{"id":"9ca41c10-9bd4-4118-8f26-6d1a3057251c","arxiv_id":"2411.13127","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Cloud-Adapter, a lightweight cross-attention adapter on a frozen DINOv2 backbone, achieves state-of-the-art cloud segmentation on six remote sensing benchmarks with a small fraction of trainable parameters.","lead":"Cloud-Adapter attaches a small trainable module to a frozen vision foundation model, letting it segment clouds in satellite images with high accuracy and few new parameters. The method reports state-of-the-art scores on six cloud segmentation benchmarks from different satellites, land covers, and annotation styles.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.6% parameter-efficiency claim excludes the trainable Mask2Former head, so the central efficiency statement is unsupported as written.","rationale":"The reader's weakest assumption is on point. The paper's strongest claim bundles a SOTA performance assertion with a parameter-efficiency assertion; the latter is the more fragile part. The ablation in Table VI is internally inconsistent: a configuration that trains a Mask2Former head cannot have 0.00M trainable parameters. This is not a matter of consensus; it is a mismatch between what is measured and what is claimed. I also noted that Table XIII shows Cloud-Adapter does not beat DBNet on every L8 Biome scene, but the overall mIoU is best and this is a secondary nuance. The empirical SOTA results are plausible and the released code and checkpoints are a genuine asset; the fix is to disclose the head parameter count and reword the efficiency claim. That is exactly the kind of clarification the CONDITIONAL verdict should require, so I keep the reader's verdict rather than moving it.","tokens_in":22822,"tokens_out":4923,"duration_ms":47099,"concrete_test":"Download the released code and checkpoints, load the Cloud-Adapter model with the DINOv2 Large backbone, and run: total_trainable = sum(p.numel() for p in model.parameters() if p.requires_grad); also print the count for the no-adapter baseline of Table VI. Report the total, the Mask2Former-head-only count, and the ratio to DINOv2 Large (304M). If the total is substantially above 1.82M (e.g., >20M), the abstract and contributions must be revised to state the adapter-only count or to include the head in the efficiency claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Cloud-Adapter's headline contribution is parameter-efficient fine-tuning: the abstract and contributions emphasize that only 0.6% of the frozen backbone's parameters are trainable (1.82M for DINOv2 Large). However, Section IV-B states that the segmentation head is Mask2Former, which is trained, and Table VI reports 0.00M parameters for the no-adapter baseline even though that baseline still trains the Mask2Former head. Thus the 1.82M count covers only the spatial perception module and the adapting modules; the head's parameters are omitted. Mask2Former heads of the scale used in dense prediction typically contain tens of millions of parameters, so the true number of trainable parameters is likely one to two orders of magnitude larger than 1.82M, and the true ratio to the 304M DINOv2 Large backbone is far above 0.6%. The central claim as stated is therefore not supported by the paper's own accounting, and comparisons to 'parameter-efficient' labels in the literature are not yet meaningful. The underlying segmentation improvements may still be real; this is an accounting and claim-support problem, not an observed performance failure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Cloud-Adapter, a parameter-efficient fine-tuning approach for cloud segmentation in remote sensing images. A vision foundation model (DINOv2-Large or SAM variants) is frozen; a lightweight spatial perception module (stem, several ConvNet blocks, and a parameter-free aggregator) extracts multi-scale context, and cross-attention adapting modules with low-rank MLP projections modulate the frozen transformer layers. The segmentation head is Mask2Former. Experiments report state-of-the-art mIoU, mAcc, aAcc, and mDice on six subsets of four datasets (HRC WHU, GF12MS WHU GF1/GF2, CloudSEN12 High L1C/L2A, L8 Biome), with a claimed 0.6% trainable parameter ratio relative to the backbone, and the code and checkpoints are released.","tokens_in":23060,"tokens_out":5569,"duration_ms":49940,"significance":"The empirical scope is a strength: the method is evaluated across multiple satellite sources, spatial resolutions, processing levels, and annotation granularities, and the released code and checkpoints support reproducibility. If the reported gains are robust, Cloud-Adapter is a practically useful instantiation of adapter-based VFM fine-tuning for cloud segmentation. However, the headline parameter-efficiency claim is not supported by the paper's own accounting, and some evaluation claims are stronger than the tables justify. The architectural contribution is incremental relative to existing adapters such as ViT-Adapter and SAM-Adapter, but the systematic comparison and multi-dataset evidence are valuable.","major_comments":[{"comment":"The paper's central efficiency claim, expressed in the abstract and contributions as 'only 0.6% of the trainable parameters of the frozen backbone' (1.82M for DINOv2-Large), counts only the spatial perception and adapting modules