{"id":"7944ef99-f6a9-41e8-9316-d8f87e849c55","arxiv_id":"2501.01132","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Training multi-view EO models on all combinations of missing views with dynamic fusion improves robustness to moderate missingness, but does not consistently improve full-view accuracy.","lead":"This paper trains multi-view Earth observation models on every possible combination of missing satellite or sensor views, so the model learns to make predictions from whatever data is available. The method improves robustness when one view is missing, though full-view gains are inconsistent across datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported robustness gains may not transfer to real EO missingness because evaluation only drops entire views at inference from full-view-trained models.","rationale":"The reader's weakest assumption correctly identifies the evaluation protocol as the point where the central claim is least secure. I agree that random whole-view dropout at inference from a full-view-trained model is not representative of real EO missingness, which is often partial, temporally correlated, and present in training data. This is the most load-bearing concern because it challenges the external validity of the entire empirical contribution: even if all tables are internally consistent, the claim that the method 'offers a single adaptive solution' in EO requires evidence under realistic missingness. The paper's own limitation statement (Sec. 5) concedes that only inference-time missingness is assessed. A concrete partial-missingness experiment would settle whether the robustness gain survives when views are degraded rather than absent. Since this concern is exactly the reader's flagged assumption and is addressable by an additional experiment, the conditional verdict remains appropriate; no change to the reader's verdict is needed.","tokens_in":24884,"tokens_out":10882,"duration_ms":106443,"concrete_test":"Re-run the main comparison (Tables 1 and 2) on CropHarvest and LFMC with a partial-missingness protocol: for each sample, do not drop the entire optical view; instead randomly mask 50% of the temporal steps within the optical view to simulate cloud cover, feed the available steps to the encoder (with a missingness indicator or per-step imputation), and use the published fusion methods. If FCoM-av/FCoM-ga no longer outperform ITempD-co and ISensD-co under moderate missingness, the reported robustness is specific to whole-view dropout and does not extend to the partial, correlated missingness that dominates real EO data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim hinges on the evaluation protocol of Sec. 4.2/4.3: models are trained on full-view data and tested by randomly dropping entire views at inference. This does not match how missingness actually occurs in EO. Cloud cover produces partial, temporally correlated gaps within a view (e.g., some months of an optical time series are cloudy), and missing views can also be present in the training distribution due to sensor failures or historical data availability. Under the paper's protocol, the model only needs to learn to ignore whole views; it never has to handle a partially corrupted view that is still present. The authors explicitly acknowledge this limitation in Sec. 5 ('we assess the effect of missing views only at inference, assuming a full-view training dataset'), but the abstract's claim of 'a single adaptive solution to operate effectively with any combination of available views' requires that the method work under realistic missingness, including partial within-view gaps and train-time missingness. If the simulated protocol is not representative, the reported robustness improvements may be an artifact of the evaluation setup rather than a property of CoM + dynamic fusion that transfers to operational EO settings.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-augmentation strategy called Combinations of Missing views (CoM) for multi-view learning (MVL) in the Earth Observation (EO) domain. Instead of imputing missing views, the method trains a feature-level fusion model on all non-empty subsets of views, and uses dynamic merge functions (average, gated fusion, cross-attention, memory fusion) whose output dimension is independent of the number of available views. The authors evaluate on four EO datasets (CropHarvest binary and multi-class classification, LFMC and PM25 regression), comparing against several baselines under simulated missing-view scenarios at inference. They report improved robustness under moderate missingness and, in some cases, improved full-view predictive performance.","tokens_in":25153,"tokens_out":9638,"duration_ms":82920,"significance":"If the results hold, the proposed CoM plus dynamic fusion provides a practical single-model solution that can adapt to arbitrary combinations of available views in EO time series, a setting where sensor failures and cloud cover are common. The paper's strengths include a broad empirical comparison across four datasets, both classification and regression tasks, ablations of augmentation levels and fusion architectures (Table 3 and Appendix B.2), and an execution-time comparison. However, the central novelty relative to the authors' prior work is incremental, and the strength of the claims about full-view improvements and the evaluation protocol need qualification. The moderate-missingness robustness claims are reasonably supported for CropH and LFMC by Tables 1 and 2, but the abstract overstates full-view gains and the evaluation does not cover partial within-view missingness.","major_comments":[{"comment":"The abstract states that the methods 'improve the predictive performance when all views are