{"id":"a3f1d00b-960a-4905-897c-9ceb2ced443f","arxiv_id":"2505.14088","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A parameter-efficient adapter with low-rank token experts and frequency-aware filtering improves domain generalization for multispectral land cover classification with frozen vision foundation models.","lead":"Land-MoE is a new adapter that helps large vision models classify land cover in satellite images even when the images come from different sensors or different cities. It reports accuracy gains over existing methods on multispectral and RGB benchmarks, offering a parameter-efficient route for generalization across domains.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never specifies how 4-band multispectral images are mapped to DINOv2's 3-channel patch embedding; if NIR is discarded or the mapping differs across sensors, the reported cross-sensor gains may reflect preprocessing rather than the adapter design.","rationale":"I read the paper as claiming that a PEFT adapter on frozen VFMs generalizes across multispectral sensors and regions. The architecture (MoLTE + FAF) is plausible, and the ablations and multi-backbone experiments provide some support. The most load-bearing gap is not the adapter math but the data interface: DINOv2 cannot accept 4 channels, and the paper never defines the 3-channel projection. This is an internal inconsistency in the implementation description, and it is externally verifiable from the missing code or preprocessing specification. The reader's weakest assumption points to the same issue. I do not see a more fundamental flaw in Eqs. (1)-(11); the routing, low-rank token formulation, and frequency filtering are described consistently. The right response is to keep the conditional verdict and require the authors to disclose the band mapping and reproduce at least the cross-sensor numbers under an identical, fixed input protocol.","tokens_in":20599,"tokens_out":4080,"duration_ms":38128,"concrete_test":"Open the supplementary code's data loader and find the exact 4-band-to-3-channel transformation for each sensor. Then rerun the cross-sensor evaluation (Table 1) with all methods (DSTC, frozen DINOv2, SET, Rein, FADA, Land-MoE) receiving exactly the same 3-channel input, reporting mIoU separately for GF-1, PlanetScope, and Sentinel-2. If Land-MoE's margins over SET/Rein/FADA persist under identical input on each sensor, the concern is resolved; if they shrink or reverse, the headline gains are artifacts of the unspecified band mapping.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim--that Land-MoE's adapters, not some input choice, drive the large cross-sensor gains--rests on an unstated preprocessing step. DINOv2's patch embedding is a 3-channel Conv2d, yet all four sensors in Appendix A.1 provide 4-band (B,G,R,NIR) images. Section 4.1 says only 'MSIs are preprocessed by cropping to 512x512'; nowhere does the paper say whether NIR is dropped, which three bands are used, whether channels are replicated, or whether a false-color composite is formed. If NIR is discarded, the cross-sensor task is effectively RGB-only and the 'multispectral spectral-shift' story is weakened. If different band mappings are used for different target sensors, the comparison across methods is uncontrolled and the per-sensor margins in Table 1 could reflect the chosen band order rather than MoLTE/FAF. Because this affects every multispectral experiment (Tables 1-4), the main quantitative claim cannot be independently checked without the omitted input specification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Land-MoE, a parameter-efficient fine-tuning approach for multispectral land cover classification under domain shift. The method inserts two types of adapters into a frozen vision foundation model (DINOv2 by default): a Mixture of Low-rank Token Experts (MoLTE) that routes each token to a learnable low-rank token expert with different rank, and frequency-aware filters (FAF) that modulate features in the real-FFT domain. The authors evaluate on cross-sensor (GF-2 to GF-1, PlanetScope, Sentinel-2) and cross-geospatial (disjoint Chinese cities) tasks built from the Five-Billion-Pixels dataset, plus RGB cross-scene tasks from LoveDA and Potsdam/Vaihingen. They report consistent improvements over frozen VFM baselines and over SET, Rein, and FADA, with the largest gains on the cross-sensor task and smaller gains on cross-geospatial and RGB tasks.","tokens_in":20960,"tokens_out":6829,"duration_ms":68053,"significance":"If the reported results are reproducible, the paper makes a useful contribution: it is among the first to apply VFM-style parameter-efficient fine-tuning to multispectral land cover classification, and the proposed MoLTE/FAF design is coherent and parameter-efficient. The evaluation covers multiple sensors and geospatial regions, uses the same backbone and decoder for all compared VFM methods, and includes ablations and sensitivity analyses for expert count, rank, token length, learning rate, and batch size. However, the central quantitative claims are currently supported too weakly: the multispectral input preprocessing is not specified, all numbers are single-run with no error bars, and the model configuration is tuned separately for each benchmark. These issues must be addressed before the claims of large-margin state-of-the-art performance can be accepted.","major_comments":[{"comment":"The manuscript never states how 4-band multispectral images (B, G, R, NIR) from GF-2, GF-1, PlanetScope, and Sentinel-2 are mapped to the 3-channel patch embedding of DINOv2. Section 4.1 only says that MSIs are preprocessed by cropping to 512x512, and Appendix A.1 confirms that all four sensors provide four spectral bands, while