{"id":"80abcda8-8d3a-4e7e-998b-561c4aa9185d","arxiv_id":"2509.09823","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A smartphone app estimates soil moisture from acoustic reflections during a vertical scan and reports 2.39% MAE in outdoor tests without soil disturbance or per-location calibration.","lead":"SoilSound measures soil moisture by pointing a smartphone at the ground, sweeping it upward while playing inaudible chirps, and analyzing the echoes with a neural network. The paper reports a mean absolute error of 2.39% across 10 outdoor locations without per-location calibration, potentially giving home gardeners and small farmers a free alternative to expensive probes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2.39% MAE is range-aware: Eq. 10 zeros any prediction inside the five-reading ground-truth spread, so the headline accuracy is not a conventional MAE and Table 1 comparisons may be overstated.","rationale":"I read the paper in good faith. The core idea—vertical-scan acoustic FMCW with a CNN mapping range-profile structure to soil moisture—is plausible, and the authors acknowledge important limitations (depth penetration, grass coverage, community scaling). The physical argument in Section 3 is internally coherent, and the fixed-height ablation in Section 6.2.1 is a genuine piece of supporting evidence for the multi-height design. However, the main advertised number is not a conventional MAE. Equation 10 zeros errors inside the measured ground-truth range, and the paper consistently reports this range-aware value as 'MAE.' Because the TEROS-12 sensor has its own calibration MAE of 3% and the paper does not report the widths of the five-reading ranges, the 2.39% field result and the Table 1 comparison to conventional sensors are not yet auditable. The reader's weakest_assumption focused on generalizability of the roughness model; my concern is more immediate: the quantitative accuracy claim is metric-dependent. This does not change the conditional verdict, but it identifies the specific condition that should be attached: the authors should provide conventional errors and per-location ground-truth spreads. I am not alleging any impropriety; the limitation section shows the authors are aware of boundary conditions, and the missing information may simply reflect space constraints.","tokens_in":16041,"tokens_out":7879,"duration_ms":70263,"concrete_test":"Release per-scan predictions and the five raw TEROS-12 readings for each of the 10 field locations (plus lab and user data), then recompute the field error as conventional MAE and RMSE of each prediction against the mean of the five readings, and report the width of each ground-truth range. If conventional MAE stays near 2.4%, the concern is resolved; if it rises materially (e.g., above 4-5%), the headline accuracy and Table 1 comparison need to be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim rests on the range-aware VWC-error (Eq. 10), which assigns zero error to any prediction inside [θ_min, θ_max] from five TEROS-12 readings. Every reported MAE—lab 5.01%, cross-soil 3.65%, user 1.14%, field 2.39%—uses this metric. This is not the same quantity as the conventional MAEs listed in Table 1 (TDR 1-4%, CoMEt 1.1%, SoilSound 2.39%). The ground-truth sensor itself has a calibration MAE of 3% [8]; five readings at nearby points can produce a wide range, especially in heterogeneous field soil. A prediction that is actually several percent away from the true mean can be counted as zero error whenever it falls inside the measured spread. The paper never reports the widths of these ground-truth ranges, so the reported 2.39% cannot be compared with prior work or audited. This directly weakens the headline claim and the cross-location generalizability claim, since the same lenient metric is used throughout. This is a methodological concern rather than a claim about the physical model; the roughness-scattering story may be correct, but the empirical evidence for the advertised accuracy is currently under-specified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents SoilSound, a smartphone-based acoustic sensing system for estimating volumetric soil water content without inserting probes. The phone transmits 7-22 kHz FMCW chirps while the user moves it upward over about 15 cm; reflection range profiles at successive heights are stacked into a 64x64 image and processed by a CNN that regresses moisture. The manuscript proposes a surface-roughness explanation for why reflections grow with moisture (Section 3), describes synchronization, direct-path cancellation, mode-based noise removal, and an Android implementation, and reports a lab LOOCV MAE of 5.01%, a cross-soil MAE of 3.65%, a five-user MAE of 1.14%, and