and excludes the trainable Mask2Former head. Section IV-B.1 states that the segmentation head is Mask2Former, and Table VI reports 0.00M parameters for the no-adapter baseline even though that baseline still trains the head. Because a Mask2Former head for dense prediction contains a substantial number of parameters, the true total trainable parameter count is far larger than 1.82M and the true ratio to the backbone is far above 0.6%. Please either report total trainable parameters including the head, or explicitly restrict the parameter-efficiency claim to the adapter modules and adjust the abstract and conclusion accordingly.","section":"Section IV-B, Tables II and VI"},{"comment":"The L8 Biome dataset is cropped and split into 6:2:2 by the authors, but the paper does not specify how each baseline method was trained or evaluated on this custom split. Without a common training protocol (optimizer, iterations, patch size, preprocessing, validation selection) or a reference to the source of the baseline numbers, the comparisons in Tables XII and XIII are not verifiable and may not be controlled. Please document the protocol for all methods, or clearly state the provenance of each baseline result.","section":"Section IV-A.2, Tables XII and XIII"},{"comment":"The text claims Cloud-Adapter 'consistently outperforms competitors such as DBNet and UNetMobv2 across diverse land scenes,' but Table XIII shows DBNet with higher mIoU on Grass/Crops (59.08 vs 41.60), Urban (59.68 vs 51.76), and Water (52.42 vs 46.43), and higher aAcc on Grass/Crops, Urban, Barren, Forest, and Water. The claim should be limited to the overall L8 Biome metrics, and the per-scene results should be discussed honestly, since they show the method is not uniformly better across land cover types.","section":"Section IV.D.2.a, Table XIII"},{"comment":"No variance or repeated-run statistics are reported, and several SOTA margins are very small (e.g., GF1 mIoU 92.55 vs 92.42 and aAcc 98.92 vs 98.91 in Table VIII). The main 'consistently achieves SOTA' claim should either be accompanied by standard deviations across multiple seeds or by a statistical test, or be softened to 'competitive with' where margins are within expected run-to-run variation.","section":"Tables VII-XI"}],"minor_comments":[{"comment":"The phrase '0.6% of the trainable parameters of the frozen backbone' is internally inconsistent because a frozen backbone has no trainable parameters; please rephrase as '0.6% of the backbone parameter count.'","section":"Abstract and Section I"},{"comment":"The stem parameter count '3e-4M' (300 parameters) appears surprisingly small for two depthwise-separable convolutions with a channel mapping; please verify and report the exact stem parameter count.","section":"Table VI"},{"comment":"Training time and GPU memory are only given in the text for DINOv2-Large and SAM-Huge; consider adding these values to Table II for all variants.","section":"Section IV.C.1"},{"comment":"It would help to mark the best per-scene result in bold and to indicate the number of scenes won or lost by Cloud-Adapter, rather than only reporting overall metrics.","section":"Table XIII"}],"recommendation":"major_revision","confidential_remarks":"The underlying segmentation results may be sound, but the parameter-efficiency accounting and the per-scene robustness claims need substantial revision, and the L8 Biome baseline protocol must be documented. These issues are fixable within the manuscript's scope, and the released code should allow the authors to supply the missing details. I do not see evidence of circularity or fabrication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Cloud-Adapter is a solid applied paper rather than a conceptual leap. The genuinely new piece is the specific recipe: freeze DINOv2 Large, bolt on a multi-scale ConvNet spatial perception module whose features are fed through low-rank cross-attention adapters into frozen transformer layers, and run it on six cloud segmentation benchmarks. The combination is not in the cited prior work, and the empirical work is broad: consistent gains over prior specialized cloud segmentation models, especially large on HRC WHU (roughly 5.6 mIoU over the prior best). The ablations are thorough, covering the stem, ConvNet blocks, aggregator, context dimension, rank, and interaction frequency, and the code and checkpoints are public. That is real evidence and earns credit.\n\nThe main soft spot is the parameter-efficiency claim. The abstract and contributions say 0.6% of the backbone is trainable, but that count omits the Mask2Former segmentation head. Section IV-B says the head is Mask2Former, and Table VI reports 0.00M parameters for the no-adapter baseline even though the head is trained in every configuration. A typical dense-prediction Mask2Former head has tens of millions of parameters, so the true trainable fraction is likely well above 0.6%. The 1.82M number is a real count of the adapter modules, but it is not the full trainable count, and the paper does not say so. This needs fixing before the efficiency narrative can be compared against other PEFT methods.