SoilSound: Smartphone Soil Moisture Sensing

• SoilSound is a smartphone-based, non-contact system that estimates near-surface soil moisture using acoustic chirps and inertial sensor data.
• It transmits frequency-modulated chirps and records echo reflections during a vertical scan to create a structured height–range image processed by a lightweight CNN.
• Field evaluations across 10 locations report a mean absolute error of 2.39% in diverse soil types, demonstrating effective, calibration-free performance.

SoilSound is a smartphone-based soil moisture sensing system that uses a phone’s built-in speaker, microphone, camera, and inertial sensors to estimate near-surface soil moisture without disturbing the soil. It operates by transmitting frequency-modulated acoustic chirps toward the ground, recording reflections during a short vertical scan, converting the measurements into a structured height–range representation, and running a lightweight convolutional neural network entirely on-device. In outdoor evaluation across 10 locations, it achieved a mean absolute error of 2.39%, with total processing time after the scan of $745 \pm 89$ ms and a deployed model size of 2.2 MB (Gao et al., 11 Sep 2025).

1. Problem domain and system definition

SoilSound is framed as a response to the practical limits of existing soil moisture measurement. Probe-based systems such as TDR, FDR, capacitance probes, and neutron probes can be accurate, but they disturb the soil during measurement because they must be inserted into it, often to depths of 10–20 cm. Non-contact alternatives generally depend on specialized gear such as ground-penetrating radar, SDR-based RF systems, buried antennas, RFID tags, radar reflectors, or custom acoustic hardware. Smartphone camera approaches are more accessible, but color varies strongly with soil organic matter and mineral composition, making it a poor proxy for moisture (Gao et al., 11 Sep 2025).

Within that context, SoilSound defines a commodity-phone pathway to non-contact, non-invasive, calibration-free moisture estimation. The measurement is active rather than passive: the phone emits chirps, records echoes, and exploits the user’s motion during a short scan. “Calibration-free” is used in the specific sense that end users do not need to calibrate the phone to each field, soil type, or location before use. “Non-invasive” means the phone does not require inserting probes, burying tags, placing reflectors, digging soil, or moving soil into a container (Gao et al., 11 Sep 2025).

The system targets near-surface moisture rather than deep profiling. That scope is central to its design: the method is intended for accessible monitoring of top-layer soil state with ordinary mobile hardware, not for root-zone tomography or deep subsurface inversion.

2. Reflection model and moisture-dependent roughness

The physical basis of SoilSound departs from earlier acoustic soil sensing work that relied on transmissive properties of sound in soil. Instead of coupling sound into the soil mass, SoilSound focuses on reflected sound from the air–soil interface. The paper argues that moisture information is not primarily carried by a large change in the classical planar-interface reflection coefficient, but by roughness-mediated changes in the balance between diffuse scattering and specular reflection (Gao et al., 11 Sep 2025).

The paper quantifies the limitation of the classical impedance-mismatch view. Using $Z_{\text{air}} = 415$ Rayls, $Z_{\text{dry}} \approx 5.85 \times 10^5$ Rayls, and $Z_{\text{sat}} \approx 3.0 \times 10^6$ Rayls, the planar-interface model yields $R_0 = 0.9972$ for dry soil and $R_0 = 0.9994$ for saturated soil, only a 0.22% change. SoilSound’s measurements contradict the implication that moisture should barely affect reflection amplitude. The proposed explanation is that dry soil is granular and rough, causing significant scattering, whereas increased moisture reduces surface roughness through capillary binding and cohesion, thereby strengthening specular reflection (Gao et al., 11 Sep 2025).

The rough-surface model is expressed with the Kirchhoff approximation. At normal incidence, specular reflection power is modeled as

$$P_{\text{spec}} = P_{\text{inc}} \cdot R_0 \cdot \exp(-g^2),$$

with Rayleigh roughness parameter

$g = \frac{2\pi \sigma_h}{\lambda}.$

For sandy soil at 15 kHz, with wavelength approximately $\lambda = 23$ mm, the paper gives example roughness values of 5–10 mm in dry soil and 0.5–1 mm in wet soil. Under those assumptions, $g$ can drop from roughly 1.9 to 0.4, producing about a $Z_{\text{air}} = 415$0 increase in specular reflection power (Gao et al., 11 Sep 2025).

