It is usually a mistake to estimate correlations between grids when one or both have been interpolated: the estimate often reflects the interpolation method rather than the data themselves. Moreover, you will always grossly over-represent the degrees of freedom, typically leading to falsely high estimates of confidence and precision.
When one dataset is a grid of measurements, or derived from measurements (as NDVI is) and the other is a point dataset, it's both better and simpler to extract the grid values at the point locations, forming an (x,y) dataset. In this case x would be the temperature at one of the points where temperature actually is observed and y would be the NDVI value at that same location. There will therefore be one (x,y) datum for each temperature datum available to you.
ArcGIS is not the software to perform a rigorous analysis of an (x,y) dataset: as @Mike Toews suggests, use statistical software. There are Web apps that will compute correlations, perform ordinary least squares regression, draw scatterplots, and give you confidence intervals for the parameters. Even handheld calculators will do this... In general, though, for anything except self-instruction, you're best off using well-tested software such as
R or a commercial statistics package.