Comparison of in-situ and ex-situ portable x-ray fluorescence model accuracy and prediction mechanisms for soil properties across a region

Portable X-ray fluorescence models predicting soil properties affecting fertility have shown promising results, but in-situ studies for this application are lacking and require comparison to ex-situ models. For this, spectra were recorded in-situ (average gravimetric soil water content = 11%) and on dried and dried/ground soils from seven arable sites in Germany (n = 279). Measurements included texture, cation exchange capacity (CEC), organic carbon (OC) content, pH, and mineralogy by X-ray diffraction. Modelling was carried out with partial least squares regression (PLSR), PLSR with memory-based learning (MBL), and support vector machine regression (SVMR) following a 5-fold random partitioning of the dataset into training (80% of samples) and testing (20%) sets. Across all soil preparations and algorithms, validation R 2 for clay, silt, sand and CEC was >0.93, OC ranged from 0.65 to 0.97, and pH from 0.74 to 0.96. Compared to in-situ models, models with dried and dried/ground soils had 8% and 13% relative improvement (RI) to validation root mean squared error (RMSE V ), respectively, and one-way randomization tests found improvements to RMSE V were significant (α = 0.05) in 7 of 18 and 9 of 18 tested models for dried and dried/ground soils, respectively. Compared to PLSR models, MBL and SVMR models had 14% and 9% RI, respectively, with significant effects on RMSE V in 9 of 18 and 8 of 18 tested models, respectively. Variable importance in the projection analysis found prediction mechanisms across all soil properties relied primarily on the peaks of elements composing soil minerals, resulting in more direct prediction mechanisms for texture and CEC than OC and pH. Soil moisture and structure caused spectral intensity loss for lighter elements, whereby scattering regions became more useful predictors in models for moist and unground soils. If soil moisture contents are low, the trade-off of efficiency gains for minor accuracy losses with in-situ measurement may be acceptable for many applications.