{"id":"3d2d1230-b598-4943-9cd3-2ed12c5c72dd","arxiv_id":"2607.20386","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Modulation tensorgrams — modulation spectra with the time axis preserved — fed to a CRDNN predict honey-bee colony strength more accurately (MAE 3.31, r=0.78, hive-independent) than spectrogram- or MFCC-based deep models.","lead":"This paper finds that keeping the time dimension in a modulation-spectrum representation of beehive audio, and feeding it to convolutional and recurrent neural networks, improves colony-strength prediction and generalization to unseen hives versus spectrogram and MFCC baselines on the public UrBAN dataset. It matters because accurate, generalizable acoustic hive monitoring is what would make IoT-based precision apiculture practical at scale.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control for label interpolation/seasonal confounding; reported cross-hive gains may reflect tracking a smooth time curve rather than acoustic colony-strength discrimination.","rationale":"The reader's weakest assumption matches the single most load-bearing concern. The paper's own limitations section admits the issue, but no experiment addresses it. That justifies the CONDITIONAL verdict: the central comparison between modulation features and baselines is credible only if the sensitivity analysis shows the gains are not driven by label interpolation or seasonal confounding. If the proposed test shows the gains are robust to near-inspection labels, the main claim would be substantially stronger; if not, the reported MAE/r numbers would need to be reinterpreted. No further verdict adjustment is needed beyond the reader's CONDITIONAL, because the paper can still be accepted provisionally pending this analysis.","tokens_in":14929,"tokens_out":3798,"duration_ms":36085,"concrete_test":"Run a controlled experiment on the UrBAN dataset: keep only audio segments recorded within ±3 days of a manual inspection, assign the actual inspection fob as label (no interpolation), and train/evaluate the best model (CRDNN-3D on tensorgram) under the same hive-independent split. Compare MAE/r to the full interpolated-label results. If performance collapses (e.g., r drops from ≈0.78 to near chance), the reported gains are largely an artifact of label smoothing/seasonal trend. Additionally, as a control, train a model on the same tensorgram inputs plus a day-of-year/hour-of-day input feature; if the time features alone recover most of the accuracy, the acoustic features are not the informative signal.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption is that the interpolated fob labels (Section III-A) are a valid proxy for true colony strength for every 15-minute segment. Because inspections are sparse, linear temporal interpolation creates a piecewise-linear, smooth function of time. Audio features also evolve smoothly with time of day, season, and weather. Under the hive-independent split, all hives share the same seasonal calendar, so a model can achieve high correlation by recognizing time-of-year acoustic cues (e.g., ambient noise, insect activity) and predicting the interpolated trend, without ever estimating frames-of-bees. This would explain the large MAE drop (3.5→1.0 random, 4.5→3.3 hive-independent) and the r≈0.78 cross-hive result: the model is tracking a smooth label curve that is itself a function of season. The paper's Section IV-F acknowledges 'seasonal confounding effects may still remain' but provides no sensitivity analysis to distinguish acoustic colony-strength information from seasonal tracking. This is the weakest link: if the interpolation encodes mostly time, the central claim that modulation temporal dynamics carry colony-strength information is not established. The Table VI rank duplication is a reporting error, but it is secondary; the interpolation/seasonality issue directly threatens the main generalization claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a new audio representation for honey bee colony strength estimation: the modulation tensorgram, which retains the time axis of the modulation spectrum instead of averaging it away. Using the public UrBAN dataset (nine hives, two years, over 3,000 hours of audio), the authors feed spectrograms, MFCCs, modulation spectrograms, and modulation tensorgrams into a family of CNN/CRDNN architectures and evaluate them under both random 5-fold splits and hive-independent 5-fold splits. They report that modulation-based inputs consistently outperform spectrogram/MFCC baselines, with the best model (CRDNN-3D on the tensorgram) reaching MAE 1.01±0.08 and r≈0.97 in the random-split setting and MAE 3.31±1.36 and r=0.78±0.17 in the hive-independent setting, versus r≈0.5 for the baselines. The paper also includes Friedman/Nemenyi statistical tests, saliency and Grad-CAM analyses, a comparison with a random-forest baseline, and a computational-complexity assessment. The central claim is that the temporal dynamics of the modulation spectrum carry colony-strength information that static spectral or cepstral features discard.","tokens_in":15133,"tokens_out":4643,"duration_ms":46332,"significance":"If the reported results hold, the paper would make a useful contribution to precision apiculture and to the broader literature on modulation-spectral representations for acoustic monitoring. The strengths of the study are its use of a public, longitudinal dataset; the inclusion of a hive-independent evaluation protocol that is more realistic for deployment; and the transparent reporting of multiple architectures, statistical tests, and computational costs. The main comparison—modulation-based inputs outperforming