The receiver does not observe only the specular path. It receives the direct near-normal reflection together with diffuse scattering from many surface facets, and the contribution of diffuse scattering depends on phone height. For $Z_{\text{air}} = 415$1, the effective solid angle scales approximately as

$Z_{\text{air}} = 415$2

This is why the vertical scan is not a mere convenience: changing height changes the captured mixture of specular and scattered energy, yielding a structured measurement that a CNN can learn (Gao et al., 11 Sep 2025).

A common misconception is that extended range bins correspond to subsurface penetration. The paper explicitly rejects that interpretation. Because the air–soil transmission coefficient satisfies $Z_{\text{air}} = 415$3, less than 0.3% of incident power enters the soil, and soil attenuation at 10 kHz in moist soil is cited as 50–80 dB/m, subsurface echoes are treated as negligible (Gao et al., 11 Sep 2025).

3. Measurement protocol and signal processing pipeline

The transmitted waveform is a sequence of 10 ms chirps sweeping linearly from 7 to 22 kHz, followed by a 5 ms silent gap, sampled at 48 kHz. During a typical 5-second scan, the system collects about 100–200 chirps while the user moves the phone upward by about 15 cm. Height is estimated using ARCore visual-inertial odometry at about 30 Hz, and the app checks whether the scan speed stays between 1 and 5 cm/s, with 3 cm/s considered optimal. Scans can be rejected if velocity variation exceeds 2 cm/s, and the scan stops automatically after 15 cm because reflected signal power becomes too weak beyond that point (Gao et al., 11 Sep 2025).

Synchronization is based on the inserted silent gaps rather than on direct correlation with the transmitted chirp. The paper computes short-time energy with a window $Z_{\text{air}} = 415$4 samples, detects a chirp boundary when the ratio between subsequent high-energy and preceding low-energy segments exceeds 8:1, verifies that the pattern repeats every 720 samples, and obtains synchronization lock after three consecutive consistent cycles, typically within 45 ms (Gao et al., 11 Sep 2025).

After synchronization, each received chirp is dechirped by multiplying it with the complex conjugate of the transmitted chirp. A Blackman window is applied, then a 1920-point discrete Fourier transform is computed, and frequency bins are mapped to distance. With $Z_{\text{air}} = 415$5 kHz, the implementation reports approximately 1.14 cm range resolution after DFT processing. Direct-path interference from the colocated speaker and microphone is removed using a template-based cancellation method: during initialization, 50 range profiles are averaged in under 1 second to build a direct-path template, which is then subtracted from each measurement (Gao et al., 11 Sep 2025).

The scan output is organized as a fixed $Z_{\text{air}} = 415$6 height–range image. Each row is a range profile from one chirp; the rows are ordered over scan height. Because a scan may contain 100–200 chirps, SoilSound uniformly samples 64 range profiles and crops each to the first 64 range bins. The paper then applies mode-based noise removal columnwise,

$Z_{\text{air}} = 415$7

and standardizes the resulting matrix before inference (Gao et al., 11 Sep 2025).

Ground truth in the lab and field is obtained using an AROYA Solus 3-in-1 analyzer with a TEROS 12 capacitance-based probe. Because the probe is point-based and local variation is unavoidable, five nearby measurements are taken for each sample or location, producing a ground-truth range $Z_{\text{air}} = 415$8 rather than a single scalar label (Gao et al., 11 Sep 2025).

4. Learning architecture, loss design, and reported performance

The model is a compact CNN for regression from a single-channel $Z_{\text{air}} = 415$9 height–range image to volumetric water content. It contains three convolutional blocks, each composed of a convolution, batch normalization, and LeakyReLU activation with slope $Z_{\text{dry}} \approx 5.85 \times 10^5$0, followed by $Z_{\text{dry}} \approx 5.85 \times 10^5$1 max pooling. Channel depth increases from 1 to 16 to 32 to 64. After flattening, the network uses fully connected layers of sizes $Z_{\text{dry}} \approx 5.85 \times 10^5$2 with dropout regularization (Gao et al., 11 Sep 2025).

