{"id":"b4fe1b0c-7d56-4fc2-b19d-d30004fbe8bf","arxiv_id":"2506.23872","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Ivy plant electrical signals, recorded outdoors for five months, let machine learning models classify day/night, rain/dry, warm/cold, and windy/calm conditions with up to 95% macro F1.","lead":"Scientists attached sensors to ivy plants outdoors and recorded their electrical signals for five months. Machine learning classified weather conditions from those signals with up to 95 percent accuracy, showing plants can work as living environmental sensors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported F1 scores are based on a random 80/20 split of autocorrelated 1-hour windows; temporal leakage likely inflates the headline 95% macro F1, so the generalizable mapping claim is not yet established.","rationale":"Temporal leakage is the single most load-bearing concern because it directly affects the numerical headline (up to 95% macro F1) and the central claim of mapping plant signals to environmental conditions. The reader's weakest_assumption is exactly this: random split assumes window independence, which is false for autocorrelated plant recordings. I agree with that diagnosis; the released data make the issue testable, which is why the verdict should remain CONDITIONAL rather than REJECT. Two secondary issues reinforce caution but do not change the main verdict: (i) the abstract says AutoML outperformed manual tuning, but Table 2 shows manual RF averaging 90.7% vs. AutoML 89.6%, an internal inconsistency in a supporting claim; (ii) the paper reports 216 days of data although the stated collection period July 5 - November 18, 2024 spans roughly 136 days, and a date typo appears in Sec. 4 ('2025'), indicating reporting imprecision. Neither is fatal, but they support the need for independent re-analysis. If a temporal split reproduces high F1, the central claim is substantially strengthened; if not, the paper would require major revision rather than minor edits.","tokens_in":9156,"tokens_out":6271,"duration_ms":66974,"concrete_test":"Using the published dataset (Zenodo DOI in Ref. [3]), re-run the best RF pipeline from Sec. 4 (rain-dry with the 49-feature subset) under blocked temporal cross-validation: assign all windows from a contiguous day to the same fold and use GroupKFold, or train on an early contiguous block (July-September) and test on a later block (October-November), separately for each plant and channel. Compare macro F1 to the random-split value of 95.5%. Also run day-night as a positive control (expected to remain high) and wind-calm as an additional rare-event check. If rain-dry or wind-calm F1 drops by more than roughly 10 points or falls toward the minority-class baseline, the random-split evaluation is the source of the claimed performance and the generalizability claim is unsubstantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Reported macro F1-scores of up to 95% (Fig. 4a, rain-dry) are computed under an 80/20 split of 1-hour windows described in Sec. 3.2, but the split is random rather than temporal. The underlying electrophysiological signal is a slowly varying, strongly autocorrelated time series collected from only four plants over about five months. A random split places neighboring windows from the same plant, same day, and same weather episode into both training and test sets. The classifier can therefore exploit plant-specific electrode offsets, circadian and seasonal drift, and the persistence of rare events (rain, wind) across adjacent windows. The macro F1 then measures interpolation between interleaved windows, not generalization to unseen time periods, which is the relevant capability for a deployed monitoring system. This is especially damaging for the minority classes (rain 4.1%, wind 6.4%), which occur in multi-hour episodes; under a random split, the model can effectively memorize each episode's surrounding context. The paper's limitation list (Sec. 5) acknowledges small sample size and arbitrary thresholds but omits this temporal leakage. The central claim is therefore supported only if the model retains high F1 under a temporal split; currently the evaluative setup does not demonstrate that.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports a five-month outdoor deployment of four PhytoNode-instrumented Hedera helix plants, with electrophysiological potentials sampled at roughly 200 Hz, downsampled, segmented into 1-hour windows, labeled from nearby weather-station thresholds into four binary tasks (day/night, rain/dry, warm/cold, windy/calm), and classified with five manually configured classifiers and an AutoML framework. The authors report macro F1-scores up to 95% for rain-dry after mutual-information feature selection, and they show precision-recall curves supporting the ranking of tasks. The central claim is that plant electrical signals can be mapped to environmental conditions under uncontrolled outdoor conditions.","tokens_in":9373,"tokens_out":3831,"duration_ms":44364,"significance":"If the performance estimates were unbiased, this would be a useful demonstration of long-term outdoor phytosensing with a self-powered wearable, including an open dataset and a fully described preprocessing and modeling pipeline. The qualitative day-night signal is convincing, the AutoML-versus-manual comparison is informative, and