Advanced interpolation of physical and hydrological soil properties using machine learning techniques: Poster presented at: Eurosoil 2021 Online Conference, 23-27 August 2021

Genf

2021

1 S.

Sonderdruck

12907S

40002730

A digital soil mapping approach was developed to predict physicalsoil properties, which are relevant for water storage capacity. Model resultswill be used for assessing the current and future water budget of 1 Mio haforested area in the province of Styria/Austria in the framework of thedynamic forest classification project FORSITE. Based on area-wide explanatory data and more than 2000 soil samplingsites, texture, soil organic matter content, bulk density and soil depth wereestimated for a 10m resolution grid with a neural network based model. Available soil data consists of several distinct datasets with different soilproperty variables from different sampling locations. The data include bothfield surveys and laboratory measurements. Topographic variables derivedfrom the digital elevation model, geological substrate classification andclimatological data were used as explanatory data layers. For each input soil data set, a separate neural network was created andtrained. For each network, the respective variables of the correspondingdata set were estimated at the grid points of the other data sets. This outputwas processed by a subsequent neural network to produce an overallestimate of the target variables.