Why is Landsat used for long-term change detection?

Landsat is the longest continuous civilian Earth observation record, running since 1972, with consistent 30 m multispectral data since 1982. No other open archive offers that time depth at moderate resolution.

Which Landsat product should I use for change analysis?

Use Collection 2 Level-2 Surface Reflectance. It is atmospherically corrected and terrain-corrected, which makes scenes from different dates directly comparable without you doing the correction yourself.

How do I make two Landsat dates comparable?

Use surface reflectance products, match the season (same phenological window), screen with the QA_PIXEL band for clouds and shadow, and compare derived indices like NDVI or NBR rather than raw digital numbers.

What are the limits of Landsat for terrain change?

At 30 m resolution Landsat cannot resolve small features, and optical sensors miss change under cloud, vegetation, or at night. It detects surface-cover change well but is weak for direct elevation change.

Landsat for Long-Term Terrain Change

Short answer

Landsat is the tool of choice for multi-decadal land-surface change because it is the only open archive with a continuous, calibrated record stretching back to 1972, and to consistent 30 m multispectral imagery since Landsat 4/5 in 1982. For terrain and land-cover work the practical product is Collection 2 Level-2 Surface Reflectance, which is already atmospherically and terrain corrected so scenes from 1990 and 2025 can be compared directly. The skill is not downloading the data; it is controlling the variables (season, atmosphere, sun angle, sensor differences) so that the change you measure is real change on the ground, not an artefact of the acquisition.

Why Landsat for the long view

Landsat's value is time depth at a usable resolution. The sensors evolved but the program preserved continuity:

Landsat 4/5 TM (1982–2013): 30 m in 6 reflective bands, 16-day revisit.
Landsat 7 ETM+ (1999–): added a 15 m panchromatic band; note the 2003 Scan Line Corrector (SLC) failure that leaves wedge-shaped data gaps in every scene after May 2003.
Landsat 8/9 OLI/TIRS (2013–, 2021–): 30 m, narrower and better-placed bands, 12-bit radiometry, and with two satellites an effective 8-day revisit.

That overlap lets you build a defensible 40-year time series. Sentinel-2 (10–20 m, since 2015) is sharper and revisits faster, but it cannot reach back before 2015. The common pattern is Landsat for the historical baseline and trend, Sentinel-2 for recent detail.

Choosing the right product

USGS distributes Landsat as Collections (currently Collection 2) at three levels:

Level-1 is top-of-atmosphere, geometrically corrected. Avoid for multi-date comparison; atmosphere varies scene to scene.
Level-2 Surface Reflectance is corrected to surface reflectance using LEDAPS (TM/ETM+) or LaSRC (OLI). This is what makes dates comparable. It ships with a QA_PIXEL bitmask band that flags cloud, cloud shadow, snow, and water per pixel.
Level-3 / analysis-ready science products (for example burned area) are derived for specific themes.

For change work, download Level-2 from USGS EarthExplorer or pull it from a cloud catalog (Google Earth Engine LANDSAT/LT05/C02/T1_L2, LANDSAT/LC08/C02/T1_L2, or AWS/Microsoft Planetary Computer STAC).

Making two dates truly comparable

This is where most analyses go wrong. Five controls matter:

Match the season. Compare June to June, not June to October. Vegetation phenology, soil moisture, and sun elevation change the signal far more than slow terrain change does. Same phenological window removes most of that noise.
Use surface reflectance, not digital numbers. Raw DN values are not physically comparable across sensors or atmospheres.
Screen clouds with QA_PIXEL. Build a mask from the cloud, shadow, and cirrus bits and apply it to both dates so a "change" is never just a cloud edge.
Account for sun angle. Even at the same date-of-year the solar elevation differs slightly across decades; surface reflectance products normalize most of this, but in steep terrain residual illumination differences remain.
Cross-sensor care. OLI band response differs from TM. When you bridge a TM scene to an OLI scene, prefer derived indices (NDVI, NBR), which are more stable across sensors than individual band reflectances, and apply published cross-sensor harmonization coefficients if you need raw-band continuity.

A worked change-detection workflow

Goal: map vegetation loss and surface disturbance between 1995 and 2025 over a catchment.

Select scenes. One low-cloud Level-2 scene per epoch in the same month. Note the path/row so both cover the area.

Stack and clip. Subset to the catchment to keep processing light:

gdalwarp -cutline catchment.gpkg -crop_to_cutline -t_srs EPSG:32630 \
  LC09_..._SR_B5.TIF nir_2025.tif

Compute NDVI for each date: NDVI = (NIR - Red) / (NIR + Red). In QGIS the Raster Calculator expression is ("nir@1" - "red@1") / ("nir@1" + "red@1"), with both bands as Level-2 surface reflectance.
Compute NBR if fire or clearing is involved: NBR = (NIR - SWIR2) / (NIR + SWIR2). The differenced version, dNBR = NBR_pre - NBR_post, is the standard burn-severity metric.
Difference the indices: dNDVI = NDVI_1995 - NDVI_2025. Strong positive values flag vegetation loss.
Threshold and validate. Pick a threshold from the histogram, then check candidate change polygons against high-resolution imagery or field knowledge. Report the threshold; it is an interpretive choice.

For trend rather than two-date change, stack annual cloud-free composites and fit a per-pixel linear trend, or use an established algorithm such as LandTrendr (designed for the Landsat record) to segment each pixel's spectral trajectory.

What Landsat cannot tell you

Be explicit about limits in any deliverable:

30 m pixels average over 900 m². Narrow roads, small slumps, and individual outcrops are subpixel.
Optical sensors see reflected light only. Cloud, dense canopy, and darkness block them. For all-weather or surface-deformation work, pair with radar (Sentinel-1, InSAR).
Landsat measures surface cover, not elevation. It tells you that a hillside changed (vegetation gone, fresh soil exposed), not how many metres of material moved. Volumetric terrain change needs DEM differencing or LiDAR.
The SLC-off gap (post-2003 Landsat 7) must be gap-filled or avoided; do not let the wedge stripes masquerade as change.

Validation and QA

Confirm both scenes share the same CRS and pixel grid before differencing; a half-pixel misregistration produces false change along every edge. Verify the QA mask actually removed cloud (overlay it and inspect). Compare your change map against an independent reference, such as a known logging boundary, a documented fire perimeter from a fire agency, or recent high-resolution basemap imagery. State the index, threshold, dates, and sensors used, so the result can be reproduced.

Bathyl perspective

We treat the Landsat archive as a memory, not a snapshot. Its strength is establishing a credible baseline and trend that pre-date any drone or LiDAR survey a client commissioned last year. We always pair it with a clear note on what 30 m optical data can and cannot resolve, so a change map is read as evidence to investigate, not as a settled measurement.