Effect of sample size, sampling design and calibration model on generating soil maps from proximal sensing data for precision liming

For site-specific estimation of lime requirement, high-resolution soil maps of clay, soil organic carbon (SOC) and pH are required. These can be generated using digital soil mapping (DSM), in which prediction models are fitted on covariates from proximal soil sensors. However, the quality of the maps derived may differ significantly depending on the methodology applied. Hence, we assessed effects of (i) calibration sample size (5-100), (ii) sampling design (simple random sampling (SRS), conditioned Latin Hypercube sampling (cLHS) and k-means sampling (KM)) and (iii) prediction model (linear regression (LR) and Random Forest (RF)) on the prediction performance for the above mentioned three soil properties using data from two multi-sensor platforms. The present case study is based on a geostatistical simulation using 250 soil samples from a 51 ha field in Germany. Among others, Lin’s concordance correlation coefficient (CCC) and root-mean-square error (RMSE) were used to evaluate model performances. Results show that with increasing sample size, improvements of RMSE and CCC decreased exponentially. We found best median RMSE values at 100 calibration soil samples, i.e. 1.73%, 0.3 and 0.21% for Clay, pH and SOC, respectively. However, already with 10 samples, models of moderate quality (CCC > 0.65) can be obtained for all three soil properties. Both, cLHS and KM obtained significantly better results than SRS. At smaller sample sizes, LR showed lower median RMSE values than RF for SOC and pH. Nonetheless, with at least 75-100 and 25-30 samples, RF eventually outperformed LR. For clay, median RMSE was lower with RF, regardless of sample size.