{"id":"988d3fc2-568d-426f-9147-32b032664aac","arxiv_id":"2507.05849","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"DFYP is a crop yield prediction framework combining resolution-aware channel attention, validation-driven edge operator selection, and CNN-ViT fusion, but the method's internal contradictions and test-set-tuned evaluation prevent the reported state-of-the-art performance from being credible.","lead":"This paper proposes DFYP, a deep learning framework that combines channel attention, adaptive edge operators, and a CNN-ViT dual branch with learned fusion to predict crop yields from satellite imagery. The reported gains over nine baselines on MODIS and Sentinel-2 data are undermined by an internally inconsistent method description and a selection-bias-prone evaluation protocol.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table II appears to select DFYP hyperparameters using the same 2009–2015 test years as the evaluation, so the reported 'consistent outperformance' may be an artifact of test-set tuning rather than the method.","rationale":"I read the paper in good faith: the proposed framework has a plausible motivation, the dual-branch design is standard, and the ablation ordering in Table III is internally consistent with the narrative that AOL and RCA each add value. The reader is right that the AOL-Net description is incoherent as written—Eq. (3) presents a continuous backprop-trained lambda, while Eq. (5) imposes a hard one-hot selection and Eq. (4) bases selection on validation accuracy rather than gradients—and this is a legitimate reproducibility defect. However, I do not think this is the single most load-bearing threat to the central claim. AOL could be repaired, or even removed, and the paper's empirical assertion could still be evaluated. The load-bearing condition is that the reported gains come from a fair comparison. The evidence in Section IV.E undercuts that condition: hyperparameters are analyzed on the exact evaluation years 2009–2015, and the 'optimal' configuration is identified from those numbers. Unless the table reports only validation-fold RMSE from data preceding each test year—which the text does not say—the final model has been selected using test labels. This alone invalidates the claim of consistent superiority, independent of the AOL ambiguity. The absence of error bars, the lack of numeric tables for baseline comparisons, and the reliance on figures without accessible values further reduce confidence, but the test-set-selection issue is decisive and concrete.","tokens_in":21592,"tokens_out":5197,"duration_ms":61640,"concrete_test":"Obtain the training logs or code and re-run the MODIS protocol with a nested temporal split: select hyperparameters (CNN layers, attention heads, ViT layers) using only data from 2003–2008, or using the 10% validation split of each training period, freeze the chosen configuration, then evaluate on 2009–2015 against the same baselines. Report whether Table II's RMSE values are test-year errors rather than validation errors. If DFYP's number of winning years shrinks or its average RMSE gap narrows substantially, the test-set tuning concern lands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—consistent state-of-the-art outperformance—requires an honest out-of-sample evaluation. Section IV.A states that models are trained on data from t_start to t-1 and then evaluated on t = 2009–2015 for MODIS. Section IV.E then presents Table II, 'Prediction Error (RMSE) Under Different Model Parameters on MODIS Dataset,' with columns for years 2009–2015, and uses it to declare the optimal configuration (6 CNN layers, 8 attention heads, 4 ViT layers). As written, the RMSE values in that table are computed on the same evaluation years, meaning the final DFYP configuration was selected using test labels. That makes the comparison against fixed baselines circular and invalidates the headline empirical claim. This concern is more fundamental than the AOL-Net trainability contradiction: even a correctly implemented AOL module cannot rescue a result whose architecture was chosen on the test set. The paper also overstates 'consistent outperformance' against its own results (CNN+LSTM wins RMSE/MAE/R² on MODIS in 2013; MMST-ViT wins R² on cotton), but the test-set tuning issue is the primary threat to the argument as written.