Now let’s assume that yellow points (land) are healthy ionomes and that blue points (water) are unhealthy. Indeed, there is no reason why the delimitation of a health status should return a regular shape.
Of course, health is not binary. In particular, if we use yield as a health index, the problem could be approached as a regression rather than a classification. A regression would add some flexibility to the approach, because the threshold that triggers an action could be decided by the professional. Although they should not be ruled out in the first place, in my experience regressions have often returned inaccurate predictions.
I thus prefered to use a yield threshold: when a predicted yield-category is low, a red flag is raised and correction measures should be applied. Hence, the boolean approach with healthy/unhealthy becomes analogous to the land/water dichotomy.
Because I work in prediction mode, I randomly assigned the points to training (70%) or testing (30%) sets to respectively fit and evaluate our future models.

Using univariate ranges

Ionomic ranges are commonly used to differentiate groups of organisms. They are usually reported in terms of concentrations \cite{Neugebauer2018}, unorganized ratios \cite{Bitts_nszky_2016} or log-ratios \cite{Prater_2018},  centered log-ratios \cite{Bar_g_2014} or isometric log-ratio balances \cite{Parent2013}. Balance ranges to diagnose health (of land) with 95% confidence level on the training set would look like the shade region in figure \ref{235902}.