present' without qualification. Table 2 contradicts this for PM25, where all four FCoM variants (best R2=0.660) are below ITempD-co (R2=0.866), and for LFMC only FCoM-ga (0.700) exceeds ITempD-co (0.691). The introduction correctly says 'in some cases', and the abstract should be reworded to match the evidence.","section":"Abstract / Table 2"},{"comment":"The ITempD-co row contains suspicious duplicated values: in LFMC, 'Missing Radar' and 'Missing Optical' are both 0.036±0.100, and 'Only available Optical' and 'Only available Radar' are both -0.036±0.06; in PM25, 'Missing Dynamic' and 'Only available Condition' are both -0.124±0.14, and 'Missing Condition' and 'Only available Dynamic' are effectively identical (0.074±0.035 vs 0.073±0.035). These coincidences are unlikely to be genuine and suggest a data handling or reporting error. Since Table 2 underpins the claim that ITempD-co has poor regression robustness, the authors should verify the baseline numbers and correct the table or explain the duplication.","section":"Table 2"},{"comment":"The gated fusion is described as a dynamic merge function that 'ignores' missing views, but the implementation zero-imputes missing features and computes the softmax weights over the full view set before masking the missing views' weights to zero. Because the softmax normalization includes the logits of the missing views, the fused vector's scale depends on the number of available views even when the missing features are zero. Thus the missing views are not fully ignored; they affect the normalization and hence the available views' effective weights. The authors should renormalize over the available views only, or explicitly state that the scale dependence on m(i) is intended and characterize its effect.","section":"Section 3.3, Eqs. (4)-(5)"},{"comment":"The evaluation protocol, as stated in Sec 4.2, simulates missingness only by dropping entire views at inference from models trained on full-view data. The abstract's claim of 'a single adaptive solution to operate effectively with any combination of available views' is therefore only supported for whole-view dropout, not for the partial within-view missingness (e.g., cloud-corrupted optical time series with some valid dates) or train-time missingness that commonly occur in EO. Although Sec 5 acknowledges this limitation, the abstract and the final conclusions present the result without that qualifier. Either add experiments with partial missingness or narrow the claims accordingly.","section":"Abstract / Sections 4.2-4.3 and 5"}],"minor_comments":[{"comment":"The paper does not report paired significance tests or effect sizes. Many of the 'best' results in Table 1 overlap with baselines within one standard deviation (e.g., CropH-b full-view, FCoM-ga 0.839±0.005 vs FCoMl-co 0.832±0.007). Consider adding a paired test such as Wilcoxon signed-rank across folds, or at least discussing significance.","section":"General / Tables 1-2"},{"comment":"The performance curves in Figure 2 and Figure 7 are plotted without error bars or confidence intervals, making it difficult to assess the reliability of differences between methods.","section":"Figure 2"},{"comment":"The selection of the two top views is based on individually trained models. Please clarify that this selection is performed within each cross-validation fold (or otherwise state why it does not leak test information).","section":"Section 4.3"},{"comment":"The table headers read '(4/4) No' and '(6/6) No'; the word 'Missing' is missing. Please correct.","section":"Table 1 / Table 2"},{"comment":"The paper states the code 'will be released' and provides a GitHub URL. Since reproducibility is a key strength, please make the code available at review time or include complete pseudocode for the masking step in the gated fusion.","section":"Section 4.2 / Data and code availability"}],"recommendation":"major_revision","confidential_remarks":"The paper builds heavily on the authors' own prior work (Mena et al. 2024a, 2024b, 2024c, 2024d), and the incremental novelty of feature-level CoM with dynamic fusion should be clearly positioned. The duplicated baseline values in Table 2 and the abstract's overstatement of full-view gains are the most pressing concerns, but they are fixable by revising text and re-running the affected baseline numbers. The gated-fusion normalization issue is a technical point that should also be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, incremental paper. The idea is to train multi-view models with all non-empty subsets of views at feature level, using dynamic merge functions (average, gated, cross-attention, memory fusion) that naturally ignore missing views. The combination of all-combination masking with feature-level dynamic fusion, plus the comparison across four EO datasets, is genuinely new relative to the cited work. The moderate-missingness robustness claim is supported by the tables, especially FCoM-av on CropH and LFMC. The design is also computationally sensible: encode each view once, then run the cheap fusion multiple times.\n\nThe main soft spots are three. First, the abstract's 'improve the predictive performance when all views are present' is overstated. On PM25, FCoM-av (0.660) and FCoM-ga (0.046) are below ITempD-co (0.866) in the full-view case. The robust-accuracy tradeoff is real and should be stated as such. Second, Table 2 appears to have a copy-paste error: ITempD-co shows identical values (0.036±0.100) for both 'missing radar' and 'missing optical' in LFMC, and similarly -0.036±0.06 for the single-view cases. That needs a fix before publication. Third, the evaluation only drops entire views at inference from full-view-trained models. The stress-test note and the authors' own Sec. 5 acknowledge this. Real EO missingness is often partial (cloudy months within an optical time series) and can also appear in training data. That is a genuine limitation of the protocol, but it is also the standard protocol in this literature, so I would not over-weight it. The authors are upfront about it.