DINOv2's patch embed is a 3-channel convolution. Because every multispectral experiment (Tables 1-4) depends on this unstated band selection or channel mapping, the central cross-sensor claim cannot be independently checked. Please specify the exact preprocessing (e.g., which three bands are used, or how NIR is folded in), and ideally include an ablation that isolates the contribution of the NIR channel.","section":"Section 4.1, Appendix A.1"},{"comment":"All accuracy numbers are reported from single runs, with no error bars, number of seeds, or significance tests. The cross-geospatial margins over SET, Rein, and FADA are only 1.59, 1.99, and 1.77 mIoU (Table 2), and the LoveDA margin over Rein is 1.40 mIoU (Table 10); differences of this size can easily be within run-to-run variation. Please provide multiple seeds with standard deviations or an appropriate significance test, or soften the 'large margin' wording for the smaller-gain settings.","section":"Tables 1, 2, 10, 11"},{"comment":"The optimal expert count and rank set are selected separately for each benchmark: Ne=3 with rk in {8,16,32} for cross-sensor but Ne=2 with rk in {8,16} for cross-geospatial, and Tables 1 and 2 use these different configurations. If target-task performance influenced this selection, the comparison is not a clean domain-generalization evaluation; at minimum, the selection criterion (e.g., source-domain validation) must be stated, and results for a single fixed configuration should be reported to support the generalization claim.","section":"Appendix B.2, Table 7"},{"comment":"The paper claims state-of-the-art MLCC performance but compares only against DSTC and three VFM-based DG segmentation methods (SET, Rein, FADA). It does not compare with the spectral-shift-specific DG methods it cites, such as FDGNet [60] or frequency-based optimal style mix [33], even though these are directly relevant to the cross-sensor/cross-geospatial setting. Without such comparisons, the 'state-of-the-art' claim is not fully supported.","section":"Section 4.2, Related Work"}],"minor_comments":[{"comment":"The weight lambda of the expert balancing loss is never given a value or ablated; please report it in the implementation details.","section":"Section 3.3, Eq. (11)"},{"comment":"The relationship between the two refinement terms is unclear: MoLTE produces Delta-X-bar_i, which is added to X_i to form Z_i for FAF, but Eq. (8) states that the input to the next layer is X_i + Delta-X_i, so the MoLTE adjustment only affects the next layer through the FAF output. Please clarify the data flow and the exact residual structure, and align the notation L_N versus N_L.","section":"Section 3.2 and Eq. (8)"},{"comment":"The header uses 'EV A02' while the text and reference list use 'EVA02'; please unify the naming.","section":"Table 4"},{"comment":"Several classes show exactly 0.00% accuracy for all methods (e.g., class C4 in Table 1), which suggests those classes may be absent from the target-domain ground truth; please explain how mAcc and mIoU aggregate classes that do not appear in the target split.","section":"Tables 1 and 2"},{"comment":"The cross-geospatial task is described as a subset of 150 GF-2 MSIs, but the number of training and test images in the source and target splits is not stated; please report the exact split sizes and the number of images per target city.","section":"Appendix A.3"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on 2505.14088. The core idea is clean: insert low-rank token experts with top-k routing plus a layer-shared frequency filter into frozen DINOv2, and fine-tune on Five-Billion-Pixels. That is a genuine new configuration, not just a pile of existing parts. The paper is also well organized, with ablations on component contribution, expert counts, ranks, token length, module placement, and four VFM backbones. That is more than many submissions bother to do, and the gains over SET, Rein, and FADA look consistent across tables.\n\nBut there is a load-bearing gap: all four sensors are 4-band (B, G, R, NIR), DINOv2's patch embedding expects 3 channels, and the paper never says how the mapping is done. Section 4.1 only says MSIs are cropped to 512x512. If NIR is dropped, the cross-sensor experiment is effectively RGB-only and the spectral-shift story weakens. If different band selections are used for different sensors, the per-sensor margins in Table 1 could reflect preprocessing, not MoLTE/FAF. This has to be disclosed before the numbers can be trusted.\n\nOther soft spots: all tables report single runs with no error bars or significance tests. Table 7 shows the optimal expert count and ranks are tuned per benchmark (3 experts with ranks {8,16,32} for cross-sensor, 2 experts with {8,16} for cross-geospatial), which is acceptable for an ablation but means the headline configuration is benchmark-specialized. The cross-geospatial gains are modest, around 1.6–2.0 mIoU, despite the title's larger-margin language. Code is promised but not released.\n\nBottom line: this is a serious paper with a reproducible-recipe gap. The architecture is worth a referee's time, and the authors are clearly competent. But the band-mapping omission is central; it must be fixed in revision. I would send it to review, with a request for the input specification, multiple seeds, and ideally a hold-out validation for hyperparameter selection.","headline":"A useful adapter recipe for multispectral land cover, but the missing 4-band-to-3-channel input specification undermines the main cross-sensor claim until it is disclosed.","tokens_in":21365,"tokens_out":1854,"would_cite":false,"duration_ms":31956,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-07T15:40:03.857543+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}