Discussion
VNIR hyperspectral imaging to map POM molecular composition
The successful modelling results for both the alkyl ratio and C/N of POM we achieved by linking VNIR with C/N measurements and 13C-CPMAS-NMR spectroscopic data in machine learning ensemble models validates this method as appropriate for studying the decomposition of organic amendments and plant residues in soils at a sub-millimetre scale. \cite{steffensFineSpatialResolution2014} were already able to distinguish three different classes of OM at different stage of decomposition based on the composition determined by 13C-CPMAS-NMR spectroscopy and on their VNIR reflectance spectra, but we went further by predicting molecular indicators, i.e. alkyl ratio and C/N, of POM at a very fine scale (Figure 3 and 4). \cite{xuHyperspectralImagingHighresolution2020} achieved the mapping of so-called different soil OM fractions, namely soil organic carbon, salt-extractable organic carbon and readily oxidizable organic C. These variables are strongly correlated with total organic carbon contents in the soil and do not reveal much about the composition of soil OM. To our knowledge, our work is the first to focus on POM molecular characterisation by VNIR hyperspectral imaging and to predict relevant proxies, C/N and alkyl/O-N alkyl ratio, for decomposition stage of organic particles at such a high resolution (Figure 2).
Classical VNIR spectroscopy has been used to estimate the contents of organic C, CaCO3, total N, Fe oxides and clays among other soil parameters \cite{ben-dorNearInfraredAnalysisRapid1995,changNearinfraredReflectanceSpectroscopy2001,viscarrarosselVisibleInfraredMid2006}, but also the type of soil OM. For example to differentiate between POM, mineral-associated organic matter (MAOM) and pyrogenic organic carbon \cite{hobleySoilCharcoalPrediction2016,sandermanSoilOrganicCarbon2021}. Depending on the wavelength range of sensors, VNIR imaging methods have the potential to predict the contents of these soil parameters and the detailed mapping made possible by this technique provides a substantial advantage for the understanding of soil functioning by taking into account the spatial arrangement of soil components. Our results indicate that VNIR imaging combined with machine learning can be used to differentiate POM from the mineral soil and to predict molecular composition of the identified POM particles, revealing the heterogeneity in their molecular composition at a sub-millimetre scale (Table 2 and Figure 2).
However there are limitations that still require further consequent developments of this approach. Firstly, the predictions are dependent on the choice of model. This has been already shown in many studies before, and the choice of the most appropriate modelling algorithm, or combination of algorithms, varies between studies and modelled parameters \cite{hobleyHotspotsSoilOrganic2018,xuHyperspectralImagingHighresolution2020,granlundIdentificationPeatType2021}. By design, RF does not extrapolate the predictions beyond the calibration range. What initially appears to be an advantage, since extrapolation is not desirable in the use of most machine learning algorithms, leads to a compression of the number of predictions in the marginal areas of the calibration range (Figure S1). The ensemble ANN on the other hand has better goodness-of-fit parameters and tuning of the number of layers was not needed, only the number of models was. This reduces the chance of over-fitting a single ANN by constraining it to a certain architecture. We obtained excellent predictions inside the calibration range, but extrapolation was not constrained to plausible results, indicated by the very large proportion of nonsensical and impossible predictions outside the calibration range (Table S2).
Multilayer perceptrons like ANN algorithms are known to perform well in generalisation but perform poorly with extrapolation. ANN ensemble and model averaging overcomes these shortfalls to some extent \cite{hobleyHotspotsSoilOrganic2018} and delivered the best and most plausible results in our study, but the values beyond the calibration range must be interpreted with great care (Figure 2).