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes DFYP, a dual-branch crop yield prediction framework combining a Resolution-aware Channel Attention (RCA) module, an Adaptive Operator Learning Network (AOL-Net) for edge-operator selection, and a learnable fusion of CNN and ViT branches. Experiments on MODIS and Sentinel-2 report RMSE, MAE, and R² improvements over nine baselines. The authors claim consistent state-of-the-art performance across spatial resolutions, crop types, and years, attributing the gains to dynamic operator selection, resolution-aware spectral reweighting, and adaptive fusion.","tokens_in":21873,"tokens_out":6783,"duration_ms":75342,"significance":"Crop yield prediction from remote sensing is an important applied problem, and the paper addresses a real limitation of fixed edge operators in CNN-based models. The empirical comparison of eight handcrafted edge operators across years and crops is a useful motivation for adaptive operator selection. If the method were validated with a clean evaluation protocol, the dual-branch fusion and AOL ideas could be a practical engineering contribution. However, the manuscript provides no code, no random seeds, and no statistical tests, and the reported evidence is compromised by model selection on test years. As written, the significance of the claimed improvements cannot be assessed.","major_comments":[{"comment":"The hyperparameter analysis in Table II reports RMSE on the same years (2009–2015) that Section IV.A defines as the evaluation years for MODIS, and the final configuration (6 CNN layers, 8 attention heads, 4 ViT layers) is selected from this table. This means the architecture used in the headline comparison of Section IV.B was chosen using test labels, making the comparison circular. Any advantage of DFYP over fixed baselines could be an artifact of test-set tuning. The authors must select hyperparameters on a separate validation split (e.g., a subset of the training years) and then evaluate once on held-out years, or use nested cross-validation, and report results for the final model without any test-set feedback.","section":"Section IV.E, Table II"},{"comment":"The paper states in Eq. (3) that λ is optimized by backpropagation to interpolate between Sobel and Scharr, but Eq. (5) imposes hard selection with λ ∈ {0,1} and Eq. (4) selects the operator by validation accuracy scores. No mechanism is provided for gradients through the arg max or through the hard binary gate, so the joint end-to-end training claimed in Section III.F and Fig. 1 is not actually specified. If the operator is chosen by validation accuracy, AOL-Net is a model-selection rule rather than a learned component; if λ is learned continuously, Eq. (5) is violated. The authors need to reconcile these definitions, for example with a Gumbel-softmax or straight-through estimator and a clear description of how validation accuracy enters the training objective.","section":"Section III.D, Eqs. (3)–(6)"},{"comment":"The abstract and conclusion claim DFYP “consistently outperforms” all baselines across all metrics, crops, and years, but the paper’s own results show that CNN+LSTM beats DFYP on all three metrics on MODIS in 2013 and that MMST-ViT has a higher R² for cotton on Sentinel-2. These exceptions are acknowledged in Section IV.B, so the global claim should be qualified to “best in most settings” or the evaluation should be restricted to the settings where the claim actually holds.","section":"Section IV.B vs. Abstract"},{"comment":"No variance or significance information is reported. All comparisons are single point estimates with no standard deviations, confidence intervals, or tests across random seeds. Given the small number of evaluation years (7 for MODIS) and crops (4 for Sentinel-2), the reported differences, especially the marginal cotton R² gap, cannot be distinguished from noise. The authors should report mean ± standard deviation over multiple runs and perform paired significance tests.","section":"Section IV.A and Fig. 2"},{"comment":"For Sentinel-2, evaluation is performed on a single test year (2022), yet the paper claims robust performance “across time periods.” A one-year evaluation cannot support a temporal-robustness claim. The authors should either evaluate on multiple held-out years or remove the temporal-consistency claim for Sentinel-2.","section":"Section IV.A"}],"minor_comments":[{"comment":"The manuscript contains numerous typos and grammatical errors, including “Yied” in the Section II.A heading, “repid development,” “comparision” in the Fig. 2 caption and Section IV.B, “abation” in Table III, and “DYFP” in Section III.B. A careful proofreading pass is needed.","section":"Throughout"},{"comment":"The notation in Eqs. (4)–(6) is confusing: Eq. (4) defines K_t as an arg max over K, but Eq. (6) defines K_t as a convex combination with coefficients λ_Sobel, λ_Scharr, λ_Learnable. These two definitions are not obviously equivalent, and the relationship between the validation-based score S(K,t) and the coefficients should be made explicit.","section":"Section III.D"},{"comment":"Table II is difficult to read because the columns are years and the rows are parameter values without a clear indication of which configuration each number belongs to. Presenting the results as a table with one row per configuration and columns for RMSE per year (or aggregated RMSE) would greatly improve clarity.","section":"Table II"},{"comment":"The choice of max pooling for