Modelling organic carbon content of southwestern German soils using visible near-infrared reflectance spectra and multi-temporal Sentinel-2 data

The sustainable management of agricultural land requires reliable information about soil physical and chemical properties. Among these properties, soil organic carbon (SOC) is a key attribute, as it serves as an important indicator for soil health and helps fighting climate change through carbon storage in soils. Since direct measurements are costly, visible near-infrared spectroscopy (VIS-NIR-SWIR) from 400 to 2500 nm is often used to estimate SOC, leveraging a statistical model which relates SOC analytical data to the spectral information obtained in the laboratory from a collection of sieved, air-dry samples. This study evaluates VIS-NIR-SWIR to predict SOC content of southwestern German soils after resampling the recorded soil spectral library (SSL) to match Sentinel-2 bands. It also examines whether these prediction models can then be applied to Sentinel-2 satellite imagery for rapid mapping of topsoil SOC content at a state-wide scale.

A suite of 1500 VIS-NIR-MIR soil spectra, recorded from air-dried, 2-mm, sieved soil samples, were associated with SOC analytical data obtained from different soil surveys done by the State Authority for Geology, Resources and Mining (LGRB) in Baden-Wuerttemberg, Germany. Partial least squares (PLS) regression and support vector machines on PLS latent variables (PLS-SVM) were used for spectroscopic modelling. Final estimates showed good results with regards to PLS-SVM with a ratio of performance to deviation (RPD) of 1.96, while slightly less accurate predictions were found for calibration models based on resampled spectra with a RPD of 1.64. The successful spectral prediction model for SOC from resampled spectra was subsequently used to produce a high-resolution map of topsoil SOC content on croplands for entire Baden-Wuerttemberg, Germany. To identify bare dry soil pixels a worfklow was established that creates per-pixel composites utilizing three years of Sentinel-2 satellite imagery and spectral indices. Direct standardization (DS) was used for the correction of environmental factors such as variable moisture conditions using a set of representative locations with both dry spectra and Sentinel-2 band values.

Preliminary results indicate that a calibration model based on resampled spectra from a region-specific SSL can be applied to multi-temporal Sentinel-2 data for rapidly estimating the spatial distribution of topsoil SOC content. Unlike official SOC products currently available for Baden-Wuerttemberg, Germany, the given approach can easily be updated if additional data becomes available or new sensors emerge, for instance, from hyperspectral satellite missions such as EnMAP or CHIME.