Large-scale subsurface characterization using Multi-Configuration EMI and image classification

An appropriate characterization of the spatial variability of soil properties and layering isvital when modelling the relationships between soil and vegetation. At the field scale, suchsubsurface structures can be investigated with a combination of soil sampling and noninvasivegeophysical measurements (e.g. electromagnetic induction EMI) within aquantitative inversion framework. It is not feasible to apply such quantitativeinterpretation at the farm scale where heterogeneous agricultural fields with variablemanagement are present. This because the required EMI calibration is currently notfeasible and a complex and large study area cannot be measured during the same periodwithout interfering with the agricultural activity. In this study, we used an imageclassification method to analyze high-resolution multi-configuration EMI measurementsand characterize patterns of soil structural organization (layering and texture) and theirimpact on crop productivity at the km2 scale. We collected EMI data on an agricultural areaof 102 ha near Selhausen (NRW, Germany). The area consists of 51 agricultural fieldscropped in rotation. Measurements were collected between April and December 2016,preferably within a few days after harvest. EMI data were automatically filtered,temperature corrected, and interpolated onto a common grid of 1 m resolution. The ECamaps indicated three main sub-areas with different subsurface heterogeneity. Small-scalegeomorphological structures as well as anthropogenic activities such as soil managementand buried drainage networks could be also identified. To obtain areas with similarsubsurface structures, we applied an image classification technique to the EMI data. Forthis, we fused the ECa maps obtained with different coil configurations in a multibandimage and applied supervised and unsupervised classification methodologies. Both showedgood results in reconstructing observed patterns in plant productivity and the subsurfacestructures associated with them. However, the supervised methodology proved moreefficient in classifying the whole study area. In a second step, we selected hundred locationswithin the study area and obtained a soil profile description with type, depth, and thicknessof the soil horizons. Each layer with characteristic texture was sampled and a total of 570samples were analyzed to obtain information on grain size distribution. Using this groundtruth data, it was possible to assign a typical soil profile to each of the main classes obtainedfrom the classification. The proposed methodology was effective in producing a highresolution subsurface model in a large and complex study area that extends well beyond thefield scale