MoistureMapper: Autonomous Soil Moisture Mapper

• MoistureMapper is an autonomous mobile robot that integrates TDR sensing, RTK-GPS localization, and adaptive Gaussian Process mapping to reconstruct high-resolution soil moisture distributions.
• It employs a two-phase workflow with coarse grid sampling followed by uncertainty-driven, iterative measurement to optimize the balance between map fidelity and traversal cost.
• Quantitative evaluations reveal enhanced mapping accuracy with RMSE reduction to 12.2% and efficient adaptive routing, demonstrating its potential as a field-scale validation tool.

MoistureMapper is an autonomous mobile robot for high-resolution soil moisture mapping at scale. The system combines an unmanned ground vehicle, Time Domain Reflectometry (TDR) sensing, a direct push drill mechanism for soil insertion, RTK-GPS localization, and Gaussian Process-based adaptive sampling to reconstruct the spatial distribution of volumetric water content from sparse in situ measurements. Its stated motivation is that existing methods are not suitable for scale applications due to large deployment costs in high-resolution sensing applications such as variable irrigation, and its reported contributions span hardware design, probabilistic mapping, large-scale simulations, and proof-of-concept field deployment (Rose et al., 17 Jul 2025).

1. Operational setting and problem formulation

MoistureMapper addresses a classical tension in soil moisture sensing: dense measurements improve map fidelity, but each physical sample incurs traversal and sensing cost. In the MoistureMapper formulation, the objective is not merely to collect measurements, but to learn a field-scale function $f(x) \approx \theta_v(x)$ over a continuous domain $X \subset \mathbb{R}^2$, where $\theta_v$ denotes volumetric water content. The robot therefore treats sampling as a coupled sensing-and-motion problem rather than as uniform survey coverage (Rose et al., 17 Jul 2025).

The system formalizes this with a cost functional

$$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$

while seeking to reduce posterior uncertainty across the field. This framing is significant because it places MoistureMapper between two established extremes: dense manual TDR campaigns, which can deliver high-fidelity local measurements but scale poorly, and passive remote products, which scale well but often lack farm-level spatial granularity. A plausible implication is that the platform is best understood as a targeted, adaptive ground-truthing instrument rather than as a replacement for all other moisture products.

The robot’s workflow is explicitly two-phase. Phase I performs coarse grid sampling $M_0$ over the field to fit an initial Gaussian Process (GP). Phase II performs iterative sampling using an acquisition function, so that subsequent waypoints are chosen from the evolving uncertainty structure of the inferred moisture map rather than from a fixed survey pattern. This makes map construction inseparable from sequential experimental design.

2. Robotic platform and sensing subsystem

The physical platform is built around an AgileX Robotics Scout 2.0 unmanned ground vehicle, a four-wheeled, off-road capable platform with skid steering. Onboard compute is provided by an NVIDIA Jetson AGX Orin with 64 GB RAM, and localization is provided by dual SwiftNav Duro Inertial RTK-GPS modules with 0.01 m positional accuracy, communicating via 902–928 MHz serial radios with 6 dBi omnidirectional antennas. Waypoint following is implemented via onboard PID controllers (Rose et al., 17 Jul 2025).

Subsystem	Specification	Function
Base vehicle	AgileX Robotics Scout 2.0; skid steering; 30 Ah lithium-ion drive battery; ≈3.5 hours of continuous locomotion	Off-road mobile platform
Compute	NVIDIA Jetson AGX Orin, 64 GB RAM	Autonomy stack and data logging
Localization	Dual SwiftNav Duro Inertial RTK-GPS; 0.01 m positional accuracy	Georeferencing and waypoint navigation
Moisture sensor	SoilMoisture Equipment Corp. MiniTrase mobile TDR; 0–100 % VWC; 10 ps resolution; ±2 % accuracy; 8–20 cm waveguides	In situ volumetric water content sensing
Probe deployment	Two synchronized DC linear actuators; 300 mm stroke; 3 000 N push force each	Direct-push drill insertion and retraction

The TDR subsystem is central to the design. An electromagnetic pulse is sent down metallic prongs, and the round-trip time $\Delta t$ is proportional to the bulk dielectric permittivity $\epsilon_b$ of the soil, which in turn correlates with volumetric water content $\theta_v$. MoistureMapper does not rely on the raw instrument output alone; instead, it calibrates the mobile TDR against a stationary hand-held TDR in laboratory soil buckets using the quadratic model

$$\theta_v = a + b \cdot \mathrm{TDR}_{\mathrm{meas}} + c \cdot \mathrm{TDR}_{\mathrm{meas}}^2.$$

The coefficients $a,b,c$ are obtained by least-squares regression, and the reported best-fit parameters yielded an average VWC error of $X \subset \mathbb{R}^2$0.

