{"id":"9f271a16-799a-4d90-a2ed-cb127e249574","arxiv_id":"2607.07834","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"low","formal_verification":"none","parameter_count":8,"one_line_summary":"Ensemble models using only satellite SST, upwelling, chlorophyll-a and PFTs predict Pseudo-nitzschia HAB occurrence at ROC–AUC 0.77±0.06 under year×cluster cross-validation on the Portuguese L1–L2 coast.","lead":"Satellite-only machine learning predicts Pseudo-nitzschia bloom occurrence on Portugal’s northern coast with ROC–AUC about 0.77 under strict year-and-cluster holdout. The work offers a realistic early-warning template for shellfish zones in eastern-boundary upwelling systems.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged threshold proxy.","rationale":"The strongest claim is an empirical performance statement under a leakage-resistant protocol, not a claim that the model predicts toxicity or replaces IPMA monitoring. The manuscript supplies the necessary caveats, open-code pointers, and conservative AUC language. The fixed 10k threshold is the softest premise for translating occurrence skill into management action, exactly as the reader notes, but it is not an internal inconsistency that collapses the reported numbers. A threshold-sensitivity re-run is the cleanest check; absent a large AUC collapse, the CONDITIONAL verdict with high confidence remains appropriate.","tokens_in":23328,"tokens_out":454,"duration_ms":5360,"concrete_test":"Re-label HAB with two alternative thresholds (e.g. 5,000 and 20,000 cells L⁻¹) and recompute the Extra Trees Env+Bio ROC–AUC under the same 66 year∪cluster folds; if AUC remains in the 0.70–0.80 band the occurrence claim is robust to threshold choice, while a large drop would tighten the operational caveat.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is carefully scoped: under year∪ cluster CV on L1–L2, HAB occurrence (PN > 10,000 cells L⁻¹) is moderately predictable from satellite-only predictors (Extra Trees 0.77±0.06). The paper itself repeatedly states that toxicity and bans can occur at low cell counts and that toxicity is not the target (Introduction; §2.4; Discussion 4.3; Conclusion limitations). The fixed threshold is therefore a definitional choice for occurrence, not a hidden claim that cell count equals public-health risk. No circular target construction, no leakage in the stated CV design, and no derivation failure undermines the reported discrimination numbers. The reader's weakest_assumption is real for operational uptake but is already disclosed and does not falsify the empirical claim as written.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript develops a satellite-only machine-learning framework to predict Pseudo-nitzschia HAB occurrence (PN > 10,000 cells L⁻¹) along the northern Portuguese coast (zones L1–L2, 2013–2023; 1,440 observations). Predictors include SST, a Bakun-style upwelling index, chlorophyll-a, plankton functional types, seasonal harmonics, spatial coordinates, and lags up to 45 days. Sampling sites are partitioned by a river-aware (ROFI) clustering scheme, and models are evaluated under a stringent year∪cluster cross-validation (66 folds) that withholds entire years and spatial clusters simultaneously. Across linear, margin-based, and tree ensembles, HAB occurrence is moderately predictable: Random Forest reaches ROC–AUC 0.74±0.05 with environmental features; Extra Trees reaches 0.77±0.06 when biological predictors are added. Feature importance attributes skill primarily to seasonality, location, and multi-week SST/UI lags, with short-lag Chl-a refining predictions. Confusion matrices at fixed FPRs and spatial probability maps illustrate operational threshold trade-offs and adaptive sampling use.","tokens_in":23602,"tokens_out":1382,"duration_ms":17488,"significance":"If the reported discrimination holds under the stated protocol, the work provides a carefully scoped, leakage-resistant demonstration that satellite-only predictors can support early-warning of Pseudo-nitzschia bloom occurrence on an eastern-boundary upwelling coast. Strengths include strict separation of IPMA labels from CMEMS predictors, multi-model and nested feature-set comparison, fold-wise mean±std over 66 year×cluster splits, open code for CMEMS matchup and ROFI clustering, and explicit operational framing (threshold trade-offs, adaptive sampling maps). The contribution is methodological and regional rather than a universal HAB solution, but it is a credible baseline for satellite-supported monitoring support where in situ coverage is sparse.","major_comments":[{"comment":"§2.4 and Table 1: The supervised target is a fixed threshold HAB = 1 if PN > 10,000 cells L⁻¹ (~14% positives). The paper repeatedly notes that DA toxicity and harvest bans can occur at low cell abundances and that toxicity is not the modelling target (Introduction; §2.4; Discussion 4.3; Conclusion limitations). That disclosure is appropriate, but the operational claim in the abstract and highlights (“operationally relevant skill for … early-warning systems”) still rests on occurrence as a risk proxy. A sensitivity analysis over alternative thresholds (e.g. 5×10³, 2×10⁴, or a percentile-based definition) and/or a secondary analysis against ban-day labels would show whether the reported AUC range is robust to the definitional choice that is load-bearing for uptake.","section":null},{"comment":"§2.7 