Find the question you need
Choose a route or begin at Chapter 1. Teaching chapters stay open; the chapter index expands by section.
Define the objects
Chapters 1–2. Define the intended ground population
Build the argument
Chapters 3–5. Build a comparison that reverses direction
Compare and interpret
Chapters 6–8. Create apparent change without surface change
Check and practise
Chapters 9–10. Distinguish spatial and temporal coverage
Questions and next routes
Chapters 11–12. Questions readers ask
Full chapter index · Worked practice · Remote Sensing subject root
A minimum-observation rule can remove unreliable pixels while changing the land area represented by a regional average. Two maps can use the same threshold and still summarise different populations of ground locations.
This guide follows four equal-area pixels through a coverage filter. It first compares two regions, then shows an apparent change over time in a surface that has not changed at all. All values and valid-observation counts are invented to isolate spatial selection.
Chapter index
Define the objects · Chapters 1–2
1. Define the intended ground population
2. Put the threshold into the regional formula
Build the argument · Chapters 3–5
3. Build a comparison that reverses direction
4. Report retained area alongside the estimate
5. Change the threshold and track the estimand
Compare and interpret · Chapters 6–8
6. Create apparent change without surface change
7. Use common support and state its limits
8. Bound a full-area mean with missing pixels
Check and practise · Chapters 9–10
9. Distinguish spatial and temporal coverage
10. Practise with unequal areas
Questions and next routes · Chapters 11–12
11. Questions readers ask
12. Sources and connected learning routes
Suppose a region contains four equal-area pixels and the target is the arithmetic mean of one specified surface variable across all four. Each pixel represents the same area, so the full-region mean is the sum of its four values divided by four.
A temporal processing step estimates a value for each pixel. A quality rule then retains only pixels with at least k valid observations during the specified window. The regional mean over retained pixels is a different statistic from the full-region target whenever any pixels are excluded.
For the initial example, assume each retained pixel estimate equals its true target value exactly. This removes per-pixel estimation error so that the effect of changing the spatial footprint can be inspected on its own.
CHAPTER 2 OF 12 · Define the objects
2. Put the threshold into the regional formula
Let xᵢ be the target value at pixel i, nᵢ its valid-observation count and aᵢ its ground area. Define Iᵢ = 1 when nᵢ ≥ k and zero otherwise. The retained-area mean is ΣaᵢIᵢxᵢ / ΣaᵢIᵢ, provided retained area is nonzero.
The full-area target is Σaᵢxᵢ / Σaᵢ. Its denominator includes the complete intended region. A pixel excluded by the threshold is missing from the first denominator; its value is not set to zero.
For equal-area cells, the area factors cancel. For unequal ground areas, merely counting cells can introduce a second weighting problem. State the grid and area convention alongside the coverage rule.
Region A has target values 0.1, 0.1, 0.7 and 0.7, with valid counts 2, 2, 8 and 8. Region B has four values of 0.5 and valid counts of eight at every pixel. Use k = 5.
Region A retains only its last two pixels, so its reported retained-area mean is 0.7. Region B retains all four and reports 0.5. On that filtered basis, A exceeds B by 0.2.
The full-region mean of A is (0.1 + 0.1 + 0.7 + 0.7)/4 = 0.4. B’s full-region mean is 0.5. Thus the intended comparison is A − B = −0.1, the opposite sign. All retained values are accurate; the footprint has changed the comparison.
| Region | Four target values | Valid counts | Passing pixels at k = 5 | Retained mean | Full mean |
|---|---|---|---|---|---|
| A | 0.1, 0.1, 0.7, 0.7 | 2, 2, 8, 8 | 3 and 4 | 0.7 | 0.4 |
| B | 0.5, 0.5, 0.5, 0.5 | 8, 8, 8, 8 | All four | 0.5 | 0.5 |
CHAPTER 4 OF 12 · Build the argument
4. Report retained area alongside the estimate
The retained area fraction is C = ΣaᵢIᵢ/Σaᵢ. In the example, C is 50% for A and 100% for B. Those coverage figures make the asymmetry visible, but do not themselves recover the missing values.
Calling both results regional means without the coverage qualifier invites readers to interpret them as summaries of the complete regions. A clearer label is mean over pixels meeting the stated observation threshold, followed by the retained area fraction.
Coverage percentage alone does not describe spatial placement. Half the region retained in a coastal strip can represent a different population from half retained as scattered cells. Keep the inclusion mask available and inspect whether exclusions align with land cover, terrain or other relevant features.
CHAPTER 5 OF 12 · Build the argument
5. Change the threshold and track the estimand
At k = 1 or k = 2, all four A pixels pass, giving its full-region mean of 0.4. At any integer threshold from three through eight, only its two high-value cells pass, giving 0.7. At k = 9, no A pixels pass and the retained-area mean is undefined.
The threshold does not reveal which answer is scientifically preferable. It controls eligibility under a measurement rule. The resulting population and evidence strength both need to match the intended question.
Do not treat the highest threshold as automatically the most accurate regional answer. It can improve a per-pixel coverage requirement while excluding more land. A sensitivity table should report the estimate, retained area and pattern of exclusions together.
