Soil analysis with machine-learning-based processing of stepped-frequency GPR field measurements: Preliminary study

TL;DR

The paper investigates end-to-end ML processing of high-resolution air-coupled SFCW GPR data to predict EMI-derived apparent conductivity (ECaR) for soil analysis in precision agriculture. It reports a large field campaign on a golf course with 3472 co-registered samples over roughly 6600 m^2, using EMI as a proxy ground truth. Among regression models, Random Forest consistently performs best under spatially dense data, with a notable Pearson correlation of $r=0.425$ in the best case, and the nugget-to-sill ratio NSR correlating with all performance metrics. The study highlights NSR as a practical, ground-truth-free metric for model assessment in field surveys and recommends expanding multi-sensor data fusion and remote sensing to improve generalization for precision agriculture.

Abstract

Ground Penetrating Radar (GPR) has been widely studied as a tool for extracting soil parameters relevant to agriculture and horticulture. When combined with Machine Learning (ML) methods, air-coupled Stepped Frequency Continuous Wave Ground Penetrating Radar (SFCW GPR) measurements could offer a cost-effective way to obtain depth-resolved soil data. As a first step of our study in this direction, we conducted an extensive field survey using a tractor-mounted air-coupled SFCW GPR instrument. Leveraging ML-based data processing, we evaluate the GPR instrument's ability by predicting the apparent electrical conductivity (ECaR) measured by a co-recorded Electromagnetic Induction (EMI) instrument. The large-scale field measurement campaign with 3472 co-registered and geo-located GPR and EMI data samples distributed over approximately 6600 square meters was performed on a golf course. This terrain offers high surface homogeneity but also presents the challenge of subtle soil parameter variations. Based on the results, we discuss challenges in this multi-sensor regression setting and propose the use of the nugget-to-sill ratio as a performance metric for evaluating ML models in agricultural field survey applications.

Figures (5)

• Figure 1: Experimental setup of the air-coupled SFCW GPR with a fixed offset and Vivaldi antennas (in yellow color) mounted directly behind the transversal EMI instrument bar on a Toro Reelmaster 5510 tractor.
• Figure 2: Information of the field campaign. a) Overview satellite map and b) detailed satellite imagery of the Fontana Golf Club, Austria. c) Velocity map of the tractor on FWY14 (left) and FWY16 (right). d) Weather conditions in the week leading up to the date that measurements were taken.
• Figure 3: Variograms of Scenario 1 with training and testing on FWY16 a) and Scenario 2 with training on FWY14 and testing on FWY16 b).
• Figure 4: Results of predicted ECaR values from SFCW GPR data in Scenario 1. A geo-randomized five-fold cross-validation is applied to measurements of FWY16 with measured ECaR values plotted in heatmap a). Results of RFR, Linear regression and KNR models are shown in b)-d). The top row presents model predictions in heatmaps and scatter plots between measured ECaR values and predictions (with linear fit and Pearson correlation coefficient $r$), while the bottom row displays model prediction errors in heatmaps and histograms. The histograms of baseline prediction errors represent from a uniform prediction using the mean of ECaR values of a training set.
• Figure 5: Results of predicted ECaR values from SFCW GPR data in Scenario 2. Models are trained on FWY14 and tested on FWY16 with measured ECaR values plotted in heatmap a). Results for RFR b), Linear regression c), and KNR d) follow the description in Fig. \ref{['FIG:GTvsModel16']}.