Training does not use ordinary MSE against a single scalar label. Instead, the paper adopts a range-aware error borrowed from CoMEt. If prediction $Z_{\text{dry}} \approx 5.85 \times 10^5$3 lies within the measured range $Z_{\text{dry}} \approx 5.85 \times 10^5$4, the error is zero; otherwise it is the distance to the nearest boundary. Training minimizes the mean squared VWC-error. Raw data collected by the Android app is streamed to a laptop over TCP via ADB reverse tunneling; training is done in PyTorch on an NVIDIA RTX 3090 GPU, exported to TorchScript, quantized, and deployed via PyTorch Mobile on a Samsung Galaxy S10e CPU (Gao et al., 11 Sep 2025).

In within-soil laboratory evaluation, the authors used leave-one-out cross-validation across nine loamy sand moisture levels. Over moisture levels spanning roughly 12.7% to 37.2% VWC, SoilSound achieved an overall MAE of 5.01%. For moisture levels inside the training range, average MAE was 3.78%; for extrapolation beyond the observed range, error increased to 9.32%. In cross-soil generalization, training on all loamy sand samples and testing directly on unseen organic potting soil with no calibration or fine-tuning yielded an average MAE of 3.65% across five moisture levels (Gao et al., 11 Sep 2025).

The vertical scan mechanism is empirically decisive. A fixed-height alternative trained only on data from 3 cm above the soil performed badly on unseen organic soil, with an average MAE of 6.81% and nearly constant predictions around 16% VWC despite actual moisture ranging from 8.5% to 32.9%. This is presented as evidence that the multi-height scan is central to generalization rather than a secondary design choice (Gao et al., 11 Sep 2025).

Field evaluation provides the headline result. In one outdoor location with progressive watering, ground-truth moisture values of 15.9%, 18.9%, and 33.3% VWC produced errors of 0.3%, 0.0%, and 4.1%, respectively. Across 10 distinct outdoor locations, SoilSound achieved an average MAE of 2.39%. Nine locations with ground-truth moisture between 10.9% and 25.8% had errors from 0.1% to 3.3%; one location, L2, at 27.47% VWC had a 9.1% error. The paper summarizes the operational range as 15.9% to 34.0% across multiple soil types, environments, and users. In a user study with five untrained participants on an unseen organic soil sample with ground-truth moisture $Z_{\text{dry}} \approx 5.85 \times 10^5$5, the average range-aware MAE was 1.14%, with the worst user at 1.6% (Gao et al., 11 Sep 2025).

On-device overhead is small relative to scan time. On the Galaxy S10e, preprocessing took $Z_{\text{dry}} \approx 5.85 \times 10^5$6 ms, inference took $Z_{\text{dry}} \approx 5.85 \times 10^5$7 ms, total post-scan processing took $Z_{\text{dry}} \approx 5.85 \times 10^5$8 ms, and reported energy per measurement was less than 0.01 mJ for the model itself (Gao et al., 11 Sep 2025).

5. Boundary conditions, failure modes, and interpretation

SoilSound measures only the top few centimeters of soil over a small area, less than about $Z_{\text{dry}} \approx 5.85 \times 10^5$9. It cannot profile deeper root-zone moisture in the manner of GPR or inserted probes. Its scan must be approximately vertical, over about 15 cm, and at a controlled speed between 1 and 5 cm/s, ideally around 3 cm/s. The implementation demonstrated in the paper is Android-only and depends on ARCore-compatible visual-inertial tracking; the reported hardware target is a Galaxy S10e (Gao et al., 11 Sep 2025).

Performance degrades at the edges of the tested moisture range, especially at high moisture or near saturation. The strongest reported failure case is thick grass coverage: in a grass-covered area with true moisture 46.73% VWC, SoilSound underestimated moisture severely, yielding an MAE of 18.52%. The explanation given is that the learned mapping assumes a bare-soil rough surface whose moisture changes alter the specular/scattering balance; thick grass introduces its own moisture, geometry, and scattering structure, so a separate model may be necessary (Gao et al., 11 Sep 2025).

