{"id":"98e3aa9a-c7b5-47be-86f8-f6f5095a9b2b","arxiv_id":"2504.19598","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"CANet trains a shared change detection backbone on one dataset, then adapts to new datasets by updating only lightweight adapters with per-dataset batch normalization and an attention mask.","lead":"This paper proposes CANet, a change detection network that shares most of its weights across datasets and learns a small per-dataset adapter, updating only 4 to 8 percent of parameters when moving to a new dataset. It tests this approach on four public remote sensing change detection benchmarks and reports competitive F1 scores with much lower training cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-dataset BN is shown necessary but not sufficient; no test with a large sensor/spectral shift supports the claim that a 4.1%–7.7% adapter alone generalizes CD across datasets.","rationale":"I read the paper as an engineering contribution: a parameter-efficient adapter for multi-dataset change detection, with a shared encoder-decoder trained on CDD and per-dataset lightweight adapters. The central claim is credible within the tested regime: the adapter updates few parameters and the reported F1 gaps to full retraining are modest (1.1–3.4 F1 on the three target datasets for the MobileNetV2/ResNet18 variants). The strongest part of the evidence is the consistent ablation story: removing ICM or unique BN degrades performance, and the parameter sweep in Table 5 shows a clear trade-off. The weakest point is the mechanism for distribution alignment. Unique BN layers are shown to help (Table 4), but the paper never probes whether BN alone can handle a large domain shift, which is exactly what 'more generalized' promises. The three target datasets are all RGB VHR aerial/satellite images, so the empirical support for the central claim is narrower than the title/abstract imply. I agree with the reader's weakest_assumption; a cross-sensor experiment would settle whether the concern lands. A secondary reproducibility concern (adapter depth η selected using test-set numbers in Table 5, no code) reinforces the conditional verdict but is not the primary logical gap.","tokens_in":12452,"tokens_out":12047,"duration_ms":130721,"concrete_test":"Take the CDD-trained CANet[M] with η=5 and adapt it to a held-out dataset with a substantially different distribution, e.g., OSCD (Sentinel-2, 10 m, multispectral) or a SAR-optical change detection benchmark, following exactly the Sec. 3.4 protocol: freeze the shared module, optimize only the adapter and per-dataset BN layers, report F1. Compare this against (a) the same architecture fully fine-tuned from the CDD checkpoint and (b) CANet-O trained from scratch on that dataset. If the adapter's F1 gap to the full fine-tune is more than ~5 points (the observed gap on SYSU/LEVIR/WHU is 1.1–3.4), then per-dataset BN alone cannot bridge large sensor/spectral shifts and the generalization claim must be scoped down.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CANet's central claim (Eq. 5, Sec. 3.3)—that freezing the CDD-trained shared module and updating only the adapter (4.1%–7.7% of parameters) generalizes to new CD datasets—rests on the assumption that per-dataset batch-normalization layers (Sec. 3.4) are sufficient to absorb all data-distribution differences between datasets. The evidence for this is the W/O BN ablation (Table 4), which shows that removing per-dataset BN costs up to 6.85 F1 on WHU. But that only demonstrates BN is necessary, not sufficient. Every evaluation dataset (SYSU, LEVIR, WHU) is very-high-resolution RGB optical imagery at 0.2–0.5 m resolution with the same two-date image-pair format; the shifts are sensor/scene style and labeling granularity, not fundamental spectral or modality changes. No experiment tests a dataset with different spectral bands, much coarser resolution, or SAR/optical modality. In such shifts, BN affine parameters and running statistics cannot repair features that the fixed shared convolution weights were never aligned to see, so the adapter would have no mechanism to recover performance. The title and abstract claim 'more generalized' CD, but the experiments only support generalization within a narrow family of optical VHR datasets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CANet, a 'change adapter network' for generalized remote sensing change detection. A shared encoder-decoder is trained on a historical dataset (CDD); for each new dataset, only