spectrogram/MFCC inputs in the hive-independent split—is the kind of empirical evidence that could move the field toward richer time-preserving representations. However, the central empirical claim is currently threatened by a label-construction issue: colony-strength labels are linearly interpolated between sparse manual inspections. Because interpolated labels are smooth functions of calendar time, and because all hives share the same seasonal calendar, the reported cross-hive correlation gains could partly reflect tracking of a seasonal/time trend rather than acoustic discrimination of true colony strength. The paper acknowledges seasonal confounding in Section IV-F but","major_comments":[{"comment":"The label for every 15-minute audio segment is obtained by linear temporal interpolation between sparse manual inspections (Section III-A). The resulting target is a piecewise-linear, smooth function of calendar time. In the hive-independent split, all hives share the same seasonal calendar, so a model could achieve high correlation by recognizing time-of-year acoustic cues (ambient noise, insect activity, weather-correlated sounds) and predicting the interpolated trend, without estimating frames-of-bees. Section IV-F acknowledges that 'temporal and seasonal confounding effects may still remain,' but no sensitivity analysis is provided. To support the load-bearing claim that modulation temporal dynamics carry colony-strength information, please add at least one of the following controls: (a) a time-only baseline (e.g., day-of-year, or a smoothed inspection curve) compared directly with t","section":"IV-F"},{"comment":"Several modulation-specific design choices were made empirically on the UrBAN data: the STFT window/hop lengths (100 ms/12.5 ms), the second-stage 60 s window, the 1 kHz frequency cap, and the bin-merging geometry. Section II-A states these were 'optimized empirically based on pilot experimentation,' and Section II-B uses the same dataset's labels to justify the 1 kHz cap and bin geometry. If these choices were tuned on the same data that is later used for evaluation, the comparison against spectrogram/MFCC baselines is optimistically biased in favor of the modulation representation. Please clarify whether all such choices were fixed before model evaluation, e.g., by selecting them on training folds only or on a separate development set, and ideally report sensitivity of the main hive-independent results to these choices (e.g., 50/100/200 ms window, 30 s vs 5 min aggregation). If the cho","section":"II-A/II-B"},{"comment":"The Friedman average ranks in Table VI contain a clear error: under Random-Split, the MAE ranks are identical for every model between 'No pre-processing' and 'Spectral amplitude subtraction' (e.g., Spec-CNN-2D 6.6/6.6, CRDNN-3D 1.4/1.4, and all other entries), while the corresponding r ranks differ. This duplication is implausible because spectral subtraction changes the input and the MAE values in Table IV. Since the Friedman/Nemenyi analysis is used to support the recommendation of CRDNN-3D, please correct the table, recompute the CD diagrams, and report whether the ranks are pooled over the four scenarios or computed separately. If the corrected ranks alter the statistical conclusions (e.g., if CRDNN-3D no longer separates from the modulation-spectrogram group), revise the claims in Section IV-A accordingly.","section":"Table VI"},{"comment":"The statistical comparison is reported only as average ranks across the five folds, without confidence intervals or a direct treatment of the fact that the Friedman test is computed on ranks from the same data used to select hyperparameters. More importantly, the Nemenyi CD diagrams in Figure 5 are the main support for the claim that the proposed method is significantly better than the baselines in the hive-independent setting, but the diagrams appear to group most modulation-based methods together. Please clarify whether the claimed advantage of CRDNN-3D over all baselines is statistically significant after correcting Table VI, and report the actual Friedman/Nemenyi p-values for the pairwise comparisons that are highlighted in the text.","section":"IV-B"}],"minor_comments":[{"comment":"The description of the tensor dimensions is confusing. The paper states that after the two STFT stages the tensor is (801×801×14), but with a 60 s no-overlap window and 15-minute segments one would expect 15 frames rather than 14. Please clarify the frame count and whether the second-stage window covers the full segment length or is computed with a leftover partial frame.","section":"II-A"},{"comment":"Figure 4 shows inspection measurements over time, but it is not indicated which points correspond to actual manual inspections and which are interpolated. Please add markers or a note explaining that the plotted curves are the interpolated labels used for training.","section":"III-A"},{"comment":"In Table VII the abbreviation 'MSAB' (presumably 'Modulation Spectral Analysis Baseline') is undefined. Please spell it out in the caption or text.","section":"IV-D"},{"comment":"The text says 'the modulation-based methods are capable of reducing the MAE to around 1.5' in the random-split setting, but Table IV shows a range from 1.01 to 1.61. Please use a range or a representative value to avoid overstating the improvement.","section":"IV-A"},{"comment":"The inference times are measured on a MacBook Pro with an Apple M1 chip, but the paper does not specify the software/version or whether the timings are CPU-only. Since the deployment discussion depends on these numbers, please provide the measurement conditions and, if possible, report on an edge-class CPU as well.","section":"IV-E"},{"comment":"The notation for the target variable is inconsistent: 'frames of bees (fob)', 'fobs', and 'fob' are all used. Define the abbreviation once and use it consistently.