What is Landsat? Satellites, Bands, Data Access

Landsat is the joint U.S. Geological Survey (USGS) and NASA Earth observation program that has imaged the Earth's land surface continuously since July 1972. It is the longest record of its kind. Landsat 8 and 9 image every land pixel every 8 days, with 30 m color and infrared bands and 100 m thermal bands. The whole archive has been free to use since 2008. This article tests one claim: Landsat's main value is continuity more than resolution, and even a cloud-free scene needs a per-pixel mask before its values can be compared.

Key takeaways

• Landsat is a series of nine Earth observation satellites built by NASA and operated by the USGS. Landsat 8 and 9 are in orbit today.
• Each mission was designed to match the one before it, so a 30 m pixel can be compared across five decades.
• Landsat 8 and 9 carry the Operational Land Imager (OLI), with nine reflective bands, and the Thermal Infrared Sensor (TIRS), with two thermal bands.
• Collection 2 is the USGS reprocessing of the whole archive. Its Level-2 products give surface reflectance and surface temperature as Cloud Optimized GeoTIFFs with a SpatioTemporal Asset Catalog (STAC), a searchable index of the files.
• Free data since 2008 made Landsat the basis for global, annual maps of forests, water, and crops.
• The open problems for the next mission are clouds, 100 m thermal pixels, continuity, and data volume. Landsat 10 is expected to add about 3.3 TB a day.
• In Wherobots, one SQL statement averages Landsat surface temperature over each census tract. This article follows Dallas, Texas, in the summer of 2023.

What is Landsat?

NASA launched the first satellite, then called the Earth Resources Technology Satellite (ERTS-1) and later renamed Landsat 1, on July 23, 1972. Eight more followed. Landsat 6 failed to reach orbit in 1993, Landsat 9 launched on September 27, 2021, and the USGS decommissioned Landsat 7 on June 4, 2025, after 26 years in orbit. NASA builds and launches the satellites, and the USGS operates them and keeps the archive at its Earth Resources Observation and Science (EROS) Center.

Landsat's defining property is continuity. Each mission was specified to overlap and calibrate against the last. The Multispectral Scanner (MSS) on Landsats 1 to 5 recorded four bands at about 80 m. The Thematic Mapper (TM) on Landsats 4 and 5 recorded seven bands at 30 m. The Enhanced Thematic Mapper Plus (ETM+) on Landsat 7 added a 15 m panchromatic band, one black-and-white band with finer pixels.

Wulder and colleagues' Current status of Landsat program, science, and applications (Remote Sensing of Environment, 2019) summarizes the program's science base. The fiftieth-anniversary review, Fifty years of Landsat science and impacts (2022), traces how the missions were kept continuous. It groups the main impacts into crop mapping and water use, climate drivers and impacts, ecosystem and land cover monitoring, and the changing human footprint.

Roy and colleagues defined the goals for Landsat 8 in Landsat-8: Science and product vision for terrestrial global change research (2014). Landsat 9 is a near copy of Landsat 8, and Masek and colleagues made the case for that choice in Landsat 9: Empowering open science and applications through continuity (2020). The main change is its OLI-2 imager, the second Operational Land Imager. It records 14-bit values, up from 12 bits, which gives more brightness levels and separates small differences over water and dense forest.

The next mission was planned as Landsat Next, a constellation of three satellites. In May 2026 NASA and the USGS replaced that design with a single 26-band observatory, renamed Landsat 10.

How Landsat works

Landsat 8 and 9 fly in the same orbit at 705 km, eight days apart. This sun-synchronous orbit crosses each place at the same local time. Each satellite repeats its ground track every 16 days, so together they image every land pixel every 8 days. Scenes are indexed by the Worldwide Reference System (WRS-2), a fixed grid of 233 paths and 248 rows. Each scene covers about 185 by 180 km. A path and row always name the same footprint. Dallas sits in path 27, row 37. Scenes are delivered as a set of GeoTIFFs, one per band.

OLI/TIRS band	Wavelength	Pixel size	Common use
1 (coastal aerosol)	0.43 to 0.45 µm	30 m	Coastal water, aerosols
2, 3, 4 (blue, green, red)	0.45 to 0.67 µm	30 m	True color, turbidity
5 (near-infrared)	0.85 to 0.88 µm	30 m	Vegetation, NDVI
6, 7 (shortwave infrared)	1.57 to 2.29 µm	30 m	Moisture, burn scars, geology
8 (panchromatic)	0.50 to 0.68 µm	15 m	Sharpening
9 (cirrus)	1.36 to 1.38 µm	30 m	Cloud detection
10, 11 (thermal)	10.6 to 12.5 µm	100 m	Surface temperature

Collection 2 and processing levels

The USGS processes the archive in collections. Each collection reprocesses every scene to one standard for pixel location and one for converting pixel values to measured light. Crawford and colleagues describe the current one in The 50-year Landsat collection 2 archive (Science of Remote Sensing, 2023). Collection 2 was released in 2020. It improved geolocation and, for the first time, added a global inventory of Level-2 surface reflectance and surface temperature.