a lightweight adapter is optimized, consisting of the last η feature-fusion blocks and a convolutional block, augmented by an 'interesting change region mask' (ICM) and per-dataset batch-normalization layers. Experiments on CDD, SYSU, LEVIR, and WHU claim that updating 4.1%-7.7% of the parameters yields F1 scores close to fully retrained networks (CANet-O) and better than most existing dataset-specific baselines. Additional experiments address training with limited labels and online fine-tuning.","tokens_in":12782,"tokens_out":6684,"duration_ms":65102,"significance":"If the central claim holds, the work would offer a practical way to share a change-detection model across datasets with substantially reduced training cost and parameter storage. The paper has real strengths: evaluation on four public datasets, four different CNN backbones, ablations that isolate the contributions of ICM and per-dataset BN, experiments with 10% training labels, and a comparison against online fine-tuning. However, the 'more generalized' claim is only demonstrated within a narrow family of very-high-resolution optical RGB change-detection datasets, and several methodological issues—selecting the adapter depth on the test sets, not reporting CANet's performance on the historical dataset, and omitting variance or significance information—weaken the quantitative conclusions. The core idea is plausible and worth publishing after the claims are scoped to the demonstrated setting and the experimental protocol is tightened.","major_comments":[{"comment":"The paper claims 'more generalized' and 'universal' CD, but all four evaluated datasets (CDD, SYSU, LEVIR, WHU) are very-high-resolution optical RGB image pairs at 0.2-0.5 m (CDD 0.03-1 m). No experiment involves a dataset with different spectral bands, SAR/optical modality, or substantially coarser resolution. The W/O BN ablation (Table 4) shows that per-dataset BN is necessary within this family, but it does not demonstrate that BN statistics plus a lightweight adapter can absorb large distribution shifts. Please either add experiments on a more diverse dataset (e.g., multi-spectral or SAR change detection) or explicitly scope the claims to adaptation across VHR optical CD datasets.","section":"§4.1, Table 1; §3.5"},{"comment":"The adapter depth η is selected based on F1 on the SYSU, LEVIR, and WHU test sets, and the same test sets are later reported as the main results in Table 3. This is a form of test-set fitting and may make the reported F1 values optimistic. Use a held-out validation split for model selection and report test results for the selected η, or report all η values with uncertainty so readers can judge the sensitivity.","section":"§4.5, Table 5"},{"comment":"The paper states that CANet performs well on the historical and new datasets simultaneously, but Table 3 reports results for SYSU, LEVIR, and WHU only; no CDD F1 is given for the adapter-based CANet. The online-training comparison in Fig. 4 is qualitative. Please add CDD results for CANet (with the adapter) and provide quantitative fine-tuning baselines to substantiate the 'simultaneously' claim.","section":"§3.5, §4.4, Table 3"},{"comment":"All reported metrics are single-run numbers without error bars or significance tests. Several comparisons are close: on LEVIR, CANet[M] F1 is 89.83 versus 90.70 for BAN and 90.91 for ChangeClip; on WHU, CANet[M] is 91.17 versus 92.17 and 92.25. The statement that CANet 'achieves better CD results on various datasets simultaneously' is not supported against every baseline, and the 'average improvement of 3.84%' compares the best CANet-O (a fully trained model) to the best other network rather than the adapter-based CANet. Please provide multiple runs with standard deviations or paired significance tests, and clarify the comparison protocol.","section":"§4.2, Tables 2-3"},{"comment":"The training-efficiency claim is not benchmarked against other parameter-efficient adaptation methods (e.g., fine-tuning only the last blocks or LoRA-style adapters). Fig. 4 shows fine-tuning results but without quantitative values. A quantitative comparison to fine-tuning the shared network on new datasets with a comparable parameter budget is needed to support the claim that the adapter is a uniquely efficient mechanism for generalization.","section":"§4.4, §3.5"}],"minor_comments":[{"comment":"The dimensions of the mask m in Eqs. (6)-(7) are not stated. If m is a two-channel mask, please explain how the final binary change