Probe deployment is performed by a direct-push drill (DPD) mechanism. Two synchronized DC linear actuators, mounted in parallel on a 2020 aluminum extrusion frame, push the TDR waveguides to the target depth, trigger a waveform measurement, and then retract the probes. The insertion/retraction sequence is explicit: position the robot over a sample waypoint, energize the actuators, measure the TDR waveform and compute VWC via calibration, then reverse the actuators. This distinguishes MoistureMapper from non-contact platforms; its sensing accuracy derives from direct subsurface contact, but so do some of its mechanical constraints.

3. Gaussian Process map construction and adaptive sampling

MoistureMapper models soil moisture as a GP over the field. The prior uses zero mean,

$X \subset \mathbb{R}^2$1

and a squared exponential or radial basis function kernel,

$X \subset \mathbb{R}^2$2

where $X \subset \mathbb{R}^2$3 is the signal variance and $X \subset \mathbb{R}^2$4 is the length scale. Given a dataset $X \subset \mathbb{R}^2$5, noise variance $X \subset \mathbb{R}^2$6, kernel matrix $X \subset \mathbb{R}^2$7, and test-point kernel vector $X \subset \mathbb{R}^2$8, the posterior predictive mean and variance are

$X \subset \mathbb{R}^2$9

$\theta_v$0

Hyperparameters, including $\theta_v$1 and $\theta_v$2, are tuned by maximizing the log marginal likelihood of the initial dataset $\theta_v$3 (Rose et al., 17 Jul 2025).

The adaptive planner uses the GP variance as a proxy for information gain, but it augments pure uncertainty with travel cost. Two cost-aware acquisition functions are implemented:

$\theta_v$4

$\theta_v$5

The next sample is chosen as

$\theta_v$6

A randomized variant selects uniformly among the top 5 candidates to avoid local minima. The baseline “greedy” strategy ignores travel cost and selects the point of maximal predictive variance.

Algorithmically, the loop is straightforward. After the initial coarse grid, the robot repeatedly computes $\theta_v$7 and $\theta_v$8 over a candidate set, evaluates the acquisition function, selects the next point, plans a path from the current pose, executes a DPD measurement, updates the dataset, and retrains or updates the GP. Termination can depend on travel distance $\theta_v$9, sample budget $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$0, or maximum predictive uncertainty $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$1. The essential point is that the robot’s route is emergent from the inferred moisture field rather than pre-authored.

4. Quantitative evaluation

The simulation study spans domain sizes $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$2 m and 60 random ground-truth maps per size, including 1 uniform, 1 sloped linear-gradient, 5 pure Gaussian, and 5 hybrid maps with gradient-plus-Gaussian clusters. Cluster radii are uniformly sampled in $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$3, with 10 clusters per map. Stopping criteria include sample counts in $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$4, travel distances in $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$5 m, and maximum uncertainty $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$6 (Rose et al., 17 Jul 2025).

In these simulations, travel distance followed the ordering benchmark $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$7 $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$8 $$C(D_M) = \sum_i c_{\text{sample}}(x_i) + \sum_i c_{\text{travel}}(x_i, x_{i+1}),$$9 $M_0$0, and $M_0$1 achieved up to $M_0$2 reduction relative to the greedy benchmark. For map uncertainty, all methods converged similarly on large fields, while greedy was slightly better in small fields for maximum variance. For average variance, randomized $M_0$3 attained up to $M_0$4 lower average $M_0$5 over $M_0$6 relative to the benchmark. The reconstructed maps reportedly show that $M_0$7 and $M_0$8 better capture hot and cold spots with fewer long detours.

The field deployment was conducted at Dick Taylor Memorial Park, Reno, Nevada, on a $M_0$9 m grass plot. Phase I used 4 grid waypoints, and ground truth was obtained with an Acclima TDR-315H using 15 cm probes. Two planners were tested: the greedy benchmark and the $\Delta t$0 adaptive planner with $\Delta t$1. The benchmark collected 16 samples, traveled 153.3 m, averaged 9.58 m per sample, and took 2019 s. The $\Delta t$2 planner collected 21 samples, traveled 176.4 m, averaged 8.40 m per sample, and took 2608 s. Reconstruction RMSE was 13.1% for the benchmark and 12.2% for $\Delta t$3.

These results require careful interpretation. The simulation claim of up to $\Delta t$4 travel reduction is valid, but the proof-of-concept field run did not produce a shorter total mission for $\Delta t$5; instead, it yielded more samples, lower distance per sample, longer total travel, and modestly lower reconstruction RMSE. The paper also reports that deeper insertion reduced measurement error, which links map quality not only to planner design but also to mechanical consistency of probe deployment.

