{"id":"10e9945b-c9b1-4249-9679-82fa913cfdcd","arxiv_id":"2502.09654","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A heterogeneous mixture-of-experts upsampler with multi-level feature guidance and dual routing achieves slightly higher PSNR and SSIM than prior methods on two remote sensing super-resolution benchmarks.","lead":"The paper introduces a super-resolution model that routes each pixel of a low-resolution remote sensing image to a specialized upsampling expert, selected from heterogeneous expert groups. On the UCMerced and AID benchmarks it reports small PSNR and SSIM gains over existing methods, with the largest gain at the ×4 scale.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The empirical 'superior' claim rests on one random split with test-set hyperparameter selection; the AID x4 margin is 0.006 dB, likely within training variance.","rationale":"The paper's contribution is coherent and the ablation is internally consistent: adding MFA and dual-routing moves PSNR upward on the reported split, and I found no load-bearing mathematical error in Eqs. (2)-(7). The central claim, however, is comparative ('superior compared to state-of-the-art methods'), and that claim is only as strong as the benchmark protocol. The protocol in Section III-A uses a single random split per dataset, bilinear-only LR degradation, no repeated seeds, and no explicit validation split for selecting N, M, K, and the expert kernel mix. The reported gains over TTST are small, especially on AID x4 where the PSNR margin is 0.0059 dB and the SSIM margin is 0.0002, both well within typical run-to-run variability for SR training. Because the strongest evidence for the contribution is empirical and the empirical comparison is not statistically anchored, the appropriate verdict is the same CONDITIONAL verdict the reader gave: the method may well work, but the load-bearing superiority claim is not yet established. Releasing code and running repeated-seed, repeated-split experiments with validation-based hyperparameter selection would settle the concern.","tokens_in":8153,"tokens_out":4815,"duration_ms":45848,"concrete_test":"Run MFG-HMoE and the closest baselines (TTST, HAT, SPT) on UCMerced and AID at x2 and x4 with 5 independent random splits and 3 training seeds per split, using the paper's fixed 100,000-iteration budget and bilinear degradation, and report mean plus standard deviation of PSNR and SSIM. In addition, perform the Table III hyperparameter selection on a validation split and evaluate the chosen configuration only on the held-out test split. If the AID x4 PSNR advantage over TTST is less than one standard deviation, the 'superior accuracy' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract and conclusion assert that MFG-HMoE achieves superior SR accuracy over state-of-the-art methods. What has to be true for that claim to hold is that the PSNR/SSIM gaps in Tables I and II are larger than the noise of the evaluation protocol. Section III-A reports one random train/test split per dataset, bilinear-only LR generation, a fixed 100,000-iteration budget, and 'best-performing model' selection, with no repeated seeds and no error bars; the hyperparameters N=2, M=8, K=1 and the 1x1+3x3 expert combination are chosen using UCMerced test-set numbers (Table III and Section III-A), which biases the reported UCMerced results upward. The margins over the strongest baseline TTST are: UCMerced x2 +0.105 dB and x4 +0.143 dB PSNR; AID x2 +0.026 dB and x4 +0.0059 dB PSNR, with SSIM differences as small as 0.0002. The AID x4 margin is well inside typical seed-to-seed variation for SR training, so the comparative claim is not yet established by the evidence as reported. This is an evaluation-protocol concern, not a mathematical error; the architecture may still be sound.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes MFG-HMoE, a super-resolution model for remote sensing images that replaces the uniform upsampling layer of an existing backbone (RHAG/HAT) with a heterogeneous mixture-of-experts head. The head organizes upsampling experts into N groups with different kernel sizes, routes each pixel first to an expert group and then to top-K experts using a multi-level feature aggregation signal (Eqs. (5)-(7)), and is trained end-to-end. Experiments are reported on UCMerced and AID at ×2 and ×4 scales, comparing with eight published SR methods; the proposed model shows the highest PSNR and SSIM in Tables I-II. Ablations on UCMerced ×4 (Table III) attribute