prediction is obtained from the masked output and how the cross-entropy loss in Eq. (2) is applied.","section":"§3.4, Eq. (7)"},{"comment":"In the contributions list, 'caused by excellent differences in the data distribution and labeling' should read 'caused by significant differences.'","section":"§1"},{"comment":"The text 'the best F1 result of CAMNet-O is 97.95%' should read 'CANet-O.'","section":"§4.2"},{"comment":"The citation '[12, 12, 44]' in the sentence about multi-temporal interaction contains a duplicate reference [12]; please fix.","section":"§2.2"},{"comment":"The limited-training-sample results are presented only as a figure; include numeric values in a table to improve reproducibility and to support the claim that CANet degrades less than other methods.","section":"§4.3, Fig. 3"},{"comment":"The online-training comparison should include a table with F1 values for CDD and each new dataset, since the figure alone does not allow quantitative assessment.","section":"§4.4, Fig. 4"},{"comment":"Please clarify the evaluation protocol for the historical dataset: after training the shared module on CDD, is an adapter also optimized for CDD, and are those results reported anywhere?","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper would be a stronger fit if the claims are scoped to the demonstrated setting; the title and abstract currently overpromise relative to the experimental coverage. I would also flag the test-set-based selection of η (Table 5) as a protocol issue that should be fixed before acceptance, and I suggest the editor verify whether the journal requires error bars or repeated runs for empirical claims of this type."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is worth refereeing. It proposes CANet, a shared backbone plus per-dataset lightweight adapters for remote sensing change detection, with two concrete new ingredients: an interesting change region mask (ICM) and per-dataset batch normalization layers. The core empirical claim holds up well: updating only 4.1–7.7% of the parameters gets within about 1–1.8 F1 of fully retrained models and beats most single-dataset baselines across four public benchmarks. The ablations are honest and useful—they show ICM and per-dataset BN each contribute, and the limited-label experiments (10% training data) are a nice practical touch.\n\nThe soft spots are real but not fatal. Most importantly, the title and abstract say “more generalized,” but every test dataset is very-high-resolution RGB optical imagery at 0.2–0.5 m with two-date image pairs. Per-dataset BN is shown to be necessary (the W/O BN ablation costs up to 6.85 F1 on WHU), but that does not prove it is sufficient for a serious domain shift like different spectral bands, much coarser resolution, or SAR/optical fusion. Without an experiment like that, “more generalized” overclaims. Second, there are no error bars or significance tests; the ~1 F1 gaps between CANet and CANet-O could easily be within run-to-run noise, especially since no seeds are reported. Third, the adapter depth η is tuned on the same test sets reported as main results (Table 5), which is mild but real selection-on-test. Fourth, the paper never cites the NLP adapter literature or per-dataset batch normalization (conditional normalization), so the novelty framing undersells the prior art. Fifth, no code is released, which makes reproducibility harder.\n\nThat said, the central mechanism—a shared feature extractor plus per-dataset adapters with an attention-style mask—is clearly described and empirically supported within the tested domain. The math is straightforward and there is no circular derivation. The paper is a legitimate engineering contribution, not a breakthrough.\n\nFor peer review: yes, send it out. The reviewers should push for a cross-domain experiment (e.g., Sentinel-2 multispectral or an SAR dataset), error bars or at least multiple seeds, and clearer positioning against adapter and conditional normalization literature. If the authors temper the generalization claim or add such an experiment, it would be a solid publication.","headline":"A useful parameter-efficient adapter for multi-dataset change detection, but the 'more generalized' claim outruns the evidence.","tokens_in":13229,"tokens_out":1697,"would_cite":true,"duration_ms":20021,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CANet performs change detection on four datasets simultaneously by updating only 4.1–7.7% of network parameters, keeping F1 scores close to fully