Soil organic carbon and particle sizes mapping using vis–NIR, EC and temperature mobile sensor platform

Abstract

Soil organic carbon (SOC) is an important parameter in the climate change mitigation strategies and it is crucial for the function of ecosystems and agriculture. Particle size fractions affect strongly the physical and chemical properties of soil and thus also SOC. Conventional analyses of SOC and particle sizes are costly limiting the detailed characterization of soil spatial variability and fine resolution mapping. Mobile sensors provide an alternative approach to soil analysis. They offer densely spaced georeferenced data in a cost-effective manner. In this study, two agricultural fields (Voulund1 and Voulund2) in Denmark were mapped with the Veris mobile sensor platform (MSP). MSP collected simultaneously visible near infrared spectra (vis–NIR; 350–2200nm), electrical conductivity (EC: shallow; 0–30cm, deep; 0–90cm), and temperature measurements. Fuzzy k-means clustering was applied to the obtained spectra to partition the fields and to select representative samples for calibration purposes. Calibration samples were analyzed for SOC and particle sizes (clay, silt and sand) using conventional wet chemistry analysis. The objectives of this study were to determine whether it is the single sensors or the fusion of sensor data that provides the best predictive ability of the soil properties in question. Using partial least square regression (PLS) excellent calibration results were generated for all soil properties with a ratio of performance to deviation (RPD) values above 2. The best predictive ability for SOC was obtained using a fusion of sensor data. The calibration models based on vis–NIR spectra and temperature resulted in RMSECV=0.14% and R2=0.94 in Voulund1. In Voulund2, the combination of EC, temperature and spectral data generated a SOC model with RMSECV=0.17% and R2=0.93. The highest predictive ability for clay was obtained using spectral data only in Voulund1 (RMSECV=0.34% and R2=0.76). Whereas in Voulund2, improved results were obtained after combining spectral and temperature data RMSECV=0.20% and R2=0.92. The best predictions of silt and sand were obtained when using spectral data only and resulted in RMSECV=0.35%, R2=0.82 and RMSECV=0.85%, R2=0.81, respectively, in Voulund1 and RMSECV=0.31%, R2=0.86 and RMSECV=0.74%, R2=0.92, respectively, in Voulund2.The best models were used to predict soil properties from the field spectra collected by the MSP. Maps of predicted soil properties were generated using ordinary kriging. Results from this study indicate that robust calibration models can be developed on the basis of the MSP and that high resolution field maps of soil properties can be compiled in a cost-effective manner.