I have a time series with two maps that represent forest cover in two different years. I have created a continues raster surface that represents deforestation rates based on zonal (Moving window) statics calculation using different windows sizes for testing. I have divided the raster map into several grids of 10x10 kms. I want to use the RMSE to evaluate my model performance.
When the surface is divided into grids, I can obtain a sample of thousands of grids. I want to evaluate the accuracy of the maps with another historical deforestation period.
Do I need to have samples that are not spatially autocorrelated, instead of using all the grids that I have to evaluate the RMSE of the models?
Or can I calculate the RMSE of the models with all the grids and without considering that the data are highly spatially autocorrelated?
In other words, does the RMSE require non spatially autocorrelated samples?