retrained models.","keywords":["change detection","remote sensing","adapter learning","parameter-efficient fine-tuning","batch normalization","change region mask","generalization","multi-dataset learning"],"falsifier":"Train CANet on CDD, freeze the shared module, and optimize only the adapter for a dataset with a large sensor or spectral shift, such as multispectral versus RGB imagery or optical versus SAR, then compare F1 with a fully fine-tuned network; if the gap exceeds the roughly 2 to 6.85 percentage-point gaps reported here, the assumption that per-dataset batch normalization absorbs distribution shift fails.","tokens_in":12298,"feed_emoji":"🛰️","tokens_out":5548,"duration_ms":49414,"temperature":0.7,"pith_summary":"The paper tries to solve a practical problem: deep change-detection networks trained on one remote sensing dataset perform poorly on others, so each dataset normally needs its own full network. It proposes CANet, in which most parameters are shared across datasets and only a lightweight adapter is trained for each new dataset. The paper reports that updating 4.1%–7.7% of the parameters is enough to reach F1 scores close to full retraining on three target datasets, and that the same network keeps working on the original dataset. If true, this makes multi-dataset change detection cheaper in storage, compute, and labels.","feed_headline":"One change-detection network adapts to four datasets with tiny updates","feed_subtitle":"Adapter retrains 4.1–7.7 percent of parameters yet F1 stays close to full retraining.","key_machinery":"The central object is the change adapter network (CANet), split into a dataset-shared encoder-decoder and a lightweight per-dataset adapter. The paper expresses the split as $\\hat{Y}=M_a(M_s(X_1,X_2))$, where $M_s$ is shared and $M_a$ is the adapter. The adapter contains an interesting change region mask (ICM), an attention-style mask generated from the raw prediction through convolutions, pooling, concatenation, and a Squeeze-and-Excitation block, plus a unique batch-normalization layer per dataset. The mask steers predictions toward the objects a dataset labels as changed, while the per-dataset batch-normalization layers absorb appearance differences. Only the last few decoder fusion blocks and a convolutional block are updated for a new dataset, amounting to 4.1%–7.7% of the network parameters.","core_discovery":"CANet claims that a single shared change-detection network can serve multiple remote sensing datasets simultaneously if each dataset is given a small adapter containing an interesting change region mask and dataset-specific batch normalization. Trained first on one historical dataset (CDD), then adapted to SYSU, LEVIR, and WHU by updating only the adapter, CANet reaches F1 scores of 81.57%, 89.83%, and 91.17% with MobileNetV2, within 1–3 points of the same architecture fully retrained per dataset, while updating 4.1%–7.7% of parameters. The paper further claims this beats existing dataset-specific baselines on SYSU and WHU by 7.23% and 5.85% F1, and degrades much less under 10% training data.","pith_inferences":["Beyond the paper, the same adapter recipe could be tested across a larger sensor shift, such as optical to SAR or multispectral to hyperspectral, where per-dataset batch normalization may not absorb the distribution gap.","Beyond the paper, the ablation's sensitivity to batch normalization suggests a measurable predictor: compute the batch-statistics divergence between source and target datasets and correlate it with the adapter-only performance drop; larger divergence should predict larger F1 loss.","Beyond the paper, the ICM's role implies a targeted experiment: pair two datasets that share a sensor but differ in labeling granularity; if ICM specifically handles labeling differences, it should matter more there than on sensor-shifted pairs."],"forward_implications":["Training a new dataset adapter costs 4.1%–7.7% of the network parameters, so adding a new change-detection dataset becomes cheap in storage and compute.","Because the shared module stays fixed, CANet keeps its performance on the historical dataset while also detecting changes on new datasets, which fine-tuned models lose.","Under 10% training data, CANet's F1 degrades far less than dataset-specific networks, so the adapter approach lowers the label requirement for new change-detection tasks.","The adapter can be attached to different backbones such as MobileNetV2, ResNet18, and IBN-Net variants, so the mechanism is