Landsat 5, 7, 8, or 9 in Earth Engine: Select Collection According to Data Year

In the Google Earth Engine catalog, Landsat is not a single dataset, but rather several collections, each containing data for specific year ranges. Choosing the wrong collection does not produce an error message. The result is simply a blank map or an index number that is quietly incorrect. This article summarizes the year ranges, band names, and scale factors that need to be checked before conducting analysis.

Which collection for which year

The Earth Engine catalog lists the availability range for each Collection 2 Level 2 (surface reflectance, 30-meter resolution) collection:

For example, data from the year 2000 can be obtained from either Landsat 5 or Landsat 7. Data from 2010 can be obtained from both satellites. Data from 2012 is only safe from Landsat 7, as the Landsat 5 collection stopped in May 2012 and Landsat 8 only started in March 2013. Data from 2015 onwards comes from Landsat 8, and since late 2021, Landsat 9 complements it, allowing the same location to be captured more frequently.

Trap 1: band numbers shift

Landsat 8 and 9 add one band at the beginning of the spectrum, causing the band numbers for the same function to shift:

The NDVI formula written for Landsat 8, which is (SR_B5 - SR_B4) / (SR_B5 + SR_B4), if run as is on Landsat 5 will use the wrong band and still produce a number between -1 and 1. Therefore, cross-satellite time series analysis should encapsulate band selection in a single function per satellite, rather than copying the formula.

Trap 2: pixel values are not yet reflectance

In the Level 2 collections above, the catalog mentions a scale factor of 0.0000275 and an offset of -0.2 for surface reflectance bands (SR_B*). Reflectance values are obtained by multiplying the raw values by 0.0000275 and then adding -0.2. The catalog code example does this with multiply(0.0000275).add(-0.2). Skipping this step makes thresholds like "NDVI above 0.5" meaningless, as the raw numbers are scaled in the tens of thousands.

Trap 3: Landsat 7 after 2003 has gaps

According to USGS, the Scan Line Corrector (SLC) of Landsat 7 was permanently damaged on May 31, 2003. The sensor then scanned the Earth in a zig-zag pattern, and after the overlapping parts were removed, about 78 percent of pixels per scene remain. Landsat 7 images after that date therefore have blank stripes. For 2004 to 2011, Landsat 5 is usually a more complete choice, while Landsat 7 is useful for patching 2012.

USGS also notes that Landsat 7 completed its nominal science mission on April 6, 2022, then orbited lower and produced images until January 19, 2024. The Earth Engine catalog adds a note about the shifted orbit to earlier acquisition times since 2017, so seasonal comparisons across years need to be done carefully.

Safe workflow sequence

1. Determine the analysis year range, then match it with the collection table.
2. Filter the collection by date and area boundaries, then count the remaining images before proceeding.
3. Apply the scale factor and offset to the SR_B* bands.
4. Mask cloudy pixels using the QA_PIXEL band. For Landsat 8 and 9, USGS defines bit 3 as cloud, bit 4 as cloud shadow, and bit 5 as snow with high confidence.
5. Select the red and NIR bands according to the satellite, then calculate the index.

The steps for selecting the workflow from the catalog can be viewed directly in the explanatory clip. To understand the threshold values for vegetation indices after the data is ready, see the NDVI article.

Sources

• Google for Developers, USGS Landsat 5 Level 2, Collection 2, Tier 1
• Google for Developers, USGS Landsat 7 Level 2, Collection 2, Tier 1
• Google for Developers, USGS Landsat 8 Level 2, Collection 2, Tier 1
• Google for Developers, USGS Landsat 9 Level 2, Collection 2, Tier 1
• USGS, Landsat 7
• USGS, Landsat Collection 2 Quality Assessment Bands