The Gold Standard of Zonal Statistics

Zonal statistics summarises the values of a raster within regions called zones. A zone may represent a watershed, school catchment, district, habitat polygon or integer raster category. A common output is the mean, sum, count, minimum, maximum or variability of values associated with each region.

The ArcGIS Pro reference makes an important distinction: a zone defines which locations are grouped, while a value raster supplies the quantities to summarise. The output inherits the meaning and limitations of both inputs.

Contents · Next

A polygon boundary describes an area, but the raster is a grid of finite-sized cells. A zone may overlap cells partially, completely or not at all. Which cells enter an analysis depends on the software, selected method, spatial reference, grid size and alignment.

Do not assume a software tool performs perfect continuous-area integration by default. In ArcGIS Pro, vector zones can be rasterised internally using cell centres; small polygons may disappear from the output when no cell centre falls inside. Name the inclusion rule.

Contents · Previous · Next

A raster zone is commonly formed by cells sharing an integer category value. Areas need not be contiguous; two distant clusters carrying the same zone identifier may be summarised together. That can be desirable for a land-cover category but wrong if the intention was one statistic per separate park.

Choose the zone field or identifier so that it matches the actual reporting unit. A school ID used across several campuses and an individual campus ID are different entities. Summation across all matching cells will follow the IDs, not the analyst’s unspoken expectations.

Contents · Previous · Next

Raster resolution determines how much spatial detail is present. A 1-km cell and a 10-metre cell cover very different ground areas. Summarising a narrow river corridor with kilometre cells can fail to capture the corridor’s true footprint.

Resampling to a smaller cell size creates more grid cells but does not automatically create new underlying information. Esri’s methods explanation warns that finer internal grids can improve representation of small zones without increasing the source raster’s intrinsic detail.

Contents · Previous · Next

When a vector polygon is converted to a raster zone using cell centres, a cell is assigned according to whether its centre falls inside the polygon. A very narrow polygon can touch multiple cells yet include none of their centres. The output may contain no record for such a region.

An empty result therefore is not proof that the geographic zone has no environmental values. It may indicate that rasterisation excluded it. Check footprint, cell centres and any rasterisation or weighting options before concluding an area was empty.

Contents · Previous · Next

Two raster layers can have the same cell size yet different grid origins. Their cell boundaries may not line up, so overlay requires resampling or regridding. Coordinate-system differences can complicate the calculation further, especially if projected and geographic units are mixed.

Document grid origin, snap raster, coordinate reference, cell size and transformation choices. ArcGIS Pro’s current tool reference explains that a mismatch can trigger internal resampling. If you rerun the analysis with different environment settings, results may change without any real-world environmental change.

Contents · Previous · Next

A raster value of zero is a valid measurement when the phenomenon permits zero. NoData means the cell has no valid supplied value under the dataset’s rules. Mixing the two can systematically lower averages or invent false absence.

ArcGIS Pro allows NoData cells to be ignored within a zone or to make the zone’s summary NoData. The chosen setting affects the denominator and should be reported with the data completeness of each zone. A blank sensor reading is not an observed rainfall of zero millimetres.

Contents · Previous · Next

A mean temperature raster measured in degrees Celsius yields a mean in the same unit. A sum over temperature pixels has units of degree-Celsius cells, which usually does not represent a meaningful physical amount. By contrast, an appropriate rainfall-depth surface might be used with area and units to estimate a volume under stated assumptions.

The mathematical operation must match the physical quantity. Do not call a raw sum of cell values a regional energy or mass without multiplying by valid cell areas and other required conversion factors.

Contents · Previous · Next

For n valid cells with equal contribution, the arithmetic zone mean is their sum divided by n. Suppose a fictional zone includes temperatures 25, 27, 28 and 30°C. The mean is 110/4=27.5°C. NoData should not silently add a fifth zero to the denominator.

A mean describes the included raster cells, not necessarily every actual point within a polygon. When cells cross the zone boundary or have unequal represented areas, the interpretation may require an explicitly area-weighted alternative.

Contents · Previous · Next

Count reports how many valid raster cells entered an analysis. Sum adds values. Variety counts unique raster classes. Majority reports the most frequent class and may have software-specific tie rules. These summaries are not interchangeable.

A categorical land-cover map is generally ill suited to a mean of category codes if numbers are labels rather than quantities. Majority or class percentages may be more meaningful. Keep class definitions, NoData and cell-inclusion rules visible.

Contents · Previous · Next

An area-weighted mean uses weights proportional to each included cell’s contribution to the target polygon or to another stated weighting measure. It differs from a simple arithmetic mean when cells have unequal areas or fractional overlap.

A GIS tool’s reported mean should not be labelled ‘area weighted by polygon intersection’ unless that is the implemented method. The chosen rasterisation and weighting determine which area the statistic actually represents. Recalculate a small boundary case by hand to validate the claim.

Contents · Previous · Next

On a regular geographic latitude-longitude grid, cell area varies with latitude. A degree of longitude spans less distance near the poles than near the equator. Taking an unweighted global mean across all latitude-longitude pixels may overrepresent higher latitudes.