• Level-1 products are calibrated radiance and reflectance at the top of the atmosphere.
• Level-2 products correct for the atmosphere and give surface reflectance and surface temperature.
• Level-3 products are science variables such as fractional snow cover and burned area.

Level-2 values are stored as scaled integers. Surface temperature in Kelvin is the pixel value multiplied by 0.00341802, plus 149.0, as the USGS explains in its guide to Level-2 scale factors. A quality assessment (QA) band named QA_PIXEL flags fill, cloud, cirrus, cloud shadow, snow, clear, and water per pixel.

Level-2 products come from these algorithms:

• Surface reflectance. Vermote and colleagues assessed the LaSRC (Land Surface Reflectance Code) atmospheric correction for OLI in Preliminary analysis of the performance of the Landsat 8/OLI land surface reflectance product (2016).
• Surface temperature. Cook, Schott, and colleagues developed the single-channel method for the thermal archive in Development of an operational calibration methodology for the Landsat thermal data archive (Remote Sensing, 2014).
• Clouds and shadows. Zhe Zhu and Curtis Woodcock introduced Fmask in Object-based cloud and cloud shadow detection in Landsat imagery (2012), the basis for automated cloud masking of Landsat scenes.

Indices build on the reflectances. Compton Tucker's Red and photographic infrared linear combinations for monitoring vegetation (1979) formalized the red and near-infrared index now known as the Normalized Difference Vegetation Index (NDVI).

Where to get Landsat data

The USGS lists every route on its Landsat data access page. EarthExplorer serves single scenes. For analysis, Collection 2 is hosted as Cloud Optimized GeoTIFFs in the usgs-landsat bucket on AWS. A STAC catalog indexes the scenes so a client can search them by place and date. Microsoft Planetary Computer and Google Earth Engine host copies. A Cloud Optimized GeoTIFF stores internal tiles and lower-resolution overviews. A client can read only the parts it needs over HTTP range requests, web requests for part of a file. The Open Geospatial Consortium's OGC Cloud Optimized GeoTIFF Standard (2023) sets out this way of reading.

Landsat vs Sentinel-2

	Landsat 8 and 9	Sentinel-2A, 2B, 2C
Operator	USGS and NASA	European Space Agency (ESA), for the EU Copernicus programme
Archive starts	1972	2015
Reflective pixel size	30 m (15 m panchromatic)	10, 20, and 60 m
Thermal bands	Two, 100 m	None
Revisit	8 days with two satellites	5 days with two satellites
Data policy	Free and open	Free and open

Landsat suits long time series and temperature. Sentinel-2 suits finer detail and more frequent revisits. NASA's Harmonized Landsat and Sentinel-2 product resamples both to one 30 m grid with matched bands. Claverie and colleagues describe it in The Harmonized Landsat and Sentinel-2 surface reflectance data set (2018). The two then work as one denser time series.

Landsat examples

Well-known projects built on the 30 m record:

• Global forest change. Hansen and colleagues' High-resolution global maps of 21st-century forest cover change (Science, 2013) classified Landsat imagery. They mapped 2.3 million km² of forest lost and 0.8 million km² gained from 2000 to 2012. The University of Maryland updates the dataset each year as the tree cover loss layer on Global Forest Watch.
• Global surface water. Pekel and colleagues' High-resolution mapping of global surface water and its long-term changes (Nature, 2016) analyzed more than three million Landsat scenes. They mapped monthly water at 30 m over 32 years.
• Crop maps. The U.S. Department of Agriculture (USDA) Cropland Data Layer maps crop type across the United States each year from Landsat and Sentinel-2.
• Urban heat. Weng, Lu, and Schubring's Estimation of land surface temperature-vegetation abundance relationship for urban heat island studies (2004) related Landsat surface temperature to vegetation cover. The Dallas example below maps that signal by census tract.
• Disaster and insurance. Analysts reconstruct pre-event conditions and post-event damage for wildfires and floods from the same 30 m record, often with change detection.
• Outreach. NASA's Your Name in Landsat spells a name with Landsat images of landscapes shaped like letters.

