Interpretative Machine Learning as a Key in Recognizing the Variability of Lakes Trophy Patterns

Abstract

The paper presents an application of interpretative machine learning to identify groups of lakes not with similar features but with similar potential factors influencing the content of total phosphorus – Ptot. The method was developed on a sample of 60 lakes from North-Eastern Poland and used 25 external explanatory variables. Selected variables are stable over a long time, first group includes morphometric parameters of lakes and the second group encompass watershed geometry geology and land use. Our method involves building a regression model, creating an explainer, finding a set of mapping functions describing how each variable influences the outcome, and finally clustering objects by ’the influence’. The influence is a non-linear and non-parametric transformation of the explanatory variables into a form describing a given variable impact on the modeled feature. Such a transformation makes group data on the functional relations between the explanatory variables and the explained variable possible. The study reveals that there are five clusters where the concentration of Ptot is shaped similarly. We compared our method with other numerical analyses and showed that it provides new information on the catchment area and lake trophy relationship.

Keywords

• total phosphorus
• interpretative machine learning
• random forest
• masurian lakes