Commodity RF Sensing of Belowground Tuber Growth

Belowground yield-forming organs of root and tuber crops are difficult to measure during growth, and management therefore relies on aboveground proxies and destructive sampling. Aboveground wireless links could provide a low-cost, non-invasive alternative, but strong attenuation and soil-dependent variability make repeatable subsurface sensing challenging. In a controlled greenhouse pot study of sweet potato, we deploy aboveground antennas in a line-of-sight-suppressed geometry and collect daily swept-frequency channel spectra together with standardized cellular link indicators, revealing consistent frequency-dependent attenuation and rippling as tubers develop. Here, we show that swept-frequency measurements in the 2.0-3.5 gigahertz band yield four interpretable spectral features that classify day-indexed growth stages with up to 87.5% accuracy across two soil recipes and two moisture regimes, and that fusing cellular link-quality indicators enables 5-centimeter-grid tuber localization with up to 95.0% accuracy, providing a proof-of-concept for subsurface crop monitoring without buried sensors, and motivating validation across cultivars and larger soil volumes.

💡 Research Summary

This paper introduces a novel, low‑cost, non‑invasive approach for monitoring the development of below‑ground organs in root and tuber crops, using only above‑ground commodity radio hardware. In a controlled greenhouse experiment with sweet potato (Ipomoea batatas) grown in pots, the authors deployed two software‑defined radio (SDR) antennas in a line‑of‑sight‑blocked geometry and collected daily swept‑frequency channel frequency response (CFR) measurements across the 2.0–3.5 GHz band. Simultaneously, they recorded five standardized cellular link‑quality indicators—Reference Signal Received Power (RSRP), Signal‑to‑Interference‑plus‑Noise Ratio (SINR), Modulation and Coding Scheme (MCS), data rate, and Block Error Rate (BLER)—available from Long‑Term Evolution (LTE) equipment.

Band selection and physical insight
Through empirical analysis, the authors identified the 2.0–3.5 GHz range as the sweet spot where the contrast between soil‑only and soil‑plus‑tuber conditions is most repeatable across two soil recipes (sand‑based and loam‑based) and two watering regimes. A pronounced absorption trough near 2.68 GHz indicates strong interaction with emerging tuber tissue, while frequencies above 3.5 GHz show diminishing returns due to excessive attenuation. A simple multilayer attenuation model corroborates that this band balances sufficient penetration depth with heightened sensitivity to fine dielectric discontinuities introduced by growing tubers.

CFR feature engineering and growth‑stage classification
Four interpretable spectral features are extracted from each daily CFR sweep: (1) Broadband Attenuation Integral (BAI) – total loss across the band; (2) High‑to‑Low ratio (H/L) – differential attenuation between the upper and lower halves of the band; (3) Slope – linear fit to the log‑magnitude spectrum; and (4) Ripple Variance – variance of high‑frequency rippling caused by multipath scattering. These features evolve consistently over the 45‑day growth cycle, reflecting increasing biomass (higher BAI and Ripple Var) and shifting frequency dependence (decreasing H/L). Using these four features as inputs to a supervised classifier (the paper reports a multi‑class model, likely a random forest or SVM), the authors achieve up to 87.5 % accuracy in assigning each day’s measurement to its correct day‑index, regardless of soil type or moisture level. This demonstrates that the CFR captures a robust temporal signature of tuber development that is not confounded by the tested environmental variability.

Cellular link‑quality fusion for spatial localization
A single LTE metric proved insufficient to reconstruct the spatial extent of tubers; each metric (RSRP, SINR, MCS, data rate, BLER) showed localized degradations over the true tuber region but differed in magnitude and position. The authors therefore fused the five metrics into a single “Fusion Score” using a constrained linear combination (non‑negative weights summing to one). The fused map dramatically improves agreement with ground‑truth occupancy maps, raising the Structural Similarity Index (SSIM) and lowering Mean Squared Error (MSE) compared with any individual metric. When evaluated on an independent scan, the fused score yields up to 95 % accuracy for detecting tuber presence on a 5 cm grid, with the lowest condition still reaching 87.5 % accuracy. This demonstrates that spatial diversity (a raster scan) combined with multi‑metric fusion can compensate for the limited information content of any single link‑quality indicator.

Cost, setup, and maintenance considerations
The hardware consists of two SDR nodes and a laptop, costing roughly US $1,000–$3,000, far cheaper than buried probes (≈$5,000–$10,000) or ground‑penetrating radar systems (≈$15,000–$50,000). Setup time is under 20 minutes, and annual maintenance is estimated at $100–$200, mainly for periodic recalibration. This positions the system as a practical, scalable solution for research stations or farms that cannot afford invasive sensor networks.

Limitations and future work
The study is confined to a controlled pot environment with fixed antenna geometry, homogeneous soil volumes, and a single crop species. Real field conditions will introduce heterogeneous soil textures, variable canopy cover, dynamic moisture gradients, and larger depths, all of which could obscure the subtle CFR signatures observed here. Scaling up will likely require multi‑antenna arrays, adaptive beamforming, and more sophisticated calibration models that account for environmental dynamics. Moreover, validation across diverse tuber crops (potato, carrot, radish) and larger plots is essential before commercial deployment.

Conclusion
The authors provide compelling proof‑of‑concept evidence that commodity above‑ground RF links, when combined with swept‑frequency channel measurements and standard LTE link‑quality metrics, can simultaneously monitor temporal growth stages and generate coarse spatial maps of below‑ground tuber development. The approach offers a non‑invasive, low‑cost, and potentially scalable alternative to traditional soil probes or radar imaging, opening new avenues for precision agriculture and breeding programs that require longitudinal phenotyping of root and tuber traits. Further field validation and algorithmic refinement will be key to translating this promising laboratory result into a robust field‑ready technology.