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The label interpolation/seasonal confounding issue is the main barrier to acceptance. The authors already acknowledge the possibility in Section IV-F, but the manuscript does not provide the controls that would let a reader distinguish acoustic colony-strength estimation from time-of-year tracking. This is fixable: a time-only baseline and a season-stratified evaluation would substantially strengthen the central claim. Please also verify Table VI, since the duplicated ranks undermine the statistical support. If the authors can provide corrected ranks and the sensitivity analyses, the paper would be a solid contribution; in its current form, the main empirical claim is not yet established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my take. The new thing is concrete: instead of handcrafting features from modulation spectra, the authors feed the raw modulation spectrogram and the time-preserving tensorgram to CNN/CRDNN models. On the 3,000-hour UrBAN dataset, under a hive-independent split, all modulation-input models beat all spectrogram/MFCC baselines, with the tensorgram CRDNN-3D hitting r≈0.78 vs ≈0.5 for the baselines. The hive-independent protocol is the right call, and the authors deserve credit for not overclaiming: they explicitly note that the CRDNN-3D's MAE edge in the cross-hive setting is within variability and they lean on rank statistics and model size instead. The explainability section is suggestive, not load-bearing.\n\nThe soft spot is the label construction. Frames-of-bees are assessed sparsely, then linearly interpolated across time. That makes the target a piecewise-linear function of the calendar. Audio also drifts with season, weather, time of day. So a model could be learning to track the smooth seasonal curve from ambient acoustic cues, not to estimate colony strength directly. The hive-independent split removes hive-specific signatures but not the shared seasonal calendar. The authors acknowledge the risk but provide no sensitivity analysis — e.g., training only on segments near inspection dates, or season-stratified folds, or conditioning on explicit time covariates. Without that, the cross-hive generalization numbers are suggestive but not airtight. This is the main thing a referee should push on.\n\nTwo smaller issues. The representation geometry (STFT windows, bin merging, frequency cap) was tuned on the same data in pilot experiments and via the F-ratio analysis. That is a mild circularity, probably not fatal given the head-to-head comparison, but it should be stated more plainly. And Table VI has an obvious mechanical error: the random-split rank columns are identical for the two enhancement conditions even though Table IV shows different MAE values. That needs correcting. Also, the authors should release code or at least fold assignments.\n\nOverall, the central finding — modulation temporal dynamics help — holds up as a claim about this dataset and task. The paper is honest, the evaluation is mostly careful, and the gap is real. I would send it to peer review and ask for the sensitivity analysis and table fix before acceptance.","headline":"Feeding raw modulation tensorgrams to deep nets gives a clear accuracy gain over spectrogram/MFCC baselines on UrBAN; the main weakness is that interpolated labels leave a seasonal-confounding hole that needs a sensitivity analysis.","tokens_in":15767,"tokens_out":3366,"would_cite":true,"duration_ms":33093,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Keeping the time axis in a modulation spectrogram—a 'modulation tensorgram'—lets deep audio models predict honey-bee colony strength more accurately and generalize across hives than standard spectrograms or MFCCs.","keywords":["honey bee colony strength","modulation tensorgram","modulation spectrum","audio IoT monitoring","deep learning","CNN","CRDNN","beehive acoustics"],"falsifier":"Compute the model's prediction error separately for segments recorded close to manual inspections versus far from them: if error rises systematically with time since the nearest inspection, the model is tracking interpolated labels. Alternatively, retrain with explicit time-of-day and calendar features and see whether the modulation tensorgram still improves correlation beyond those covariates; if it does not, the temporal dynamics may be seasonal rather than acoustic.","tokens_in":14726,"feed_emoji":"🐝","tokens_out":3403,"duration_ms":32011,"temperature":0.7,"pith_summary":"The paper claims that the temporal dynamics of the modulation spectrum—how the rate of amplitude modulation changes over time—carry information about honey-bee colony strength that prior averaged representations discard. To capture that information, the authors build a modulation tensorgram that preserves the time dimension, then feed it to convolutional and convolutional-recurrent neural networks. On a public dataset with more than 3,000 hours of beehive audio, they report that this approach outperforms spectrogram and MFCC baselines in both random splits and hive-independent splits, with the best model reaching a mean absolute error of about 1.0 frame of bees in random splits and about 3.3 frames for unseen hives. The paper also shows, via saliency maps and Grad-CAM, that the models rely on structured modulation-frequency patterns that shift with colony strength. A sympathetic reader would take the central message as: acoustic monitoring