Introduction
Figure S1 shows that the relevance scores computed by LRP are proportional to the derivative of the output variable with respect to the input variable, providing an intuitive way to interpret relevance scores. Figure S2 shows that fitting a linear model on the relevance scores produced by LRP are able to reproduce the output of the neural network. Figure S3 shows a full breakdown of how well a linear model at each site can reproduce the turbulent heat fluxes at each other site.