and Figure 10: The year∪cluster blocking is a genuine strength, but the manuscript does not quantify residual leakage risk from (i) multi-week lagged predictors that may still share large-scale forcing with held-out years/clusters, or (ii) the fact that a held-out cluster’s other years remain in training (and a held-out year’s other clusters remain in training). Reporting a pure leave-one-year-out and pure leave-one-cluster-out ablation alongside the union design would make the “realistic forecasting” claim more transparent and would show how much of the 0.74–0.77 AUC is carried by each blocking axis.","section":null},{"comment":"§2.7.1 / Table 2: Hyperparameters for the winning RF (env-only) and Extra Trees (full) configurations are stated as selected within training folds, but the selection procedure (grid vs. default, search space, whether nested CV was used) is not fully specified, and only the final settings are given. Given class imbalance and 66 folds, a brief nested-CV or fixed-hyperparameter sensitivity check is needed to confirm that the 0.03 AUC gain from adding biological predictors is not partly an artefact of unequal tuning effort across feature sets.","section":null}],"minor_comments":[{"comment":"Appendix A vs. main text §2.1/Figure 7: Main text refers to “river-aware K-means” and K=6, while Appendix A primarily describes DBSCAN with an augmented distance and notes k-means as an alternative. Please state unambiguously which algorithm produced the six clusters used for CV blocking.","section":null},{"comment":"Table 1: “Remote-sensed biological uncertainty estimates” list CHL and PFT uncertainty ranges but it is unclear whether these uncertainty fields enter any of the three model configurations or are only descriptive. Clarify in §2.3/§2.6.","section":null},{"comment":"Figure 9 caption and §2.5: Toxicity-related ban periods are highlighted, but ban labels are not used as a modelling target. A short note in the figure caption that bans are shown for context only would avoid reader confusion.","section":null},{"comment":"§2.3 / Appendix B: Coastline angle defaults (−32°) and zone-specific adjustments for L1/L2 are mentioned; give the numerical angles used for L1 and L2 so the UI is fully reproducible from the text.","section":null},{"comment":"Typographical/consistency: “Pseudo-nitzschia” is occasionally concatenated without space or italics in headings; “Downwelling/Upwelling Index” units and sign convention should be stated once in the main text as well as Appendix B.","section":null},{"comment":"References: Several recent HAB ML and Portuguese Pseudo-nitzschia monitoring papers are cited appropriately; ensure all arXiv/software DOIs (CMEMS_Data_Analysis, ROFI_Clustering) remain accessible at publication.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The central empirical claim is carefully scoped and the leakage-resistant design is above average for this literature. The main residual risk for the journal is over-reading of “operational early warning” when the target is occurrence, not toxicity; the authors already disclose this, so minor revision with a threshold-sensitivity check should suffice. Scope is a good fit for an applied ML / environmental informatics venue; less so for a pure oceanography journal unless the ROFI clustering and UI derivation are foregrounded more as physical contributions."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that this is a careful satellite-only HAB-occurrence study on the Portuguese L1–L2 hotspot, not a toxin model and not a claim of breakthrough accuracy. Under year∪ cluster blocking they get Extra Trees at 0.77±0.06 and RF at 0.74±0.05. Those numbers are moderate and they say so.\n\nWhat is actually new is the decade-long IPMA–CMEMS matchup for this coast, the river-aware ROFI clustering used as spatial blocks, and the simultaneous year-and-cluster hold-out that kills the usual leakage. Nested feature sets, multi-model comparison, fixed-FPR confusion matrices, and out-of-fold importance are all done cleanly. Predictors and labels are strictly separated. Code and data pointers are there. That is real craft for an applied coastal ML paper.\n\nSoft spots are real but already mostly disclosed. The fixed 10k cells L⁻¹ threshold is a definitional choice for occurrence, not a hidden claim that cell count equals public-health risk; the paper itself flags that DA and bans can occur at low abundance and that toxicity is out of scope. Coastal ocean-colour uncertainty for PFTs is acknowledged. Scope is L1–L2 only. None of that falsifies the reported discrimination numbers. Free parameters (K, lag depth, α, coastline angle, tree hyperparameters) are standard for this genre and not hidden.\n\nMath and citation pattern look fine: Bakun UI, seasonal harmonics, and the C-HARM / Monterey / Florida / Yellow Sea comparisons are the right literature. No circular target construction.