Pooled Accuracy, Hidden Fragility: Reproduction Package

Data and code reproducing every table and figure in "Pooled Accuracy, Hidden Fragility: A Spatial Cross-Validation Diagnostic for Socioeconomic Augmentation of Satellite Cropland-Use Models in Rural Thailand" The paper tests whether household-survey-derived socioeconomic variables improve satellite-based prediction of cropland-use weakening in rural Thailand. Under spatial leave-one-block-out cross-validation, pooled AUC-ROC suggests a strong result (+0.215); evaluated the way the cross-validation literature recommends — per block, then averaged — the gain collapses to +0.015, indistinguishable from noise under a 500-permutation test, a positive control, cardinality-matched noise controls, three calibrated label thresholds, and standard permutation feature importance. This repository contains the full diagnostic chain as a reproducible package.