Another limitation of the modelling approach is that predictions are applied pixel wise (53 x 53 µm² per pixel) whilst calibration is bound to ROI with larger areas (up to centimetre scale). The ROI areas must be large enough to sample the thinnest layer as possible to obtain the minimum quantities of material (usually a minimum of 100 mg) needed for C, N and 13C-CPMAS-NMR spectroscopy measurements of the material measured by VNIR, which is only based on surface reflectance, resulting in the need for very shallow sampling. This is a classical limitation we noticed in other studies on SOM prediction based on hyperspectral imaging. Improvement of the calibration procedure requires methods producing spatially-resolved molecular information at a lower scale, such as the combination of scanning electron microscope with energy dispersive X-ray spectroscopy (SEM-EDX, e.g. \cite{hapcaThreeDimensionalMappingSoil2015}) or scanning transmission X-ray microscopy.
Molecular composition of straw and green manure at different stages of decomposition
The dominance of a single but shifting molecular signature in the wheat straw before and after incubation for both alkyl ratio and C/N is consistent with studies reporting decreasing proportions of straw carbon as carbohydrate and increasing proportions of aromatic compounds during straw decomposition \cite{cogleUse13CNMRStudies1989,gaoDecompositionDynamicsChanges2016}. The alkyl ratio increase indicates the decomposition of straw compounds, likely by a shift from substituted aliphatic alcohols and ethers to unsubstituted C in paraffinic structures \cite{wilsonStudiesLitterAcid1983,kogel-knabnerMacromolecularOrganicComposition2002}, whereas the decrease in C/N is consistent with a loss of C caused by microbial respiration \cite{geissenDecompositionRatesFeeding1999} and the preservation of N by decomposers in their tissues and by-products. However, the greater heterogeneity of both alkyl ratio and C/N of straw at the end of the incubation (Table 2) indicates that decomposition was spatially heterogeneous and that some regions of the straw amendment decomposed faster than others. Although this could be explained by preferential decomposition of specific plant-derived structural compounds by microorganisms during the first stages of decomposition \cite{golchinModelLinkingOrganic1997}, this spatially heterogeneous degradation of the chemically homogeneous straw amendment is more likely attributable to abiotic factors, i.e. the microscale conditions of the microbial habitats surrounding POM that regulates soil moisture and microbial accessibility of the OM \cite{dungait_2012_SoilOrganicMatter}.
In contrast to the clear shift in POM composition during incubation of straw, changes in the composition of green manure were less clear, a result of the greater heterogeneity of green manure before and after incubation. Indeed, the general increase in C/N and decrease in alkyl ratio after incubation are contrary to typical changes induced by decomposition of organic materials \cite{golchinStudyFreeOccluded1994,baldockAssessingExtentDecomposition1997,stoneParticulateOrganicMatter2001,kogel-knabnerMacromolecularOrganicComposition2002}. We attribute this to a shift in the exposed surface of the material before and after incubation. The green manure material contained highly different types of plant tissues, such as leaves, woody parts and bark material and was composted before we used it. While pieces of bark and partially decomposed green manure were observed at the beginning of the experiment, these particles may have significantly peeled off and dispersed during the incubation, exposing internal woody structures of green manure material. To better explain the changes in green manure chemistry, further research into the molecular composition of the various types of organic material present in the mixture and their decomposition is required.
Straw amendments are more decomposed in topsoils than in subsoils
The lower C/N and larger alkyl ratio of straw particles (Table 2), and the stronger shifts in distribution in the topsoil at the end of the incubation indicate a more advanced degradation than in the subsoil (Figure 2), which we attribute to the differences in the initial conditions of the two soil matrices that drive decomposition processes and rates. Generally, topsoils accommodate more diverse microbial communities and more microbial biomass than subsoils \cite{taylorComparisonMicrobialNumbers2002}, and contain more SOC. The production of dissolved organic matter (DOM) is correlated with high SOC contents and abundance of microorganisms \cite{guigueWaterextractableOrganicMatter2015} and larger quantities of DOM are measured in topsoils than in subsoils \cite{kalbitzControlsDynamicsDissolved2000,kaiserCyclingDownwardsDissolved2012}. DOM is is composed of labile and energy-rich compounds easily converted either by microbial resynthesis or respiration \cite{straussEffectDissolvedOrganic2002,guggenbergerDissolvedOrganicMatter2003,kaiserCyclingDownwardsDissolved2012}. This fuels the production of extracellular hydrolytic enzymes that contribute to POM decomposition \cite{bergPlantLitterDecomposition2014}, and enables the faster decomposition of the added fresh organic matter in topsoils than subsoils.
