No abstract

No abstract

The development of accurate calibration models for selected soil properties is a crucial prerequisite for successful implementation of visible and near infrared (Vis-NIR) spectroscopy for soil analysis. This paper compares the performance of calibration models developed for individual farms with that of general models valid for three farms in three European countries. Fresh soil samples collected from farms in the Czech Republic, Germany and Denmark were scanned with a fibre-type Vis-NIR spectrophotometer. After dividing spectra into calibration (70%) and validation (30%) sets, spectra in the calibration set were subjected to partial least squares regression (PLSR) with leave-one-out cross-validation to establish calibration models of soil properties. Except for the Czech Republic farm, individual farm models provided successful calibration for total carbon (TC), total nitrogen (TN) and organic carbon (OC), with coefficients of determination (R 2 ) of 0.85-0.93 and 0.74-0.96 and residual prediction deviations (RPD) of 2.61-3.96 and 2.00-4.95 for the cross-validation and independent validation respectively. General calibration models gave improved prediction accuracies compared with models of farms in the Czech Republic and Germany, which was attributed to larger ranges in the variation of soil properties in general models compared with those in individual farm models. The results revealed that larger standard deviations (SDs) and wider variation ranges have resulted in larger R 2 and RPD, but also larger root mean square errors of prediction (RMSEP). Therefore, a compromise solution, which also results in small RMSEP values, should be found when selecting soil samples for Vis-NIR calibration to cover a wide variation range.

Summary Spiking is a useful approach to improve the accuracy of regional or national calibrations when they are used to predict at local scales. To do this, a small subset of local samples (spiking subset) is added to recalibrate the initial calibration. If the spiking subset is small in comparison with the size of the initial calibration set, then it could have little noticeable effect and a small improvement can be expected. For these reasons, we hypothesized that the accuracy of the spiked calibrations can be improved when the spiking subset is extra‐weighted. We also hypothesized that the spiking subset selection and the initial calibration size could affect the accuracy of the recalibrated models. To test these hypotheses, we evaluated different strategies to select the best spiking subset, with and without extra‐weighting, to spike three different‐sized initial calibrations. These calibrations were used to predict the soil organic carbon (SOC) content in samples from four target sites. Our results confirmed that spiking improved the prediction accuracy of the initial calibrations, with any differences depending on the spiking subset used. The best results were obtained when the spiking subset contained local samples evenly distributed in the spectral space, regardless of the initial calibration's characteristics. The accuracy was improved significantly when the spiking subset was extra‐weighted. For medium‐ and large‐sized initial calibrations, the improvement from extra‐weighting was larger than that caused by the increase in spiking subset size. Similar accuracies were obtained using small‐ and large‐sized calibrations, suggesting that incipient spectral libraries could be useful if the spiking subset is properly selected and extra‐weighted. When small‐sized spiking subsets were used, the predictions were more accurate than those obtained with ‘geographically‐local’ models. Overall, our results indicate that we can minimize the efforts needed to use near‐infrared (NIR) spectroscopy effectively for SOC assessment at local scales.

No abstract