Prediction of soil organic carbon at the country scale: stratification strategies for near‐infrared data

Abstract

SummaryResearch has shown that the application of near‐infrared (NIR) spectroscopy can be used to predict soil attributes, in particular for regional to continental scales. However, there are challenges when NIR is used at the regional scale because of the considerable spatial variation. This study has predicted SOC at the country scale (German agricultural soil inventory) with different stratification strategies for NIR data: (i) calibration with memory‐based learning (MBL) algorithms that use spectral similarity and (ii) simple stratification based on soil properties (depth, pH and soil texture) and land use. To optimize calibration models, this study aimed to predict soil organic carbon (SOC) determined by these three strategies for 1410 soil profiles selected from the German agricultural soil inventory. The profiles covered a wide range of soil types and characteristics. The calibration procedures were based on complete soil profile data of two‐thirds of the dataset and one‐third of the dataset was used for independent validation (prediction); the profiles were selected randomly. Available soil properties for stratifying the datasets were: soil depth (topsoil 0–30 cm and subsoil 31–100 cm), pH and texture class (silty, clayey, sandy and loamy). The profiles were also stratified by land use (cropland and grassland) and with the MBL method. The calibrations were carried out by partial least‐squares regression (PLSR), and each stratification model was compared with the global model. The root mean square error of cross‐validation (RMSECV) for the global model was 4.2 g SOC kg−1. Stratification according to soil depth reduced the error by 10% (RMSECV 3.8 g kg−1). The best stratification by soil texture was when sandy soil samples were separated from the other samples, which reduced the RMSECV by 14%. Calibration with MBL provided the most accurate predictions of SOC, with an error reduction of 25% (RMSECV 3.2 g kg−1). Thus, calibrations with NIR of country‐scale datasets can be improved easily by stratification or application of the MBL algorithm.Highlights Large country‐scale soil dataset of near‐infrared with > 1400 soil profiles was used. Stratification of NIR data by soil properties increased the accuracy of SOC calibration at country scale. The best stratification design involved calibrating sandy soil separately from other texture classes. A decrease of 25% in calibration error and 22% in prediction error with the MBL model compared with global model.