{"id":"816f8e74-2741-495c-be32-56ce1cc27765","arxiv_id":"2506.01040","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"ECP-Mamba, a Mamba-based network with a spiral scan and multi-scale self-distillation, reports state-of-the-art PolSAR image classification accuracy at label rates as low as 0.2%.","lead":"This paper presents a deep learning method for classifying radar satellite images of the ground using very few labeled examples. It combines self-supervised learning with a Mamba state space model and reports top accuracy on four benchmark radar datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim is not yet secure: training and test pixels overlap within 32-pixel patches with no spatial buffer or multiple-seed error bars, so reported 99.70% OA may reflect memorization rather than generalization.","rationale":"The paper's central claim is state-of-the-art accuracy on four PolSAR benchmarks at extremely low label rates (0.05-0.2%). The architecture is plausible: DINO-style self-distillation with local/global branches, a spiral scan that places the central pixel near the class token, and a lightweight Cross Mamba fusion. The ablation study shows coherent improvement from each component, and the efficiency numbers (31.44M FLOPs for the non-CL version) support the speed claim. However, the entire empirical edifice rests on a single evaluation protocol: a random 0.2% sample of labeled pixels is used for training, and all labeled pixels—including the training pixels and their spatial neighbors—are used for evaluation. Because each pixel is classified from a 16×16 or 32×32 patch centered on it, a test pixel one pixel away from a training pixel shares nearly all of its receptive field with the training patch. Under this protocol, a model can memorize the training patches and propagate labels to adjacent pixels, inflating OA/AA/Kappa. The paper does not describe any spatial buffer, does not exclude training pixels from the test set, and does not report error bars over multiple random splits. The reported margins over PiCL (4.25% OA on Flevoland 1989) could be genuine, but they could also be an artifact of this protocol, especially since ECP-Mamba's larger global receptive field (32×32) may benefit more from overlap than baselines. The reader's weakest assumption identifies exactly this issue, so I agree. Secondary issues such as the EMA momentum inconsistency and tuning on the headline dataset add to uncertainty but are not independently fatal. The concrete test I propose—evaluating on pixels at least 32 pixels away from any training pixel—would directly measure the leakage effect. If the accuracy drop is small, the SOTA claim stands; if large, the claim needs substantial qualification. Thus the conditional verdict is appropriate and the reader's concern is confirmed.","tokens_in":24267,"tokens_out":9146,"duration_ms":92624,"concrete_test":"Re-run the Flevoland 1989 experiment with the exact hyperparameters, but evaluate only on pixels whose Chebyshev distance to every training pixel is at least 32 (the global patch radius). If OA/AA/Kappa on this disjoint test set fall by more than 1 percentage point from the reported 99.70/99.64/99.62, the SOTA claim is materially inflated by receptive-field overlap. Also repeat the 0.2% random label sampling with five seeds and report mean ± std to quantify protocol variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—state-of-the-art accuracy on Flevoland 1989 (OA 99.70%, AA 99.64%, Kappa 99.62e-2) at a 0.2% label rate—depends entirely on the evaluation protocol. The paper trains on a random 0.2% subset of labeled pixels (345 pixels on Flevoland 1989) and then classifies every pixel in the image (Algorithm 2 returns O for the complete dataset D). It never states that training pixels are excluded from evaluation, and it applies no spatial buffer between training and test pixels. Each prediction uses a 16×16 local patch and a 32×32 global patch centered on the target pixel. A test pixel just one pixel away from a training pixel shares 31 of 32 patch columns/rows with the training patch; even a test pixel 16 pixels away still overlaps the training patch substantially. The model can therefore memorize the labels of the few hundred training patches and propagate them to neighboring pixels by patch similarity. This inflates all metrics, and the effect is likely stronger for ECP-Mamba than for baselines because of its larger receptive field. With no error bars and no multiple-seed variation reported, the 99.70% figure is a single draw from an unknown distribution under a protocol that does not measure true generalization. Secondary issues—the EMA momentum inconsistency between Section IV-D (cosine from 0.9995 to 1) and Section VII-A (fixed 0.996), and hyperparameters tuned on the headline dataset—reduce confidence further but are not the crux. The crux is the unaddressed overlap between training and evaluation pixels.