INTRODUCTION

Soils provide many services in terrestrial ecosystems and adequate management of agricultural soils support food security, water quality, biodiversity and climate regulation \cite{lalSoilHealthCarbon2016}. Soil organic matter (SOM) quantity influences soil fertility and water holding capacity, and is associated with more diverse and resilient soil fauna and flora \cite{bossuytProtectionSoilCarbon2005,verbruggenArbuscularMycorrhizalFungi2016,kravchenkoMicrobialSpatialFootprint2019}. Maintaining and increasing SOM quantity in agricultural soils is therefore often considered a keystone for their sustainable use \cite{chenuIncreasingOrganicStocks2019}.
The addition of organic residues into soil, either by direct return of crop residues or from exogenous sources, is a common practice in agricultural management aiming to enhance SOM quantity \cite{sisouvanhCanOrganicAmendments2021,maillardAnimalManureApplication2014}. SOM also controls other important soil parameters such as soil structural development or activity of (micro)organisms \cite{pascaultStimulationDifferentFunctional2013,baiEffectsAgriculturalManagement2018}. Organic fertilization through crop residue addition mainly results in the accumulation of coarse organic material at the annual to decadal timescale \cite{vanwesemaelIndicatorOrganicMatter2019} while changes in the quantity of mineral-associated organic matter (MAOM) occur generally on the scale of decades and more \cite{cambardellaParticulateSoilOrganicmatter1992,cotrufoSoilCarbonStorage2019}. The coarse fraction of SOM, corresponding to particulate organic matter (POM), is composed of partially decomposed plant fragments, whose association with the soil mineral matrix is restricted to surface coatings of the organic particles by minerals or occlusion within aggregates \cite{golchinStudyFreeOccluded1994,wagaiNatureOccludedLowdensity2009}. This fraction contrasts with the MAOM, which is composed of organic compounds with lower molecular weight and strongly attached to mineral phases (mostly to phyllosilicates and oxides) or protected within micro aggregates \cite{vonlutzowStabilizationMechanismsOrganic2008}. The accrual of SOM through POM accumulation has been proposed as a possible C sequestration strategy \cite{lavalleeConceptualizingSoilOrganic2020} having the advantages, in contrast to MAOM, of not being constrained by saturation of mineral surfaces or C/N stoichiometry \cite{cotrufoSoilCarbonStorage2019}.
The composition of POM reflects both its source material and its degradation processes during decomposition. Initially, this fraction is primarily composed of two types of plant tissues, namely parenchymatic tissues (e.g. leaves and bark) and woody tissues, the former being mainly made up of (hemi-)cellulose walls and rich in proteins while the latter contains more lignin \cite{kogel-knabnerMacromolecularOrganicComposition2002}. During decomposition, POM serves as an energy source for microorganisms which take up C and other nutrients and respire a considerable fraction of the processed C. Under oxic conditions, the N used by microorganisms tends to be recycled, resulting in a decrease of the POM C to N ratio (C/N) during initial decomposition stages. Other indications of changes in POM properties during decomposition can be obtained from 13C nuclear magnetic resonance spectroscopy., Studies conducted on a diverse range of natural organic materials revealed that O/N alkyl structures are preferentially taken up by microorganisms, and that the relative proportion of alkyl C generally increases as decomposition progresses \cite{baldockAssessingExtentDecomposition1997,kogel-knabner13C15NNMR1997,gaoDecompositionDynamicsChanges2016}. The ratio of alkyl C to O/N-alkyl C (hereafter noted alkyl ratio) is thus used as an indicator of degree of decomposition, with higher ratios indicating a greater degree of decomposition. These general dynamics of decomposition are modulated by the composition of the original material and soil conditions \cite{wilsonStudiesLitterAcid1983}.
Water and oxygen availability are important for POM decomposition and can strongly vary with soil depth. In deeper horizons, decomposition might be restricted by lower temperature and oxygen availability in comparison to topsoils, even if there is little direct evidence of these constraints and that specific microbial taxa have temperature optima below 10 °C \cite{rumpelDeepSoilOrganic2011,gaoDecompositionDynamicsChanges2016}. In addition, the spatial heterogeneity of soil components differs between topsoil and subsoil in agricultural soils , especially in fields where regular ploughing of the topsoil takes place \cite{chabbiStabilisedCarbonSubsoil2009,hobleyHotspotsSoilOrganic2018}. SOM spatial distribution is more heterogeneous in the subsoil and this patchy distribution results in hotspots where SOM decomposition, and potentially stabilisation, as well as nutrient cycling is centralised \cite{heinzeFactorsControllingVariability2018,heitkotterThereAnybodyOut2018,hobleyHotspotsSoilOrganic2018}. This highlights the importance of the location of organic substrates in niches, i.e. microdomains with various dynamics for their decomposition, and of techniques able to resolve such small scale heterogeneity.
Techniques to analyse the fine-scale spatial heterogeneity of SOM distribution and composition can help to elucidate processes of SOM decomposition. Laboratory VNIR imaging spectroscopy applied to Histosols has been shown to enable the identification of OM with different compositions \cite{steffensFineSpatialResolution2014,granlundIdentificationPeatType2021}. This method was also successfully used for the classification of diagnostic soil horizons and the mapping of several elements (C, N, Fe) at the pedon scale \cite{steffensLaboratoryImagingSpectroscopy2013,hobleyHotspotsSoilOrganic2018,sorensonDistributionMappingSoil2020}, as well as for the determination of soil structure arrangement and mapping of soil components (POM, Fe oxides, mineral matrix) at a sub-millimetre scale \cite{muellerPermafrostSoilComplexity2021,lucasCombinationImagingInfrared2020}. Combined with modern prediction methods based upon machine learning and artificial intelligence (e.g. random forest or artificial neural networks), this high spatial-resolution analytical technique enables mapping of chemical composition in soil profiles at a sub-millimetre scale.
In this study we hypothesized that both soil depth and quality of organic amendment are strong determinants of the molecular composition of POM and its decomposition, and aimed at characterising in-situ the small scale molecular heterogeneity of organic amendments. Decomposition of POM in topsoil will be faster than in the subsoil, where nutrient and oxygen availability is limited, thereby retarding decomposition. Additionally, organic amendments high in N will be preferentially decomposed by soil microorganisms, whereas decomposition of organic amendments dominated by C compounds will be relatively slower due to N limitation during the decomposition process. To test these hypotheses we incubated soil samples from the surface (0-30 cm) and upper subsoil (30-60 cm) of a Luvisol and added either green manure with a low C/N or straw amendment with a high C/N for a period of 180 days. We used VNIR hyperspectral imaging combined with artificial intelligence modelling to map the composition of POM prior to and after incubation.