Landsat in Wherobots

WherobotsDB reads Landsat Collection 2 scenes as out-db rasters. Each row holds a reference to a Cloud Optimized GeoTIFF on S3. Pixels are read only when a function needs them. The STAC reader searches the Landsat catalog by area and date. It returns one row per scene, with each band registered as an out-db raster.

For paid editions, the Wherobots Data Hub includes a ready Landsat surface temperature table in wherobots_pro_data.landsat, built from Landsat 8 and 9 Collection 2 surface temperature (band ST_B10). The scene-level table holds 82,960 scenes acquired from August 1 to October 11, 2023, and a second table cuts each scene into 1024 by 1024 pixel tiles.

Dallas County surface temperature by census tract

Nine scenes in the table cover Dallas County. The Landsat 9 scene from August 19, 2023 reports no cloud. This query averages that scene's surface temperature over each Dallas County census tract, from two sources:

Source	Format	Size	What the query read
wherobots_pro_data.landsat.landsat_surface_temperature_1024x1024_outdb	Out-db raster tiles that point to a Cloud Optimized GeoTIFF on S3	One scene of 7,591 by 7,721 pixels, stored as 64 tiles of 1024 by 1024 pixels	Band ST_B10 surface temperature tiles for scene LC09_L2SP_027037_20230819_20230821_02_T1_ST
wherobots_open_data.us_census.tiger_tract	Vector table of US Census TIGER tract polygons	85,529 US tracts, 645 of them in Dallas County	Tract ID (GEOID) and boundary for Dallas County, Texas

The query joins the tiles to the tract polygons where they overlap (ST_Intersects). RS_ZonalStatsAll returns the pixel count and pixel sum for each tile and tract pair. The outer query adds up the tiles per tract and applies the USGS scale factor:

WITH t AS (
  SELECT GEOID, geometry FROM wherobots_open_data.us_census.tiger_tract
  WHERE STATEFP = '48' AND COUNTYFP = '113'          -- Dallas County, Texas
),
r AS (
  SELECT rast, geometry
  FROM wherobots_pro_data.landsat.landsat_surface_temperature_1024x1024_outdb
  WHERE id = 'LC09_L2SP_027037_20230819_20230821_02_T1_ST'
),
z AS (
  SELECT t.GEOID, RS_ZonalStatsAll(r.rast, t.geometry, 1, false, true) AS s
  FROM t JOIN r ON ST_Intersects(r.geometry, t.geometry)
)
SELECT GEOID,
       SUM(s.count) AS pixels,
       SUM(s.sum) / SUM(s.count) * 0.00341802 + 149.0 - 273.15 AS lst_c
FROM z
WHERE s.count > 0
GROUP BY GEOID

Raster and vector run in one engine; most stacks pick one. The query returned all 645 tracts in the county and summarized 2.6 million 30 m pixels. The median tract averaged 52.4 °C, and 550 of the 645 tracts averaged 50 °C or more. The five coolest tracts ranged from 37.9 to 43.5 °C. Each includes lake water inside its TIGER boundary, from 0.8 to 11.0 km² of Joe Pool Lake, Lake Ray Hubbard, or White Rock Lake. Open water stays far cooler than roofs and pavement. The hottest tract, at 56.4 °C, sits beside Dallas Love Field, on the airport's west side.

Reading the Dallas results

Scene choice. The nine scenes over WRS-2 path 27, row 37 alternate between Landsat 8 and Landsat 9 every 8 days:

Date (2023)	Satellite	Cloud cover
August 3	Landsat 9	not reported
August 11	Landsat 8	24.2%
August 19	Landsat 9	0%
August 27	Landsat 8	37.9%
September 4	Landsat 9	9.8%
September 12	Landsat 8	86.1%
September 20	Landsat 9	36.8%
September 28	Landsat 8	3.4%
October 6	Landsat 9	0.2%

Three of the nine scenes report under 5% cloud. Scene-level cloud cover is a first filter. Ranking tracts still needs a per-pixel QA_PIXEL mask.

Distribution. The tract means cluster tightly. The 10th to 90th percentile runs from 49.2 to 54.5 °C, and 36 tracts reached 55 °C or more. The pixel-weighted county mean was 50.8 °C, below the 52.4 °C median tract. Pixel weighting counts large tracts for more, and several of the largest tracts include lake water.

Water. As the five lake tracts show, a tract mean that mixes lake and streets understates the heat on land. One fix drops water pixels with the QA_PIXEL water flag. Another clips tracts to land before the zonal statistics. The surface temperature table in the Data Hub holds only band ST_B10. The water mask has to come from the matching QA_PIXEL file in the STAC catalog.