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ECP-Mamba, a framework for PolSAR image classification that combines multi-scale self-supervised contrastive learning with a state space model (Mamba) backbone. The method introduces a spiral scan strategy to order patch tokens by causal relevance to the central pixel, a bidirectional selective scanning block, and a Cross Mamba module for fusing local and global features. The authors report state-of-the-art results on four PolSAR benchmarks, with a headline overall accuracy of 99.70% on Flevoland 1989 using only 0.2% labeled pixels. Ablation studies and parameter analyses are provided, and the code is promised.","tokens_in":24657,"tokens_out":2663,"duration_ms":28625,"significance":"If the reported results are trustworthy, ECP-Mamba would be a meaningful advance for label-scarce PolSAR classification, being the first adaptation of Mamba/SSM to this task and showing large gains over prior contrastive and supervised baselines at extremely low sampling rates. The proposed spiral scan and Cross Mamba are intuitively motivated and the reported FLOPs are low. However, the central empirical claim rests on an evaluation protocol that may allow training/test leakage through overlapping receptive fields, and the results lack error bars or multiple-seed variation. The significance of the contribution is therefore currently conditional on the protocol being corrected.","major_comments":[{"comment":"The evaluation protocol does not prevent overlap between training and test patches. Algorithm 2 states that the classifier is applied to the complete dataset D and outputs O, while fine-tuning uses a randomly selected labeled subset D′. With local patch size k=16 and global patch size K=32 centered on each pixel, a test pixel located even one pixel away from a training pixel shares 31 of 32 columns/rows of its global patch and large parts of its local patch. If the training pixels themselves are also evaluated, or if no spatial buffer is used, the model can memorize the few hundred training patches and propagate their labels to neighboring pixels by patch similarity. The paper does not state that training pixels are excluded from evaluation, nor does it describe any patch-disjoint split or spatial buffer. This threatens the validity of the headline accuracies, including the 99.70% OA on Flevoland 1989. Please clarify whether training pixels are included in the evaluation set and report results with a spatial buffer (e.g., excluding all pixels whose receptive field overlaps a training pixel) and with training pixels removed from the test set.","section":"Section VII-A and Algorithm 2"},{"comment":"The network hyperparameters (patch sizes, hidden dimension, number of BPSS layers, etc.) are selected on the Flevoland 1989 dataset in the parameter analysis, and the same dataset is then used for the headline SOTA claim in Table III. This creates a risk of overfitting to a single test set. While the method is also evaluated on three other datasets, the default settings are not validated for sensitivity across all datasets. Please either perform parameter selection on a validation split of Flevoland 1989 or on a different dataset, and report results for the other datasets with per-dataset tuned parameters alongside the fixed-default results.","section":"Section VII-D1 and Table III"},{"comment":"All experimental numbers are given without error bars or repeated runs. At sampling rates as low as 0.2% (and 0.05% for San Francisco), the random choice of training pixels can cause large variance across seeds; for example, the Flevoland 1989 buildings class has only 2 labeled instances, so a single random draw may not be representative. The paper should report the mean and standard deviation over at least 5 independent random splits, and ideally over multiple training seeds, for both the proposed method and the main baselines. Without this, the claimed margins over PiCL (e.g., 4.25% OA on Flevoland 1989) cannot be assessed statistically.","section":"Tables III, IV, V, VI and Table VIII"},{"comment":"There is a direct inconsistency in the EMA momentum parameter λ: Section IV-D states that λ follows a cosine schedule increasing from 0.9995 to 1 during pre-training, while Section VII-A states 'The momentum parameter λ is 0.996.' The paper must state which schedule is actually used, since this affects reproducibility and possibly the reported results.","section":"Section IV-D versus Section VII-A"}],"minor_comments":[{"comment":"In the paragraph preceding Table VII, the text says 'the number of BBPS blocks' but the block is defined as BPSS in Section IV-C2. Please fix the typo.","section":"Section