How does extreme point sampling affect non-extreme simulation in geographical random forest?

Spatial heterogeneity brings numerous uncertainties to training datasets in the modeling process. An arbitrary selection of training samples can result in a biased simulation. Although previous research provides a chance of reducing the degree of spatial variance through homogeneous divisions, detailed information regarding the impact of the configuration of divisions for training remains unknown. Moreover, few studies investigate the cross impact of extreme sampling on non-extreme simulation. Therefore, we extend previous research to investigate the cross impact and further examine whether the divisions of extremely high (EXH) and low (EXL) quantiles contribute equally to the simulation bias when employing the spatial stratified sampling. Statistical assessment demonstrates that the selection of extreme training sample does affect the non-extreme simulation. The model has the best performance (RMSE: 2.735, VE: 7.481, Bias: -0.033) when the least proportion (25%) of EXH and EXL was selected for training. Further analysis also indicated that the EXH and EXL divisions contribute unequally to the process. Particularly, the non-extreme simulation is more sensitive to the EXH training data with a steeper change rate of 0.043. This research provides a critical insight into the extreme point sampling for a machine learning process. Different sensitivity of division calls upon that extreme training sample should be adjusted on a basis of percentage rather than their amounts when applying stratified sampling in Geographical Random Forest.