With progressive implementation of precision agriculture (PA) techniques in current agricultural/ viticultural practice, the need for high-resolution information on soil properties at low effort and cost is increasing. Moreover, climate change and extended drought periods do even increase this demand. Evaluating soil fertility and carbon storage potential of arable fields and vineyards, e.g. for future economic assessment of ecosystem services, requires spatially resolved soil data. Soil texture is a core soil property parameter for water and carbon storage. Proximal gamma spectroscopy (GS) is considered as appropriate tool for topsoil texture prediction, at least when prediction models are calibrated site specifically. However, for a broader use of GS, prediction model transferability between widely varying sites is a prerequisite. To now, the diversity of geopedological conditions, i.e. the sum of petrological-mineralogical and pedological settings, impedes calibration of universally valid, precise linear models. In this respect, prediction models based on support vector machines (SVM) are proven to have advantages over linear models. We therefore tested the performance of site-independently calibrated SVM-based prediction models across widely differing geopedological conditions in Germany.

In general, prediction quality of site-independent models was poor when applied to sites that were not in the calibration dataset. For one third of the sites under study, the mean absolute errors (MAE) for sand, silt, and clay prediction were greater than 10 % and consequently not sufficient for PA.  Nevertheless, for “neighboring” sites, i.e. sites located in areas with similar geopedological conditions, MAE was below 5% for all three texture fractions. This result points out that GS has not to be calibrated strictly site-specific (as proposed by some studies) but according to specific geopedological units.

Prediction quality at unknown sites was improved by spiking the calibration sample set with few samples of the respective study site. Therefore, SVM calibrated prediction models can be trained to specific geopedological conditions with little effort.

However, there is still a way to go for a broad and universal application of GS in terms of PA. Future research efforts will be to search for suitable co-variables for the underlying geopedological conditions and to enlarge the database with new study sites. This research is still going on and more sites will be included before the conference.