VII-D1"},{"comment":"The first expression for the PolyLoss uses (1+ε1)(1−ȳ(i)) and a summation over j, while the second simplifies to −y(i)log(ȳ(i))+ε(1−ȳ(i)). The equivalence is not shown and the first form appears to contain a typo (ε1 vs ε, and the y(i) term). Please rewrite the equation and show the expansion.","section":"Equation (18)"},{"comment":"The sentence 'Only the final tokens from w′ and W′, denoted as ct and Ct, are used' uses lowercase ct for both the class token and the token from the local branch; in Section IV-C1 ct is defined as the learnable tail class token. Consider renaming the fine-tuning tokens (e.g., c_t and C_t) to avoid ambiguity.","section":"Section V-B"},{"comment":"The paper says 'The learning rate and momentum parameters undergo decay following a cosine annealing strategy [60]' but the EMA momentum λ is described as fixed at 0.996 in the same paragraph. Please clarify which momentum is decayed and which is fixed.","section":"Section VII-A"},{"comment":"The FLOPs for CL + Multi-scale Efficient Mamba are listed as '-' with the note that FLOPs do not apply to the self-supervised method. This is confusing because the fine-tuning stage still uses the same architecture; if FLOPs are only reported for the fine-tuning forward pass, please state that explicitly.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a technically interesting adaptation of Mamba to PolSAR classification and the qualitative motivations are sound, but the evaluation protocol is a serious concern that must be addressed before publication. The overlap between training and test receptive fields and the absence of error bars are fixable with additional experiments, so I would not reject the manuscript outright. I also suggest the editor verify the 'first SSM in PolSAR' novelty claim against recent literature (e.g., RSMamba for remote sensing classification and other SSM-based hyperspectral/PoLSAR works), as the related-work section may understate concurrent developments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read, but I would not take the 99.70% OA at face value yet. The paper is the first to bring a selective state space model to PolSAR classification, and the adaptation is genuinely task-specific: the spiral scan plus tail class token is a reasonable way to respect the centrality of the target pixel, and the Cross Mamba fusion module is light and sensible. The multi-scale self-distillation without negative pairs is also a clean pretext task for low-label scenarios. The ablation study on Flevoland 1989 is thorough and shows each component buys something. Four datasets and a comparison that includes recent SSL methods, including their own PiCL, is a fair empirical scope.\n\nThe soft spot is the evaluation protocol, and it is a load-bearing one. They randomly sample 0.2% of labeled pixels for training, then classify every pixel in the image—including those same training pixels and their immediate neighbors. With a 16x16 local patch and a 32x32 global patch centered on each test pixel, a test pixel one pixel away from a training pixel shares 31 of 32 patch columns/rows with the training patch, and even one 16 pixels away still overlaps substantially. There is no spatial buffer and no statement that training pixels are excluded. Under that setup, the model can memorize a few hundred training patches and propagate their labels outward by patch similarity, and the effect is plausibly stronger for ECP-Mamba than for smaller-receptive-field baselines. That directly undermines the claimed state-of-the-art margins.\n\nThe secondary issues are real but minor by comparison: no error bars or multiple seeds, hyperparameters selected on Flevoland 1989 and then showcased on the same dataset, and an internal inconsistency in the EMA momentum (cosine from 0.9995 to 1 in Section IV-D vs. fixed 0.996 in VII-A). Also, some baseline numbers look suspiciously low (e.g., PolSF at 0% on bare soil), which suggests the comparisons may not be apples-to-apples.\n\nThe method itself is not invalidated. The architecture and pretext task are plausible and clearly described. But the empirical claim, as it stands, is conditional on fixing the evaluation: report multiple seeds, exclude training pixels or add a spatial buffer, and ideally release code so the community can rerun the protocol.