High resolution soil data are an essential prerequisite for the application of precision farming techniques. Sensor-based evaluation of soil properties may replace or at least reduce laborious, time-consuming and expensive soil sampling with subsequent measurements in the lab.

Gamma spectrometry usually provides information that can be translated into topsoil texture data after calibration. This is because the natural content of the radioactive isotopes 40-K, 232-Th, and 238-U as well as the Total Counts, i.e., the sum of gamma emitting nuclides, is correlated with soil mineralogy and, consequently, with soil texture. Yet, preliminary studies at different study sites revealed that transferability of local or regional calibration models is limited mostly because of different parent material and pedogenesis. Predominantly mineralogical soil composition and weathering processes lead to varying element (nuclide) contents and, in consequence, to highly variable gamma spectra. However, nonlinear models using support vector machines (SVM) have been successfully applied in order to predict soil properties. Such models are able to accommodate a variety of different landscape settings and can thus successfully reduce the need for sampling, analysis and local calibration. In this regard, our recently published results rely on spectra from ten German arable fields in different geological and pedological environments and were recorded directly in the field. To calibrate a model with high generalization abilities for texture prediction from gamma spectra using SVM, we used a tractor-based gamma spectrometer in a stationary stop-and-go mode.

In the ongoing project “I4S - Intelligence for Soil” in the frame of the German Federal BonaRes program we broadened the database to more parent materials and environmental settings. Further, we evaluated the preciseness and usability of spectra and optimized models recorded on-the-go, i.e. from a driving tractor. Promising results were achieved inter alia on the central I4S-project test site Görzig, established for the evaluation of a range of soil sensors for precision farming applications. Prediction errors for soil texture as derived from on-the-go gamma data were small and, consequently, distinction of soil texture classes was successful. Texture classes are the basis for variable rate application adapted to crop demand, at least for P and K. In many cases, prediction of sand, silt, and clay was even as precise as conventional lab analyses. This approach enables us to directly adapt texture-related agronomic measures such as lime and fertilizer application on-the-go. Examples for such adapted variable rate treatments will be given.