A computational audit of spatial aggregation workflows has revealed that software routines extracting zonal summaries from satellite data can generate silent discrepancies in district areas even when reported temperature moments match to eight decimal places. Researchers Mona Eisa, Youssef M. Youssef, Omar I. Aboulola, Amnah Aldohan, Dominika Dąbrowska, and Islam M. Hamdi evaluated the interface where continuous satellite observations are aggregated across thematic boundaries produced by machine-learning models. Evaluating land surface temperature retrievals stratified by Local Climate Zone classes across metropolitan Riyadh, Saudi Arabia, the team discovered that standard commercial and open-source geospatial tools produced complete, plausible outputs while diverging on the total cell counts assigned to over half of the evaluated zones.
The study examined the stage of Earth observation pipelines where raster data intersect thematic zones. Local Climate Zone classifications partition urban terrain into distinct morphological and thermal categories, ranging from dense built fabrics to unpaved natural surfaces. While machine-learning classifiers routinely generate these categorical maps from satellite archives, quantitative environmental assessments require summary operators to aggregate continuous measurements such as land surface temperatures within each zone. Urban heat studies and local mitigation assessments routinely rely on these zonal statistics to calculate class-level thermal contrasts, rank surface types by warming intensity, and determine the spatial footprint of heat exposure across metropolitan populations.
To audit computational reliability at this interface, the investigators established a controlled comparative framework applying two separate processing engines to identical data layers. The team utilized the ArcGIS Pro Zonal Statistics as Table tool alongside an open-source Python processing chain constructed with GeoPandas, rasterstats, rasterio, and NumPy. Both systems processed a twelve-class Local Climate Zone layer aligned to a thirty-meter grid alongside a Landsat 8 and Landsat 9 mean-composite land surface temperature raster spanning the winter compositing window from December 1, 2024, to February 28, 2025. Both workflows enforced identical analytical parameters, using class-dissolved polygon boundaries, a cell-center inclusion criterion where cells are assigned when their centers fall inside the boundary, and explicit filtering of missing data.
The test geography in Riyadh provided a rigorous experimental environment because arid metropolitan thermal patterns defy conventional temperate assumptions. Urban climatology studies in hot-desert regions have consistently demonstrated a pronounced daytime surface urban cool island during warm seasons, wherein built structures exhibit lower surface temperatures than surrounding desert sands, alternating with nocturnal heat islands. During winter periods, surface thermal contrasts moderate, compressing class temperatures into a narrow twenty-six-degree span. In this compressed thermal regime, modest calculation shifts can alter relative ranks among neighborhood classifications, preventing domain experts from spotting corrupted estimates through visual inspection alone.
The researchers audited the resulting outputs against three automated validation rules computable entirely from the exported summary tables without accessing the underlying imagery or running recomputations: table completeness, quantile monotonicity, and moment-count consistency. Completeness checking verified that every submitted geographic class returned all requested summary fields. Quantile monotonicity verified that estimated order statistics obeyed necessary ascending sequences, confirming that fifth, tenth, twenty-fifth, median, seventy-fifth, and ninetieth percentiles increased monotonically. The moment-count consistency test examined the mathematical identity connecting the sample mean, cell count, and total sum, assessing whether the reported mean multiplied by the cell count reproduced the reported sum within acceptable rounding limits.
Initial completeness and quantile checks gave no indication of analytical failure in either workflow. Both implementations returned complete, error-free records across all twelve climate zone categories, generating no runtime warnings, null cells, or omitted classes. Furthermore, the reported percentile ladders satisfied monotonicity across every category in both tables. When comparing thermal metrics directly, both platforms agreed across all twelve classes to within 5 × 10−9 degrees Celsius for minimum, maximum, mean, and median values, and within 4.5 × 10−10 degrees Celsius for standard deviations. Both platforms yielded identical winter thermal rankings, producing Spearman and Kendall correlation coefficients of 1.000 across the twelve classes.
Despite this near-perfect agreement on thermal metrics, the moment-count consistency audit exposed a fundamental structural breakdown in the open-source pipeline. While the ArcGIS Pro output satisfied the arithmetic identity exactly across all classes, the Python workflow exhibited non-zero residuals across seven of the twelve zones, accounting for 58.3 percent of the evaluated classes. In the affected classes, the product of the reported mean and reported cell count departed from the reported sum by between -0.50 percent and +0.75 percent. Inverting the identity by dividing the Python workflow’s reported sum by its reported mean exactly recovered the cell counts reported by ArcGIS Pro down to the individual cell in every single case.