the feature-selection analysis is presented in a reproducible way. However, the headline F1 numbers rely on an evaluation design that has not yet established generalization to unseen time periods or unseen plants, so the main quantitative claim is conditional on re-analysis.","major_comments":[{"comment":"The 80/20 split is described as a random split over 1-hour windows, not as a temporal or plant-blocked split. Because the electrophysiological signal is slowly varying and autocorrelated (as the daily traces in Fig. 3 indicate), and because rain and wind occur in multi-hour episodes, adjacent windows from the same plant, day, and weather event appear in both training and test sets. The reported macro F1 values (e.g., 93.78% for rain-dry with the manual RF, 95.5% after feature selection) therefore measure interpolation among interleaved windows rather than generalization to unseen periods, which is the capability a deployed monitoring system would require. I ask the authors to re-evaluate with a temporal split (e.g., train on earlier months and test on later months) or with blocked cross-validation by day or by plant, and to report both random-split and temporal-split results; the central claim should be based on the temporal-split numbers.","section":"Sec. 3.2 (Step 4), Table 2, Fig. 4a"},{"comment":"The mutual-information feature selection is applied on the training portion of the same random split, but the text does not specify how the number k of selected features was chosen. If k was selected by inspecting Fig. 4a across values of k on the test set, then the 'best subset' F1 values are optimistically biased; if it was selected on a validation fold or by nested cross-validation, that procedure should be stated explicitly. Please clarify the protocol or report a nested evaluation that keeps feature selection inside the training loop.","section":"Sec. 3.2 (Step 4) and Sec. 4"},{"comment":"The paper motivates the use of four plants as a way to 'identify species-related patterns that extend individual plant characteristics,' but all reported evaluations pool the four plants and use random splits. No result is shown for held-out plants. A per-plant or per-plant-group cross-validation would provide evidence that the learned patterns are not dominated by plant-specific electrode offsets or individual baselines; without such an evaluation, the species-level generalization claim is not supported by the current experiments.","section":"Sec. 3.1 and Sec. 4"}],"minor_comments":[{"comment":"The sentence 'the classifier of wind-calm achieved a recall of 0.79 and a recall of 0.83' should read 'a precision of 0.83 and a recall of 0.79' (or the appropriate ordering), since the following explanation describes 83% of predictions being correct and 79% of actual instances being captured.","section":"Sec. 4, paragraph after Fig. 4b"},{"comment":"The date range 'November 5, 2024, to November 12, 2025' appears to be a typo; given the stated collection period from July to November 2024, it should likely read 'November 5, 2024, to November 12, 2024.'","section":"Sec. 4, first paragraph"},{"comment":"There are several typographical errors: 'classifer' in the Table 2 caption, 'tow classes' in Sec. 4, and 'Principle component analysis' in the Table 2 abbreviations should be 'principal component analysis.'","section":"Table 2 caption and Sec. 4"},{"comment":"The sentence 'Additional visualizations of differential potential dynamics are available in [4,8]' cites [8], which is a reference on ozone detection by Dolfi et al. and is unlikely to contain the authors' visualizations; please verify the intended citation.","section":"Sec. 4, first paragraph"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope, and the main issues are methodological rather than conceptual. The temporal-leakage concern raised in the stress-test note is real and lands on the central claim, but it is addressable with a re-analysis using temporal or blocked splits. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine outdoor deployment with a public dataset, but the headline F1 numbers should not be trusted as evidence of generalization. The train/test split is random rather than temporal, and on autocorrelated 1-hour windows that is a real leak.\n\nWhat the paper does well: it moves the PhytoNode hardware out of the lab, runs four ivy plants for five months next to a weather station, and releases the data. The qualitative day-night pattern in the electrical potential is clear. The ManualML vs AutoML comparison is reasonable, and AutoML matching a tuned random forest is a normal result. The feature-selection analysis shows you can get good performance with far fewer than 700 features. The authors are also upfront about small sample size and arbitrary class thresholds.\n\nThe soft spots, in order of severity. First, the evaluation split. Section 3.2 says '80%/20% split' but not temporal. The windows are 1-hour segments from a slowly varying signal, so adjacent windows are correlated. A random split places neighboring windows from the same weather episode in both training and test. That lets the classifier effectively memorize the context around each rain or wind event, and the minority classes (rain 4.1%, wind 6.4%) are exactly where this inflates macro F1. The claimed 95% for rain-dry is therefore likely optimistic. Without a temporal split, the generalization claim is not established. This is the main issue and it is fixable.