General Method for Predicting Soil Data Via Pattern Matching on Pedotransfer Functions

Abstract

This thesis describes a generalized methodology for predicting numerical data based on patternmatching over sets of pedotransfer functions (PTFs). The test apparatus developed for this work was Katoomba 1, an instance of a Soil Inferencing System or a SINFERS (McBratney and Minasny 2004; Tranter 2009). Specifically, Katoomba is a rule-based expert system, deployed as a web-application, that is modelled on a hypothetical human soil scientist who possesses expert knowledge for predicting soil properties. Katoomba’s knowledge is stored primarily as a set of rules about soil property PTFs. Given any initial set of input soil properties, Katoomba uses its knowledgebase of soil property PTF rules and that initial input set to infer new soil property values and their uncertainties. Each new output is fed back as a new input, producing a new cycle of inference until no more properties can be inferred from the original input set. This thesis is significant because there is a growing need for inexpensive and reliable soil property data, and systems like Katoomba can meet that need. The generalization presented herein extends the methodology used in Katoomba to any domain that uses similar mathematics to predict data, such as hydrology. Furthermore, the work is original in that no general purpose soil property inferencing system has ever been built.A large portion of this thesis concerns the designing, implementing, and testing of Katoomba. Unlike most research-oriented theses that adhere purely to the scientific method for their structure and flow, this thesis also follows well-established software engineering and knowledge engineering practices for overcoming the challenge of converting the abstract notion of a SINFERS into a functioning one. To be sure, there are tables and graphs of experimental results and discussions of the predictions made by Katoomba, both in terms of soil property values and their uncertainties. However, building such a system is a lot like constructing a proof by mathematical induction. If the proof holds for 1 and it holds for (n + 1), then it is safe to assume that it works for any number, n – there is no need to keep trying different values of n. Thus, this thesis is more about the specific sub-problems encountered whilst building Katoomba, the reasoning and procedures used to solve them, and the resulting implementation decisions and their implications. As the heavy technical exposition gives way to more placid reflection, the thesis culminates in a discussion about how the SINFERS approach could be generalized to other domains that use PTF-like functions.