a field MAE of 2.39% across 10 outdoor locations. The central claim is accurate, calibration-free, non-invasive soil moisture estimation across soil types, environments, and users.","tokens_in":16343,"tokens_out":5233,"duration_ms":43512,"significance":"If the reported accuracy were measured on the same basis as prior work, SoilSound would be a noteworthy demonstration of commodity-smartphone acoustic reflectometry for soil moisture, with practical accessibility advantages. The paper deserves credit for implementing a complete on-device pipeline, for the vertical-scan ablation in Section 6.2.1 showing that single-height features overfit soil type, and for candidly reporting limitations such as grass coverage and depth penetration. However, the central quantitative claim is weakened by the range-aware error metric and by the small number of soil types and field locations; the physical model in Section 3.2 is plausible but not directly validated. The contribution is promising, but the evidence as reported does not yet substantiate the headline 2.39% MAE in a way that is comparable with Table 1.","major_comments":[{"comment":"The reported MAE values, including the headline 2.39% field figure, use a range-aware VWC error that assigns zero error whenever the prediction falls inside the interval spanned by five TEROS-12 readings. This is not the same quantity as the conventional MAEs listed in Table 1; because the TEROS-12 itself has a calibration MAE of about 3% [8], wide ground-truth intervals can make predictions several percent from the true mean count as zero error. The paper never reports the widths of the ground-truth ranges, so the reader cannot compare SoilSound with prior systems or audit the headline claim. Please report conventional MAE against the mean (or midpoint) of the ground-truth readings, together with the distribution of range widths, and use that metric in Table 1.","section":"Section 4.4.2, Eq. (10)"},{"comment":"The generalization claims rest on very small samples: nine loamy-sand samples for LOOCV, five potting-soil samples for cross-soil transfer, five users, and ten field locations with no reported soil classification. Only two base soil types are prepared in the lab, and the outdoor locations are not described in terms of texture or organic matter. With this sample size, the claim in Section 6.1.2 that the model learned a moisture-invariant mapping across soil types is not strongly supported; the fixed-height ablation demonstrates the value of vertical scans but does not identify the acoustic features that generalize. Please provide per-location soil descriptions, confidence intervals for the reported MAEs, and ideally an independent held-out soil type from a different textural class.","section":"Sections 6.1 and 6.4"},{"comment":"The mechanistic claim that moisture-dependent surface roughness is the dominant cause of the observed reflection changes is supported only by illustrative parameter values (alpha = 3-5, beta = 1.5, sigma_h_dry = 5-10 mm, sigma_h_sat = 0.5-1 mm) taken from the soil-mechanics literature; these parameters are not measured for the tested soils, nor is the roughness-moisture relation directly validated. Since the CNN is trained on labeled samples, the learned mapping could in principle exploit soil-specific acoustic signatures rather than the hypothesized roughness effect. I am not asking for a full physical inversion, but the paper should provide either direct surface-roughness measurements or a synthetic-data test using Eq. (5) to show that the model's learned features are consistent with the proposed mechanism.","section":"Section 3.2.1, Eq. (5)"}],"minor_comments":[{"comment":"The phrase \"an xx camera\" is an unresolved placeholder and should be replaced with the actual camera specification.","section":"Section 5"},{"comment":"The sentence \"We then freeze this model and use it for training for all subsequent experiments\" is confusing; presumably the frozen model is used for evaluation, not for training, in later sections.","section":"Section 6.1"},{"comment":"The statement that the system achieved \"0.0% error\" should be clarified as meaning the prediction fell inside the range-aware ground-truth interval, and the width of that interval should be given.","section":"Section 6.4.1"},{"comment":"\"Taro probe\" appears to be a typo; please clarify which probe was used to verify moisture content before each experiment.","section":"Section 6.1"},{"comment":"There is a typo \"surafces\" for \"surfaces,\" and in several places \"Soilsound\" should be \"SoilSound.