\n\nOne more issue: the code availability statement promises release, but the reviewer at this point can't check it. Not a reason to reject, but the authors should actually ship it.\n\nOn the citation pattern: heavy self-citation, but the cited results (their prior work on dynamic fusion and SimSensD) are directly relevant. Not a problem.\n\nBottom line: worth a serious referee. The corrected paper will be a useful reference for people working on missing sensors in EO and multi-view robustness generally. Send it to review; the revisions are mechanical (data error, abstract rewrite) and one substantive discussion (limitation of the missingness simulation).","headline":"Solid, incremental robustness trick for multi-view EO models; the abstract oversells full-view gains, but the moderate-missingness claim holds.","tokens_in":25623,"tokens_out":2238,"would_cite":true,"duration_ms":20847,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Training a multi-view Earth-observation model on every non-empty combination of available views, fused with a dynamic merge function that ignores missing views, makes predictions robust to missing sensors and can slightly improve…","keywords":["multi-view learning","earth observation","missing data","data augmentation","feature-level fusion","robustness","dynamic merge functions","satellite time series"],"falsifier":"Run the same CoM models on the same datasets but simulate missingness from cloud masks or actual sensor outages (such as the Sentinel-1B failure window) instead of random whole-view dropping, and test both full-view-trained and partially-missing-trained scenarios; if the robustness advantage of FCoM methods over imputation baselines disappears or reverses under these realistic masks, the central claim is not general. A controlled comparison that applies missingness to random time steps within a view rather than the whole view, at matched overall missing fractions, would also settle whether the benefit comes from ignoring views or from the augmentation effect.","tokens_in":24701,"feed_emoji":"🛰️","tokens_out":4353,"duration_ms":38202,"temperature":0.7,"pith_summary":"This paper claims that training a multi-view Earth-observation model on every non-empty combination of available views, and fusing encoded features with a merge function whose output size is independent of how many views are present, makes the model robust to missing views without imputation. The authors test this on four EO datasets with temporal and static views, comparing average, gated, cross-attention, and memory fusions against five baseline methods that impute missing data or use random sensor dropout. Their central result is that under moderate missingness (one view missing), the CoM-based methods keep predictive performance nearly intact and beat the baselines, and in some classification cases they also improve full-view accuracy. A sympathetic reader would care because it points toward a single adaptive model that works with any subset of available sensors, avoiding the need for separate imputation pipelines.","feed_headline":"Training on every missing-view combo hardens EO models","feed_subtitle":"Multi-view Earth-observation models that fuse only available views keep accuracy when sensors fail.","key_machinery":"The central object is the CoM (Combinations of Missing views) augmentation: from the power set of the view set minus the empty set, each training sample generates 2^m - 1 prediction tasks, one for each combination of available views. The complementary mechanism is the dynamic merge function, a fusion that maps any number of encoded view representations to the same dimension d; the paper uses average, gated softmax-weighted fusion, cross-attention with a learnable fusion token, and LSTM memory fusion. Together they allow the model to ignore missing views at inference without any imputation or forced view alignment.","core_discovery":"The paper introduces Combinations of Missing views (CoM), a data-augmentation method that, at the feature level, exposes the model to every non-empty subset of views during training, and combines it with a dynamic merge function that produces a fixed-size fused representation regardless of the number of available views, so missing views are ignored rather than imputed. Across the four datasets, the methods FCoM-av, FCoM-ga, FCoM-cr, and FCoM-me improve robustness to moderate missingness and, in some cases, exceed the full-view performance of the baselines. The authors conclude that integrating CoM with a merge function that ignores the missing views enhances both the predictions and the robustness of multi-view learning models, delivering a single adaptive solution that operates with any combination of available views.","pith_inferences":["The simulated missingness drops entire views wholesale, whereas real EO failures often corrupt patches, orbits, or time steps within a view; a natural test is whether CoM retains its advantage when missingness is partial within a view and correlated with cloud cover or season.","Because CoM operates at the feature level, it can be layered on top of pre-trained encoders, suggesting a cheap way to retrofit existing multi-view models for robustness without retraining from scratch.","The paper's observation that simple average fusion is a strong baseline hints that most of the benefit comes from the CoM augmentation itself, and that complex fusions add robustness mainly in