C/N spatial distribution and coupling of C and N cycling
The C/N of above 40 after incubation (Figure 2) are well above typical lower bounds for SOM in central European agricultural soils, which rarely exceed C/N of 15 \cite{matschullatGEMASCNSConcentrations2018}. This indicates that the decomposition processes of the POM was not completed during the period of incubation and we assume that C/N of POM would continue to decrease if the incubation was prolonged. The initially lower C/N of green manure in the topsoil is likely the result of the prior composting of this material. It may also result from a greater proportion of N-rich compounds, such as bark, on the surface of the composted material when it was introduced to the topsoil for incubation.
The C/N of the POM fraction has been reported as being greater than 20 in soils under cropland, forest and grassland \cite{warrenAvailableSoilNitrogen1988,meijboomDensityFractionationSoil1995,hassinkDensityFractionsSoil1995,baldockOrganicMatterSeen2003,bimullerDecoupledCarbonNitrogen2014}. In addition, C/N ranging from 20 to 25 are generally accepted to be the thresholds for the shift of microbial N immobilisation to N mineralisation \cite{nicolardotSimulationMineralisationCrop2001,robertsonNitrogenTransformations2015} that stabilises the C/N by balancing C and N losses and further leading to a rapid mineralisation of POM. Straw decomposition is reported to start rapidly before slowing down during the process \cite{dewillingenDecompositionAccumulationOrganic2008}, with decomposition mostly fuelled by hemicellulose while more recalcitrant ligneous materials are decomposed slower. Thus, the relative contribution of ligneous material increases during decomposition \cite{cogleUse13CNMRStudies1989,gaoDecompositionDynamicsChanges2016}, concurrent with decreases in the decomposition rate, corresponding to the preservation of remaining straw residues with a high C/N. Furthermore, lignin macromolecules decomposition is promoted when the macrostructures are firstly shredded by macro fauna \cite{scheuDecompositionLigninSoil1992}. The absence of such organisms in our experiment likely supported the preservation of large lignin-like moieties with high C/N.
Conclusion
The coupling of VNIR imaging with machine learning modelling was successful for the sub-millimetre scale mapping of molecular composition of various types of POM at distinct decomposition stages. Our novel approach based on model averaging of a random forest with an ensemble ANN overcomes issues relating to training of a single ANN and extrapolation of results beyond calibration ranges, presenting a new direction in machine learning and spectroscopy in soils. With this technique, we demonstrated the spatially heterogeneous changes in alkyl ratio and C/N of POM during decomposition with an overall increase in alkyl ratio and decrease in C/N of straw as a result of decomposition. The changes of the more heterogeneous green manure showed an opposite trend, likely associated with the preferential decomposition of N-rich bark tissues together with the preservation of less decomposable C-rich plant residues enriched in ligneous material. The decomposition of both straw and green manure was retarded in the subsoil compared to topsoil, as highlighted by smaller changes in the POM in the subsoil after the 6-month incubation in the subsoil. The visualisation approach presented has a great potential for applications aiming to investigate the spatial heterogeneity in molecular changes of organic particles during decomposition, and can help to disentangle the concurrent roles of accessibility and recalcitrance during the first steps of the decomposition cascade of organic matter in soils.
Acknowledgements
This work was financially supported by the German Federal Ministry for Education and Research, through the BonaRes initiative (BMBF, grant FKZ 031B0026B, Soil3 project and 031B0511C BonaRes centre). As a part of the BonaRes initiative, the authors would like to thank all contributors to the Soil3 project and the BonaRes centre.
Supplementary files