Answers are grounded in this paper and cite their sources.
Thematic maps generated by machine-learning classifiers, including Local Climate Zone (LCZ) maps, enter quantitative Earth observation science only after geospatial operators aggregate raster observations within the mapped zones. Zonal statistics constitute this interface, yet the internal coherence of their outputs is assumed rather than tested. This study audits it for LCZ-stratified land surface temperature (LST) analysis in Riyadh, Saudi Arabia. The ArcGIS Pro Zonal Statistics as Table tool and a scripted open-source Python workflow (GeoPandas, rasterstats, NumPy) were executed on a twelve-class LCZ map and a Landsat 8/9 winter 2024/2025 mean-composite LST raster, and their outputs were audited for completeness, quantile monotonicity, moment–count consistency, and agreement. The results indicate that both implementations returned complete records for all twelve classes and agreed on the minimum, maximum, mean, standard deviation, median, and sum to within 5 × 10 −9 °C, so that class-level thermal statistics, thermal ranking (Spearman’s ρ = 1.000, Kendall’s τ = 1.000), and inter-class contrasts are identical between platforms. The two outputs nevertheless disagree on the cell count of seven of the twelve zones (58.3%), by −0.50% to +0.75%; these are exactly the zones adjoining the NoData margin, and the five interior zones agree exactly. A moment–count test locates the defect: the product of the reported mean and count reproduces the reported sum exactly throughout the ArcGIS table but departs from it in precisely the seven affected classes of the Python table, and dividing that table’s sum by its mean recovers the ArcGIS count in every case. The reported count and the reported moments of the open workflow therefore derive from different cell sets, an inconsistency that is invisible in the table, raises no error, and propagates to every area-derived quantity: class extents are displaced by up to 4.4 km 2 and the analyzed extent by 7.6 km 2 . Three inexpensive validation rules, completeness auditing, quantile monotonicity, and moment–count consistency, are proposed as automated components for AI-era Earth observation pipelines; only the last detected the defect. Future work should extend the audit to further platforms, patch-level zones, and in-situ validation.
No takes yet. Share an insight, caveat, or question.
Eisa et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: