Evaluation of Mid-Infrared and X-ray Fluorescence Data Fusion Approaches for Prediction of Soil Properties at the Field Scale

Previous studies investigating multi-sensor fusion for the collection of soil information have shown variable improvements, and the underlying prediction mechanisms are not sufficiently understood for spectrally-active and -inactive properties. Our objective was to study prediction mechanisms and benefits of model fusion by measuring mid-infrared (MIR) and X-ray fluorescence (XRF) spectra, texture, total and labile organic carbon (OC) and nitrogen (N) content, pH, and cation exchange capacity (CEC) for n = 117 soils from an arable field in Germany. Partial least squares regression models underwent a three-fold training/testing procedure using MIR spectra or elemental concentrations derived from XRF spectra. Additionally, two sequential hybrid and two high-level fusion approaches were tested. For the studied field, MIR was superior for organic properties (ratio of prediction to interquartile distance of validation (RPIQV) for total OC = 7.7 and N = 5.0)), while XRF was superior for inorganic properties (RPIQV for clay = 3.4, silt = 3.0, and sand = 1.8). Even the optimal fusion approach brought little to no accuracy improvement for these properties. The high XRF accuracy for clay and silt is explained by the large number of elements with variable importance in the projection scores >1 (Fe ≈ Ni > Si ≈ Al ≈ Mg > Mn ≈ K ≈ Pb (clay only) ≈ Cr) with strong spearman correlations (±0.57 < r(s) < ±0.90) with clay and silt. For spectrally-inactive properties relying on indirect prediction mechanisms, the relative improvements from the optimal fusion approach compared to the best single spectrometer were marginal for pH (3.2% increase in RPIQV versus MIR alone) but more pronounced for labile OC (9.3% versus MIR) and CEC (12% versus XRF). Dominance of a suboptimal spectrometer in a fusion approach worsened performance compared to the best single spectrometer. Granger-Ramanathan averaging, which weights predictions according to accuracy in training, is therefore recommended as a robust approach to capturing the potential benefits of multiple sensors.