We explore a methodology to identify spatial boundaries of non-observed spatial factors. Insurance data has heavy tailed severity and scarce count frequency, making it difficult to separate contextual (spatial) and systematic (random) risk. Parametric models kriging and non-parametric models such as Kernel smoothing and Local Polynomial Regression are adapted to the insurance context. The smoothing surfaces are used in a binning method to adjust premiums. Applying this methodology can also be used in a Big Data context to correlate the higher or lower risk areas with new explanatory variables and reduce spatial uncertainty.