This dataset is part of both deliverable 4.2 and 4.3 and was produced by the WP4 team of the Landmark H2020 project. It contains the Bayesian networks for the following crops:
Barley
Silage maize
Grain maize
Intensive grasslands - hay
Intensive grasslands - grazed
Extensive grasslands - grazed
Rapeseed
Peas
Potato
Rye
Sugar beet
Sunflower
Spring wheat
Spring durum wheat
Soybean
Wheat
Durum wheat
For each crop, a Bayesian network was derived from a DayCent crop simulation. DayCent, the daily time-step version of the CENTURY biogeochemical model, simulates fluxes of carbon (C) and nitrogen (N) among the atmosphere, vegetation, and soil. It incorporates a wide range of sub models including soil water content and temperature by layer, plant production and allocation of net primary production (NPP), etc. The DayCent model was applied on a 12-year simulation period following the procedure in Lugato et al. (2017).
From this simulation only the last 4-years of the model run were used to develop the data table for the machine learning. The daily time-step output from DayCent was converted into a data table with seasonal and yearly averages and totals. This table includes a wide range of explaining variables from DayCent (Table 1). In total 14 indicators were derived from the DayCent modelling results to describe different aspects of the four soil functions.
To develop the BNs, each explaining variable and soil function indicator from the data table, which was derived from the DayCent dataset, was discretized into 5 classes using the Jenks natural breaks classification method. For each crop, a BN was derived from these discretized datasets with a Bayesian Search algorithm. During the Bayesian Search calculation, connections between explaining variables were forbidden, while connections between explaining variables and soil function variables were always starting from the explaining variables. The crop BNs were validated with a k-fold cross-validation during which the network is tested in its ability to predict indicator values from the dataset. The BNs were built and validated using SMILE Academic 2.2 software (BayesFusion LLC, University of Pittsburgh, PA, USA) (Druzdzel, 1999).
These BNs were later used to evaluate the soil functions under current conditions as well as a range of policy options. The results of these calculations can be found in https://data.inra.fr/dataverse/LandmarkH2020.
More information regarding calculation and interpretation of both this dataset and the soil function maps used to calculate the z-scores can be found in:
Vrebos D., F. Bampa, R. Creamer, A. Jones, E. Lugato, L. O’Sullivan, P. Meire, R.P.O. Schulte, J. Schröder and J. Staes (2018). Scenarios maps: visualizing optimized scenarios where supply of soil functions matches demands. LANDMARK Report 4.3.
Available from www.landmark2020.eu.