Investigating the Multi-Scale Reproducibility of GeoAI Models for Urban Perception: Prediction and Metadata Packages

This record provides the prediction and derived metadata packages accompanying the SIGSPATIAL study on multi-scale reproducibility of GeoAI models for urban perception. It includes per-seed test predictions for 328 reported training runs and derived metadata used for scale-sensitive stability analysis, source decomposition, ensembling evaluation, multi-split robustness checks, and ICC robustness checks. Raw Place Pulse 2.0 imagery is not redistributed.