\n\nTwo smaller issues. There are no error bars or multi-seed results, which matters when the reported gains on some datasets are under a point. And the baselines are not described as retrained under the same protocol for the custom splits, notably L8 Biome; a reader cannot tell whether the comparison is protocol-matched. Also, the L8 Biome per-scene table (Table XIII) shows Cloud-Adapter losing to DBNet on several land cover types (Grass/Crops, Urban, Forest, Water), so the 'consistently superior' wording is too strong even though the overall mIoU is leading. The limitations section candidly notes the L8 Biome gap, which is fair. The citation pattern is fine; the reuse of the authors' PMAA block is a legitimate building block, not a hidden dependency.\n\nWho this is for: people building operational cloud masks or studying parameter-efficient adaptation for remote sensing. It deserves a serious referee; the experiments are extensive and the code is public. My recommendation is to engage, with the parameter accounting as a required revision rather than a desk-reject issue.","headline":"Solid adapter-for-cloud-segmentation study with extensive benchmarks, but the headline 0.6% parameter-efficiency claim omits the trained segmentation head.","tokens_in":23569,"tokens_out":4100,"would_cite":true,"duration_ms":33751,"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":"Cloud-Adapter freezes a DINOv2 backbone and adds 1.82M trainable parameters through a spatial-perception plus cross-attention adapter, reporting state-of-the-art cloud segmentation on all six benchmark subsets tested.","keywords":["cloud segmentation","remote sensing","vision foundation models","parameter-efficient fine-tuning","cross-attention adapter","DINOv2","low-rank adaptation","multi-scale spatial context"],"falsifier":"Count every trainable parameter in the full pipeline, including the segmentation head, and compare the total with the compared methods; if the total is not far smaller, the paper's parameter-efficiency claim is refuted. A second check is to retrain with a randomly initialized backbone of the same size; if the reported gains vanish, the frozen foundation model is doing the work.","tokens_in":22657,"feed_emoji":"☁️","tokens_out":10092,"duration_ms":87354,"temperature":0.7,"pith_summary":"The paper tries to show that a frozen, general-purpose vision foundation model can be turned into a strong cloud segmenter with a small trainable add-on. Cloud-Adapter keeps DINOv2 Large frozen and attaches a spatial perception module plus low-rank cross-attention adapting modules, 1.82M trainable parameters reported as 0.6% of the backbone. Across six benchmark subsets from Sentinel-2, Landsat 8, Gaofen-1/2, and Google Earth imagery, the method reports the best mIoU, mAcc, aAcc, and mDice in every comparison, including 89.05 mIoU on HRC WHU versus 83.44 for the previous best. If these results hold, cloud detection can inherit general visual knowledge without retraining a large backbone. The authors also state that the method inherits the foundation model's deployment cost and still leaves a substantial accuracy gap on the hardest dataset, L8 Biome.","feed_headline":"A 1.82M-parameter adapter tops six cloud-segmentation benchmarks","feed_subtitle":"Freezing a vision backbone plus a cross-attention adapter beats prior cloud segmenters on all six benchmarks.","key_machinery":"The adapting module is the mechanism that carries the argument: each frozen transformer layer provides queries while the aggregated multi-scale spatial context provides keys and values, followed by a low-rank MLP and a residual addition. The spatial perception module, built from two depthwise-separable-convolution stem layers, four stride-2 ConvNet blocks, and a parameter-free adaptive-average-pooling aggregator, supplies the dense multi-scale cloud context that lets the frozen backbone adapt to satellite imagery without any weight updates.","core_discovery":"The central claim is that a frozen DINOv2 Large backbone, modulated by (i) a spatial perception module that extracts multi-scale convolutional features and pools them into a dense context and (ii) adapting modules that use cross-attention with low-rank MLP projections to inject that context into every transformer layer, reaches state of the art on cloud segmentation. The paper reports consistent improvements over prior methods on all six tested subsets, with particularly large gains in difficult classes such as thin clouds and cloud shadows on CloudSEN12 High and on the high-resolution HRC WHU benchmark.","pith_inferences":["The reported 0.6% figure counts only the adapter and spatial-perception modules; including the trainable Mask2Former head that produces the final masks would put the total trainable parameters far above 1.82M, so the efficiency advantage over fully fine-tuned models is smaller than the headline suggests.","The aggregate L8 Biome lead is not uniform across scenes: the per-land-cover table shows Cloud-Adapter trailing DBNet on Grass/Crops and Urban, so the dataset-level robustness claim is driven largely by Snow/Ice and Shrubland gains.","The same frozen-backbone-plus-cross-attention recipe could apply to other dense remote-sensing tasks such as snow cover mapping, land cover segmentation, or burn scar detection, since the spatial perception module injects task-local context without touching the VFM weights.","An ablation with a randomly initialized backbone of the same size would separate how much of the gain comes from DINOv2's pretrained features versus from the adapter's spatial context; the paper only