space, improves detection accuracy and speeds convergence on four building-change benchmarks. It also introduces DVCD, a 13,800-pair drone-view dataset with textual instructions, filling a gap for semantic-level change detection from UAV perspectives. The reported tables show SegChange-R1 with the best F1, IoU, and overall accuracy on DSIFN-CD, CDD, and DVCD; on WHU-CD, ChangeCLIP has a higher F1 (0.982 versus 0.968) while SegChange-R1 has higher IoU and overall accuracy. If the architecture's contribution is real, it offers a reusable recipe for text-conditioned change detectors and a benchmark for drone-view building change.","feed_headline":"Text-aware detector tops three of four change benchmarks","feed_subtitle":"SegChange-R1 fuses LLM text with a linear-attention BEV module and adds a drone-view dataset.","key_machinery":"The central machinery is the 'BEV Space Converter,' defined by a learned linear projection per time phase, $z_t = W_z x_t + b_z$, and linear-attention scores $A_{ij} = w_a^\\top \\mathrm{ReLU}(W_{a1} z_i + W_{a2} z_j)$, softmax-normalized to aggregate features. The paper calls this a spatial transformation that unifies features from different times into a shared BEV space and thereby resolves modal misalignment. Around it sit the LLM text encoder (Microsoft/Phi-1.5) whose embeddings are projected into the visual space by the D-Projector, and a masked decoder with a learnable-query transformer and channel-attention mask head. The DVCD dataset, 13,800 drone-view building-change image pairs with text instructions, is the benchmark contribution that makes the text-conditioned setup testable.","core_discovery":"The discovery the paper is trying to establish is that semantic guidance from an LLM, fused through a text-conditioned mask decoder, changes the behavior of a remote-sensing change detector: the model concentrates on described change types, converges faster, and reaches higher F1, IoU, and overall accuracy than visual-only and static-prompt variants. The paper further claims that its linear-attention BEV converter aligns the two temporal feature maps in a shared spatial representation, improving edge and spatial consistency. On the reported tables this package beats prior methods on DSIFN-CD, CDD, and the new DVCD dataset; on WHU-CD, ChangeCLIP retains the higher F1 (0.982 versus 0.968), while SegChange-R1 reports the higher IoU and overall accuracy. A new contribution is DVCD, a UAV-captured building-change dataset of 13,800 image pairs with fine-grained text descriptions, designed to support semantic-level change detection.","pith_inferences":["If the BEV converter really is just a learned linear projection with attention, then the paper's central architectural novelty reduces to linear attention plus text fusion; the name 'BEV' does no geometric work, and the DVCD gains would come from the LLM guidance and decoder design.","The DVCD dataset, with its natural-language instructions, could support few-shot or zero-shot change detection by prompting an LLM to describe arbitrary change types, a direction the paper mentions but does not test.","The reported convergence in 64 epochs versus the default 128 suggests text conditioning acts as a regularizer; plotting training loss curves for the no-text and LLM variants would make this effect measurable.","The same architecture could be applied to non-building change types, such as vegetation, water, or infrastructure, by changing the prompt text, which would test whether the semantic guidance generalizes beyond the benchmark categories."],"forward_implications":["Adding text descriptions guides the model to focus on changed regions, improving F1, IoU, and overall accuracy over visual-only and static-prompt variants.","The linear-attention BEV converter reduces parameters and FLOPs compared with transformer attention while improving or matching accuracy, making the method cheaper to train.","The DVCD dataset provides a drone-view benchmark with fine-grained textual descriptions for building change detection.","LLM-enhanced text reasoning converges faster (64 epochs) than the default training schedule, implying lower compute cost.","Because SegChange-R1 reports best results on DSIFN-CD, CDD, and DVCD, the recipe appears to transfer across resolutions, seasons, and imaging platforms."],"supporting_citations":[{"why":"Provides the ChangeCLIP baseline that SegChange-R1 compares against and the instruction-guided visual-language learning approach the paper extends.","marker":"[51]"},{"why":"Supplies the transformer-based siamese change detection baseline (ChangeFormer) used in the comparison.","marker":"[3]"},{"why":"Supplies the transformer change detection baseline (BIT) used in the comparison.","marker":"[11]"},{"why":"Supplies the Mamba state-space-model baseline (ChangeMamba) used in the comparison.","marker":"[26]"},{"why":"Defines the Swin Transformer backbone adopted by SegChange-R1 for multi-scale feature extraction.","marker":"[34]"},{"why":"Establishes the BEV representation via depth-based lifting that motivates the paper's BEV Space Converter.","marker":"[37]"},{"why":"Provides the DSIFN-CD dataset on which SegChange-R1 reports its best F1 score.","marker":"[49]"},{"why":"Provides the CDD dataset on which SegChange-R1 reports its best F1, IoU, and overall accuracy.","marker":"[25]"},{"why":"Provides the WHU-CD dataset and early fully convolutional siamese baselines used in comparison.","marker":"[16]"}],"fun_headline_variants":["LLM text guidance boosts change detection on 3 of 4 sets","Text-augmented detector wins 3 benchmarks, adds UAV dataset","SegChange-R1: LLM focus speeds convergence, tops 3 datasets","BEV + LLM text sharpen change maps, beat baseline on 3 sets","New UAV change dataset plus LLM fusion lifts accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the learned linear projection and attention scores of Eqs. (1)-(2) really place multi-temporal features into a shared bird's-eye-view space, even though the equations contain no camera model, depth estimate, homography, or coordinate grid.","fun_headline_variants_meta":{"raw":{"variants":["LLM text guidance boosts change detection on 3 of 4 sets","Text-augmented detector wins 3 benchmarks, adds UAV dataset","SegChange-R1: LLM focus speeds convergence, tops 3 datasets","BEV + LLM text sharpen change maps, beat baseline on 3 sets","New UAV change dataset plus LLM fusion lifts accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000843,"raw_usage":{"total_tokens":3631,"prompt_tokens":866,"completion_tokens":2765,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":2671}},"tokens_in":482,"tokens_out":2765,"duration_ms":18312,"temperature":1.0,"reasoning_tokens":2671,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:23:35.036177+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the model on DVCD with the BEV Space Converter replaced by an identical linear-attention layer that makes no claim of spatial transformation; if F1 and IoU do not drop, the BEV claim is empty, and one can also inspect the learned matrix $W_z$ to see whether it encodes any position-dependent geometric mapping.","supporting_citations":[{"cited_title":"Zhang, X","cited_arxiv_id":null,"evidence_quote":"Provides the ChangeCLIP baseline that SegChange-R1 compares against and the instruction-guided visual-language learning approach the paper extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the transformer-based siamese change detection baseline (ChangeFormer) used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the transformer change detection baseline (BIT) used in the comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Swin Transformer backbone adopted by SegChange-R1 for multi-scale feature extraction."},{"cited_title":"Philion and S","cited_arxiv_id":null,"evidence_quote":"Establishes the BEV representation via depth-based lifting that motivates the paper's BEV Space Converter."},{"cited_title":"Zhang, C","cited_arxiv_id":null,"evidence_quote":"Provides the DSIFN-CD dataset on which SegChange-R1 reports its best F1 score."},{"cited_title":"Lebedev, M","cited_arxiv_id":null,"evidence_quote":"Provides the CDD dataset on which SegChange-R1 reports its best F1, IoU, and overall accuracy."},{"cited_title":"Daudt, R","cited_arxiv_id":null,"evidence_quote":"Provides the WHU-CD dataset and early fully convolutional siamese baselines used in comparison."}],"review_version":1}