Discussion Questions

Although machine learning algorithms (MLAs) have demonstrated strong potential in improving the accuracy of spatial predictions, several methodological challenges still need to be addressed before they can be fully and reliably implemented in spatial modeling. Key open questions include:

1. How can spatial simulations be generated that accurately reproduce the underlying spatial autocorrelation structures when using Random Forest and related ML models?
2. How can predictions be effectively produced across different spatial supports, such as translating from point-based data to block-level estimates and vice versa?
3. How can extrapolation issues be mitigated, both in the feature space (beyond observed variable ranges) and in the geographical space (beyond sampled areas)?
4. How can models appropriately account for spatial and spatiotemporal clustering of observations to ensure unbiased and realistic predictions?