low-resolution images and average pooling for high-resolution images is stated without empirical justification or a reference. A short experiment or a more detailed rationale would help the reader understand why this resolution-dependent pooling is preferable.","section":"Section III.C, Eq. (2)"}],"recommendation":"reject","confidential_remarks":"The test-set model-selection issue in Table II is the primary reason for rejection; it invalidates the headline empirical comparison. The AOL-Net gradient contradiction in Section III.D is a second serious flaw. If the authors re-run all experiments with a clean validation protocol, add variance reporting, and resolve the operator-selection gradient issue, a resubmission could be considered, but the current manuscript does not support its central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you want a clean example of test-set tuning masquerading as hyperparameter analysis. The architecture is a mashup of SE-style channel attention, a validation-selected edge operator, and CNN-ViT fusion; the evaluation claims are not supported because Table II selects the final configuration using the same 2009–2015 MODIS years that are then reported as test results. That is circular, and it alone would invalidate the headline of consistent state-of-the-art performance.\n\nWhat is actually new: the paper runs a useful empirical comparison of eight fixed edge operators and shows that no single one wins across years and crops, which motivates the adaptive operator idea. The ablation structure is thorough in covering CNN/ViT/RCA/AOL components, and the writing is clear.\n\nSoft spots: first, the AOL-Net description contradicts itself. Eq. (3) trains a continuous interpolation lambda, Eq. (5) imposes a hard one-hot selection with lambda in {0,1}, and Eq. (4) selects the operator by validation accuracy. No mechanism is given for backpropagating through the hard arg max, so the 'end-to-end learned selection' claim is not actually defined. If the selection is validation-based, that is model selection, not learning. Second, the evaluation protocol: Table II reports RMSE for each hyperparameter setting on the same test years, and the 'optimal' configuration is chosen from that table. That is a textbook leak. Third, the abstract claims 'consistently outperforms,' but the paper's own results show CNN+LSTM wins all metrics on MODIS in 2013 and MMST-ViT wins R² for cotton. That is not consistent. The lack of error bars or code makes it worse.\n\nNone of these flaws are fatal to the underlying idea — an adaptive edge operator plus fusion might help — but the current evidence is not sufficient. The paper deserves a serious referee only if the authors redo the hyperparameter selection with a proper validation split and reconcile the operator selection story. As written, I would not accept it.","headline":"A clear case of test-set tuning invalidates the headline performance claims; the AOL-Net description is also internally inconsistent, so the paper needs major revision before it is credible.","tokens_in":22387,"tokens_out":2610,"would_cite":false,"duration_ms":29760,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DFYP, a dual-branch deep network that couples resolution-aware spectral channel attention with a learned operator-selection module for edge features and a learnable fusion of CNN and Vision Transformer outputs, claims consistent…","keywords":["crop yield prediction","remote sensing","dynamic fusion","spectral channel attention","adaptive operator learning","vision transformer","concept drift","multispectral imagery"],"falsifier":"Inspect the operator-selection gate's outputs for every county in the test set. If the gate selects the same operator for every county within a given crop and year (i.e., the choice does not vary across inputs), then AOL-Net is effectively a validation-chosen fixed operator, and a fair comparison with the best fixed operator chosen on validation would reproduce the results. A model whose selected operator actually varies with input features would provide evidence for the claimed dynamic mechanism.","tokens_in":21371,"feed_emoji":"🌾","tokens_out":7654,"duration_ms":72763,"temperature":0.7,"pith_summary":"The paper aims to show that a single architecture can outperform specialized baselines for crop yield prediction across different satellite resolutions, crop types, and years by making three parts of the model adaptive: spectral channel weighting, edge-operator choice, and the fusion of local and global features. On the coarse-resolution MODIS dataset it reports the best RMSE, MAE, and R² in six of seven test years, and on the high-resolution Sentinel-2 dataset it reports the best RMSE and MAE for all four crops, with the best R² except for cotton where MMST-ViT is marginally ahead. The central claim is that fixed handcrafted edge operators (Sobel, Scharr, etc.) and static fusion strategies leave