Rethinking Validation for Spatial Machine Learning: Takeaways from the Talk

This article summarizes a keynote talk and workshop from the Machine Learning for Earth Observation 2026 conference. It challenges common assumptions in spatial machine learning validation, such as the ability to predict everywhere, the existence of a single correct validation approach, and the equality of all validation points. The talk introduces concepts like Area of Applicability (AoA), Local Point Density (LPD), and k-Nearest Neighbor Distance Matching (kNNDM) to align validation with the actual prediction domain. The workshop provides practical R workflows for implementing these ideas, emphasizing prediction-domain adaptive evaluation to ensure model reliability in spatial contexts.