\"","section":"Section 6.4.3"},{"comment":"Add a footnote to the SoilSound row stating that the 2.39% figure is a range-aware MAE defined by Eq. (10), not a conventional MAE, so readers are not misled by the comparison.","section":"Table 1"},{"comment":"The sentence \"Several works have showed that the soil moisture can be mapped by accurately by determining the velocity\" contains a grammatical error; please revise.","section":"Section 2.3"},{"comment":"Reference [4] is incomplete and reference [22] lacks full publication details; please complete all bibliographic entries.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The range-aware metric issue is the main gate for this paper: without a conventional MAE and ground-truth range widths, the headline accuracy cannot be compared with Table 1. If the authors cannot provide those numbers, the paper should not be accepted. The paper also releases no data or code; given the small sample sizes, I would encourage the authors to share at least range-profile examples and per-location errors to make the claims auditable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me cut to the chase. SoilSound is the first smartphone-only reflective acoustic soil moisture sensor, using a vertical scan and a CNN on the 2D height-range image. That part is real and worth taking seriously. But the headline 2.39% MAE is not the same quantity as the conventional MAE numbers in Table 1. Equation 10 defines a range-aware error that counts any prediction inside the spread of five TEROS-12 readings as zero, and those five readings are themselves noisy — the sensor's own calibration MAE is about 3% [8]. So the sub-3% claims are less impressive than they look, and the comparison to prior work is apples-to-oranges.\n\nWhat the paper does well: the fixed-height ablation is the strongest piece of evidence. A model trained on a single height fails on unseen soil, while the vertical-scan model generalizes. That supports the core design. The roughness-scattering story (Section 3) is plausible and nicely motivated by the impedance-mismatch wall. The user study, though small, is honest. The limitations section actually names grass, salt, frozen soil, and depth penetration, which is more than many systems papers do.\n\nThe soft spots beyond the metric: the evaluation is tiny. Nine lab samples, one unseen soil type (potting soil), ten outdoor points, five users. The cross-soil claim rests on a single transfer test. No code or data is released, so the CNN's moisture-invariant features can't be audited. Eq. 5's parameters come from the literature, not from fitting; the physical model is a motivating mechanism, not a validated predictor. That's fine, but it should be stated more sharply.\n\nBottom line: this is a decent systems paper with a real novelty, a plausible mechanism, and an honest limitations section. The evaluation design is the weak link. A referee should push for a conventional error metric (or at least report the ground-truth range widths), more soil types, and ideally release the data. I'd send it to peer review rather than desk reject, but it needs work before the 2.39% number can be taken at face value.\n\nFor your reading group: maybe. I'd cite the vertical-scan approach if I worked in acoustic or mobile sensing.","headline":"SoilSound's vertical-scan reflective acoustic approach is a genuine novelty, but its headline 2.39% MAE rests on a lenient range-aware metric that overstates accuracy relative to prior work.","tokens_in":16870,"tokens_out":2323,"would_cite":true,"duration_ms":19955,"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":"SoilSound claims that a commodity smartphone, by emitting FMCW acoustic chirps toward the soil and recording their reflections during a five-second vertical lift, can estimate volumetric soil moisture without any calibration or soil…","keywords":["soil moisture sensing","smartphone acoustic sensing","FMCW chirp","surface roughness model","specular reflection","convolutional neural network","vertical scan","non-invasive agriculture"],"falsifier":"A direct test would be to measure the specular reflection strength and the surface roughness simultaneously on the same soil sample as it dries, using a profilometer or confocal microscopy for $\\sigma_h$ and the FMCW setup for $P_{\\mathrm{spec}}$, and check whether $\\ln(P_{\\mathrm{spec}})$ declines linearly with $\\sigma_h^2$ across moisture levels and soil types. If the relationship breaks when soil texture or organic matter changes, then the Eq. 5 model, not moisture, is what the CNN has learned.","tokens_in":15854,"feed_emoji":"📱","tokens_out":4990,"duration_ms":393486,"temperature":0.7,"pith_summary":"SoilSound proposes that a commodity smartphone can estimate soil moisture by emitting FMCW acoustic chirps toward the ground and recording their reflections while the phone is lifted slowly. The paper argues the physical link is