DFG SFB TR32 - Patterns in Soil-Vegetation-Atmosphere-Systems
- +49 241 80 49881
- Send Email
The heterogeneous distribution of soil physical properties (e.g. soil structure, porosity, permeability) in the subsurface plays a significant role in the upscaling of fluxes from the vadose zone to larger-scale regional models. These heterogeneities are dominated by different patterns on different scales of subsurface structures (i.e. configuration of soil layers or texture).
Estimates about the relevant spatial and vertical distribution has been non-invasively obtained from geophysical measurements using electromagnetic induction (EMI) (project B6). It has been shown that soil patterns in the subsurface correlate with crop growth, and a clear link is present with remote sensing information, for example Leaf Area Index (LAI), investigated in detail in project B6 and D2. The observed correlations provide a potential pathway to extend the locally available high-resolution analysis of subsurface patterns to larger areas where only remote sensing data is available.
To reach this goal, we have employed established geostatistical methods in combination with novel unsupervised machine learning approaches for cluster analysis and pattern detection to create a knowledge-driven simulation method to predict soil patterns from remote sensing information. In this way, we attempt to link existing and on-going efforts in other projects to characterise subsurface heterogeneities, and the analysis of large-scale geophysical process simulation and parameter inversion.
The detailed geophysical investigations in cluster B on the field scale of the available test sites (e.g. Selhausen), which are aimed at reaching a model extent of up to one kilometer in Phase III, are used in combination with repeated LAI images to determine spatial patterns and parameter correlations using an unsupervised machine learning algorithm. The determined correlations are then the basis for an advanced investigation of soil heterogeneities on the basis of remote sensing information alone. With this project, we contribute to to link to the next level of process simulations on the kilometer scale by developing cutting edge pattern extraction algorithms and applying it to the field and regional scale. Very promising preliminary results to date have been achieved in training our model of pattern recognition on the available data sets, and the next step is to validate the pattern simulation technique with newly acquired geophysical survey information.