of hive strength can be both accurate and generalizable if the temporal structure of modulation is preserved rather than averaged away.","feed_headline":"Audio time dynamics predict bee colony strength across hives","feed_subtitle":"Keeping the modulation spectrum's time axis lifts accuracy and cross-hive generalization over spectrogram benchmarks.","key_machinery":"The modulation tensorgram: a 3D array (acoustic frequency × modulation frequency × time) computed by taking a short-time Fourier transform of the audio, then a second STFT along the time axis of each frequency bin, with the time axis preserved instead of averaged away. This is the object that carries the argument: it provides deep networks with both spectral structure and temporal evolution, and the paper's F-ratio analysis identifies which frequency-modulation regions are most discriminative across colony-strength categories.","core_discovery":"The central discovery is that modulation-spectrum temporal dynamics are discriminative for honey-bee colony strength and that a modulation tensorgram—a 3D representation of acoustic frequency, modulation frequency, and time—lets deep networks exploit them. In hive-independent evaluation, where entire hives are held out, the best model (a convolutional-recurrent network with 3D tensorgram input) achieves mean absolute error 3.31±1.36 and Pearson correlation r=0.78±0.17, while spectrogram and MFCC baselines stay around r≈0.5. In random splits, the same model reaches MAE 1.01±0.08 and r=0.97±0.01. The paper further shows that the modulation-based inputs benefit less from spectral-subtraction pr","pith_inferences":["The largest numerical gap over baselines appears in random splits, where temporally adjacent segments share acoustic conditions; the more meaningful result is the hive-independent gap, and that gap is smaller in absolute terms, so deployment claims should be weighed against the hive-independent numbers.","A concrete stress test of the interpolation assumption: evaluate prediction error as a function of time since the nearest manual inspection; if errors grow sharply with distance from inspections, the model is partly tracking a linearly interpolated label curve rather than true colony strength.","Another test: add time-of-day, date, and weather as auxiliary inputs and check whether the tensorgram still adds predictive value; this would separate acoustic temporal dynamics from seasonal confounding, which the paper acknowledges as a remaining limitation.","The same tensorgram architecture could transfer to other bioacoustic monitoring tasks—such as varroa infestation or queen presence—especially if datasets with per-segment noise annotations become available to test robustness directly."],"forward_implications":["Hive-strength monitoring could shift from sparse manual inspections to continuous acoustic estimates with errors around one frame of bees in familiar settings and about three frames on entirely unseen hives.","Models trained on modulation tensorgrams transfer to new hives better than spectrogram or MFCC models, which matters for deploying a single system across many apiaries.","Because the representation tolerates background noise without aggressive preprocessing, in-the-wild deployments may require less audio enhancement.","The best-performing recurrent model is light enough (roughly 4.5 MB and a few milliseconds of inference on a consumer CPU) to make on-hive edge deployment plausible."],"fun_headline_variants":["Modulation tensorgrams boost bee-hive acoustic monitoring","Bee colony strength from acoustic time dynamics not just spectra","Keeping time axis in audio features sharpens bee hive predictions","Recurrent nets on modulation tensors generalize across bee hives","Audio temporal dynamics lift bee colony strength prediction"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper assigns every 15-minute audio segment a colony-strength label by linearly interpolating manual 'frames of bees' counts between sparse inspections (Section III-A), and it acknowledges in Section IV-F that seasonality and environmental factors may confound the acoustic signal; if true colony strength changes nonlinearly between inspections, or if the interpolation mostly encodes time-of-year trends, the reported accuracy and cross-hive gains could partly measure track","fun_headline_variants_meta":{"raw":{"variants":["Modulation tensorgrams boost bee-hive acoustic monitoring","Bee colony strength from acoustic time dynamics not just spectra","Keeping time axis in audio features sharpens bee hive predictions","Recurrent nets on modulation tensors generalize across bee hives","Audio temporal dynamics lift bee colony strength prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000451,"raw_usage":{"total_tokens":2129,"prompt_tokens":789,"completion_tokens":1340,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":533,"completion_tokens_details":{"reasoning_tokens":1277}},"tokens_in":533,"tokens_out":1340,"duration_ms":8900,"temperature":1.0,"reasoning_tokens":1277,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T09:58:27.776554+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the model's prediction error separately for segments recorded close to manual inspections versus far from them: if error rises systematically with time since the nearest inspection, the model is tracking interpolated labels. Alternatively, retrain with explicit time-of-day and calendar features and see whether the modulation tensorgram still improves correlation beyond those covariates; if it does not, the temporal dynamics may be seasonal rather than acoustic.","supporting_citations":[],"review_version":1}