surface roughness: wetter soil is smoother, so it reflects more sound specularly, and a CNN trained on 2D range-profiles from vertical scans learns to read moisture from this signature. The system requires no calibration, no probes in the soil, and no specialized hardware, and the field evaluation reports a mean absolute error of 2.39% across ten outdoor locations and tracks moisture from 15.9% to 34.0%.","feed_headline":"Your phone's chirp can gauge soil moisture without digging","feed_subtitle":"No probes, no calibration: a 5-second vertical scan reads soil moisture with a mean error of 2.39%.","key_machinery":"The central object is the moisture-dependent surface roughness model, Eq. (5): $\\sigma_h(\\theta_v) = \\sigma_{h,\\mathrm{dry}}\\exp(-\\alpha S_e^\\beta) + \\sigma_{h,\\mathrm{sat}}$, where $S_e$ is effective saturation, combined with the Rayleigh roughness parameter $g = 2\\pi\\sigma_h/\\lambda$ and the Kirchhoff-approximation specular reflection law $P_{\\mathrm{spec}} = P_{\\mathrm{inc}} R_0 \\exp(-g^2)$. This pair turns moisture content into a measurable acoustic reflection amplitude. The vertical scan adds the second ingredient: because the collection solid angle $\\Omega_{\\mathrm{eff}}(h) \\propto (a/h)^2$ narrows with height, multi-height range profiles re-weight the specular versus diffuse components, giving the CNN independent constraints and preventing overfit to a single height.","core_discovery":"The central claim is that soil moisture can be measured from the reflected acoustic signal at the air-soil interface, using the roughness-dependent specular reflection rather than the transmission of sound through soil. A dry soil surface is rough and scatters sound diffusely; as water binds particles through capillary forces, the surface smooths and the specular reflection strengthens. The paper models the moisture-dependent roughness with an exponential form and shows that at common audio wavelengths the Rayleigh roughness parameter drops from about 1.9 to 0.4, multiplying the specular power by a large factor while the plane-wave reflection coefficient stays almost constant. A convolutional network trained on the combined range-bin and device-height structure of the reflections then estimates volumetric water content. The paper reports an MAE of 2.39% across ten outdoor locations and argues the system generalizes across soil types because the vertical scan supplies geometric diversity that prevents the model from latching onto soil-specific signatures.","pith_inferences":["The same specular-versus-diffuse reasoning could extend to other granular surfaces (sand, snow, mulch) where water or cohesion changes roughness, giving a general phone-based surface-moisture probe.","A testable extension is to use the model's activation map directly to estimate the Rayleigh parameter $g$ per height, potentially yielding a physics-grounded moisture estimate that needs no training data at all.","The observed 9.1% error at one high-moisture outdoor location and degraded extrapolation performance suggest the CNN may encode the training moisture distribution rather than the full $\\exp(-g^2)$ curve; a dataset spanning saturation for each soil type would show whether the model is truly learning the physics.","Because transfer to high-organic soil worked without fine-tuning, collecting crowdsourced scans with the app itself could let the model adapt to new soil classes through federated learning, which the paper mentions as future work."],"forward_implications":["A user can point a phone at bare soil, lift it steadily for five seconds, and get a moisture reading in under one second of processing, with no calibration or soil disturbance.","Because the model relies on reflection, it reads only the top few centimeters of soil; subsurface layers remain effectively invisible.","The system works across soil types without retraining: a model trained on loamy sand transferred to organic potting soil, and field errors stayed low at nine of ten outdoor locations.","The approach opens the 1–22 kHz audio band of commodity phones to material characterization, not just ranging.","Grass cover breaks the assumption: predictions on thick grass consistently underestimate moisture (MAE 18.52%)."],"supporting_citations":[{"why":"provides the Kirchhoff-approximation specular reflection law with the Rayleigh roughness parameter that turns surface roughness into an expected reflection power.","marker":"[9]"},{"why":"supplies the original rough-surface scattering theory from which the Kirchhoff approximation for specular reflection is taken.","marker":"[5]"},{"why":"supplies the unsaturated-soil mechanics result that capillary forces bind particles and generate cohesion as water content rises, the