Autor:innen:
Michael Seidel | Institut für Geographie, Universität Leipzig | Germany
Christopher Hutengs | Institut für Geographie, Universität Leipzig | Germany
Dr. Isabel Greenberg | Ökologische Agrarwissenschaften - Umweltchemie / Universität Kassel | Germany
Prof. Dr. Bernard Ludwig | Ökologische Agrarwissenschaften - Umweltchemie / Universität Kassel | Germany
Prof. Dr. Michael Vohland | Institut für Geographie, Universität Leipzig | Germany
The recent development of extensive soil spectral libraries in the visible to near-infrared (VNIR) and mid-infrared (MIR), and the availability of portable spectrometers in both ranges have opened up new opportunities for the application of soil reflectance spectroscopy in soil survey and monitoring. Compared to laboratory measurements on pre-treated soil material, field recordings of reflectance spectra are affected by in situ soil conditions, such as variable soil moisture contents, that modify and degrade the measured reflectance signal. These conditions prevent leveraging available SSL to build predictive models of soil properties for in situ or on-site applications.
The aim of this study was to test and compare different strategies of integrating a regional SSL (n = 300) in the modeling workflow to predict soil organic carbon (SOC) at field sites with variable soil moisture contents. For each sample in the SSL, reflectance spectra had been acquired both in the field, i.e., measured in situ on the soil surface with known moisture contents, and in the laboratory on pre-treated soil material (dried and sieved to ≤ 2 mm). The spectral measurements were carried out with portable spectrometers in the VNIR range (ASD FieldSpec 4, 350 - 2500 nm) and in the MIR range (Agilent 4300 FTIR Handheld, 4000 - 650 cm-1). The different modeling strategies included the separate application of the two spectral ranges, data fusion approaches, methods of compensating for the spectral effect of soil moisture (external parameter orthogonalization - EPO, global modeling with varying moisture contents - GMM), and spiking. Model validation was carried out on the data sets of six different independent field sites, where only in situ VNIR and MIR spectral measurements with varying moisture contents were available.
Our preliminary results indicate the complementary use of both spectral ranges to be the best approach for developing moisture-robust predictive models, compared to the use of individual spectral ranges. Additionally, spiking in combination with both the EPO approach or the robust calibration (GMM) significantly improved prediction accuracy across all target sites. These findings suggest that integrating spectra of different moisture contents into available VNIR and MIR spectral libraries would offer potential for building predictive models that are more robust to variations in soil moisture in the target data sets.