the gains to the MoE head, the multi-level feature aggregation, and the dual-routing mechanism. The code is promised at a public GitHub URL.","tokens_in":8419,"tokens_out":7676,"duration_ms":61467,"significance":"The idea of routing pixels to heterogeneous upsampling experts is sensible and the paper provides a concrete instantiation with a public code link. If the reported gains are reproducible, the contribution is a useful modular upsampling head that can be attached to existing remote-sensing SR backbones. However, the empirical support is currently too thin to establish the advertised 'superior' claim: the reported margins over the strongest baseline are often smaller than typical training noise (e.g., AID ×4 PSNR +0.0059 dB vs. TTST), and the evaluation protocol uses one split, no repeated seeds, no validation-based model selection, and test-set hyperparameter choice. The architectural components are validated in the expected direction in the ablations, which is a strength, and the paper is clearly written.","major_comments":[{"comment":"The central 'superior' claim is not supported by the reported statistics because each model is trained once on a single random split with no repeated seeds or confidence intervals. The margins over the strongest baseline TTST are as small as +0.0059 dB PSNR and +0.0002 SSIM on AID ×4 (Table II), and +0.0258 dB on AID ×2; such differences may be within typical seed-to-seed variation in SR training, so the ranking is not demonstrably stable. Please provide repeated-seed experiments with mean and standard deviation (or error bars) for the proposed method and the strongest baselines, and adjust the abstract and conclusion claims accordingly.","section":"Section III-A and Tables I-II"},{"comment":"Hyperparameters N=2, M=8, K=1 and the 1×1+3×3 expert combination are selected based on UCMerced ×4 test-set numbers in Table III, and the same UCMerced test set is then used for the final comparison in Table I. In addition, the text states that 'the best-performing model was selected for evaluation' without describing a validation split, so model selection appears to use test data as well. This makes the UCMerced results optimistically biased and does not provide an honest out-of-sample comparison. Please introduce a held-out validation set for model and hyperparameter selection, or explicitly report the selection procedure and present unbiased test-set results.","section":"Section III-A and Table III"},{"comment":"The ablation table is internally inconsistent in its current form: the first row has fewer entries than the number of columns, the row '16 0 16 0 0.7924 29.1938' appears twice, and several Group 4 and Group 5 rows do not align with the column headers. Because this table is the evidence for the MFA, dual-routing, expert-number, and kernel-size conclusions, it must be corrected and rechecked before the design claims can be assessed.","section":"Section III-C and Table III"}],"minor_comments":[{"comment":"The notation W_i ∈ R^{C×N} appears inconsistent with a linear map from a C-dimensional pixel feature to N group scores; each group should have a weight vector W_i ∈ R^C, or the matrix should be W ∈ R^{N×C}. The same issue appears for W_i^j in Eq. (6).","section":"Section II-C, Eq. (5)"},{"comment":"The baseline list says 'SPT [16]', but the reference list and Tables I-II identify SPT as reference [15]; reference [16] is the saliency-aware dynamic routing method, so the citation should be corrected.","section":"Section III-A"},{"comment":"The metric protocol is incomplete: please specify whether PSNR and SSIM are computed on the Y channel and whether a border is cropped before evaluation, as is standard in the SR literature.","section":"Section III-A"},{"comment":"If K > 1, the gating weights are sums of p_i p_j over the selected experts and are not renormalized to 1; please clarify whether this is intentional and whether the implementation uses exactly this weighting.","section":"Section II-C, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the architecture is plausible, but the empirical evidence is currently borderline because of the single-split, no-validation, test-set-selection protocol and the very small margins on AID. I would be willing to re-review after the authors add repeated-seed experiments, correct the evaluation protocol, and fix the ablation table."