Global environmental changes have resulted in changes in key ecosystem services that soils provide. It is necessary to have up to date soil information on regional and global scales to ensure that these services continue to be provided. As a result, Digital Soil Mapping (DSM) research priorities are among others, advancing methods for data collection and analyses tailored towards large-scale mapping of soil properties. Scientifically, this thesis contributed to the development of methodologies, which aim to optimally use remote and proximal sensing (RS and PS) for DSM to facilitate regional soil mapping. The main contributions of this work with respect to the latter are (I) the critical evaluation of recent research achievements and identification of knowledge gaps for large-scale DSM using RS and PS data, (II) the development of a sparse RS-based sampling approach to represent major soil variability at regional scale, (III) the evaluation and development of different state-of-the-art methods to retrieve soil mineral information from PS, (IV) the improvement of spatially explicit soil prediction models and (V) the integration of RS and PS methods with geostatistical and DSM methods.
A review on existing literature about the use of RS and PS for soil and terrain mapping was presented in Chapter 2. Recent work indicated the large potential of using RS and PS methods for DSM. However, for large-scale mapping, current methods will need to be extended beyond the plot. Improvements may be expected in the fields of developing more quantitative methods, enhanced geostatistical analysis and improved transferability to other areas. From these findings, three major research interests were selected: (I) soil sampling strategies, (II) retrieval of soil information from PS and (III) spatially continuous mapping of soil properties at larger scales using RS.
Budgetary constraints, limited time and available soil legacy data restricted the soil data acquisition, presented in Chapter 3. A 15.000 km2 area located in Northern Morocco served as test case. Here, a sample was collected using constrained Latin Hypercube Sampling (cLHS) of RS and elevation data. The RS data served as proxy for soil variability, as alternative for the required soil legacy data supporting the sampling strategy. The sampling aim was to optimally sample the variability in the RS data while minimizing the acquisition efforts. This sample resulted in a dataset representing major soil variability. The cLHS sample failed to express spatial correlation; constraining the LHS by a distance criterion favoured large spatial variability over short distances. The absence of spatial correlation in the sampled soil variability precludes the use of additional geostatistical analyses to spatially predict soil properties. Predicting soil properties using the cLHS sample is thus restricted to a modelled statistical relation between the sample and exhaustive predictor variables. For this, the RS data provided the necessary spatial information because of the strong spatial correlation while the spectral information provided the variability of the environment (Chapter 3 and 6). Concluding, the RS-based cLHS approach is considered a time and cost efficient method for acquiring information on soil resources over extended areas.
This sample was further used for developing methods to derive soil mineral information from PS, and to characterize regional soil mineralogy using RS. In Chapter 4, the influences of complex scattering within the mixture and overlapping absorption features were investigated. This was done by comparing the success of PRISM’s MICA in determining mineralogy of natural samples and modelled spectra. The modelled spectra were developed by a linearly forward model of reflectance spectra, using the fraction of known constituents within the sample. The modelled spectra accounted for the co-occurrence of absorption features but eluded the complex interaction between the components. It was found that more minerals could be determined with higher accuracy using modelled reflectance. The absorption features in the natural samples were less distinct or even absent, which hampered the classification routine. Nevertheless, grouping the individual minerals into mineral categories significantly improved the classification accuracy. These mineral categories are particularly useful for regional scale studies, as key soil property for parent material characterization and soil formation. Characterizing regional soil mineralogy by mineral categories was further described in Chapter 6. Retrieval of refined information from natural samples, such as mineral abundances, is more complex; estimating abundances requires a method that accounts for the interaction between minerals within the intimate mixture. This can be done by addressing the interaction with a non-linear model (Chapter 5).
Chapter 5 showed that mineral abundances in complex mixtures could be estimated using absorption features in the 2.1–2.4 µm wavelength region. First, the absorption behaviour of mineral mixtures was parameterized by exponential Gaussian optimization (EGO). Next, mineral abundances were successfully predicted by regression tree analysis, using these parameters as inputs. Estimating mineral abundances using prepared mixes of calcite, kaolinite, montmorillonite and dioctahedral mica or field samples proved the validity of the proposed method. Estimating mineral abundances of field samples showed the necessity to deconvolve spectra by EGO. Due to the nature of the field samples, the simple representation of the complex scattering behaviour by a few Gaussian bands required the parameters asymmetry and saturation to accurately deconvolve the spectra. Also, asymmetry of the EGO profiles showed to be an important parameter for estimating the abundances of the field samples. The robustness of the method in handling the omission of minerals during the training phase was tested by replacing part of the quartz with chlorite. It was found that the accuracy of the predicted mineral content was hardly affected. Concluding, the proposed method allowed for estimating more than two minerals within a mixture. This approach advances existing PS methods and has the potential to quantify a wider set of soil properties. With this method the soil science community was provided an improved inference method to derive and quantify soil properties
The final challenge of this thesis was to spatially explicit model regional soil mineralogy using the sparse sample from Chapter 3. Prediction models have especially difficulties relating predictor variables to sampled properties having high spatial correlation. Chapter 6 presented a methodology that improved prediction models by using scale-dependent spatial variability observed in RS data. Mineral predictions were made using the abundances from X-ray diffraction analysis and mineral categories determined by PRISM. The models indicated that using the original RS data resulted in lower model performance than those models using scaled RS data. Key to the improved predictions was representing the variability of the RS data at the same scale as the sampled soil variability. This was realized by considering the medium and long-range spatial variability in the RS data. Using Fixed Rank Kriging allowed smoothing the massive RS datasets to these ranges. The resulting images resembled more closely the regional spatial variability of soil and environmental properties. Further improvements resulted from using multi-scale soil-landscape relationships to predict mineralogy. The maps of predicted mineralogy showed agreement between the mineral categories and abundances. Using a geostatistical approach in combination with a small sample, substantially improves the feasibility to quantitatively map regional mineralogy. Moreover, the spectroscopic method appeared sufficiently detailed to map major mineral variability. Finally, this approach has the potential for modelling various natural resources and thereby enhances the perspective of a global system for inventorying and monitoring the earth’s soil resources.
With this thesis it is demonstrated that RS and PS methods are an important but also an essential source for regional-scale DSM. Following the main findings from this thesis, it can be concluded that: Improvements in regional-scale DSM result from the integrated use of RS and PS with geostatistical methods. In every step of the soil mapping process, spectroscopy can play a key role and can deliver data in a time and cost efficient manner. Nevertheless, there are issues that need to be resolved in the near future. Research priorities involve the development of operational tools to quantify soil properties, sensor integration, spatiotemporal modelling and the use of geostatistical methods that allow working with massive RS datasets. This will allow us in the near future to deliver more accurate and comprehensive information about soils, soil resources and ecosystem services provided by soils at regional and, ultimately, global scale.