PedoScale - Hyperskalige Bodenprognose und Analyse der Bodenbildung
Zusammenfassung der Projektergebnisse
Soil is an indispensable and complex environmental resource, formed by the influence and interactions between all other environmental covariates such as relief, climate, parent material and organisms. Soil is vital and a non-renewable resource under increasing pressure. Furthermore, soils are among the largest carbon reservoirs in the world and offer potential for increased carbon storage to mitigate climate change. Accurate and high-resolution soil maps are therefore urgently needed. Soil maps are usually created by soil scientists during a field survey. Based on the visual and haptic analysis of soil profiles or soil cores, existing maps of environmental covariates and experience, the expert develops a mental model of soil genesis and thus of the spatial distribution of soils. This is the basis to delineate the soil mapping units and to create a classical soil map. This process is time-consuming, expensive and often subjective. Over the last two decades, research branches with a strong overlap between soil science and statistics have developed that focus on the statistical and mathematical description of soils. These are pedometrics and in particular digital soil mapping. The basic concept of digital soil mapping is environmental correlation, i.e. the identification and use of statistical dependencies between soil units/properties and environmental covariates. The spatial distribution of soil properties shows structural dependencies on environmental covariates as well as spatial dependencies such as spatial autocorrelation, which Tobler expressed in his formulation of the first law of geography, namely that "everything is related to everything else, but near things are more related than distant things". Although there are geostatistical models that use spatial autocorrelation to interpolate soil properties and thus produce soil property maps, spatial dependence is a symptom of structural dependencies that occur at multiple scales. Therefore, the focus of the Pedoscale project was on the development of new multi-scale and hyperscale modelling approaches that provide high modelling accuracies. In addition to the fact that the new methods are based on structural relationships and eliminate spatial autocorrelation in the model residuals, they allow the interpretation of the data and thus of soil formation. Such an interpretation is not limited to the assessment of the importance of a covariate, e.g. of slope or curvature on soil thickness, but also allows the analysis of the scales on which a covariate has an influence. Other methods developed within the Pedoscale project even allow the spatial localization of a particular landscape element, such as a peak, which may affect the soil property value of a particular sample, even if it is several kilometres away. This influence may be due, for example, to the fact that the peak influences the wind speed and thus the sedimentation of a pollutant at the sample location. The Pedoscale project is situated at the interface between soil science, pedology and geomorphology on the one hand and machine learning research on the other hand. The methods used originate mainly from the fields of image processing, computer vision and in particular feature engineering, which have been applied to environmental data in this project. Although the development carried out in the Pedoscale project can be seen as multidisciplinary basic research, the proposed approaches are straightforward to implement and to apply and therefore have great potential for wider application and audience. With four new methods for multi-scale environmental modelling and several new concepts for pedological interpretation of the models, the project can lay the foundation for a new field of soil science and environmental modelling, namely contextual spatial modelling. The approaches developed in Pedoscale show that multiscale modeling can increase accuracy by up to 0.5 with respect to the R², which means that an additional 50% of the variance in the data can be explained if relevant multiscale features are available. This is a strong argument to investigate and develop future multi-scale and hyperscale modelling approaches. The paper on Multi-scale Digital Soil Mapping with deep learning by Thorsten Behrens, Karsten Schmidt, Robert MacMillan, Raphael Viscarra Rossel, which was published in Scientific Reports (8, 15244) in 2018, was awarded Best Paper 2018 in Pedometrics by the Pedometrics Commission of the International Union of Soil Sciences (IUSS).
Projektbezogene Publikationen (Auswahl)
- (2019) Teleconnections in spatial modelling. Geoderma 354 113854
Behrens, Thorsten; MacMillan, Robert A.; Viscarra Rossel, Raphael A.; Schmidt, Karsten; Lee, Juhwan
(Siehe online unter https://doi.org/10.1016/j.geoderma.2019.07.012) - 2014. Hyperscale digital soil mapping and soil formation analysis. Geoderma 213, 578-588
Behrens, T., Schmidt, K., Ramirez-Lopez, L., Gallant, J., Zhu, A-X., Scholten, T.
(Siehe online unter https://doi.org/10.1016/j.geoderma.2013.07.031) - 2017. Multiscale contextual spatial modelling with the Gaussian scale space. Geoderma 310: 128-137
Behrens, T., Schmidt, K., MacMillan, R.A., Viscarra Rossel, R.A.
(Siehe online unter https://doi.org/10.1016/j.geoderma.2017.09.015) - 2018. Multiscale Digital Soil Mapping with deep learning. Scientific Reports 8,15244. (Awarded the Best Paper 2018 in Pedometrics)
Behrens, T., Schmidt, K., MacMillan, R.A., Viscarra Rossel, R.
(Siehe online unter https://doi.org/10.1038/s41598-018-33516-6) - 2018. Spatial modelling with Euclidean distance fields and machine learning. European Journal of Soil Science 69/5
Behrens, T., Schmidt, K., Rossel, R.A., Gries, P., Scholten, T., MacMillan, R.A.
(Siehe online unter https://doi.org/10.1111/ejss.12687)
