Studying The Spatial Prediction Method Based On Conformal Prediction

Spatial prediction has become a topic of great scientific significance and application in many research fields such as earth science,social economy,and health care,etc.Accurately predicting unknown values and quantifying their uncertainty can provide a important reference for decision making on key issues.In fact,the actual change processes behind the spatial data are very complex,but the current spatial prediction methods are often model-based,which can lead to large prediction errors when model assumptions are not valid.Therefore,it is important to explore model-free spatial prediction methods with a wider range of applicability.In this thesis,we combine the spatial interpolation based on forest algorithm and the conformal prediction method and propose a forest-based spatial conformal prediction method to construct a more suitable prediction interval for general cases.In order to adapt to the heterogeneity of the data and construct a shorter prediction interval,the conformal quantile regression method was used.In addition,for more accurate quantile estimation,continuous ranked probability score(CRPS)was used as the loss function for variable selection and adjustment of forest parameters.Finaly,the proposed method is compared with the existing spatial conformal prediction methods based on kriging in simulation studies and real data analysis.The results show that the proposed method outperforms the existing spatial conformal prediction methods.The proposed method can guarantee a finite sample coverage without model assumptions and obtain a shorter interval length.