Quantification of Soil Organic Carbon at Regional Scale: Benefits of Fusing Vis-NIR and MIR Diffuse Reflectance Data Are Greater for in Situ than for Laboratory-Based Modelling Approaches

Abstract

Benefits of fusion approaches for visible to near (vis-NIR) and mid-infrared (MIR) chemometric modelling have been studied to some extent for laboratory-based soil studies, but little is known about the usefulness and limitations for in situ studies. Objectives were to compare laboratory-based and in situ vis-NIR and MIR partial least squares (PLS) and bagging-PLS regression approaches and to explore the potentials of combining both types of spectral data for the quantification of soil organic carbon (SOC). We applied different established low-level (spectra concatenation, outer product fusion approach) and high-level (averaging of vis-NIR and MIR modelling results) data fusion methods. The studied set comprised a total of 186 soil samples collected in Saxony-Anhalt and northern Saxony, Central Germany. One subset (Querfurt Plateau) covered 90 finely-textured soils originating from the Chernozem soil region, another (Düben Heath) with 96 samples was characterized by a wider pedological variety. Vis-NIR and MIR diffuse reflectance spectra were measured in situ on the soil surface and in the laboratory on pre-treated (dried and finely ground) soil material with the ASDFieldSpec4 and the Agilent 4300 Handheld FTIR instruments. We found a regionally stratified approach to be beneficial for accurate estimations for both laboratory and in situ data. For laboratory spectra, MIR outperformed vis-NIR data in both regions (Querfurt Plateau: r2 = 0.85 vs. 0.65, RMSE = 0.11 (in % SOC) vs. 0.17; Db̈en Heath: 0.77 vs. 0.69 (r2) and 0.27 vs. 0.40 (RMSE)). Ranking for in situ data was the same, but accuracies decreased markedly. With MIR, r2 amounted to 0.58 and RMSE was 0.20 for the Querfurt Plateau (vis-NIR: r2 = 0.26, RMSE = 0.27); for Dë̈n Heath, r2 was 0.60 and RMSE was 0.39 for MIR data, while vis-NIR resulted in an r2 of 0.53 and an RMSE of 0.43. For the studied samples, which had medium to low water contents (0.68 to 16.8 wt%, median at 5.4 wt%), we found accuracies with both spectral datasets to be similarly affected by in situ conditions. Model ensemble averaging based on bagging-PLS regression was the most efficient approach to improve SOC estimation accuracies with in situ spectral data, whereas model averaging was in general of little effect for laboratory data. Improvements were most marked for the in situ data of the Dn̈̈ Heath region, where r2 increased to a value of 0.77 and RMSE decreased to 0.28. Low-level data fusion methods did not yield any improvements compared to model ensemble averaging. For the latter, we identified averaging with weights derived sample-wise from uncertainties in the bootstrap-based modelling as being most accurate, but with little benefits compared to a simple (unweighted) averaging of vis-NIR and MIR estimates. Our results suggest that already a simple averaging procedure has the potential to advance multi-sensor applications integrating vis-NIR and MIR data for in situ or on-site soil spectroscopy. This applies especially to regions with heterogeneous soil conditions, tied to spectral variablity, as this increases the probability of complementary vis-NIR and MIR information and their prospective fusion.