\n\nSecond, the class labels depend on arbitrary thresholds. The authors acknowledge this, and it means the tasks are setup-dependent, but they are transparent. Third, the small sample: four plants, one location, five months of one season. Also acknowledged.\n\nThere are minor text errors: the results mention data through November 12, 2025, contradicting the July–November 2024 window given earlier; and 'recall of 0.79 and a recall of 0.83' should be precision and recall. Cosmetic, but worth fixing.\n\nWho this is for: people working on biohybrid sensors and anyone applying ML to time-series physiological data. It is a useful case study of the temporal-split pitfall, and the data release is a plus. I would bring it to a reading group.\n\nRecommendation: this deserves peer review, not desk rejection. Send it out. The reviewers should require a temporal split and a report of what the F1 becomes under true temporal generalization. Even if the numbers drop, the paper still stands as a hardware demonstration and dataset contribution, just with more honest claims.","headline":"Genuine outdoor deployment and a useful dataset, but the reported F1 scores are probably inflated by a random split on autocorrelated windows.","tokens_in":9889,"tokens_out":4740,"would_cite":false,"duration_ms":46466,"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":"Ivy, fitted with a solar-powered wearable node, carries enough information in its electrical potential to classify day/night, rain/dry, wind/calm, and warm/cold with macro F1-scores up to 95 percent.","keywords":["plant electrophysiology","phytosensing","Hedera helix","plant wearable","automated machine learning","environmental monitoring","biohybrid systems","time-series classification"],"falsifier":"Retrain the same classifiers on the first four months of the five-month record and test only on the final month, keeping the same features and thresholds. If macro F1 stays near the reported values, the generalization claim is supported; if it falls toward the minority-class baseline, the random 80/20 split is the source of the high scores.","tokens_in":8923,"feed_emoji":"🌿","tokens_out":12451,"duration_ms":120563,"temperature":0.7,"pith_summary":"This paper reports that a solar-powered wearable device called PhytoNode can record the electrical potential of common ivy (Hedera helix) outdoors for five months, and that machine-learning classifiers can map those recordings to environmental conditions. In four binary tasks—day versus night, rain versus dry, windy versus calm, and warm versus cold—the best models reach macro F1-scores up to 95 percent. The authors also report that automated machine-learning pipelines consistently produce high scores and argue that they are preferable to manual tuning, while selecting a few dozen statistical features performs as well as the full set of over 700. If the result holds, plants could serve as low-cost, self-sustaining environmental sensors that need no grid power and can be read wirelessly.","feed_headline":"95% F1: ivy voltages classify day, rain, wind, and warmth","feed_subtitle":"A five-month outdoor record shows a solar-powered plant wearable can map electrical activity to environmental conditions.","key_machinery":"The central object is PhytoNode, a self-sustaining plant-wearable node that samples the electrical potential at roughly 200 Hz between two silver-coated electrodes, one on the stem and one on a leaf. The raw stream is downsampled to 1 Hz, cut into one-hour windows, and reduced to over 700 statistical features by the tsfresh library. Class labels come from environmental thresholds, class imbalance is handled with SMOTE upsampling, and five classifiers (random forest, SVM, k-nearest neighbours, naive Bayes, multi-layer perceptron) plus AutoML pipelines built on scikit-learn map the features to labels. Mutual-information feature selection then identifies smaller feature subsets that match or beat the full set. This machinery converts a continuous voltage stream into labeled, classifiable units.","core_discovery":"On its own terms, the paper establishes that the differential electrical potential of Hedera helix, measured between stem and leaf electrodes by the PhytoNode in an uncontrolled outdoor garden, contains recurring patterns that track slow environmental state. Threshold labels built from irradiance, precipitation, temperature, and wind speed define four binary tasks; random-forest classifiers and AutoML-selected pipelines achieve macro F1 values between about 85 and 95 percent depending on task and channel, with rain-versus-dry reaching 95.5 percent after feature selection. The paper's stated contribution is moving phytosensing out of controlled laboratories into harsh, real-world conditions and validating the whole acquisition-to-classification pipeline there.","pith_inferences":["Beyond the paper, a strict temporal validation—training on the first months and testing on the final month—would probably lower the reported F1-scores, but it would also clarify whether the classifiers track slow seasonal drifts or only short-term weather; the random split cannot distinguish these.","The day/night task may be driven partly by circadian rhythms or by the irradiance threshold itself rather than by a plant physiological response; a comparison