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on MFG-HMoE. It is a solid, incremental architecture paper with a genuine, if modest, idea: route each pixel to a specialized upsampler, with experts grouped by kernel size and a two-level router. That specific combination -- group router followed by expert router, with multi-level features feeding both -- is not in the papers they cite. The ablations line up as expected: MoE beats a single upsampler, MFA and dual routing each add a bit. The expert-selection visualization is a nice sanity check that the router is doing something semantic. If you work on remote sensing SR, this is worth a look.\n\nThe soft spot is exactly what the stress-test flags: the empirical \"superior\" claim is built on a single random split per dataset, no repeated seeds, no error bars, and hyperparameters (N, M, K, expert kernel mix) chosen using the UCMerced test set. That selection bias inflates the UCMerced numbers, and the AID margins are tiny -- 0.0059 dB PSNR at x4, with SSIM differences at the fourth decimal. Those are inside typical seed-to-seed variation for SR training, so the claim of state-of-the-art superiority over TTST is not actually established by the evidence as reported. The paper also says \"best-performing model was selected\" without saying on which split; if it is the test set, that is leakage. Minor issues: SPT is cited as [16] in the text but [15] in the tables, and the ablation table is formatted awkwardly. No code is available yet despite the promise, and there are no derivations, so there is nothing to formally check.\n\nNone of this breaks the architecture. The design is reasonable, the math is simple and correct, and the ablations are internally consistent. The problem is the evaluation protocol, not the method. The authors could fix this with multi-seed runs, a proper validation split for ablations, error bars, and code release. If they do, the paper would be a decent venue-level contribution; as it stands, the headline claim overreaches.\n\nFor review: I would send this to referees. It is a legitimate architectural variant with a clear story, and the protocol weakness is fixable. But I would expect major revision, and I would push hard on the evaluation before accepting.","headline":"Incremental but sensible MoE-based upsampling for remote sensing SR; the architecture is plausible, but the reported superiority over SOTA is not established because the evaluation rests on one split with test-set tuning and tiny margins.","tokens_in":8959,"tokens_out":2737,"would_cite":false,"duration_ms":30565,"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":"This paper claims that a heterogeneous mixture-of-experts upsampling head with dual routing reconstructs remote sensing images more accurately than uniform upsampling heads, reporting higher PSNR and SSIM than prior state-of-the-art…","keywords":["remote sensing image super-resolution","mixture of experts","heterogeneous experts","dual routing","multi-level feature aggregation","pixel-level selection","UCMerced","AID"],"falsifier":"Re-run the x4 UCMerced comparison with three or more different random splits of the 21 scene categories into training and testing, keeping the same 3:1 ratio and the same 100,000-iteration budget, and record the PSNR gap between MFG-HMoE and TTST; if the gap changes sign under any split, or if a test set built with bicubic downsampling instead of bilinear downsampling reverses the ranking, the paper's central superiority claim would be refuted.","tokens_in":7940,"feed_emoji":"🛰️","tokens_out":7802,"duration_ms":472625,"temperature":0.7,"pith_summary":"This paper argues that the final upsampling stage of a remote sensing super-resolution network should not treat every pixel the same. It introduces MFG-HMoE, which replaces a single fixed upsampling layer with a mixture of expert upsamplers: small convolution-plus-pixel-shuffle heads with different kernel sizes, grouped so experts within a group share an architecture while groups differ. A multi-level feature aggregation module pools shallow and deep backbone features to drive two routing steps, first selecting an expert group for each pixel and then selecting the best expert inside that group. The paper reports that this raises PSNR and SSIM over a uniform upsampling head and over prior state-of-the-art models at 2x and 4x magnification on UCMerced and AID. The intended significance is that specialized per-pixel reconstruction patterns, rather than only a stronger uniform network, are what let SR models cope with mixed ground objects such as buildings, vegetation, and vehicles.","feed_headline":"Per-pixel