Sampling agricultural imagery with zonal statistics

Start with the operational question

Define the polygon, pixel inclusion rule, nodata handling, and summary statistics before comparing fields. Inspect distributions and valid coverage rather than relying on a mean alone.

Separate observation from explanation

Averages can hide a small damaged patch or be dominated by mixed edge pixels. Different valid-pixel counts can make apparently comparable block summaries represent different amounts of crop.

Design reference observations that represent the field

State the population that the sample should represent and the question it must answer. Select locations across relevant crop stages, terrain, management, and image conditions. Include ordinary areas and difficult cases, not only conspicuous anomalies or convenient roadside stops. Record selection rules so that later users understand where the reference is representative.

Define the observation unit, area or depth, timing, instrument or laboratory method, and required replicates using the appropriate discipline's protocol. A point observation and a satellite pixel have different spatial support. Record enough context to evaluate that mismatch. Keep inaccessible sites in the sampling log and explain any replacement rather than silently dropping them.

Separate reference collection from the result being evaluated when feasible. Review observer consistency and missing data before calculating agreement or error. Retain raw observations, corrections, and exclusions with their reasons. More samples can improve precision, but a large biased sample still fails to represent excluded conditions. Use the results within the population and period that the sampling design supports.

Match spatial detail to the question

Agricultural GIS connects field boundaries, observations and imagery, but each dataset has its own spatial support. A point marks a location, a polygon describes an area and a raster pixel summarises a footprint. They should not be treated as equivalent measurements. Before overlaying layers, check their coordinate reference systems, dates and units. A visible alignment problem can arise from a coordinate or georeferencing issue rather than movement or change in the crop itself.

Sentinel-2's bands have different native spatial resolutions. Its visible and broad near-infrared bands are available at finer resolution than its red-edge and shortwave-infrared bands. Resampling can put those bands onto a common grid, but it does not create new observed detail. When a field is narrow or small, boundary pixels may mix crop with roads, trees, water or neighbouring land. Explain that mixture before interpreting a field average as a pure crop measurement.

Preserve quality and time information through the analysis

Optical imagery requires attention to clouds, haze and shadows. A scene with a small overall cloud percentage may still have the target field obscured. Inspect the actual area of interest and retain the masking or quality method used. Compare compatible products and document processing choices, especially when building a time series from different dates or sensors. An unexplained jump may be a data-quality issue rather than a sudden change in vegetation.

Vegetation indices describe combinations of reflected spectral signals. USGS explains that NDVI is sensitive to soil background and can saturate in dense vegetation. Therefore an index is best interpreted with crop stage, canopy structure and field evidence. A colour class is not a direct measurement of nutrient concentration, disease identity or harvestable yield. Those interpretations need suitable calibration and validation, and any prediction should retain its uncertainty and the conditions under which it was evaluated.

Make maps reproducible and useful for the next observer

Keep the original inputs, analysis date, method version and relevant parameters with the output. Use a clear legend that distinguishes observed values, derived indicators and unverified candidates. Where a statistic excludes cloudy pixels or non-crop areas, document the exclusion and the valid coverage. If the coverage changes between dates, acknowledge that the averages may represent different parts of the field. This is especially important when a report turns a sequence of maps into a trend.

Design the validation around the intended use. A model for one crop and season should be tested on independent examples that represent the conditions where it will operate. Field checks should look for missed objects as well as incorrect proposed objects, and classification checks should include each meaningful category. Retain the disagreements and review their causes rather than removing difficult cases. A map becomes operationally valuable when another person can follow its evidence, understand its limits and decide what observation is needed next.

Check the support behind every mapped value

Inspect the source over the actual field of interest. Scene-wide metadata can conceal a cloud, shadow, blurred strip, incomplete survey, or alignment problem exactly where the decision is needed. Record the date, spatial support, quality flags, and valid coverage for each field. An observation with little usable support should remain unavailable rather than being filled with an apparently plausible value.

Review missing values and exclusions separately from valid measurements. A zero, an unvisited location, a masked pixel, and a sensor error describe different circumstances. Preserve the reason for missing data through the summary stage. If an estimated value is used, mark it as estimated and explain the method so that it cannot be mistaken for an independent observation.

Choose acceptance rules before comparing outcomes. Apply them consistently to favourable and unfavourable results. Retain a small collection of difficult examples and revisit it after processing changes. Quality review should include omissions and inaccessible areas as well as the records that were successfully produced. A transparent incomplete dataset is more useful than an apparently complete map whose unsupported values cannot be identified.

Keep identifiers, history, and evidence connected

Use a stable identifier for each field or asset and store its descriptive name separately. The identifier should remain useful when a crop changes, staff move, or an operational unit is subdivided. Define relationships explicitly: one field may have many visits, one visit may have several photographs, and one harvest lot may refer to several fields. Document those relationships before combining spreadsheets or map layers.

Changes need an effective date and a reason. Preserve earlier geometry or attribute versions when a correction affects historical reports. Record who approved the change and which source supported it. Correcting a boundary should not silently rewrite the area that was used for a previous season unless the report is intentionally revised and the revision is explained.

Test an export by bringing representative records back into the receiving workflow. Check that identifiers, dates, units, text encoding, coordinate interpretation, and missing values survive. Keep an unaltered source copy and a transformation note. A simple data dictionary and a named owner for each dataset often prevent more errors than an elaborate visual interface without clear responsibilities.

Define the deliverable and approval

Deliver a zonal-statistics table that includes valid area, observation date, distribution summaries, exclusions, and notes on fields that fail the minimum-support rule.

A practical review before sharing the result

Check the observation against its source and make sure the field identifier is correct. Review dates, units, missing values, quality exclusions, and geometry versions. Ask a colleague to reproduce one representative result from the recorded inputs. This small independent check often exposes a misunderstood field code or a processing assumption before it affects an entire report.

Write down the alternative explanations for the mapped pattern and identify what evidence would distinguish them. Keep a measured observation separate from the interpretation and from any treatment or management instruction. A decision should name its responsible reviewer and account for local operating procedures. If the evidence is insufficient, the appropriate output is a targeted inspection rather than a confident claim.

Keep the original result, any reviewed version, and a short change log. State which locations and periods the output can reasonably describe. When sharing a map, provide a clear legend, source dates, and an accessible explanation of uncertainty. The recipient should know whether a feature is a confirmed observation, a candidate awaiting inspection, or a model-based estimate.

Turn the investigation into a repeatable field cycle

Prepare the visit with a map that can be used at the intended location and a concise list of questions. Confirm access, permissions, equipment, and the availability of a person familiar with the field. Choose representative stops as well as unusual areas. Record where the planned route could not be followed, because missing observations can change how the result should be interpreted.

After the visit, review the observations while the team can still clarify them. Connect photographs, measurements, and comments to the same field and visit identifiers. Assign a follow-up status and preserve unresolved cases. A correction should explain what changed and why; it should not erase the original evidence or make an earlier interpretation appear to have been certain.

At the next suitable observation date, compare conditions using the same definitions and review any management or weather events that occurred between visits. Record the action actually completed and its operational outcome. This creates an evidence trail from an initial pattern to field confirmation and later review, making the mapping process easier to improve over successive seasons.