This mathematical inversion demonstrated that the open-source workflow had evaluated its statistical moments and its cell counts across two non-identical sample populations. While the statistical moments had been calculated over the exact cell set recognized by the commercial tool, the reported cell count field had been populated from a discordant extraction pass. Spatial inspection revealed that the seven divergent classes were precisely those that adjoined the outer NoData boundary of the masked satellite scene, whereas the five interior classes exhibited zero residual and perfect cell-count concordance. The discrepancy arose from differing extraction sequences between the zone mask and the missing-data mask, allowing boundary cells holding NoData values to be recorded in count fields before being excluded from statistical moment summations.
Because the anomaly altered cell tallies without disturbing thermal moments, downstream effects propagated exclusively into spatial extent calculations. Across the entire metropolitan scene, the total analyzed footprint diverged by 8,463 cells, representing a 7.62-square-kilometer difference between the 1,655.78 square kilometers reported by the commercial tool and the 1,663.40 square kilometers recorded by the Python script. Class extents shifted by substantial margins, with large lowrise commercial fabrics displaced by 4.38 square kilometers and compact lowrise residential zones displaced by 2.34 square kilometers.
A third area metric compounded the ambiguity of spatial reporting. Summing the geometric vector attributes stored directly within the original boundary shapefile yielded a metropolitan area of 1,506.47 square kilometers, differing by 9.0 percent from the 1,655.78 square kilometers derived from the raster cell counts. For the compact lowrise class alone, available area figures ranged from 449.75 square kilometers in vector geometry to 494.86 square kilometers in commercial raster extraction and 497.21 square kilometers in open-source tallies, illustrating how standard reprojection and gridding steps introduce substantial area shifts that go unflagged in final tables.
The verified temperature data confirmed critical structural characteristics of Riyadh’s winter urban climate. Mid-morning winter observations identified water bodies as the warmest surface class with a median temperature of 29.241 degrees Celsius, reflecting high seasonal thermal inertia, while compact midrise structures formed the warmest built category at 25.673 degrees Celsius. Open highrise and compact lowrise neighborhoods represented the coolest urban fabrics at 20.924 degrees Celsius and 21.143 degrees Celsius respectively, displaying temperatures 1.777 degrees Celsius and 1.559 degrees Celsius lower than the reference low-plant baseline. Intra-class standard deviations spanning 1.5 to 1.9 degrees Celsius demonstrated that local surface variation within individual zones matched the magnitude of differences between distinct fabric types.
The study authors highlighted several methodological boundaries surrounding their findings. The investigation examined two specific software versions operating on a single winter composite, meaning numerical error rates cannot be directly extrapolated to other software releases, seasons, or geographic regions. Additionally, while the internal moment-count audit successfully pinpointed the mathematical defect, the study did not establish synthetic ground-truth benchmarks or evaluate satellite statistics against in-situ ground weather stations or mobile thermal transects. The audit focused on aggregated class-level polygons, noting that localized patch-level extractions with high perimeter-to-area ratios would likely experience proportionally greater boundary count displacements.
To prevent undetected errors from propagating through automated Earth observation systems and cloud computing platforms, the researchers recommended that operational data pipelines embed moment-count validation alongside schema and monotonicity tests after every aggregation step. Verifying that sample sizes match reported moments provides a computationally lightweight safeguard that operates directly on output tables. As automated workflows increasingly process petabyte-scale satellite archives, systematic validation of intermediate aggregation steps ensures that scientific assessments and urban heat mitigation policies rest on internally consistent spatial foundations.
Subject of Research: Audit of zonal statistics consistency in urban climate Earth observation pipelines
Article Title: 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
Article References: Eisa, M., Youssef, Y. M., Aboulola, O. I., Aldohan, A., Dąbrowska, D., & Hamdi, I. M. (2026). 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. Frontiers in Environmental Science, 14, 1971523. https://doi.org/10.3389/fenvs.2026.1971523
Image Credits: AI Generated
DOI: 10.3389/fenvs.2026.1971523
Keywords: zonal statistics, Earth observation, local climate zones, land surface temperature, Riyadh, Landsat 8, Landsat 9, ArcGIS Pro, rasterstats, data validation, spatial analysis, urban climatology, remote sensing, reproducibility, GeoAI, pipeline auditing
Cite Scienmag News
Violet Maxwell. (November 16, 2026). Earth Observation Zonal Calculations Conceal Area Errors While Temperature Statistics Match. Scienmag. https://scienmag.com/earth-observation-zonal-calculations-conceal-area-errors-while-temperature-statistics-match/
Violet Maxwell. "Earth Observation Zonal Calculations Conceal Area Errors While Temperature Statistics Match." Scienmag, 16 November 2026, https://scienmag.com/earth-observation-zonal-calculations-conceal-area-errors-while-temperature-statistics-match/. Accessed 16 November 2026.
Violet Maxwell. "Earth Observation Zonal Calculations Conceal Area Errors While Temperature Statistics Match." Scienmag. November 16, 2026. https://scienmag.com/earth-observation-zonal-calculations-conceal-area-errors-while-temperature-statistics-match/