compares different pretrained backbones."],"forward_implications":["On HRC WHU, Cloud-Adapter reports 89.05 mIoU against 83.44 for the prior best method, and it reports the highest mIoU, mAcc, aAcc, and mDice on GF1, GF2, CloudSEN12 L1C, CloudSEN12 L2A, and L8 Biome as well.","The hardest four-class categories improve substantially: on CloudSEN12 High L1C, thin-cloud IoU rises to 56.15 and cloud-shadow IoU to 65.93, above the previous bests.","The frozen-backbone design transfers across sensor families and resolutions, from 10 m Sentinel-2 to 30 m Landsat 8 to 0.5-15 m Google Earth imagery, without unfreezing the backbone.","The trainable adapter is only 1.82M parameters according to the paper's accounting, and the ablations show each added component, stem, ConvNet blocks, aggregator, and adapting module, contributes a positive mIoU increment.","The paper's own limitations note that the method depends heavily on the VFM, that the large frozen backbone hinders edge-device deployment, and that L8 Biome still shows a substantial gap to manual masks at 57.53 mIoU."],"supporting_citations":[{"why":"Supplies the frozen DINOv2 Large backbone used in all final experiments.","marker":"[23]"},{"why":"Provides the alternative SAM backbone compared in the backbone ablation.","marker":"[22]"},{"why":"The Mask2Former segmentation head that turns adapted features into masks.","marker":"[60]"},{"why":"Inspires the low-rank MLP mapping used inside the adapting module.","marker":"[46]"},{"why":"Provides the CloudSEN12 High multi-class benchmark for L1C and L2A.","marker":"[56]"},{"why":"Provides the L8 Biome multi-land-cover benchmark.","marker":"[57]"},{"why":"Provides the GF12MS WHU Gaofen-1 and Gaofen-2 benchmark.","marker":"[35]"},{"why":"Provides the HRC WHU high-resolution binary cloud segmentation benchmark.","marker":"[12]"},{"why":"Supplies the Pure-ConvNet block adopted in the spatial perception module.","marker":"[55]"}],"fun_headline_variants":["0.6% param adapter tops six cloud segmentation sets","Frozen backbone and light adapter top cloud benchmarks","Cloud-Adapter 1.82M params SOTA on six cloud sets","Freeze VFM, add tiny adapter, beat cloud segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The efficiency claim rests on counting only the extra modules and not the part of the network that produces the final cloud maps; include that part, and the trained share of the model is much larger than 0.6%.","fun_headline_variants_meta":{"raw":{"variants":["0.6% param adapter tops six cloud segmentation sets","Frozen backbone and light adapter top cloud benchmarks","Cloud-Adapter 1.82M params SOTA on six cloud sets","Freeze VFM, add tiny adapter, beat cloud segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00058,"raw_usage":{"total_tokens":2697,"prompt_tokens":878,"completion_tokens":1819,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":1749}},"tokens_in":494,"tokens_out":1819,"duration_ms":14855,"temperature":1.0,"reasoning_tokens":1749,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:48:06.074524+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count every trainable parameter in the full pipeline, including the segmentation head, and compare the total with the compared methods; if the total is not far smaller, the paper's parameter-efficiency claim is refuted. A second check is to retrain with a randomly initialized backbone of the same size; if the reported gains vanish, the frozen foundation model is doing the work.","supporting_citations":[{"cited_title":"Masked-attention mask transformer for universal image segmentation,","cited_arxiv_id":null,"evidence_quote":"The Mask2Former segmentation head that turns adapted features into masks."},{"cited_title":"LoRA: Low-rank adaptation of large language models,","cited_arxiv_id":null,"evidence_quote":"Inspires the low-rank MLP mapping used inside the adapting module."},{"cited_title":"Cloudsen12, a global dataset for semantic understanding of cloud and cloud shadow in sentinel-2,","cited_arxiv_id":null,"evidence_quote":"Provides the CloudSEN12 High multi-class benchmark for L1C and L2A."},{"cited_title":"Cloud detection algorithm comparison and validation for operational landsat data products,","cited_arxiv_id":null,"evidence_quote":"Provides the L8 Biome multi-land-cover benchmark."},{"cited_title":"Transferring deep models for cloud detection in multisensor images via weakly supervised learning,","cited_arxiv_id":null,"evidence_quote":"Provides the GF12MS WHU Gaofen-1 and Gaofen-2 benchmark."},{"cited_title":"Deep learning based cloud detection for medium and high resolution remote sensing images of different sensors,","cited_arxiv_id":null,"evidence_quote":"Provides the HRC WHU high-resolution binary cloud segmentation benchmark."},{"cited_title":"PMAA: A progressive multi- scale attention autoencoder model for high-performance cloud removal from multi-temporal satellite imagery,","cited_arxiv_id":null,"evidence_quote":"Supplies the Pure-ConvNet block adopted in the spatial perception module."},{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Provides the alternative SAM backbone compared in the backbone ablation."},{"cited_title":"DINOv2: Learning Robust Visual Features without Supervision,","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen DINOv2 Large backbone used in all final experiments."}],"review_version":1}