For a large-area physical average, use an appropriate area-weighting method or equal-area analysis where the quantity permits it. Projected coordinates alone do not solve all precision questions: projection choice, grid alignment and geometric meaning still require attention.

Contents · Previous · Next

Elevation, temperature and reflectance typically have continuous numeric interpretations, while habitat-class numbers are labels. The arithmetic mean of values 1, 2 and 3 may be meaningful for a continuous field but arbitrary for categories called 1=forest, 2=water and 3=urban.

Check the raster type and metadata first. ArcGIS restricts some statistics, such as majority and variety, to integer-valued rasters; floating rasters support a different set. A tool offering a calculation is not proof it answers a scientifically useful question.

Contents · Previous · Next

A zone’s minimum and maximum describe extremes among included cells. They can highlight potential hazards or errors but are sensitive to isolated bad values, resampling and data coverage. A maximum satellite temperature is not automatically the hottest physical point in a real region.

Retain information about pixel quality, confidence and observation time. Reporting just an extreme without verifying its origin invites a misleading local narrative. Use percentiles, distributions and spatial review when extremes drive a decision.

Contents · Previous · Next

Median summarises the middle of an ordered set and may better describe the typical included value under strong skew. It is not equivalent to an area-weighted median unless appropriate weights are used. The mean and median can both be informative when reported together.

A district with one very high pollution pixel and many low pixels can have a mean above the median. Which statistic matters depends on whether the decision concerns typical exposure, total burden or the highest-risk locations.

Contents · Previous · Next

A zone’s standard deviation describes spread among included raster values under a specified definition. High spread may reflect genuine heterogeneity, measurement noise, mixed land covers or inconsistent data acquisition. It does not directly describe the uncertainty of the regional mean.

A larger number of spatial pixels does not guarantee thousands of independent measurements because nearby cells may be highly autocorrelated. Sampling error and spatial variation should not be collapsed into the same standard-deviation figure.

Contents · Previous · Next

Directions such as compass headings wrap around 360°. A linear mean of 359° and 1° gives 180°, which points in the opposite direction from the two near-north directions. Circular statistics handle such wraparound appropriately when the data and analysis permit it.

ArcGIS Pro offers circular mean and circular standard deviation options with a wrap value. Name the variable’s cyclic nature and chosen method rather than averaging directions as ordinary linear numbers.

Contents · Previous · Next

A satellite image may contain several spectral bands representing different wavelengths. A zonal mean of one band cannot be interpreted as a mean of all bands. Derived indices such as NDVI have their own formulas, scaling and valid-value ranges.

Select the correct band or derived raster before summarising. The QGIS processing reference documents its raster-band parameter and available zonal statistics. Coordinate comparison must use compatible acquisition dates and preprocessing.

Contents · Previous · Next

A region’s zonal statistic can change because environmental conditions changed, boundaries changed, the sensor changed or data completeness shifted. A before-and-after map needs compatible dates, grid, classification and inclusion rules.

An annual mean computed from 365 daily rasters is not the same as a mean of one composite image if composites omit clouds or use a different weighting. Explicit temporal aggregations prevent calendar and sampling ambiguities.

Contents · Previous · Next

A zonal statistic describes a region’s raster values under the analysis. It cannot automatically describe every person, animal or property inside that region. A school catchment’s average temperature does not equal each student’s exposure.

If the question concerns individuals, need appropriate individual data, within-zone variability and causal or sampling design. Aggregate associations may reflect different processes from individual-level associations.

Contents · Previous · Next

Suppose a fictional zone’s raster cells contain 10, 20, 30 and NoData. Ignoring NoData yields a count of three and arithmetic mean (10+20+30)/3=20 units. Treating NoData as zero would instead give 15, a different and generally unjustified result.

The output should include the number and proportion of valid cells. If a large fraction of a zone is missing, the mean of valid cells may not be representative even though its arithmetic is correct.

Contents · Previous · Next

Consider a hypothetical polygon intersecting three valid value cells with values 2, 10 and 20, covering fractions 1.0, 0.5 and 0.25 of equal-area cells. The fractional-area-weighted mean is (2×1+10×0.5+20×0.25)/(1+0.5+0.25)=12/1.75≈6.86.

An unweighted mean of those three values is 10.67, showing that inclusion weighting matters. The fractional calculation is an explicit alternative method, not automatically the output of every software zonal-statistics command.

Contents · Previous · Next

A fictional zone contains directions 359° and 1°. The ordinary arithmetic mean is 180°, misleadingly southward. Circular averaging uses unit vectors: the sine components cancel and cosine components point nearly north, so the circular mean is 0°.

This is a methodological example for angular observations, not a demonstration that two wind vectors have equal speed or that their vector-average velocity is northward. Direction and magnitude may require distinct treatment.

Contents · Previous · Next

The gold standard of zonal statistics ties a geographic zone, original value raster, coordinate reference, alignment, NoData rules, weights, chosen statistic and validation to the question being answered. Without that chain, a colourful region map can hide arbitrary data-processing choices.

Follow Geospatial Analysis, Cartography, Remote Sensing, Statistics and Geography. These are applications of Learning Transfers into Life.

Contents · Previous · Sources