The paper is explicit that “calibration-free” does not mean universally exact. Standing water, frozen soil, high salinity, very high organic matter, or dense vegetation can violate the assumptions of the reflection model. Device-to-device variation is not deeply studied, so cross-hardware transfer remains open. These are not minor caveats: they delimit the conditions under which the roughness-mediated reflection mechanism is expected to hold (Gao et al., 11 Sep 2025).

A further interpretive point is that SoilSound is not a transmissive subsurface sensor disguised as a reflection system. The usable signal is dominated by the air–soil interface and its near-surface scattering geometry. This makes the method accessible and mobile, but it also constrains what kind of hydrologic information can be recovered.

6. Related acoustic and multimodal research around SoilSound

The name “SoilSound” also appears in a broader research context that uses soil-related acoustics, vibration, sonification, and adjacent sensing modalities as reference points for non-invasive environmental monitoring. One directly related development is device-level sonification for soil moisture monitoring in water-scarcity contexts: an Internet of Sounds implementation formalizes on-device wavetable sonification so that both pitch and timbre communicate moisture conditions in real time or from logged data, directly on embedded sensing hardware rather than only after data are sent to an application or server (Roddy, 23 Sep 2025).

At a larger spatial scale, agroseismology extends the acoustic idea from smartphone reflection sensing to distributed field monitoring. Using distributed acoustic sensing and ambient seismic interferometry, long-term agricultural plots with controlled tillage and compaction histories were monitored at minute-to-hour timescales and meter-scale spacing over 40 hours. Rainfall caused sharp velocity drops where disturbed soils developed transient near-surface saturation, while drying and evapotranspiration caused large hysteretic velocity rebounds. The paper attributes these effects to rate-dependent suction stresses and dynamic capillary effects, implying that soil acoustics can encode hydromechanical state rather than only scalar moisture (Shi et al., 11 Sep 2025).

Structure-borne acoustic learning provides another adjacent line of development. SonicBoom uses an array of six passive piezoelectric contact microphones, GCC-PHAT features, spectrograms, proprioception, and multimodal transformer fusion to localize impulsive contacts on a cylindrical robot link, reporting 0.43 cm error on an in-distribution test set and 2.22 cm in a stationary-robot, human-strike zero-shot transfer setting. The technical brief included with that work identifies rigid agricultural tools, probes, and shanks as plausible transfer targets for SoilSound-style sensing, while also noting that soil is more heterogeneous, dissipative, and dispersive than the PVC-link setting used there (Lee et al., 2024).

Complementary non-acoustic modalities address related belowground inference problems. Commodity RF sensing of belowground tuber growth uses aboveground swept-frequency measurements in the 2.0–3.5 GHz band to extract Broadband Attenuation Integral, High-to-Low band ratio, spectral slope, and RippleVar, achieving up to 87.5% accuracy for day-indexed growth-stage classification and up to 95.0% accuracy for 5 cm-grid localization in a controlled sweet potato pot study (Li et al., 31 Jan 2026). RF soil moisture sensing via buried radar backscatter tags uses dielectric-delay estimation rather than acoustic reflection, reporting average accuracy within 0.01–0.03 $Z_{\text{sat}} \approx 3.0 \times 10^6$0 of ground truth and operation at a depth of at least 30 cm and up to 75 cm (Josephson et al., 2019). These results do not belong to SoilSound proper, but they show that geometry control, baseline normalization, and explicit treatment of moisture as a confounder are recurring design principles across non-invasive soil sensing.

Older acoustics work helps situate the physical backdrop. High-speed imaging of “silo music” demonstrated that in a flowing granular column the grains at different heights do not oscillate in phase; instead, disturbances propagate upward as sound waves whose velocity varies with position in the silo, indicating that granular acoustics encode local state rather than rigid-body motion alone (Borzsonyi et al., 2011). In a different application, an integrated photogrammetric-acoustic study of acoustic screens in sandy soils showed that barrier permeability changes sand deposition patterns and, in turn, frequency-dependent insertion loss, establishing a direct link between moving granular media and acoustic performance (Bravo et al., 2024). Together with SoilSound, these studies suggest a broader interpretation: soil acoustics is not limited to one moisture-estimation application, but spans interface reflection, structure-borne vibration, distributed seismics, granular wave propagation, and sonified environmental monitoring.