I determined the vegetation type (4 types all in all) at 120 points (stratified random sampling) in my study area (650ha). I want to use the points to train a statistic model (eg. random forest) based on some environmental predictors. I also want to test the model performance against a truly independent hold-out sample.
Can I just divide my study area in half and use one part of the points for training and the other for testing? Would this be considered a true independent validation? How can I calculate the distance that the points need to have to be spatially independent? I want to reduce the effect that a validation point very close to a calibration point has probably the same vegetation type because of its spatial closeness.
Also there is a west-east gradient in some of the predictor data- should i therefor divide the study area in north and south?