How to Build and Maintain Reproducible Satellite Image Time‑Series: Workflow, Checklist & Sensor Comparison
satellite time seriesworkflow checklistremote sensingearth observation datareproducible research

How to Build and Maintain Reproducible Satellite Image Time‑Series: Workflow, Checklist & Sensor Comparison

CCaptains Space Editorial
2026-11-11
8 min read

Build a reproducible satellite image time series with sensor guidance, QA checkpoints, metadata logs, and a schedule for reprocessing.

A satellite image time series is only as reliable as its processing record. This guide presents a reproducible workflow for monitoring NDVI, land cover, glacier extent, burn severity, or another environmental indicator, with a sensor comparison, quality-control checkpoints, and a maintenance checklist for future data releases.

Overview

A time series combines observations of the same place and variable across multiple dates. In Earth observation, that sounds straightforward: download images, calculate an indicator, and compare the results. In practice, differences in sensor design, atmospheric correction, cloud masking, projection, spatial resolution, and processing collections can create changes that look environmental but are actually caused by the data pipeline.

A reproducible remote sensing time series makes each decision visible and repeatable. Someone should be able to identify the input scene or product, processing level, software environment, quality filters, formulas, resampling method, and summary statistics used to create every reported value. Reproducibility is also useful for your future self: when a data provider releases a reprocessed collection or a better quality mask, you can determine what changed and whether the historical record needs to be regenerated.

Before downloading anything, define the question in measurable terms. For example, “How did vegetation greenness change in this watershed?” is too broad for a stable workflow. A more useful specification might be “Calculate the median growing-season NDVI within the watershed boundary for each month, excluding pixels marked cloudy, snow-covered, saturated, or otherwise invalid.” The same principle applies to glacier extent, burned area, surface water, or land-cover transitions.

For background on bands, processing levels, and quality flags, see How to Read Satellite Data Metadata. The key idea is to treat metadata as part of the measurement, not as optional documentation.

What to track

1. Define the indicator and its measurement unit

Write down the indicator, formula, expected range, unit, spatial support, and temporal aggregation before processing the first image. NDVI, for instance, is commonly calculated from near-infrared and red reflectance, but the formula alone does not define the analysis. You must also state whether the result is a pixel-level raster, an area mean, a median, a percentile, or a classification derived from thresholds.

Record the study area in a versioned file rather than relying on a manually drawn map. Note the coordinate reference system, boundary source, geometry edits, and whether the analysis uses the full polygon or an inward buffer to reduce edge effects. For glacier or shoreline monitoring, document how mixed pixels, seasonal snow, shadows, and water are treated.

2. Choose a sensor for the question

Mission familyUseful roleImportant considerationsTypical processing choice
LandsatLong-running land-change records with moderate spatial detailCheck collection version, band definitions, scale factors, cloud and saturation flags, and gaps caused by clouds or acquisition conditions.Use a consistent surface-reflectance or equivalent analysis-ready product when the indicator requires comparable reflectance.
Sentinel-2Frequent optical monitoring at relatively fine spatial detailAccount for different band resolutions, cloud probability or scene classification, haze, shadows, and processor or collection changes.Use a consistently processed surface-reflectance product and document any harmonization or resampling.
MODIS or VIIRSFrequent, broad-area monitoring and ready-made vegetation or disturbance productsCoarser pixels can mix land covers; product compositing, quality flags, geolocation, and reprocessing history affect interpretation.Prefer a validated higher-level composite when its spatial and temporal characteristics fit the question.

This table is a starting point, not a ranking. A sensor with finer spatial resolution is not automatically better if cloud cover, revisit timing, missing observations, or the indicator’s scale make the record less consistent. Avoid combining sensors simply to fill gaps unless you have a documented harmonization method and have tested overlap periods.

3. Capture provenance for every input

At minimum, record the mission and instrument, product or collection name, processing level, acquisition date, publication date if available, scene or granule identifier, source URL or catalog identifier, download date, file checksum when practical, and spatial and temporal filters. Also record the software package versions, processing scripts, configuration files, and the date each result was generated.

A simple processing log can use these fields: run_id, input_id, source_collection, acquisition_date, AOI_version, CRS, cloud_rule, QA_rule, resampling_method, algorithm_version, output_path, operator, and notes. Do not overwrite old logs when a method changes. Create a new run identifier and retain the earlier output for comparison.

4. Separate the workflow into stages

  1. Discover: search the catalog using the area, date range, platform, product, and acquisition constraints.
  2. Inspect: read the product metadata and confirm bands, scale factors, units, QA layers, no-data values, and coordinate reference system.
  3. Filter: apply scene-level and pixel-level quality rules. A low scene-cloud percentage does not guarantee that your area of interest is clear.
  4. Transform: convert digital numbers to physical quantities when required, calculate the indicator, and apply masks in a documented order.
  5. Align: reproject, resample, clip, and aggregate using fixed rules. Avoid changing the grid from run to run.
  6. Summarize: calculate the selected statistic and retain pixel counts, valid-area fraction, and uncertainty or quality fields alongside the indicator.
  7. Validate: inspect maps and time-series plots, compare a sample of dates with independent observations where available, and investigate unusual jumps.