performance on the table, and that learning to select operators and to weight branches dynamically yields more stable, transferable predictions. If correct, this would offer a practical, end-to-end trainable alternative to dataset-specific agricultural monitoring pipelines.","feed_headline":"Adaptive fusion model tops crop-yield baselines on both satellites","feed_subtitle":"On MODIS it wins six of seven years; on Sentinel-2 it posts best RMSE and MAE for all four crops.","key_machinery":"The central object is the adaptive operator pool in AOL-Net: a library of Sobel, Scharr, and a learnable kernel initialized as a convex combination of the two, with a selection gate that picks exactly one operator per time step to compute an edge-enhanced map that is blended into the CNN input via a learnable scaling. RCA modifies the Squeeze-and-Excitation channel attention to use max pooling for low-resolution inputs and average pooling for high-resolution inputs, so that spectral reweighting depends on resolution. The third mechanism is the fusion function $f_\\theta(X) = \\alpha f_{\\theta_1}(X) + \\beta f_{\\theta_2}(X)$ with $\\alpha, \\beta \\in (0,1)$ learned by backpropagation; the paper deliberately relaxes the constraint $\\alpha+\\beta=1$ to let the model rescale branch contributions. All three are trained jointly by MSE backpropagation through the final fused prediction.","core_discovery":"DFYP's core discovery is that the combination of three adaptive mechanisms—resolution-aware channel attention, an operator-selection gate that picks among Sobel, Scharr, and a learnable kernel, and a loss-guided weighted fusion of a CNN branch and a Vision Transformer branch—produces consistently lower RMSE and MAE and higher R² than nine existing deep-learning baselines on two benchmark datasets. The paper argues that no single classical edge operator is best across years and crops, which it supports with experiments on eight fixed operators, and that a learnable fusion weight (not constrained to sum to one) stabilizes performance where concatenation or fixed averaging fluctuates. The claimed result is a framework that generalizes across coarse (MODIS) and fine (Sentinel-2) imagery and across soybean, corn, cotton, and winter wheat, with ablations showing each component contributes to the final gain.","pith_inferences":["The unnormalized fusion weights ($\\alpha$ and $\\beta$ free in (0,1)) mean the model can rescale its own output; a fair ablation that normalizes them ($\\alpha+\\beta=1$) would reveal whether the reported robustness comes from the dynamic weighting or simply from extra scale flexibility.","The operator-selection gate in the paper is specified as hard binary selection but trained as a continuous interpolation in Eq. (3); a testable extension is a Gumbel-Softmax relaxation, which would make the whole selection genuinely differentiable and could be compared against the current validation-score-based gate.","The paper's empirical claim that fixed edge operators fluctuate with crop and year implies that other remote sensing regression tasks (e.g., biomass, evapotranspiration) may benefit from the same adaptive-operator plugin, a hypothesis that can be tested directly on those tasks.","Because the gate is described as selecting per time step, one could examine whether the chosen operator correlates with identifiable conditions (image date, phenological stage, drought index); if such correlations exist, the module could double as an interpretability tool for when edge emphasis matters."],"forward_implications":["A single DFYP-style architecture can replace dataset-specific handcrafted pipelines for county-level yield forecasting across different satellite sensors.","Because the model uses only pre-season and in-season observations, it is suitable for operational forecasts issued before harvest.","The dynamic fusion and operator selection reduce year-to-year error fluctuations, indicating a path toward models that tolerate concept drift in agricultural remote sensing.","The AOL ablation suggests that adaptive edge processing contributes more on high-resolution imagery, while ViT global modeling matters most on low-resolution imagery, guiding where to put architectural capacity.","If replicated, the reported gains would make DFYP a strong default baseline for future crop yield prediction research on the same datasets."],"supporting_citations":[{"why":"Provides the Sentinel-2 multi-crop dataset (corn, cotton, soybean, winter wheat) used for the high-resolution experiments.","marker":"[75]"},{"why":"MMST-ViT, the strongest multimodal Vision Transformer baseline and the only method that edges out DFYP on cotton R².","marker":"[44]"},{"why":"DeepYield/DeepField, a CNN-LSTM baseline that DFYP must beat and whose lack of adaptive fusion is used to motivate the design.","marker":"[47]"},{"why":"Squeeze-and-Excitation network, the basis for the RCA module's channel