Aarhus University Seal / Aarhus Universitets segl

Mogens Humlekrog Greve

Using vis-NIR to predict soil organic carbon and clay at national scale: validation of geographically closest resampling strategy

Publikation: KonferencebidragKonferenceabstrakt til konferenceForskning

The Danish soil visible-near infrared (vis-NIR) spectral library has proved capable of predicting soil properties in Denmark such as soil organic carbon (SOC) at field scale using the geographically closest resampling strategy. However, this strategy has only been tested on one Danish local field/farm and for future application it is necessary to extend this approach to all types of Danish soils and evaluate the predictive capabilities of the library at national scale. In the present study, partial least squares regression was used to develop models and predict topsoil SOC (490 samples) and clay contents (442 samples) for soils from each 7-km grid sampling point in the country. In the resampling and modelling process, each target sample was predicted by a specific model which was calibrated using geographically closest soil spectra. The geographically closest 20, 30, 40, and 50 sampling points (profiles) were selected and used in model development. Models were evaluated on the basis of the root mean square error (RMSE), R2, ratio of performance to deviation (RPD) and the ratio of performance of interquartile distance (RPIQ). For both SOC and clay predictions the best results were obtained using the 40 geographically closest sampling points. The SOC prediction resulted in R2: 0.76; RMSE: 4.02 %; RPD: 1.59; RPIQ: 0.35. The results for clay prediction were also successful (R2: 0.84; RMSE: 2.36 %; RPD: 2.35; RPIQ: 2.88). For SOC predictions, over 90% of soil samples were well predicted compared with the uncertainties of traditional laboratory wet chemistry analysis. However, for organic soils (48 samples SOC >7%) originating from wetland or forested areas the SOC predictions were generally under-estimated and not satisfactory. For prediction of clay content, only 12 out of 442 predictions were unsatisfactory, for geological reasons. We concluded that a geographically closest resampling strategy is a promising and efficient method for creating sub-models of the Danish spectral library and predicting unknown soil samples. The accuracy of the SOC model depended on land use / land cover, whereas the accuracy of the clay model was strongly affected by soil parent material and landscape.
OriginalsprogEngelsk
Udgivelsesår2016
StatusUdgivet - 2016

Se relationer på Aarhus Universitet Citationsformater

ID: 108949903