Commodity RF Sensing of Belowground Tuber Growth

Abstract

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.

Figures (5)

• Figure 1: Overview of aboveground radio sensing of belowground tubers.a, SWAMP (Subsurface Wireless Agricultural Monitoring Platform) uses aboveground radio-frequency (RF) antennas to sense tubers of sweet potato (Ipomoea batatas) non-invasively. b, Conceptual propagation paths. An aluminium barrier blocks the aboveground line-of-sight (LoS) path so that received signals are dominated by energy that interacts with the pot and soil volume and tubers.
• Figure 2: Experimental validation of band selection, sensing metrics, localization, and growth monitoring.a, Channel frequency response (CFR) samples over antenna separation, where the most stable contrast between the soil-only case and the soil-plus-tuber case concentrates in 2.0 to 3.5 GHz (red box; color indicates normalized amplitude). b, Penetration score $\Psi_P$ versus relative activity $\Psi_A/\Psi_P$ derived from the attenuation model, illustrating the penetration and sensitivity tradeoff across frequency (color indicates frequency). c, Long Term Evolution (LTE) link quality indicators measured along the lateral scan axis $x$, reported as mean with standard deviation. (Error bars show s.d.; $n=100$ samples per position from a 20 s window at 5 Hz unless stated otherwise.) Reference signal received power (RSRP), signal to interference plus noise ratio (SINR), modulation and coding scheme (MCS), data rate, and block error rate (BLER) are shown. Vertical dashed lines mark the ground truth tuber span and red boxes highlight the region where single metrics degrade relative to the fused estimate. d, Localization agreement with ground truth quantified by structural similarity index (SSIM) and mean squared error (MSE), where linear fusion improves SSIM and reduces MSE compared with any single LTE metric. e, Representative CFR derived features over the 45 day growth cycle for two soil and moisture conditions, showing consistent temporal evolution in H/L, BAI, Slope, and RippleVar. (Each curve shows one pot; $n=45$ days per pot.)
• Figure 3: A single metric cannot fully represent the ground-truth heatmap. Using the BLER generated heatmap as an example, the blue rectangles indicate regions where the reconstructed heatmap matches the ground truth, while the red rectangles indicate regions that do not. This mismatch motivates the need for a fusion metric to generate heatmaps that better align with the ground truth.
• Figure 4: LTE condition sensitivity and tuber counting signatures.a, Distributions of standardized LTE link indicators under three high moisture cases with increasing nitrogen level (N:L, N:M, N:H). Violin plots show the empirical density and overlaid points show individual measurements for RSRP, SINR, MCS, data rate, and BLER after z-score standardization, illustrating condition dependent shifts and within case variability. b, Grid aggregated LTE signatures for tuber counting across lateral position $x$ for three occupancy settings (0.5, 1, and 3 sweet potatoes). Curves show mean trends with shaded variability over repeated scans, demonstrating separable responses for sparse versus dense occupancy while the intermediate case is less consistently separated.
• Figure 5: System overview of SWAMP for sweet potato monitoring with two wireless data streams. A cart based measurement platform performs repeated scans of potted plants in a greenhouse. Swept frequency channel frequency response (CFR) is collected using a software defined radio (SDR) link and converted into CFR derived features for longitudinal growth stage monitoring. In parallel, cellular Long Term Evolution (LTE) link quality metrics are sampled over a planar grid and fused with a constrained linear model to produce a spatial probability map for tuber counting and localization. Representative excavations and reconstructed maps illustrate the growth baseline across days.