attention mechanism.","marker":"[74]"},{"why":"Survey of hybrid CNN-ViT architectures that supplies the taxonomy of fusion strategies and the alternative fusion baselines compared in Section IV.D.","marker":"[38]"},{"why":"GNN-RNN baseline for spatial-temporal crop yield prediction that DFYP is compared against.","marker":"[30]"},{"why":"Deep Gaussian Process models (CNN+GP, LSTM+GP) used as baselines and as examples of methods without adaptive fusion.","marker":"[8]"},{"why":"CNN+LSTM baseline from UAV time series that DFYP outperforms, also used in the MODIS year-by-year comparison.","marker":"[71]"},{"why":"3DCNN+ConvLSTM baseline that DFYP is compared against on both datasets.","marker":"[72]"}],"fun_headline_variants":["Adaptive fusion framework DFYP beats nine crop-yield baselines","Dynamic operator selection boosts yield forecasts on both satellites","DFYP: Learnable fusion of CNN and ViT wins on crop yield","Spectral channel attention plus adaptive operators improve yield prediction","Cross-crop, cross-resolution yield model tops MODIS and Sentinel-2"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the adaptive operator selection is learned end-to-end by backpropagation, even though the paper specifies a hard, non-differentiable binary selection rule alongside a continuous backprop-trained interpolation and a validation-accuracy-based gate, and never reconciles these three formulations.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive fusion framework DFYP beats nine crop-yield baselines","Dynamic operator selection boosts yield forecasts on both satellites","DFYP: Learnable fusion of CNN and ViT wins on crop yield","Spectral channel attention plus adaptive operators improve yield prediction","Cross-crop, cross-resolution yield model tops MODIS and Sentinel-2"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000205,"raw_usage":{"total_tokens":1404,"prompt_tokens":969,"completion_tokens":435,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":347}},"tokens_in":585,"tokens_out":435,"duration_ms":4871,"temperature":1.0,"reasoning_tokens":347,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:16:56.956751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the operator-selection gate's outputs for every county in the test set. If the gate selects the same operator for every county within a given crop and year (i.e., the choice does not vary across inputs), then AOL-Net is effectively a validation-chosen fixed operator, and a fair comparison with the best fixed operator chosen on validation would reproduce the results. A model whose selected operator actually varies with input features would provide evidence for the claimed dynamic mechanism.","supporting_citations":[{"cited_title":"An open and large-scale dataset for multi-modal climate change-aware crop yield predictions,","cited_arxiv_id":null,"evidence_quote":"Provides the Sentinel-2 multi-crop dataset (corn, cotton, soybean, winter wheat) used for the high-resolution experiments."},{"cited_title":"Mmst-vit: Climate change- aware crop yield prediction via multi-modal spatial-temporal vision transformer,","cited_arxiv_id":null,"evidence_quote":"MMST-ViT, the strongest multimodal Vision Transformer baseline and the only method that edges out DFYP on cotton R²."},{"cited_title":"Deepyield: A com- bined convolutional neural network with long short-term memory for crop yield forecasting,","cited_arxiv_id":null,"evidence_quote":"DeepYield/DeepField, a CNN-LSTM baseline that DFYP must beat and whose lack of adaptive fusion is used to motivate the design."},{"cited_title":"Exploring the synergies of hybrid cnns and vits architectures for computer vision: A survey,","cited_arxiv_id":null,"evidence_quote":"Survey of hybrid CNN-ViT architectures that supplies the taxonomy of fusion strategies and the alternative fusion baselines compared in Section IV.D."},{"cited_title":"A gnn-rnn approach for harnessing geospatial and temporal information: applica- tion to crop yield prediction,","cited_arxiv_id":null,"evidence_quote":"GNN-RNN baseline for spatial-temporal crop yield prediction that DFYP is compared against."},{"cited_title":"Deep gaussian process for crop yield prediction based on remote sensing data,","cited_arxiv_id":null,"evidence_quote":"Deep Gaussian Process models (CNN+GP, LSTM+GP) used as baselines and as examples of methods without adaptive fusion."},{"cited_title":"Crop yield prediction using multitemporal uav data and spatio-temporal deep learning models,","cited_arxiv_id":null,"evidence_quote":"CNN+LSTM baseline from UAV time series that DFYP outperforms, also used in the MODIS year-by-year comparison."},{"cited_title":"Multispectral crop yield prediction using 3dconvolutional neural networks and attention convolutional lstm approaches,","cited_arxiv_id":null,"evidence_quote":"3DCNN+ConvLSTM baseline that DFYP is compared against on both datasets."}],"review_version":1}