extreme missingness; a cleaner ablation would isolate these two contributions.","The relative-robustness metric PRS favors a simpler ensemble baseline (ESensI-av), while absolute performance favors CoM methods; a combined metric that mixes both level and robustness would make the practical choice clearer for a practitioner."],"forward_implications":["A single model can serve any view-availability scenario, eliminating the need for separate missing-data imputation pipelines.","Training cost stays moderate because the encoders are forwarded once per sample, while the fusion and prediction head process all 2^m - 1 combinations.","Classification tasks gain a small full-view performance boost from the augmentation effect, while regression tasks gain robustness mainly under moderate missingness.","Prediction time shrinks as more views are missing, since the model only processes the available data rather than inserting fake values."],"supporting_citations":[{"why":"Introduces the idea of training on all combinations of missing views at the feature level with a concatenation-based merge, which the paper extends to dynamic merge functions and multiple views.","marker":"[Gawlikowski et al., 2023]"},{"why":"Shows that dropping sensors during training (SensD) and ignoring missing views at the feature level improves robustness; forms the direct baseline and motivation for CoM at feature level.","marker":"[Mena et al., 2024b]"},{"why":"Quantifies the negative predictive impact of missing views and identifies dynamic fusion that ignores missing views as the most robust mitigation strategy, which the paper builds on.","marker":"[Mena et al., 2024a]"},{"why":"Presents Temporal Dropout (TempD) as a missing-data augmentation technique for satellite time series, used as a key competing baseline.","marker":"[Sainte Fare Garnot et al., 2022]"},{"why":"Adapts cross-attention fusion for incomplete multimodal remote sensing data with sensor dropout, providing the basis for the FSensD-cr baseline and the cross-attention merge design.","marker":"[Chen et al., 2024]"},{"why":"Uses random view dropout in a Transformer-based EO model (Presto) to increase robustness, which the paper cites as motivation and as a related augmentation approach.","marker":"[Tseng et al., 2023]"},{"why":"Provides the gated fusion mechanism (EmbraceNet) that the paper adapts to per-dimension softmax weighting for the FCoM-ga merge function.","marker":"[Choi and Lee, 2019]"},{"why":"Supplies the permutation-invariance and set-aggregation perspective that underlies the dynamic, order-insensitive merge functions used with CoM.","marker":"[Lee et al., 2019]"}],"fun_headline_variants":["Missing-view combos become training data for EO models","Train on every missing-view combo to harden EO models","Dynamic merges ignore missing views and boost EO models","Simulating missing-view combos yields robust EO fusion","All missing-view combos as augmentation for multi-view EO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes that randomly dropping entire views at inference, from a model trained on full-view data, faithfully represents real Earth-observation missingness, which is often partial within a view, correlated with cloud cover or season, and may already be present in training data.","fun_headline_variants_meta":{"raw":{"variants":["Missing-view combos become training data for EO models","Train on every missing-view combo to harden EO models","Dynamic merges ignore missing views and boost EO models","Simulating missing-view combos yields robust EO fusion","All missing-view combos as augmentation for multi-view EO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000498,"raw_usage":{"total_tokens":2423,"prompt_tokens":909,"completion_tokens":1514,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":1436}},"tokens_in":525,"tokens_out":1514,"duration_ms":10453,"temperature":1.0,"reasoning_tokens":1436,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:34:11.234560+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same CoM models on the same datasets but simulate missingness from cloud masks or actual sensor outages (such as the Sentinel-1B failure window) instead of random whole-view dropping, and test both full-view-trained and partially-missing-trained scenarios; if the robustness advantage of FCoM methods over imputation baselines disappears or reverses under these realistic masks, the central claim is not general. A controlled comparison that applies missingness to random time steps within a view rather than the whole view, at matched overall missing fractions, would also settle whether the benefit comes from ignoring views or from the augmentation effect.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the idea of training on all combinations of missing views at the feature level with a concatenation-based merge, which the paper extends to dynamic merge functions and multiple views."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Presents Temporal Dropout (TempD) as a missing-data augmentation technique for satellite time series, used as a key competing baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Adapts cross-attention fusion for incomplete multimodal remote sensing data with sensor dropout, providing the basis for the FSensD-cr baseline and the cross-attention merge design."},{"cited_title":"and Lee, J.-S","cited_arxiv_id":null,"evidence_quote":"Provides the gated fusion mechanism (EmbraceNet) that the paper adapts to per-dimension softmax weighting for the FCoM-ga merge function."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the permutation-invariance and set-aggregation perspective that underlies the dynamic, order-insensitive merge functions used with CoM."}],"review_version":1}