{"id":"f1e3e03e-f92b-4321-813c-f1cb650f3783","arxiv_id":"2605.04548","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Reformulating disease prediction as event detection within 3-7 day windows using ML on weather data enables better short-horizon vineyard risk warnings than daily classification.","lead":"This paper reformulates vineyard disease risk prediction from daily status checks to detecting upcoming risk events in environmental time series. A smart generalist might read it to understand how changing the problem setup affects practical early warning tools for sustainable farming.","discovery_kind":"unclear","skeptic_critique":{"model":"grok-4.3","headline":"Event definition depends on unvalidated manual annotations and fixed minimum gap, risking models learning labeling artifacts instead of true precursors","rationale":"The reader's weakest assumption correctly isolates the labeling step as the least secure link in the argument. Because the provided text is the abstract only, no further internal checks (e.g., on data splits or metric definitions) are possible, so the concern stands and keeps the verdict at UNVERDICTED. No other technical inconsistency appears in the stated approach.","tokens_in":1750,"tokens_out":352,"duration_ms":46709,"concrete_test":"Vary the minimum disease-free gap across plausible values (e.g., 3, 5, 7, 10 days), re-annotate events, retrain all three model families on the same environmental features, and recompute the event-oriented metrics; if recall or lead-time rankings shift by more than 15% or false-alert rates change materially, the headline performance depends on the arbitrary gap choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that reformulating prediction as transitions into risk periods (after a minimum disease-free gap) causes models to learn environmental precursors rather than persistence. This holds only if the manual annotations accurately capture biological onset and the gap parameter does not create artificial event boundaries that models can exploit. The abstract states annotations are manual and the gap is chosen to reduce fragmentation, but supplies no annotation protocol, validation against observed disease incidence, inter-rater reliability, or sensitivity results for the gap value. Without these, reported trade-offs in recall, lead time, and false alerts could be artifacts of the chosen labeling scheme rather than evidence for the formulation's practical utility.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript presents an event-based approach to early warning of vineyard disease risk using environmental time series data. Instead of daily disease presence classification, it reformulates the task as predicting transitions into annotated risk periods within a 3-7 day horizon, defining new events only after a minimum disease-free gap to reduce fragmentation. The authors construct features from humidity, rainfall, temperature, and seasonal encoding, and compare XGBoost, LSTM, and TCN models using standard metrics and an event-oriented protocol, concluding that the formulation supports practical short-horizon warning with distinct model trade-offs.","tokens_in":1921,"tokens_out":627,"duration_ms":77374,"significance":"If the empirical results are robust, this work underscores the importance of problem formulation in time-series machine learning for environmental applications. By shifting focus to event transitions, it may enable more actionable predictions for disease management, reducing unnecessary interventions and supporting sustainable practices. The combination of classical and deep learning models with event-specific evaluation metrics provides a useful template for similar warning systems.","major_comments":[{"comment":"§3.1 (Annotation of Disease-Risk Periods): The manual annotation protocol for defining disease-risk periods is not described in detail, with no mention of validation against observed disease incidence data, inter-rater reliability, or sensitivity to annotator choices. This is load-bearing for the central claim, as the event-based formulation's advantage in capturing precursors (rather than persistence) depends on these labels accurately reflecting biological onsets; without it, reported trade-offs in recall and lead time could be labeling artifacts.","section":"§3.1 (Annotation of Disease-Risk Periods)"},{"comment":"§3.2 (Event Definition and Gap Parameter): The minimum disease-free gap length is introduced to reduce fragmentation but is treated as a fixed choice with no ablation study or sensitivity results on how different gap values affect event boundaries, model performance, or comparison to daily classification baselines. This parameter directly shapes the prediction targets and thus the claimed practical utility of the event-based approach.","section":"§3.2 (Event Definition and Gap Parameter)"}],"minor_comments":[{"comment":"§5 (Results and Evaluation): The event-oriented protocol metrics (recall, lead time, false alerts) are used but would benefit from an explicit worked example on a sample time series to clarify how they are computed from model outputs.","section":"§5 (Results and Evaluation)"},{"comment":"Figure 3 (Model Comparisons): The trade-off plots are informative but lack error bars or statistical significance tests across the multi-year folds, making it harder to assess whether observed differences between XGBoost, LSTM, and TCN are reliable.","section":"Figure 3 (Model Comparisons)"}],"recommendation":"major_revision","confidential_remarks":"The work aligns with the journal's interest in applied ML for environmental data, but the citation list could better acknowledge prior event-based formulations in time-series forecasting outside agriculture to strengthen novelty positioning."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for their constructive comments, which help improve the clarity and robustness of our work on event-based disease risk prediction. We address each major comment below, committing to revisions where appropriate.","responses":[{"response":"We acknowledge that the annotation protocol requires more detailed exposition to substantiate the central claims of the paper. We will revise §3.1 to include a comprehensive description of the annotation protocol, specifying the steps taken to define disease-risk periods based on environmental conditions and expert input. Additionally, we will add text discussing the absence of inter-rater reliability measures and sensitivity analysis to annotator choices, as well as the lack of direct validation against independent disease incidence observations. We will explain that the protocol was designed based on established agricultural practices for vineyard disease monitoring and discuss how this might affect the interpretation of the model trade-offs in recall and lead time. This will mitigate concerns about labeling artifacts.","revision_made":"yes","referee_comment":"[§3.1 (Annotation of Disease-Risk Periods)] §3.1 (Annotation of Disease-Risk Periods): The manual annotation protocol for defining disease-risk periods is not described in detail, with no mention of validation against observed disease incidence data, inter-rater reliability, or sensitivity to annotator choices. This is load-bearing for the central claim, as the event-based formulation's advantage in capturing precursors (rather than persistence) depends on these labels accurately reflecting biological onsets; without it, reported trade-offs in recall and lead time could be labeling artifacts."