\n\nI would send this to serious peer review with a request for major revision. The contribution is novel enough for the remote sensing audience, and the evaluation can be repaired. This is not a desk reject, but it is also not a paper whose headline accuracy can be cited yet.","headline":"Useful Mamba-for-PolSAR adaptation with a sensible spiral scan and a clean multi-scale pretext task, but the headline SOTA numbers are not yet trustworthy because training and test pixels overlap with no error bars.","tokens_in":25160,"tokens_out":1543,"would_cite":false,"duration_ms":18044,"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":"Radar pixel map hits 99.70% with a spiral-scan Mamba","keywords":["PolSAR image classification","self-supervised learning","contrastive learning","state space model","Mamba","spiral scan strategy","multi-scale feature fusion","self-distillation"],"falsifier":"Re-run the Flevoland 1989 experiment at a 0.2% sampling rate but evaluate each pixel only when its entire local and global patch is disjoint from every training pixel, or hold out whole contiguous regions; if overall accuracy drops well below the reported 99.70%, part of the result is carried by shared receptive fields rather than by generalization.","tokens_in":24081,"feed_emoji":"📡","tokens_out":8307,"duration_ms":76343,"temperature":0.7,"pith_summary":"ECP-Mamba claims to classify polarimetric SAR (PolSAR) imagery pixel by pixel with state-of-the-art accuracy while needing very few labels and much less computation than Transformer-based models. The method rests on two ideas: a self-supervised pre-training task that aligns local and global views of the same pixel through self-distillation, without negative samples, and a Mamba backbone whose spiral scan orders tokens so that context near the central pixel accumulates at the end of the sequence, right before a learnable class token. A lightweight Cross Mamba module then exchanges class tokens between the local and global branches so multi-scale features reinforce each other at minimal extra cost. On the Flevoland 1989 benchmark the paper reports 99.70% overall accuracy, 99.64% average accuracy, and a Kappa coefficient of 99.62e-2 at a 0.2% label sampling rate, and reports that ECP-Mamba beats competing self-supervised methods on all four datasets studied.","feed_headline":"Radar pixel map hits 99.70% with a spiral-scan Mamba","feed_subtitle":"ECP-Mamba pairs a spiral-scan state space model with multi-scale self-distillation to classify PolSAR imagery from 0.2% labels.","key_machinery":"The load-bearing mechanism is the spiral scan strategy: a patch's tokens are reordered by a spiral traversal that starts at the top-left corner and winds inward, then a learnable tail class token is appended and sinusoidal positional embeddings are added, so the central pixel ends up adjacent to the class token where the SSM's hidden state has accumulated the most context. Around this sits the Bidirectional Polarimetric Selective Scanning block, a forward-backward selective SSM with residual connection, and the Cross Mamba module, which fuses scales by exchanging tail class tokens between branches. The self-supervised loss is a negative-free self-distillation objective: softmax cross-entropy between the local branch's sharpened distribution and the global branch's centered, sharpened distribution, with the global branch updated by exponential moving average.","core_discovery":"The paper's central claim is that a selective state space model can be adapted to per-pixel PolSAR classification by making the scanning order respect the causal structure of the task: pixels closer to the central pixel carry more information about its label, so the spiral scan places them near the end of the sequence, immediately before a learnable tail class token that summarizes the whole patch. Classification fine-tuning fuses local 16×16 and global 32×32 views through a Cross Mamba module that swaps the tail class tokens of the two branches, and the contrastive pre-training stage learns without labels by predicting the global view's sharpened, centered softmax distribution from the local view, with only the local branch receiving gradients and the global branch updated as an exponential moving average. The paper reports that this combination achieves 99.70% OA, 99.64% AA, and 99.62e-2 Kappa on Flevoland 1989, and its ablations show that spiral scan, multi-scale branches, Cross Mamba, contrastive pre-training, and data augmentation each contribute to that result.","pith_inferences":["Beyond the paper: the spiral-scan principle should transfer to other per-pixel remote-sensing tasks, such as land-cover mapping, change detection, and hyperspectral classification, wherever the label belongs to the center of a window and near context matters more than far context.","Beyond the paper: a controlled comparison with a CNN or Transformer trained under the identical multi-scale self-distillation pre-text would isolate how much of the reported gain comes from the SSM and spiral scan rather than from the pre-training objective.","Beyond the paper: because the paper does not describe a patch-disjoint train/test split, the headline numbers should be re-checked under spatial separation; only then can the true gain over leakage-prone