expert routing beats uniform upsampling in remote sensing SR","feed_subtitle":"MFG-HMoE lifts x4 PSNR to 29.29 dB on UCMerced, surpassing eight prior SR baselines.","key_machinery":"The load-bearing object is the Dual-Routing Heterogeneous Mixture of Experts (DR-HMoE) upsampling head. Each expert is a convolution followed by pixel shuffle, $E_{ij}(\\cdot)=\\mathrm{PS}(\\mathrm{Conv2D}(\\cdot))$, indexed by group $i$ and in-group position $j$. The group router computes selection probabilities $p^k_i$ by softmax over the aggregated feature $x^{\\mathrm{agg}}_k$ for $N$ groups; the expert router computes $p^k_j$ using the same aggregated features plus a position encoding of the active group; and the final pixel output is the probability-weighted combination $X^k_{\\mathrm{out}}=\\sum_{j\\in G_i} p^k_i p^k_j E_{ij}(x^{\\mathrm{feat}}_k)$ over the top-$K$ selected experts. The Multi-level Feature Aggregation (MFA) module supplies the router input by summing transformed outputs of all RHAG blocks, $X_{\\mathrm{agg}}=\\mathrm{LeakyReLU}(\\mathrm{Conv2D}(\\mathrm{LN}(\\sum_i X_i')))$. The mechanism's job is to give each pixel a specialized upsampling path; the ablations credit the expert mixture, the MFA guidance, and the dual-routing step with successive gains.","core_discovery":"The central claim is that heterogeneity plus dual routing in the upsampling head produces measurable super-resolution gains on remote sensing imagery. Concretely, with $N=2$ expert groups, $M=8$ experts per group, and $K=1$ selected expert per pixel, the best configuration pairs $1\\times1$ and $3\\times3$ convolutional upsamplers; on UCMerced $\\times4$ it reaches 29.2882 dB PSNR and 0.7954 SSIM, above TTST's 29.1455 dB and 0.7896, and on AID $\\times2$ it reaches 36.5399 dB PSNR and 0.9443 SSIM. The paper also claims the routing is semantically meaningful: relatively smooth regions such as vegetation, water, and shadows, and small objects such as vehicles, select one expert group, while larger objects with richer detail such as buildings and oil tanks select the other. This division of labor is presented as evidence that heterogeneous experts are required, not merely helpful, for remote sensing scenes.","pith_inferences":["A natural test of the semantic-specialization claim would be to count, across many random seeds and splits, which expert group each UCMerced or AID scene class selects; stable class-to-group mappings would strengthen the claim that the routers encode ground-object categories rather than local texture.","Because low-resolution images are produced only by bilinear downsampling, the method's robustness to more realistic degradations such as motion blur plus sensor noise is untested; an extension would train the same routers on bicubic downsampling or on paired real-world low/high-resolution imagery and check whether the gains persist.","The paper fixes $K=1$ for the top-1 expert selection; a cheap extension is to test $K=2$ or $K=3$ with the soft weighting in Eq. (7) to see whether combining more experts per pixel yields additional accuracy.","If per-pixel routing is the source of the gain, the same dual-routing head should transfer to non-remote-sensing super-resolution benchmarks; comparing MFG-HMoE's head on natural-image SR datasets is a direct way to separate the benefit of heterogeneous routing from the benefit of the RHAG backbone."],"forward_implications":["On UCMerced and AID at both $\\times2$ and $\\times4$, MFG-HMoE reports higher PSNR and SSIM than each of the eight compared baselines: EDSR-L, RRDBNet, RCAN, SwinIR, HAT, HAUNet, SPT, and TTST.","The ablation study shows the heterogeneous mix matters: replacing 16 $1\\times1$ experts with the optimal 8 $1\\times1$ plus 8 $3\\times3$ experts, and adding dual routing, raises $\\times4$ UCMerced PSNR from 29.1938 dB to 29.2882 dB.","Adding the MFA module to a 16-expert homogeneous mixture raises $\\times4$ UCMerced PSNR from 29.1938 dB to 29.2631 dB, showing that shallow-plus-deep routing cues contribute beyond expert count.","The routing visualizations indicate that expert groups acquire a semantic division of labor, with smooth and small objects sent to one group and large detailed objects to the other.","Because the RHAG backbone is retained and only the upsampling head is changed, the proposed head can be grafted onto existing feature extraction networks without architectural retraining of the backbone."],"supporting_citations":[{"why":"The top-K token-selective transformer is the strongest prior