physical basis for moisture smoothing the surface.","marker":"[22]"},{"why":"provides experimental evidence that apparent soil cohesion increases with water content, supporting the moisture-dependent roughness model of Eq. (5).","marker":"[28]"},{"why":"gives the acoustic impedance, attenuation, and speed-of-sound values for soils used to argue that the plane-wave reflection coefficient stays nearly constant while penetration is negligible.","marker":"[24]"},{"why":"supplies typical dry-soil surface roughness values for sandy soils used in the Rayleigh parameter calculation.","marker":"[4]"},{"why":"defines the range-aware VWC error used as the training loss and evaluation metric to handle ground-truth sensor uncertainty.","marker":"[18]"},{"why":"supplies the template-based direct-path cancellation technique adapted to remove speaker-to-microphone interference in the range profiles.","marker":"[32]"},{"why":"is the commercial capacitance sensor (TEROS 12) used to establish ground-truth moisture ranges for training and evaluation.","marker":"[2]"}],"fun_headline_variants":["SoilSound: phone's chirp maps soil moisture","Acoustic reflections estimate soil moisture on smartphone","No probes: phone speaker scans soil for moisture","Chirp scan reads soil wetness, zero calibration","Smartphone audio senses soil moisture from surface"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the exponential, moisture-dependent surface roughness law (Eq. 5) with its literature-derived parameters is the dominant, generalizable mechanism by which water content changes the acoustic reflection, and that a CNN trained on nine loamy sand samples plus one transfer test on potting soil has learned this moisture-invariant mapping rather than soil-specific acoustic signatures.","fun_headline_variants_meta":{"raw":{"variants":["SoilSound: phone's chirp maps soil moisture","Acoustic reflections estimate soil moisture on smartphone","No probes: phone speaker scans soil for moisture","Chirp scan reads soil wetness, zero calibration","Smartphone audio senses soil moisture from surface"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000481,"raw_usage":{"total_tokens":2392,"prompt_tokens":972,"completion_tokens":1420,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":1348}},"tokens_in":588,"tokens_out":1420,"duration_ms":13125,"temperature":1.0,"reasoning_tokens":1348,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:57:41.545172+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would be to measure the specular reflection strength and the surface roughness simultaneously on the same soil sample as it dries, using a profilometer or confocal microscopy for $\\sigma_h$ and the FMCW setup for $P_{\\mathrm{spec}}$, and check whether $\\ln(P_{\\mathrm{spec}})$ declines linearly with $\\sigma_h^2$ across moisture levels and soil types. If the relationship breaks when soil texture or organic matter changes, then the Eq. 5 model, not moisture, is what the CNN has learned.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"is the commercial capacitance sensor (TEROS 12) used to establish ground-truth moisture ranges for training and evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides the Kirchhoff-approximation specular reflection law with the Rayleigh roughness parameter that turns surface roughness into an expected reflection power."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the original rough-surface scattering theory from which the Kirchhoff approximation for specular reflection is taken."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the unsaturated-soil mechanics result that capillary forces bind particles and generate cohesion as water content rises, the physical basis for moisture smoothing the surface."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"provides experimental evidence that apparent soil cohesion increases with water content, supporting the moisture-dependent roughness model of Eq. (5)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives the acoustic impedance, attenuation, and speed-of-sound values for soils used to argue that the plane-wave reflection coefficient stays nearly constant while penetration is negligible."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies typical dry-soil surface roughness values for sandy soils used in the Rayleigh parameter calculation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"defines the range-aware VWC error used as the training loss and evaluation metric to handle ground-truth sensor uncertainty."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the template-based direct-path cancellation technique adapted to remove speaker-to-microphone interference in the range profiles."}],"review_version":2}