evaluation be quantified.","Beyond the paper: the negative-free self-distillation recipe can be stress-tested directly by removing the EMA and centering operations and measuring whether the global branch collapses, which would show whether the stability machinery is essential to the result."],"forward_implications":["If ECP-Mamba is right, per-pixel PolSAR classification no longer needs Transformer-scale attention; in the paper's ablations the spiral-scan Mamba reaches higher accuracy than a ViT baseline with about one tenth of the FLOPs.","The reported sampling-rate experiments imply self-supervised pre-training carries most of the labeling burden: at 0.05% labels on Flevoland 1989 the full method reports 97.99% overall accuracy while the Vim baseline reports 71.45%.","The local-to-global self-distillation objective is presented as a reusable recipe for any pixel-centric task: pre-train on unlabeled image patches by predicting a global view from a local view, with no negative pairs.","The tail class token position is load-bearing: performance on Flevoland 1989 peaks near the center of the spiral and drops when the token is placed far from the central pixel, so scan order should be tuned for Mamba-based per-pixel classifiers."],"supporting_citations":[{"why":"Supplies the selective state space model (Mamba) that the paper adapts to pixel-wise classification.","marker":"[34]"},{"why":"Supplies the bidirectional selective scanning block structure and the Vim baseline that the BPSS block is based on.","marker":"[32]"},{"why":"Supplies the self-distillation pretext task and the EMA/centering mechanism that the local-to-global pre-training adapts.","marker":"[45]"},{"why":"Provides the strongest contrastive learning baseline and the label sampling-rate protocol used in the fine-tuning experiments.","marker":"[4]"},{"why":"Provides the SSM discretization and convolutional representation that the Mamba backbone relies on.","marker":"[35]"},{"why":"Provides the classic complex-valued CNN baseline that the comparisons must outperform.","marker":"[17]"},{"why":"Provides the Transformer-based supervised baseline with local window attention used in the low-label comparisons.","marker":"[21]"},{"why":"Supplies the PolyLoss objective used in the classification fine-tuning stage.","marker":"[59]"}],"fun_headline_variants":["Spiral-scan Mamba hits 99.70% OA on PolSAR","Self-supervised Mamba: 99.70% PolSAR accuracy with sparse labels","ECP-Mamba: spiral-scan SSM for PolSAR reaches 99.70%","Radar pixel mapping: Mamba spiral scan scores 99.70%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains assume that randomly choosing 0.2% of labeled pixels to train on and then predicting every pixel, with local and global windows that overlap between training and test locations, does not inflate the scores; the paper does not say that training and test patches are kept spatially separate.","fun_headline_variants_meta":{"raw":{"variants":["Spiral-scan Mamba hits 99.70% OA on PolSAR","Self-supervised Mamba: 99.70% PolSAR accuracy with sparse labels","ECP-Mamba: spiral-scan SSM for PolSAR reaches 99.70%","Radar pixel mapping: Mamba spiral scan scores 99.70%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000499,"raw_usage":{"total_tokens":2498,"prompt_tokens":1055,"completion_tokens":1443,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":671,"completion_tokens_details":{"reasoning_tokens":1353}},"tokens_in":671,"tokens_out":1443,"duration_ms":12277,"temperature":1.0,"reasoning_tokens":1353,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:52:18.829017+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Flevoland 1989 experiment at a 0.2% sampling rate but evaluate each pixel only when its entire local and global patch is disjoint from every training pixel, or hold out whole contiguous regions; if overall accuracy drops well below the reported 99.70%, part of the result is carried by shared receptive fields rather than by generalization.","supporting_citations":[{"cited_title":"Polarimetry- inspired contrastive learning for class-imbalanced polsar image classification,","cited_arxiv_id":null,"evidence_quote":"Provides the strongest contrastive learning baseline and the label sampling-rate protocol used in the fine-tuning experiments."},{"cited_title":"Complex- valued convolutional neural network and its application in polarimetric sar image classification,","cited_arxiv_id":null,"evidence_quote":"Provides the classic complex-valued CNN baseline that the comparisons must outperform."},{"cited_title":"Local window attention transformer for polarimetric sar image classification,","cited_arxiv_id":null,"evidence_quote":"Provides the Transformer-based supervised baseline with local window attention used in the low-label comparisons."}],"review_version":1}