},{"response":"We agree that an analysis of the gap parameter's sensitivity would strengthen the paper and better support the claimed practical utility. In the revised version, we will add an ablation study, varying the minimum disease-free gap length across a range of values and reporting the effects on the number of defined events, model performance metrics, and comparisons to the daily classification baseline. We will also provide a justification for the gap length selected in the primary experiments, grounded in the typical disease cycle characteristics in vineyards. These additions will demonstrate the robustness of our findings to this parameter.","revision_made":"yes","referee_comment":"[§3.2 (Event Definition and Gap Parameter)] §3.2 (Event Definition and Gap Parameter): The minimum disease-free gap length is introduced to reduce fragmentation but is treated as a fixed choice with no ablation study or sensitivity results on how different gap values affect event boundaries, model performance, or comparison to daily classification baselines. This parameter directly shapes the prediction targets and thus the claimed practical utility of the event-based approach."}],"tokens_in":1497,"tokens_out":550,"duration_ms":74878,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"This paper's main move is to stop treating vineyard disease risk as a daily yes/no call and instead predict when a new risk period is about to begin within a 3-7 day window. They add a minimum disease-free gap before counting a fresh event, which cuts down on fragmented labels from brief interruptions. The idea is that this pushes models to look for real precursors in the weather data rather than just copying yesterday's status forward.","headline":"The paper's event-based reformulation for vineyard disease warnings is a clean tweak on problem setup, but the abstract gives no numbers and the manual labels lack any validation.","tokens_in":2408,"tokens_out":166,"would_cite":false,"duration_ms":48655,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Reformulating vineyard disease risk as event onsets enables practical 3-7 day early warnings from environmental time series.","keywords":["vineyard disease prediction","event-based early warning","environmental time series","machine learning","LSTM","TCN","XGBoost","agro-meteorological data"],"falsifier":"Independent verification of actual disease outbreaks in a new vineyard season, checking whether alerts issued by the models would have prompted interventions that measurably reduce observed disease incidence.","tokens_in":2671,"feed_emoji":"🍇","tokens_out":733,"duration_ms":49551,"temperature":0.7,"pith_summary":"The paper argues that daily classification of disease status tends to reward models for repeating recent labels rather than detecting real precursors in weather patterns. Instead it defines distinct risk events only after a minimum disease-free gap and asks models to forecast the start of these events within the next 3-7 days. Input features track humidity dynamics, rainfall totals, temperature swings, and seasonal cycles through cyclic encoding. When tested on multi-year vineyard data with XGBoost, LSTM, and TCN, the event formulation yields usable short-horizon alerts while the models differ in how well they balance event detection against false alarms and warning lead time. A reader would care because such warnings can guide timely, targeted interventions that limit unnecessary pesticide applications.","feed_headline":"Event reformulation enables 3-7 day vineyard disease warnings","feed_subtitle":"Predicting risk-period onsets after a disease-free gap yields actionable alerts with model-specific trade-offs in recall and false alerts.","key_machinery":"The event-based reformulation that creates new risk events only after a minimum disease-free gap in the binary labels, turning the task into forecasting onset transitions within a 3-7 day window.","core_discovery":"The paper establishes that shifting from daily disease-presence classification to the prediction of transitions into annotated risk periods 3-7 days ahead, after enforcing a minimum disease-free gap between events, lets models focus on environmental precursors rather than persistence. This is shown by constructing time-series inputs that encode humidity dynamics, rainfall accumulation, temperature variability, and seasonal structure, then comparing XGBoost, LSTM, and TCN under both standard metrics and an event-oriented protocol that tracks recall, lead time, and false-alert behavior.","pith_inferences":["The approach could be tested on other crops by swapping the annotation source while keeping the event-gap rule fixed.","Streaming sensor data could feed the same models to produce automated daily risk-onset scores for farm management systems.","Pairing the learned precursors with existing biological disease models might expose and correct any systematic annotation biases.","Operational systems could select among the three model families according to whether their priority is high recall or low false-alarm rate."],"forward_implications":["Models trained this way issue warnings with measurable lead time instead of simply echoing recent daily status.","XGBoost, LSTM, and TCN display distinct operating points on the trade-off between catching true risk events and limiting false alerts.","Features built from accumulation metrics and cyclic seasonal encoding help surface the environmental conditions that precede risk periods.","The same event definition can be applied to other environmental time-series tasks where persistence bias distorts daily forecasts."],"fun_headline_variants":["Event detection for 3-7 day vineyard disease risk periods","Transitions to disease risk predicted from agro-meteorological inputs","Minimum disease-free gaps separate risk events in vineyards","Comparing classical and deep models on vineyard event prediction"],"cache_read_input_tokens":64,"weakest_assumption_plain":"The manually annotated disease-risk periods and the chosen minimum disease-free gap accurately reflect real biological onset without introducing labeling artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Event detection for 3-7 day vineyard disease risk periods","Transitions to disease risk predicted from agro-meteorological inputs","Minimum disease-free gaps separate risk events in vineyards","Comparing classical and deep models on vineyard event prediction"]},"model":"grok-4.3","cost_usd":0.008419,"raw_usage":{"total_tokens":3770,"prompt_tokens":751,"num_sources_used":0,"completion_tokens":56,"cost_in_usd_ticks":84190500,"prompt_tokens_details":{"text_tokens":751,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2963,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":751,"tokens_out":56,"duration_ms":26425,"temperature":1.0,"reasoning_tokens":2963,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-05-08T16:39:49.628867+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Independent verification of actual disease outbreaks in a new vineyard season, checking whether alerts issued by the models would have prompted interventions that measurably reduce observed disease incidence.","supporting_citations":[],"review_version":1}