baseline in Tables I and II; MFG-HMoE's superiority claim is evaluated against it.","marker":"[13]"},{"why":"Supplies the Residual Hybrid Attention Group backbone used as the feature extraction network, and is also a comparison baseline.","marker":"[22]"},{"why":"Supplies the UCMerced dataset used for training, testing, and all ablation studies.","marker":"[19]"},{"why":"Supplies the AID dataset, the second benchmark for the main comparison.","marker":"[20]"},{"why":"One of the eight state-of-the-art baselines in the comparisons; establishes the uniform-head baseline family.","marker":"[23]"},{"why":"RRDBNet, a generative SR baseline included in the comparison tables.","marker":"[24]"},{"why":"RCAN, a very deep residual channel attention baseline included in the comparison tables.","marker":"[25]"},{"why":"SwinIR, the transformer restoration baseline that HAT extends, included in the comparison tables.","marker":"[26]"}],"fun_headline_variants":["Heterogeneous Mixture of Experts beats uniform SR upsamplers","Dual-routing MoE improves remote sensing image super-resolution","Pixel-wise expert routing outperforms single-network SR on aerial data","MoE groups tailored for ground objects lift PSNR in remote sensing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation protocol that supports the superiority claim uses a single random train/test split per dataset, low-resolution images generated only by bilinear downsampling, and a fixed 100,000-iteration training budget with the best checkpoint selected; if the reported ranking does not survive other random splits or other degradation models, the claim of general superiority over state-of-the-art methods collapses.","fun_headline_variants_meta":{"raw":{"variants":["Heterogeneous Mixture of Experts beats uniform SR upsamplers","Dual-routing MoE improves remote sensing image super-resolution","Pixel-wise expert routing outperforms single-network SR on aerial data","MoE groups tailored for ground objects lift PSNR in remote sensing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000508,"raw_usage":{"total_tokens":2501,"prompt_tokens":999,"completion_tokens":1502,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":1427}},"tokens_in":615,"tokens_out":1502,"duration_ms":31871,"temperature":1.0,"reasoning_tokens":1427,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T10:29:42.916887+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the x4 UCMerced comparison with three or more different random splits of the 21 scene categories into training and testing, keeping the same 3:1 ratio and the same 100,000-iteration budget, and record the PSNR gap between MFG-HMoE and TTST; if the gap changes sign under any split, or if a test set built with bicubic downsampling instead of bilinear downsampling reverses the ranking, the paper's central superiority claim would be refuted.","supporting_citations":[{"cited_title":"Ttst: A top-k token selective transformer for remote sensing image super-resolution,","cited_arxiv_id":null,"evidence_quote":"The top-K token-selective transformer is the strongest prior baseline in Tables I and II; MFG-HMoE's superiority claim is evaluated against it."},{"cited_title":"Acti- vating more pixels in image super-resolution transformer,","cited_arxiv_id":null,"evidence_quote":"Supplies the Residual Hybrid Attention Group backbone used as the feature extraction network, and is also a comparison baseline."},{"cited_title":"Bag-of-visual-words and spatial extensions for land-use classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the UCMerced dataset used for training, testing, and all ablation studies."},{"cited_title":"Enhanced deep residual networks for single image super-resolution,","cited_arxiv_id":null,"evidence_quote":"One of the eight state-of-the-art baselines in the comparisons; establishes the uniform-head baseline family."},{"cited_title":"Esrgan: Enhanced super-resolution generative adversarial networks,","cited_arxiv_id":null,"evidence_quote":"RRDBNet, a generative SR baseline included in the comparison tables."},{"cited_title":"Image super-resolution using very deep residual channel attention net- works,","cited_arxiv_id":null,"evidence_quote":"RCAN, a very deep residual channel attention baseline included in the comparison tables."},{"cited_title":"Swinir: Image restoration using swin transformer,","cited_arxiv_id":null,"evidence_quote":"SwinIR, the transformer restoration baseline that HAT extends, included in the comparison tables."}],"review_version":1}