not tied to one architecture.","Per-dataset batch normalization and an unshared ICM mask each contribute to the gains; ablations show removing them drops F1 by up to 6.85%."],"supporting_citations":[{"why":"Supplies the CDD dataset, the historical dataset on which the shared module is first trained.","marker":"[23]"},{"why":"Supplies the SYSU dataset used to test generalization of the adapter to a new domain.","marker":"[36]"},{"why":"Supplies the LEVIR building-change dataset used to test generalization.","marker":"[4]"},{"why":"Supplies the WHU building dataset used to test generalization.","marker":"[20]"},{"why":"Provides batch normalization, the layer CANet makes dataset-specific to absorb appearance differences.","marker":"[19]"},{"why":"Provides the Squeeze-and-Excitation block used inside the interesting change region mask.","marker":"[17]"},{"why":"Provides the CBAM attention module used in the decoder feature fusion blocks.","marker":"[41]"},{"why":"Provides MobileNetV2, the backbone used for the main ablations and parameter counts.","marker":"[34]"},{"why":"Provides ResNet18, one of the backbones tested to show adapter flexibility.","marker":"[15]"},{"why":"Provides IBN-Net backbones tested to show adapter flexibility across architectures.","marker":"[30]"}],"fun_headline_variants":["Lightweight adapters let one CD network serve four datasets","Retrain only 4–8% of a CD network for new datasets","CANet: adapter-based change detection generalizes across datasets","Single CD model, dataset-specific adapters: 4% training cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that per-dataset batch-normalization statistics can absorb the entire appearance and distribution difference between datasets while all shared weights stay frozen.","fun_headline_variants_meta":{"raw":{"variants":["Lightweight adapters let one CD network serve four datasets","Retrain only 4–8% of a CD network for new datasets","CANet: adapter-based change detection generalizes across datasets","Single CD model, dataset-specific adapters: 4% training cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000332,"raw_usage":{"total_tokens":1865,"prompt_tokens":980,"completion_tokens":885,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":811}},"tokens_in":596,"tokens_out":885,"duration_ms":9274,"temperature":1.0,"reasoning_tokens":811,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:47:55.198638+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CANet on CDD, freeze the shared module, and optimize only the adapter for a dataset with a large sensor or spectral shift, such as multispectral versus RGB imagery or optical versus SAR, then compare F1 with a fully fine-tuned network; if the gap exceeds the roughly 2 to 6.85 percentage-point gaps reported here, the assumption that per-dataset batch normalization absorbs distribution shift fails.","supporting_citations":[{"cited_title":"A spatial-temporal attention- based method and a new dataset for remote sensing image change detection","cited_arxiv_id":null,"evidence_quote":"Supplies the LEVIR building-change dataset used to test generalization."},{"cited_title":"Batch normalization: Accelerating deep network training by reducing internal co- variate shift","cited_arxiv_id":null,"evidence_quote":"Provides batch normalization, the layer CANet makes dataset-specific to absorb appearance differences."},{"cited_title":"Mobilenetv2: Inverted residuals and linear bottlenecks, 2019","cited_arxiv_id":null,"evidence_quote":"Provides MobileNetV2, the backbone used for the main ablations and parameter counts."},{"cited_title":"Deep residual learning for image recognition, 2015","cited_arxiv_id":null,"evidence_quote":"Provides ResNet18, one of the backbones tested to show adapter flexibility."},{"cited_title":"Two at once: Enhancing learning and generalization capacities via ibn-net, 2020","cited_arxiv_id":null,"evidence_quote":"Provides IBN-Net backbones tested to show adapter flexibility across architectures."},{"cited_title":"Change detection in remote sensing images using conditional adversarial networks","cited_arxiv_id":null,"evidence_quote":"Supplies the CDD dataset, the historical dataset on which the shared module is first trained."},{"cited_title":"A deeply supervised attention metric-based network and an open aerial image dataset for remote sensing change detection","cited_arxiv_id":null,"evidence_quote":"Supplies the SYSU dataset used to test generalization of the adapter to a new domain."}],"review_version":1}