Prediction of soil organic carbon content by diffuse reflectance spectroscopy using a local partial least square regression approach

Abstract

Due to the large spatial variation of soil organic carbon (SOC) content, assessing the current state of SOC for large areas is costly and time consuming. Visible and Near Infrared Diffuse Reflectance Spectroscopy (Vis-NIR DRS) is a fast and cheap tool for measuring SOC based on empirical equations and spectral libraries. While the approach has been demonstrated to yield accurate predictions for databases containing samples belonging to soils with similar characteristics such as mineralogy, texture, iron, and CaCO3 content, spectroscopic calibrations have been less successful when applied to large and diverse soil spectral libraries. The scope of this study was to predict SOC using a local partial least square regression approach. In total, 19,969 topsoil (0–20 cm) samples collected all over the European Union were analyzed for physical and chemical properties, and scanned with a Vis-NIR spectrometer in a single laboratory. The local regression method builds a different multivariate model for each sample to predict. Each local model is trained with neighbours' samples selected from a large spectral library, based on their spectral similarity with the sample to predict. We modified the local regression procedure by including other covariates (geographical and texture information) in the computation of the distance between samples. The results showed good prediction ability for mineral soils under cropland (RMSE = 3.6 g C kg−1) and grassland (RMSE = 7.2 g C kg−1). Predictions of mineral soils under woodland (RMSE = 11.9 g C kg−1) and organic soils (RMSE = 51.1 g C kg−1) were less accurate. The use of sand content in the computation of the sample similarities provided the most accurate SOC predictions due to its influence on light scattering properties of soils. In large datasets, using additional soil or environmental information allows to select neighbours that have overall the same soil composition as the samples to predict, resulting in more accurate models. This study shows that (i) it is possible to realize low-cost estimations of SOC at continental scale using large spectral libraries with a reasonable accuracy, and (ii) the local approach is a valuable tool to deal with large datasets, especially if existing soil property maps or soil legacy data could be used as covariates in the SOC prediction models.