Lesson 6 of 6 · Earth observation
Audit a satellite time series
9 minute lesson · 3 checks along the way
Your goal for this lessonSeparate physical change from changing samples, calibration, and processing.
Words with a dotted underline open a short definition right here in the lesson.
A line on a graph hides many choices
A satellite repeatedly estimates a quantity for a defined place. The graph may look simple, but every point depends on which observations were available, which were retained, and how they were combined. A monthly might select a median, an average, or the clearest available observation. Those methods answer different questions and need not represent one simultaneous snapshot of the region.
Begin by fixing the geographic boundary, physical variable, units, and time convention. Then inspect missing observations and quality flags. Clouds can preferentially remove particular seasons or weather conditions, so remaining observations may not represent the period evenly. Filling gaps with zeros changes the statistic; interpolating creates estimates rather than new measurements. Both choices must be visible rather than silently hidden in a smooth curve.
Make sensors comparable without declaring them identical
Different instruments can have slightly different wavelength responses, viewing angles, grids, and . A jump near a sensor change may therefore be partly instrumental. NASA’s Landsat Sentinel-2 products address several differences through atmospheric correction, cloud and shadow masking, angular normalization, bandpass adjustment, and common gridding. Harmonization makes comparisons more consistent; it does not erase all uncertainty or manufacture missing observations.
Overlap periods, stable reference sites, and independent measurements help check continuity. A consistent processing version also matters: reprocessing can revise older values without the past landscape changing. Keep product identifiers and versions with the analysis. If two records disagree, investigate which quantity, spatial support, timing, and corrections each represents before deciding that one sensor must be defective.
Check 1 of 3
10 pointsUse the reading above. You can try again without losing points.
Work through a decision
Expose a seasonal sampling illusion
Imagine an unchanged landscape whose toy index is always 0.30 during spring and 0.70 during summer. Year One retains one spring observation and one summer observation, giving an unweighted mean of 0.50. Year Two retains one spring observation and three summer observations. Its four-value mean is (0.30+0.70+0.70+0.70)/4=0.60, although neither seasonal value changed. Unequal sampling alone raised the reported observation-weighted mean.
To compare the same two-season quantity, first calculate each season’s mean, then give spring and summer equal weight in both years. Both results are 0.50. Equal seasonal weighting is appropriate for this explicitly defined comparison, not a universal recipe for every climate variable. Real incomplete seasons may remain uncertain. The exercise shows why the sampling design and intended statistic must be stated before interpreting a trend.
Sampling changes the reported mean
Original hypothetical sampling comparison. Dots represent retained measurements, not a continuous trajectory. Unequal sample counts reweight the seasons without changing either seasonal index.
Open full-size diagram in a new tab: Sampling changes the reported meanCheck 2 of 3
10 pointsUse the reading above. You can try again without losing points.
Test a change against alternative explanations
Suppose another invented record rises by 0.03 when a new sensor begins. Overlapping observations of comparable targets establish that the new sensor reads 0.04 higher on the old scale in this toy case. Subtracting that known changes the apparent step to −0.01. In a real comparison, in the correction also matters. An apparent jump cannot be assigned entirely to Earth or entirely to calibration without testing both.
A reproducible time-series claim records the retained dates, coverage, masks, calibration, aggregation, and uncertainty. Compare matched seasons, test alternative reasonable filtering choices, and seek independent evidence for the proposed physical explanation. Report a persistent signal at the scale actually supported, while keeping unresolved gaps visible. The strongest story is the one another reader can recalculate from the same documented observations.
Check 3 of 3
10 pointsUse the reading above. You can try again without losing points.
Put it together
Your lesson progress
Each check connects to a part of the story. Complete all 3 to earn your 25-point lesson bonus.
Go to the source
Explore the evidence behind this lesson. These optional readings and videos open in a new tab.