A BAYESIAN BELIEF NETWORK TO OPERATIONALISE THE CONCEPTS OF SOIL QUALITY AND HEALTH
Combining quantitative and qualitative information in a formal structure to make the concepts of soil quality and soil health more concrete and measurable.
Saturday, April 1, 2017 - 14:15
- Scottish Government
- Soil Security Programme
Though soil health and soil quality are important to the concept of soil security, no strict definition exists for either. This is partly because they derive from both opinion and measurement. Bayesian Belief Networks can successfully combine these types of hard and soft data.
‘Soil Quality’ and ‘Soil Health’ are general terms for indicators that are associated with ‘Soil Security’. However, neither of these terms within quotation marks is easy to define. Neither are they easy to quantify rigorously in a way that avoids dispute. Nonetheless all three terms have traction with policy makers, and with land managers and regulators. Indicators provide benchmarks for ranking different places or practices, and deciding where to deploy effort to bring about change as effectively and economically as possible; and they provide a means to assess afterwards whether or not and to what extent this change has actually been brought about.
As a result, indicators of this kind are attractive to stakeholders. Indicators often rely on expert opinion for their derivation, but experts differ. Even apparently objective biophysical measurements are subject to error and worse, the soil itself varies from place to place and even time to time. It is not clear how to eliminate bias or how to weight the different kinds of information – opinion and measurement.
There is therefore, scope for developing a rigorous, scientific approach to soil quality and health that incorporates expert-derived opinion alongside physically-based measurements in our understanding of Soil Quality and Health (SQH) in a scientific manner.
Bayesian Belief Networks are graph-based, directional networks that can incorporate probability distributions of these various kinds of data. Essentially the directedness leads from multiple pieces of data to a conclusion – in our case a rating of SQH. The network is self-learning, in that any additional soils and data for which quality assessments are available will re-inforce the pathways that decide the quality rating. In use, SQH ratings for additional soils that contain even partial data can still be obtained if the net defaults to mean values where data is missing.
A simple example of the proposed BBN concept: functions are inferred from data including expert opinion, Soil Quality or Soil Health is inferred from this structure.