Can several satellites contribute?

Yes, provided differences are addressed appropriately.

Is an image time series necessarily an AI model?

No. It is data that statistical or AI methods may analyze.

AI Glossary / Statistics and probability / Advanced

Satellite image time series

Definition

A satellite image time series is a sequence of satellite observations of an area, ordered by acquisition time and prepared so that measurements can be compared across dates. It can retain complete images or extract measurements for pixels, fields, buildings or other regions. Its purpose is to study how conditions evolve, rather than describe an area from one isolated image.

What it means

A single image can show a field that appears less green than its neighbor. A time series can show whether that difference is persistent, seasonal or a recent departure from its previous pattern. The extra information comes from repeated observations and their timing.

The observations are not automatically comparable just because they cover the same coordinates. Sensor characteristics, atmospheric conditions, spatial alignment and processing choices can affect the recorded values. A useful series therefore includes information about both the observations and their quality.

How it works

Define the area, time period and measurement of interest. Select suitable imagery, apply the product's scaling and quality rules, and align observations to a consistent spatial reference. For optical imagery, screen unsuitable cloud, shadow and other contaminated observations before interpreting surface changes.

Surface reflectance products account for atmospheric effects to improve comparison. An index such as NDVI combines near-infrared and red reflectance into a vegetation-related measurement. The resulting sequence may still have irregular gaps.

Analysis can describe trends, seasonal cycles or abrupt changes. AI models can use temporal patterns for classification or prediction, but their inputs must preserve the distinction between an observed measurement, a missing value and an estimate inserted to fill a gap.

Three practical examples

1. Crop monitoring: compare a field's vegetation-index trajectory with its previous growing seasons. A departure can guide inspection without establishing its cause from imagery alone.

2. Construction monitoring: examine repeated observations of a development area to identify changes in surface cover. Confirm apparent changes against spatial resolution, shadows and independent project records.

3. Habitat review: track seasonal water or vegetation patterns across a defined region. Keep observation availability alongside the measurements so that a cloudy period is not mistaken for an absence of change.

Think of it this way

Think of a garden photographed repeatedly from the same position. The sequence reveals changes that one photograph cannot. However, a picture taken through fog and a missing photograph are different from clear evidence that the garden has deteriorated.

Worked example

Consider five fictional observation dates for one consistently located pixel: days 1, 11, 21, 31 and 41. These dates are chosen for arithmetic and do not describe a particular satellite's revisit schedule.

Suppose usable surface-reflectance measurements are available on days 1, 11, 31 and 41. Their near-infrared values are 0.30, 0.40, 0.50 and 0.30, respectively. Red reflectance is 0.10 on each usable date. Day 21 is cloud-contaminated and excluded.

Calculate NDVI as (near-infrared - red)/(near-infrared + red). The four values are 0.20/0.40 = 0.50; 0.30/0.50 = 0.60; 0.40/0.60 = 2/3; and 0.20/0.40 = 0.50.

Their arithmetic mean is (0.50 + 0.60 + 2/3 + 0.50)/4 = 17/30, approximately 0.5667. This is the mean of the four available measurements. It is not automatically the time-weighted average over the whole forty-day interval.

A faulty pipeline replaces the missing day-21 value with zero and divides by five. It reports 34/75, approximately 0.4533. The reduction is an artifact of the replacement rule: no valid zero measurement was observed. There are four usable observations out of five scheduled dates, giving 80-percent availability in this example.

If linear interpolation is explicitly chosen, day 21 lies halfway between days 11 and 31. Its estimated NDVI would be (0.60 + 2/3)/2 = 19/30, approximately 0.6333. The five-value mean would then be 0.58. This is another derived summary, dependent on an interpolation assumption. It does not reveal what the satellite would actually have measured through a clear atmosphere on day 21.

Finally, the usable sequence begins and ends at 0.50. That equality does not mean nothing changed in between: the intermediate measurements are higher. Keeping the dates and the full trajectory prevents an endpoint-only summary from discarding that information.

Illustrative case study

Fictional Willow Fields builds a monitoring report using the worked example. Its first dashboard draws a sharp decline to zero on day 21 and flags possible crop damage.

The analyst checks the quality information and discovers that the point represents missing data. Willow removes the false observation, displays a gap and records four usable dates. If an interpolated line is shown, its legend identifies it as an estimate.

The team compares the observed pattern with the appropriate season and arranges an independent field check before attributing changes to a particular cause. This invented example demonstrates data handling; it reports no real agricultural outcome or measured AI benefit.

Common misconceptions and mistakes

"Every scheduled revisit gives a usable observation." Conditions can make an acquired image unsuitable for the intended analysis.

"A missing pixel means a zero index." Missingness describes unavailable information, not a measured value.

"A smooth reconstructed curve is entirely observed." Smoothing and gap filling introduce modeled values.

"Higher NDVI directly measures higher yield." The index is a spectral measurement; translating it into an outcome requires additional evidence and an appropriate model.

Business use

Specify what decision the series should inform and which spatial unit matters. A pixel crossing a field boundary can mix surfaces that have different management histories. Keep the chosen boundary and aggregation rule consistent or record their changes.

For operational AI systems, retain acquisition times, quality flags and processing versions. Evaluate a forecast using only information available by its decision date. An interpolation using a later observation may be useful for retrospective analysis but would supply unavailable information to an earlier forecast.

Check your understanding

Why exclude day 21?

Answer: Its measurement was cloud-contaminated.

Why is zero filling misleading?

Answer: It invents an observation.

Is 0.6333 measured on day 21?

Answer: No; it is the stipulated linear interpolation.

Related terms

Missing data
Remote sensing
surface reflectance
NDVI
cloud masking
temporal resolution
change detection

Learn next

Plot the five dates with a visible gap, then add the interpolated value with a different symbol. Compare what each display actually supports.

Sources

U.S. Geological Survey, Landsat Collection 2 Surface Reflectance: https://www.usgs.gov/landsat-missions/landsat-collection-2-surface-reflectance

U.S. Geological Survey, Landsat Normalized Difference Vegetation Index: https://www.usgs.gov/landsat-missions/landsat-normalized-difference-vegetation-index

Earth Engine developer-community author ghidora77, Time Series Modeling, community tutorial: https://developers.google.com/earth-engine/tutorials/community/time-series-modeling

Previous: SAT Next: Satisfiability