Synergetic use of mobile and lab-based spectroscopic techniques (vis-NIR, lab and hand-held MIR, portable hyperspectral frame camera) to optimize the determination of soil properties with high variability in time and space

土壤有机碳(SOC)、氮(N)和磷(P)的总含量在土壤肥力的管理研究(如施肥或耕作)中的作用有限；土壤有机碳和N组分以及土壤微生物特性是更敏感的指标。然而，只有通过非破坏性采样技术才能获得高空间和时间密度的样本。在此背景下，该项目研究了光谱技术的潜力，通过将近中区域(Vis-NIR和MIR)的非成像光谱与高光谱成像相结合，高精度地确定关键土壤属性(SOC、N、pH、SOC和N、P、硫、钾、铁、阳离子交换量、土壤质地、微生物和热水可溶C和N)。除了实验室规模外，我们还将重点放在现场光谱测量上，这是由于新的仪器开发、便携式MIR光谱仪和便携式高光谱成帧相机所青睐的。MIR范围对于土壤光谱学是必不可少的，因为可以测量化学基团的基带(不同于仅有组合带和泛音的近红外范围)。对于总共8个具有不同质地土壤的耕地，表层土壤和土壤剖面将被取样，通过结合上述估计土壤性质的不同光谱技术，研究实验室光谱与现场光谱的潜力。为了提高获得的精度，将通过使用支持向量机或随机森林来代替PLSR，通过应用光谱变量选择技术，通过用局部校准取代全局校准(即，执行适当的校准样本的逐个样本选择)，以及通过使用尖峰方法来局部适应校准模型，来优化多变量校准方法。优化后的技术将在现有的数据集上进行验证。此外，还将分析是否以及在多大程度上可以补偿由不同土壤表面粗糙度或不同土壤水分引起的干扰。对已有的土壤光谱库(LUCAS、ICRAF-ISRIC)进行了评估，以选择可能支持定标模型定义和优化的适当样本。此外，还将分析潜在的光谱预测机制(例如，通过2D相关光谱学)，以阐明对于所研究的每一种土壤性质，直接光谱预测还是仅间接光谱预测是可行的。这是澄清预测模型一旦校准，原则上是否可以在空间和时间上转移的基础。

Total contents of soil organic carbon (SOC), nitrogen (N) and phosphorus (P) are only of limited use for studies of management (e.g. fertilizations or tillage) on soil fertility; SOC and N fractions as well as soil microbial properties are much more sensitive indicators. However, a high spatial and temporal density of samples can only be achieved with non-destructive sampling techniques. In this context, the project studies the potentials of spectroscopic techniques to determine key soil properties (SOC, N, pH, fractions of SOC and N, P, sulphur, potassium, iron, cation exchange capacity, soil texture, microbial and hot water-soluble C and N) with high accuracy by combining non-imaging spectroscopy in the near and middle (vis-NIR and MIR) domain with hyperspectral imaging. In addition to the lab scale, we focus on the field scale with on-site spectroscopic measurements, which is favoured by new instrumental developments, a portable MIR spectrometer and a portable hyperspectral frame camera. The MIR range is essential for soil spectroscopy, as fundamental bands of chemical groups can be measured (different from the NIR range with only combination bands and overtones). For a total of eight arable sites with soils of differing textures, top soils and soil profiles will be sampled to investigate the potentials of lab spectroscopy compared to on-site spectroscopy by combining the different spectroscopic techniques for the estimation of the soil properties mentioned above. To improve obtained accuracies, methods of multivariate calibration will be optimized by using e.g. Support Vector Machines or Random Forest instead of PLSR, by applying spectral variable selection techniques, by substituting global by local calibrations (i.e., a sample-wise selection of appropriate calibration samples is performed) and by using the approach of spiking to locally adapt calibration models. The optimized techniques will then be validated on existing data sets. Additionally, it will be analysed, whether and to what extent disturbances originating from different soil surface roughness or from different soil water contents can be compensated. Already existing soil spectral libraries (LUCAS, ICRAF-ISRIC) are evaluated to select appropriate samples which may support the definition and optimization of calibration models. In addition, the underlying spectral predictive mechanisms will be analysed (e.g., by 2D-correlation spectroscopy) to elucidate whether a direct or only indirect spectral prediction is feasible for each of the studied soil properties. This is fundamental to clarify whether a prediction model, once calibrated, may be in principle transferred in space and time.