5. Position within the broader moisture-mapping literature

MoistureMapper occupies a specific niche within soil-moisture sensing: direct-contact, high-resolution, mobile, and adaptively sampled. This distinguishes it from earlier robotic mapping with non-contact sensing, such as kriging-based exploration using a cosmic-ray neutron sensor, where measurements follow a Poisson model and uncertainty-aware exploration is built around Poisson-Kriging rather than probe insertion (Fentanes et al., 2018). In that earlier framework, the robot is guided by kriging variance and can use adaptive measurement intervals, but its sensing footprint and statistical observation model are fundamentally different.

It is also distinct from microwave-imaging approaches designed for subsurface drip irrigation. Laboratory-scale systems using back projection and Born approximation reconstruct leakage and local moisture near buried pipes, with reported localization to within $\Delta t$6 cm and moisture-estimation error below $\Delta t$7 for the back projection approach under cluttered conditions (Ramezaninia et al., 2024). A subsequent machine-learning-driven version feeds two-dimensional microwave images into KNN and CNN models and reports more accurate moisture estimation using CNN after clutter reduction with the back projection algorithm (Ramezaninia et al., 2024). Those systems target local moisture detection around pipes and roots; MoistureMapper instead performs field traversal to obtain sparse but direct in situ samples over larger areas.

At a larger spatial scale, remote-sensing and data-fusion systems solve a different problem. Deep fusion of Sentinel-1, Sentinel-2, SMAP, SoilGrids, and GLDAS produces nominal 320 m maps with average per-sensor correlation of 0.727 and ubRMSE of 0.054, while a pan-European 10 m framework using Sentinel-1, Sentinel-2, and ERA5 reports $\Delta t$8 under its optimal setting and finds that foundation model embeddings add negligible improvement over hand-crafted features (Batchu et al., 2022, Kontogiorgakis et al., 20 Feb 2026). Such products offer continental or global scalability, but they are not direct physical samplers. A plausible implication is that a platform like MoistureMapper is most valuable as a field-scale mapper, calibrator, or validation instrument in regimes where satellite retrievals are too coarse or insufficiently localized.

Stationary and ultra-low-power systems define another comparison class. Low-cost IoT soil probes built around Raspberry Pi and capacitive sensors report $\Delta t$9 and RMSE of about $\epsilon_b$0 against gravimetric observations (Deshpande et al., 2022), while battery-free buried nodes with vehicle-based data retrieval operate for up to 21 days on a single capacitor charge and cost less than \$35 per node (Thoene et al., 29 Apr 2026). Smartphone methods provide still another trade-off: image-based estimation can achieve strong SVR performance under indirect sunlight (Hossain et al., 2023), and acoustic sensing with vertical scans can reach a mean absolute error of 2.39% across 10 locations without disturbing the soil (Gao et al., 11 Sep 2025). Relative to these systems, MoistureMapper sacrifices the simplicity of passive or static deployment in exchange for direct, adaptive, georeferenced subsurface measurement.

6. Limitations, interpretation, and future directions

The main limitations are mechanical, statistical, and operational. Mechanically, incomplete probe insertion in dense soils can lift the UGV and bias readings. Statistically, the GP currently uses a fixed kernel length scale $\epsilon_b$1, which may not match heterogeneous soil correlation lengths. Operationally, the current system relies on manual depth marking and lacks closed-loop force or depth feedback (Rose et al., 17 Jul 2025).

These limitations bear directly on interpretation. A common misconception is that adaptive sampling is uniformly superior on every metric. The reported results are more qualified. In simulation, the benchmark collected the fewest samples, greedy was slightly better for maximum variance in small fields, and randomized $\epsilon_b$2 achieved the best average-variance reduction. In field deployment, the adaptive $\epsilon_b$3 planner improved reconstruction RMSE only modestly and required more samples, longer mission time, and greater total travel than the benchmark. This suggests that “better” depends on the objective: total distance, distance per sample, maximum variance, average variance, or final reconstruction error.

The stated future work follows naturally from those constraints. The paper proposes integrating force sensing and laser-based probe-depth measurement to ensure full insertion and auto-abort on stall, implementing TDR waveform analysis for partial-insertion correction following Dakshinamurthy et al. (2024), developing online hyperparameter adaptation for non-stationary soil textures, and scaling deployments to varied soil types, topographies, and irrigation scenarios (Rose et al., 17 Jul 2025). If these extensions are realized, the platform would move from proof-of-concept robotic mapper toward a more autonomous field instrument with tighter coupling between mechanics, probabilistic inference, and mission planning.