Geophysical and Hyperspectral Data Fusion Techniques for In-Field Estimation of Soil Properties

There is great interest in the development of inexpensive and rapid soil mapping methods for precision agriculture applications. Proximal and remote soil sensing are particularly valuable for this purpose, and there is great scope for their synergic use. The objective of this study was to compare different methods allowing the joint exploitation of hyperspectral satellite data and geophysical data for estimating soil properties at the field scale. Soil samples were collected in an agricultural field in Central Italy for the determination of several soil properties. Satellite images were acquired by the CHRIS-PROBA sensor, both under bare soil conditions and when the field was covered by a wheat crop. Geophysical data were obtained by the automatic resistivity profiling method (ARP), providing apparent soil electrical resistivity of the 0 to 50 cm layer. Regression-kriging (RK), partial least square regression (PLSR), and a combination of PLSR with geostatistics through kriging of the PLSR residuals (PLSR-K) were applied to estimate soil properties by employing all combinations of the available covariates. A multiple jack-knifing procedure was used for a statistical comparison of the ratio of performance to deviation (RPD) statistics across 300 replicates. Clay, sand, and available soil water content were estimated with a sufficient degree of accuracy (RPD > 1.4), especially when using the RK technique. PLSR-K estimated these variables with intermediate ability by using only remote sensing covariates and obtained, in most cases, better results than PLSR. For other soil variables, the prediction ability was unsatisfactory (RPD < 1.4) due to smaller sample and range and weaker correlation with the covariates.