CHAPTER 6 OF 12 · Compare and interpret
6. Create apparent change without surface change
Now examine the same four-pixel region at two windows. Its underlying values remain 0.1, 0.1, 0.7 and 0.7 in both windows. Window 1 has counts 2, 2, 8 and 8. Window 2 has eight observations at every pixel.
With k = 5, the retained-area mean falls from 0.7 to 0.4, an apparent decline of 0.3. There is no physical change in any pixel. The second window simply includes the two low-value cells that the first excluded.
The subtraction compares different spatial populations. It is numerically correct for those two retained-area summaries, but does not measure change in a fixed footprint. Before assigning a physical explanation to a regional trend, compare the inclusion masks.
| Window | Valid counts | Retained area | Retained mean | Common-support mean |
|---|---|---|---|---|
| 1 | 2, 2, 8, 8 | 50% | 0.7 | 0.7 |
| 2 | 8, 8, 8, 8 | 100% | 0.4 | 0.7 |
For the two-window example, restrict both means to pixels that pass in both windows. The common support consists of pixels 3 and 4. Their mean is 0.7 in each window, so their estimated change is zero.
This is a valid comparison for that shared half of the region. It is not an estimate of change over the excluded half. Common support fixes the comparison population by narrowing the question; it does not reconstruct unavailable surface values.
Always report common-support area as well as the change. A stable mean over a small surviving fraction can coexist with large unobserved changes elsewhere. A fixed mask chosen for a particular analysis also needs a transparent rationale and sensitivity checks.
CHAPTER 8 OF 12 · Compare and interpret
8. Bound a full-area mean with missing pixels
Suppose only the two retained A values, both 0.7, are known. If this teaching variable is constrained to [0,1], the unknown sum from the other two pixels lies between zero and two.
The full-area mean therefore lies between 1.4/4 = 0.35 and 3.4/4 = 0.85. Comparing it with B’s fully known mean of 0.5 gives a difference range from −0.15 to 0.35. Its sign cannot be determined from the retained observations alone.
These are worst-case bounds under a stated value range, not a confidence interval. Use the actual variable’s documented range. Many remotely sensed indices or physical quantities do not belong to [0,1], and an unsupported range would invalidate the bounds.
A valid-observation count is attached to each pixel and band. Eight observations can be clustered into one part of a season or spread throughout it. Equal counts do not guarantee equal temporal representation.
This article assumes the per-pixel target estimates are exact to isolate spatial selection. In practice, the cloud-mask and seasonal-mean guide shows how temporal selection can affect those estimates before the spatial threshold is even applied.
Document the product, band, time window, mask, counting unit, threshold, grid, area weights and regional boundary. Check whether overlapping scenes or repeated acquisitions contribute multiple observations on one date. A platform’s image count need not equal a count of distinct well-spaced dates.
A region has three cells with areas 1, 1 and 2 square kilometres. Their target values are 0.2, 0.4 and 0.8. Their valid counts are 6, 2 and 6. A threshold of five retains the first and third cells.
The retained area is three of four square kilometres, or 75%. Its area-weighted mean is (1×0.2 + 2×0.8)/3 = 0.6. The unweighted average of the two retained cell values would be 0.5, answering a cell-count question instead.
The full-area mean is (1×0.2 + 1×0.4 + 2×0.8)/4 = 0.55. Explain both differences: filtering changes the population, while area weighting changes the contribution of cells within a chosen population. One correction cannot substitute for the other.
Does using the same threshold make two regions comparable? It makes the eligibility rule consistent. It does not guarantee equal coverage, matching land populations or representative temporal samples.
Should excluded pixels be zero-filled? Only when zero is justified as their actual target value, not merely because the measurement is missing. Otherwise it inserts invented observations.
Does common support solve every comparison problem? No. It fixes one spatial population while leaving temporal sampling, calibration, registration and other measurement issues to be checked.
Can I publish a filtered regional mean? Yes, with a clear definition of the retained population, coverage information and limits. The claim should match what the data support rather than silently expanding to the whole region.
CHAPTER 12 OF 12 · Questions and next routes
12. Sources and connected learning routes
Google Earth Engine documents ImageCollection.count as counting valid masked observations for each pixel across matching bands. Its regional-reduction documentation explains how pixels and weights enter spatial statistics. The four-pixel reversal, time comparison and bounds above are original arithmetic examples.
Follow the cloud-mask guide for the seasonal population behind each pixel’s mean, the co-registration guide for false changes at boundaries, and the index-aggregation guide for nonlinear averaging. These mechanisms can act together but require distinct checks.
For transfer, ask whether a school or survey average changed because the population changed or because the same people changed. The remote-sensing version requires the same discipline: identify the units retained before interpreting the average. Return to Remote Sensing for the complete measurement route.
References checked 22 October 2026
Earth Engine: ImageCollection.count
Earth Engine: statistics of an image region
Continue through Remote Sensing
Cloud masks and the seasonal population behind a clear-sky mean
Co-registration offsets and false change at a boundary
Why averaging NDVI differs from averaging reflectance
Remote Sensing root and full subject directory · The Gold Standard of Education · How X Works