R Squared tells you how well the model of the prediction fits the actual. You want as high of an R Squared as possible. Its range should be between 0 and 1 (0% and 100%). Areas with low R squared, you can't trust the predicted raster. Areas with high R squared you can trust more. I don't know anything about RMSE for Temperature Harmonic, so I don't know how to tell you to interpret your result, but based off of the fact that you say for the harmonic that a lower value is a better fit, then I would say the purple is a bad fit, but I feel you need someone in your industry to truly answer that question, however R Squared is works just like it does for Linear Regression in how you interpret it.
Robert “Edward” Gause, GISP | Director of Information Services | HTC | p 843-369-8483 | www.htcinc.net | This is life. Connet with it.