 |
| Titel |
Analyzing the prediction error of large scale Vis-NIR spectroscopic models |
| VerfasserIn |
Antoine Stevens, Marco Nocita, Luca Montanarella, Bas van Wesemael |
| Konferenz |
EGU General Assembly 2013
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 15 (2013) |
| Datensatznummer |
250083874
|
|
|
|
|
|
| Zusammenfassung |
| Based on the LUCAS soil spectral library (~ 20,000 samples distributed over 23 EU countries), we developed multivariate calibration models (model trees) for estimating the SOC content from the visible and near infrared reflectance (Vis-NIR) spectra. The root mean square error of validation of these models ranged from 4 to 15 g C kg-1. The prediction accuracy is usually negatively related to samples heterogeneity in a given library, so that large scale databases typically demonstrate low prediction accuracy compared to local scale studies. This is inherent to the empirical nature of the approach that cannot accommodate well the changing and scale-dependent relationship between Vis-NIR spectra and soil properties. In our study, we analyzed the effect of key soil properties and environmental covariates (land cover) on the SOC prediction accuracy of the spectroscopic models. It is shown that mineralogy as well as soil texture have large impacts on prediction accuracy and that pedogenetic factors that are easily obtainable if the samples are geo-referenced can be used as input in the spectroscopic models to improve model accuracies. |
|
|
|
|
|
|