Silent inconsistencies at the aggregation stage of Earth observation pipelines: an internal-consistency audit of zonal statistics for local climate zone–based land surface temperature analysis

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.

Authors

Institutions

Publication Details

Journal
Frontiers in Environmental Science
Published
2026-11-16
DOI
https://doi.org/10.3389/fenvs.2026.1971523
Primary Topic
Urban Heat Island Mitigation
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Silent inconsistencies at the aggregation stage of Earth observation pipelines: an internal-consistency audit of zonal statistics for local climate zone–based land surface temperature analysis

Islam M. Hamdi, Youssef M. Youssef, Dominika Dąbrowska, Omar I. Aboulola et al.
Frontiers in Environmental Science
Urban Heat Island Mitigation
article

Silent inconsistencies at the aggregation stage of Earth observation pipelines: an internal-consistency audit of zonal statistics for local climate zone–based land surface temperature analysis

Islam M. Hamdi, Youssef M. Youssef, Dominika Dąbrowska, Omar I. Aboulola, Mona Eisa, Amnah Aldohan
article en

Abstract

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.

Frontiers in Environmental ScienceVol. 14
Princess Nourah bint Abdulrahman University (SA), Beni-Suef University (EG), Suez University (EG), University of Jeddah (SA), University of Silesia in Katowice (PL)
Princess Nourah Bint Abdulrahman University
Urban Heat Island Mitigation
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon