Error propagation in spectrometric functions of soil organic carbon

摘要

Abstract. Soil organic carbon (SOC) plays a major role concerning chemical, physical,and biological soil properties and functions. To get a better understandingof how soil management affects the SOC content, the precise monitoring ofSOC on long-term field experiments (LTFEs) is needed. Visible andnear-infrared (Vis–NIR) reflectance spectrometry provides an inexpensive andfast opportunity to complement conventional SOC analysis and has often beenused to predict SOC. For this study, 100 soil samples were collected at anLTFE in central Germany by two different sampling designs. SOC values rangedbetween 1.5 % and 2.9 %. Regression models were built using partial leastsquare regression (PLSR). In order to build robust models, a nested repeated5-fold group cross-validation (CV) approach was used, which comprised modeltuning and evaluation. Various aspects that influence the obtained errormeasure were analysed and discussed. Four pre-processing methods werecompared in order to extract information regarding SOC from the spectra.Finally, the best model performance which did not consider error propagationcorresponded to a mean RMSE _(MV) of 0.12 % SOC ( R ~(2)=0.86 ). This model performance was impaired by Δ RMSE _(MV)=0.04 % SOC while considering input data uncertainties ( ΔR ~(2)=0.09 ), and by Δ RMSE _(MV)=0.12 % SOC( ΔR ~(2)=0.17 ) considering an inappropriatepre-processing. The effect of the sampling design amounted to a Δ RMSE _(MV) of 0.02 % SOC ( ΔR ~(2)=0.05 ). Overall,we emphasize the necessity of transparent and precise documentation of themeasurement protocol, the model building, and validation procedure in orderto assess model performance in a comprehensive way and allow for acomparison between publications. The consideration of uncertaintypropagation is essential when applying Vis–NIR spectrometry for soilmonitoring.