Materials and methods

Soil sample and incubation experiment

The soil used for the incubation experiment was sampled from an agricultural trial field at ‘Campus Klein-Altendorf’ experimental research station (50°37’16”N; 6°59’53”E), University of Bonn, Germany. The mean annual air temperature at Bonn-Rohleber is 10.3 °C and the mean annual precipitation is 816 mm for the period 1991-2010. The soil is a Haplic Luvisol \cite{iussworkinggroupwrbWorldReferenceBase2015} derived from quaternary Loess deposits, with a pHwater of 6.5 and a texture dominated by silt particles, with an enrichment in clay with depth (clay/silt/sand ⋍ 25-43/50-68/4-6 %). The organic C and total N contents are respectively 8.0 and 1.1 mg g-1 in the topsoil and 3.0 and 0.5 mg g-1 in the subsoil. Two soil depths (0-30 cm, and 30-60 cm), hereafter named topsoil and subsoil corresponding to the ploughed layer and the non-ploughed upper subsoil, were collected using a backhoe and directly sieved to 5 cm.
Soil cores were prepared in 90-cm high PVC pipes (⌀=7.5 cm) by inserting, from bottom to top, 20 cm of quartz sand, 30 cm of subsoil, 5 cm of mixed topsoil and subsoil, and 25 cm of topsoil. Soils were amended with two types of organic materials, wheat straw and green manure, added either to the topsoil or to the subsoil. Straw material corresponds to harvest residues whereas the composted green manure was a mixture of green waste (trees, bushes and shrubs) from public green areas and parks.
The organic material was incorporated by mixing it with the soil (1:4 volume mixing ratio for organic:soil) before packing the cores. Sixteen soil cores were prepared, corresponding to duplicates of the two types of organic material amendments (straw or green manure), admixed into one of the two soil depths (denoted top- or sub-) for initial sampling (T0) and after 180 days (T1) of incubation at 20 °C and 60 % maximum water-holding capacity determined gravimetrically.

VNIR hyperspectral imaging

Before recording hyperspectral images, soil cores were cut lengthwise, from bottom to top, into two equal halves and then dried at room temperature. For imaging, samples were illuminated with two 150-W halogen lamps. Hyperspectral images were recorded using a Hyspex VNIR-1800 camera (Norsk elektro optikk, Norway) after automatic dark background correction. The sensor was equipped with a 30-cm lens, giving a final field of view of approximately 9 cm for the 1800 detectors (53 × 53 µm2 per pixel). For each pixel, light reflectance intensity was measured for 186 bands in the region 400-990 nm (spectral resolution of 3.17 nm per band).
To account for potential unevenness in illumination and spectral response at different horizontal locations in the core, the spectral intensity (I) of the raw images of the soil cores were normalised to the defined reflectance (R) of calibration target for each wavelength (λ) and pixel (x):