Treat noise and missing values differently

Repeated readings can differ because of instrument noise, changing observation conditions or real physical variation. Calibration and uncertainty analysis help separate these possibilities. Averaging can reduce some random variation under suitable assumptions, but it does not automatically remove a shared bias or a processing mistake.

An unusually bright pixel — open definition might be meaningful, saturated or affected by an artifact. A useful investigation consults quality information and neighboring or repeated measurements. Deleting every inconvenient value is not a valid substitute for a documented quality rule.

Science products often use a fill value — open definition or mask to mark locations without usable data. USGS quality-assessment bands encode conditions relevant to Landsat interpretation, including fill and saturation flags. Those flags must be decoded according to the product documentation rather than treated as ordinary measured quantities.

For a fictional series of 10, 12 and one missing value, the mean of the two valid readings is 11. Replacing the missing value with zero gives about 7.33, answering a different and unjustified calculation. Report both the valid mean and the fact that only two of three intended readings are available.

Missingness can itself be patterned. If clouds obscure a surface preferentially during particular conditions, clear-sky measurements may not represent the full period in the same way as a complete record. More arithmetic cannot replace observations of conditions the sample systematically omits.

Keep a record of exclusions, their reasons and their effect on coverage. A comparison between two dates should not silently compare different subsets of a region. You may need a common valid area or an explicitly modeled treatment of missing observations, with the remaining limits stated alongside the result.