For decisions about spatial and temporal resolution, consult Satellite Revisit Time and Resolution Explained. Resolution should be selected from the process being measured, not from the headline specification of a mission.

Cadence and checkpoints

Use a two-level maintenance schedule. Run routine quality checks whenever new observations are added, and conduct a more formal review monthly or quarterly, depending on how quickly the monitored feature changes.

At every processing run

  • Confirm that the expected date range and area of interest were searched.
  • Count scenes or granules found, rejected, and retained.
  • Check valid-pixel coverage after masking; a single mean is not enough.
  • Plot the indicator with observation dates, gaps, and quality flags visible.
  • Inspect representative maps from clear, cloudy, and unusual dates.
  • Save the exact configuration and output identifier.

Monthly or quarterly review

Compare recent coverage with the planned cadence. Look for a sudden change in observation counts, spatial footprints, cloud-mask behavior, scale factors, or value distributions. Review product notices and release notes for collection changes, processor updates, calibration information, or retirement of a data stream. A change notice does not automatically mean the time series is wrong, but it is a reason to identify affected dates and test a controlled reprocessing run.

Maintain a small benchmark set of dates and locations. Re-run these examples after software, algorithm, or collection changes and compare the old and new outputs. Store both results, the difference map, and an explanation of whether the change reflects improved processing, changed inputs, or an unintended regression.

How to interpret changes

Start by separating environmental signal from observation and processing effects. A sharp indicator change on one date may result from a real disturbance, but it may also reflect cloud contamination, haze, snow, shadow, a different viewing geometry, a missing mask, or a scene that barely covers the study area. Examine the map, QA layers, valid-area fraction, and neighboring dates before assigning an environmental explanation.

Use comparable summaries. A monthly median calculated from many valid pixels is not directly equivalent to a single clear-scene mean. Document whether comparisons use the same season, baseline, spatial mask, compositing window, and minimum valid-pixel threshold. Seasonal indicators should generally be compared with the same season in other years, or transformed into anomalies relative to a clearly stated baseline. For broader guidance on baselines and misleading comparisons, see How to Read a Climate Graph.

Do not hide gaps by treating missing observations as zero or by interpolating without labeling the result. If interpolation is necessary, preserve the original observations, state the method, and show which values are estimated. Report coverage and uncertainty with the indicator. A modest change supported by a large, consistent valid area may be more credible than a dramatic change based on a small number of pixels.

When combining sensors, first compare overlapping observations after applying the same area, date window, quality rules, and aggregation. Look for systematic offsets, seasonal differences, and changes in variance. If harmonization is used, save the coefficients, training period, validation results, and limitations. A sensor transition should be visible in the metadata and, where possible, marked on the time-series plot.

When to revisit

Revisit the workflow on a regular monthly or quarterly schedule while the project is active, and immediately after a major data release or processing notice. Update the sensor comparison and recommended processing levels when a collection, processor, calibration record, or quality algorithm changes. Review cloud-mask and QA recommendations when a new masking method becomes available or when benchmark scenes show a systematic problem. Check download links, CRS definitions, scale factors, and mission status at least annually or after a major mission notice.

Use this practical checklist before publishing a new update or extending the time series:

  1. Has the source collection or product version changed?
  2. Are the bands, units, scale factors, and QA flags still interpreted correctly?
  3. Did the search return the expected observations for the latest period?
  4. Are the AOI, grid, projection, resampling, mask, and aggregation rules unchanged?
  5. Do benchmark scenes produce comparable results under the current software?
  6. Are gaps, valid-pixel counts, and uncertainty retained with the published values?
  7. Can another analyst reproduce the output from the log, script, configuration, and input identifiers?

If a correction is needed, do not silently replace the historical series. Archive the prior version, document the reason for reprocessing, run the updated method across all affected dates, and compare old and new results. Clearly label the revised series and explain whether conclusions changed. This habit turns a satellite time series from a one-off download into a maintained environmental record.

For related applications, see Satellite Imagery for Wildfires and Air Quality and How to Measure Sea-Level Rise. Both illustrate why sensor choice, metadata, coverage, and uncertainty must remain visible throughout an Earth observation workflow.

Related Topics

#satellite time series#workflow checklist#remote sensing#earth observation data#reproducible research
C

Captains Space Editorial

Earth Observation Editor

Senior editor and content strategist. Writing about technology, design, and the future of digital media. Follow along for deep dives into the industry's moving parts.