Dataset and code for pseudo-space learning-based urban heat exposure risk prediction

Code description
This repository provides the implementation of a pseudo-space learning (PSL) framework for spatial prediction and uncertainty estimation. The workflow integrates machine learning models with residual kriging to improve prediction accuracy and capture spatial heterogeneity.
The code performs the following steps:
Data preprocessing:
The input geospatial data (shapefile format) are loaded and cleaned by removing missing values. Selected variables are normalized to ensure comparability across features.Model training with cross-validation:
A 5-fold cross-validation strategy is used to train and evaluate models. Two machine learning algorithms are implemented:Multi-layer perceptron (MLP)Gradient boosting machine (GBM)Pseudo-space configurations:
Three spatial configurations are constructed:Geographic space (longitude, latitude)Attribute space (e.g., population density, footprint indicators)Mixed pseudo-space (combining spatial and attribute variables)Residual kriging correction:
Model residuals are interpolated using kriging (via the automap package), and the kriging predictions are added back to the machine learning outputs to obtain corrected predictions.Model selection:
The optimal model for each configuration is selected based on a combined error metric (MAE + RMSE).Model saving and outputs:
The best-performing models are saved as .RData files for reuse.Spatial structure analysis:
Variogram analysis is conducted on residuals in each space to evaluate spatial dependence, and the results are visualized and exported as figures.