\n\nThis is for people who care about operational HAB triage on upwelling coasts and for anyone who wants a template for leakage-resistant coastal CV. It deserves a serious referee. I would engage with it, cite the validation design and the Portuguese numbers if I work in this area, and send it to review rather than desk-reject.","headline":"Solid applied ML paper: honest moderate AUCs under a real year∪ cluster CV, with the main novelty in the Portuguese matchup and river-aware blocking rather than the model class.","tokens_in":24214,"tokens_out":522,"would_cite":true,"duration_ms":6467,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Satellite-only models predict Pseudo-nitzschia bloom occurrence on the Portuguese coast with ROC-AUC up to 0.77 under strict year-and-cluster holdout.","keywords":["Pseudo-nitzschia","harmful algal blooms","remote sensing","machine learning","spatio-temporal forecasting","Portuguese Atlantic coast","upwelling","satellite predictors"],"falsifier":"Re-run the same year-and-cluster cross-validation after re-labelling HAB events from matched shellfish domoic-acid measurements (or actual harvest-ban days) instead of the fixed cell threshold; if discrimination collapses well below the reported 0.7 range, the occurrence proxy does not support operational early warning.","tokens_in":24233,"feed_emoji":"🌊","tokens_out":747,"duration_ms":7387,"temperature":0.7,"pith_summary":"The paper asks whether Pseudo-nitzschia harmful algal blooms along the northern Portuguese coast can be anticipated from satellites alone, without any local water samples as inputs, under evaluation rules that match real forecasting. Using a decade of monitoring matched to sea-surface temperature, an upwelling index, chlorophyll-a, and plankton functional types, the authors show moderate skill: tree ensembles reach ROC-AUC about 0.74 with physical predictors and about 0.77 when biological satellite products are added. The evaluation withholds entire years and entire river-influenced coastal clusters at once, so the numbers measure generalisation to new seasons and new shoreline segments rather than memorisation of nearby samples. Feature rankings put seasonal timing, location, and multi-week temperature and upwelling history first, with recent chlorophyll refining risk inside those windows. A sympathetic reader cares because shellfish closures carry real cost and cell counts alone do not always flag toxin risk; a remote early-warning layer could focus limited field sampling where it matters most.","feed_headline":"Satellites alone flag Pseudo-nitzschia blooms at ROC-AUC 0.77","feed_subtitle":"Strict year-and-cluster holdout on the Portuguese coast shows operational early-warning skill without local samples.","key_machinery":"Union-of-year-and-cluster blocking combined with river-aware spatial clustering of sampling sites: each validation fold holds out every observation from one calendar year together with every observation from one freshwater-influence cluster, so lagged satellite predictors cannot leak information from the held-out regime.","core_discovery":"Under a leakage-resistant spatio-temporal cross-validation that simultaneously withholds calendar years and river-aware spatial clusters, Pseudo-nitzschia HAB occurrence (defined as cell concentration above 10,000 cells per litre) on the L1-L2 Portuguese hotspot is moderately predictable from exclusively satellite-derived predictors. Extra Trees with environmental plus biological features reach ROC-AUC 0.77 plus or minus 0.06; Random Forest with environmental features alone reaches 0.74 plus or minus 0.05. Seasonal harmonics, spatial coordinates, and lagged physical forcing dominate, while short-lag chlorophyll and plankton functional types refine likelihood within favourable windows.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Satellites predict Pseudo-nitzschia HABs at 0.77 ROC-AUC under strict holdout","Extra Trees reach 0.77 AUC for Portuguese Pseudo-nitzschia from satellite data","Leakage-resistant CV shows satellite HAB skill of 0.77 on Portuguese coast","Seasonal and lagged satellite features drive 0.77 ROC-AUC Pseudo-nitzschia forecasts","Year-and-cluster holdouts yield 0.77 AUC satellite HAB models for L1-L2"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The claim treats a fixed cell-count threshold of 10,000 cells per litre as an adequate operational proxy for harmful risk, even though toxin and harvest-ban events can occur at lower abundances and toxicity itself is not the modelled target.","fun_headline_variants_meta":{"raw":{"variants":["Satellites predict Pseudo-nitzschia HABs at 0.77 ROC-AUC under strict holdout","Extra Trees reach 0.77 AUC for Portuguese Pseudo-nitzschia from satellite data","Leakage-resistant CV shows satellite HAB skill of 0.77 on Portuguese coast","Seasonal and lagged satellite features drive 0.77 ROC-AUC Pseudo-nitzschia forecasts","Year-and-cluster holdouts yield 0.77 AUC satellite HAB models for L1-L2"]},"model":"grok-4.5","effort":"low","cost_usd":0.01009,"raw_usage":{"total_tokens":2357,"prompt_tokens":917,"num_sources_used":0,"completion_tokens":127,"cost_in_usd_ticks":100900000,"prompt_tokens_details":{"text_tokens":917,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1313,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":917,"tokens_out":127,"duration_ms":11175,"temperature":1.0,"reasoning_tokens":1313,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T17:04:12.486022+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Re-run the same year-and-cluster cross-validation after re-labelling HAB events from matched shellfish domoic-acid measurements (or actual harvest-ban days) instead of the fixed cell threshold; if discrimination collapses well below the reported 0.7 range, the occurrence proxy does not support operational early warning.","supporting_citations":[],"review_version":1}