under artificial light or constant temperature would separate those causes.","If the informative signal is slow environmental state, the one-hour window could likely be shortened to give faster alerts; the paper says determining the shortest feasible interval is future work.","Different plant species with complementary sensitivities could be combined into a heterogeneous sensor network, but the paper only studies one species and four individuals."],"forward_implications":["PhytoNode's solar power, weatherproof housing, and Bluetooth Low Energy transmission make distributed deployment feasible in urban or agricultural settings without wired infrastructure.","Because selected subsets of features match or beat the full feature set, the final classifiers can be compact enough to run directly on the node, supporting real-time rather than offline classification.","The same ivy plant can serve as a multi-purpose sensor, simultaneously supporting day/night, rain/dry, wind/calm, and warm/cold discrimination.","The five-month outdoor record suggests the approach can be extended to year-round monitoring, although the paper lists species, season, and location as untested sources of variation."],"supporting_citations":[{"why":"Defines the PhytoNode hardware and its earlier controlled-environment validation, which this paper moves outdoors.","marker":"[4]"},{"why":"Releases the five-month dataset of electrical potential and environmental measurements that all analyses use.","marker":"[3]"},{"why":"Supplies the Naive AutoML library that builds and optimizes the comparison pipelines.","marker":"[19]"},{"why":"Provides the scikit-learn classifiers and feature-selection tools used for manual and AutoML models.","marker":"[20]"},{"why":"Introduces SMOTE, the synthetic upsampling technique used to compensate for class imbalance.","marker":"[6]"},{"why":"Implements tsfresh, which computes the 700-plus statistical features for each one-hour window.","marker":"[7]"},{"why":"Gives the earlier ozone and wind phytosensing AutoML framework and the leaf-versus-stem comparison this paper builds on.","marker":"[1]"}],"fun_headline_variants":["Ivy voltage patterns reveal weather with 95% accuracy","Five months of ivy voltages: AI decodes day, rain, wind","Plant-wearable tech: ivy's electrical signals monitor climate","Ivy as sensor: machine learning maps plant voltages to weather","Biohybrid sensing: ivy's electrical pulses classify outdoor conditions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one-hour windows can be treated as independent when split randomly into training and test sets; because consecutive windows come from the same plant and day, that split leaks information and inflates the reported macro F1-scores.","fun_headline_variants_meta":{"raw":{"variants":["Ivy voltage patterns reveal weather with 95% accuracy","Five months of ivy voltages: AI decodes day, rain, wind","Plant-wearable tech: ivy's electrical signals monitor climate","Ivy as sensor: machine learning maps plant voltages to weather","Biohybrid sensing: ivy's electrical pulses classify outdoor conditions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000573,"raw_usage":{"total_tokens":2673,"prompt_tokens":878,"completion_tokens":1795,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":1705}},"tokens_in":494,"tokens_out":1795,"duration_ms":14258,"temperature":1.0,"reasoning_tokens":1705,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:29:03.533509+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same classifiers on the first four months of the five-month record and test only on the final month, keeping the same features and thresholds. If macro F1 stays near the reported values, the generalization claim is supported; if it falls toward the minority-class baseline, the random 80/20 split is the source of the high scores.","supporting_citations":[{"cited_title":"In: Future of Information and Communication Conference","cited_arxiv_id":null,"evidence_quote":"Defines the PhytoNode hardware and its earlier controlled-environment validation, which this paper moves outdoors."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Releases the five-month dataset of electrical potential and environmental measurements that all analyses use."},{"cited_title":"Machine Learning 112(4), 1131–1170 (2023)","cited_arxiv_id":null,"evidence_quote":"Supplies the Naive AutoML library that builds and optimizes the comparison pipelines."},{"cited_title":"Journal of Machine Learning Research12, 2825–2830 (2011)","cited_arxiv_id":null,"evidence_quote":"Provides the scikit-learn classifiers and feature-selection tools used for manual and AutoML models."},{"cited_title":"Journal of artificial intelligence research 16, 321–357 (2002)","cited_arxiv_id":null,"evidence_quote":"Introduces SMOTE, the synthetic upsampling technique used to compensate for class imbalance."},{"cited_title":"Neuro- computing 307, 72–77 (2018)","cited_arxiv_id":null,"evidence_quote":"Implements tsfresh, which computes the 700-plus statistical features for each one-hour window."},{"cited_title":"In: 2025 IEEE Symposia on Computational Intelligence for Energy, Transport and Environmental Sustainabil- ity (CIETES)","cited_arxiv_id":null,"evidence_quote":"Gives the earlier ozone and wind phytosensing AutoML framework and the leaf-versus-stem comparison this paper builds on."}],"review_version":1}