Which and how many soil sensors are ideal to predict key soil properties: A case study with seven sensors

Abstract

Soil sensing enables rapid and cost-effective soil analysis. However, a single sensor often does not generate enough information to reliably predict a wide range of soil properties. Within a case-study, our objective was to identify how many and which combinations of soil sensors prove to be suitable for high-resolution soil mapping. On a subplot of an agricultural field showing a high spatial soil variability, six in-situ proximal soil sensors (PSSs) next to remote sensing (RS) data from Sentinel-2 were evaluated based on their capabilities to predict a set of soil properties including: soil organic carbon, pH, moisture as well as plant-available phosphorus, magnesium and potassium. The set of PSSs consisted of ion-selective pH electrodes, a capacitive soil moisture sensor, an apparent soil electrical conductivity measuring system as well as passive gamma-ray-, X-ray fluorescence- and near-infrared spectroscopy. All possible combinations of sensors were exhaustively evaluated and ranked based on their prediction performances using model stacking. Over all soil properties, data fusion demonstrated a considerable increase in prediction accuracy. Five out of six soil properties were predicted with an R² ≥ 0.80 with the best sensor fusion model. Nonetheless, the improvement derived from fusing an increasing number of PSSs was subject to diminishing returns. Sometimes adding more PSSs even decreased prediction performances. Gamma-ray spectroscopy and near-infrared spectroscopy demonstrated to be most effective, both as single sensors or in combination with other sensors. As a single sensor, RS outperformed three out of six PSSs. RS showed especially potential for fusion with single PSSs but was of limited benefit when multiple PSSs were fused. Model stacking proved to be more robust than using single base-models because sensor performances were less model-dependent.