Extending the query

Because raster and vector data share one engine, the same pattern runs at other scales. Join to Overture building footprints for a building footprint-level heat map, or loop over the nine scenes for a seasonal series. RS_MapAlgebra computes indices across bands, and RS_Value samples a pixel at a point. RasterFlow's bring-your-own-rasters path builds mosaics from rasters you supply, such as Landsat scenes. It then runs PyTorch models over them for land cover classification. The Wherobots MCP Server uses the Model Context Protocol (MCP), a standard way for AI tools to call outside services. It lets an AI coding tool find this table and write the query from a plain-language request.

Landsat research and what scale changes

Four changes shaped Landsat research. They are the 2008 free data policy, algorithms that fit a model to each pixel's history, computing next to the archive, and foundation models pretrained on the archive.

Free data changed the method

Until 2008 the USGS charged for scenes, so many studies used a few cloud-free images. Woodcock and colleagues described the new policy in Free access to Landsat imagery (Science, 2008), and the USGS began serving the US archive at no cost that year. Zhu and colleagues' Benefits of the free and open Landsat data policy (2019) documents the result. Studies shifted from a few selected dates to dense time series over large areas. National and global monitoring became affordable. The forest and surface water maps in the examples above depend on that shift.

Time series replaced image pairs

Classic change detection compares two dates. With every scene available, algorithms fit a model to each pixel's whole history and flag the dates where values break from it.

• LandTrendr comes from Kennedy, Yang, and Cohen in Detecting trends in forest disturbance and recovery using yearly Landsat time series (2010). It fits straight-line segments to one value per pixel per year, so a clear-cut shows as a drop and regrowth as a slow rise.
• Continuous Change Detection and Classification (CCDC) comes from Zhu and Woodcock in Continuous change detection and classification of land cover using all available Landsat data (2014). It fits seasonal curves to every clear observation of a pixel. It labels the land cover in each period between breaks. The USGS uses it for its Land Change Monitoring, Assessment, and Projection (LCMAP) product.

A missed cloud looks like a disturbance, which is why cloud masks and the Level-2 corrections come first in a time series workflow.

Compute moved to the archive

Analysis over millions of scenes needs compute next to the data. Gorelick and colleagues' Google Earth Engine: Planetary-scale geospatial analysis for everyone (2017) describes a platform built around the Landsat archive. The AWS copy is stored as Cloud Optimized GeoTIFFs with STAC, so any engine that reads that format can read scenes in place. On the vector side, Jia Yu, Jinxuan Wu, and Mohamed Sarwat introduced GeoSpark (ACM SIGSPATIAL 2015). Yu, Zongsi Zhang, and Sarwat described how it splits data by location to run joins in Spatial data management in Apache Spark (GeoInformatica, 2019). GeoSpark became Apache Sedona. Its 1.5 series made large raster and vector joins with RS_Intersects and RS_ZonalStats practical. A Landsat scene and a table of polygons can join in one distributed query. Yu and Sarwat went on to found Wherobots.

Foundation models

Landsat is a common pretraining source for geospatial foundation models, large models trained on broad data and then adapted to specific tasks. Stewart and colleagues released SSL4EO-L (NeurIPS 2023). It holds about 5 million image patches across Landsat sensors, and the first foundation models pretrained on Landsat. Jakubik and colleagues describe NASA and IBM's Prithvi model in Foundation models for generalist geospatial artificial intelligence (2023). It was pretrained on Harmonized Landsat and Sentinel-2 imagery and fine-tuned for flood mapping, burn scars, and crop segmentation.

Open problems

• Clouds and gaps. Usable observations per pixel vary widely by region and season, so a time series in the humid tropics has far fewer clear dates than one in a desert.
• Thermal resolution. TIRS measures at 100 m, coarser than many city blocks. The 30 m surface temperature product is resampled from those pixels. Values within one thermal footprint are therefore not independent.
• Harmonization. Sensors in a combined record such as Harmonized Landsat and Sentinel-2 differ slightly in band edges, viewing angle, and calibration. Every harmonization gives up some of each sensor's detail to fit a common grid.
• Continuity. With Landsat 7 retired, the record rests on Landsat 8 and 9, and Landsat 10 is the planned successor. Roy and colleagues review the Landsat Next design in The next Landsat: Mission turning point? (Remote Sensing of Environment, 2026). That design had twenty-one 10 and 20 m reflective bands, five 60 m thermal bands, and a six-day revisit. The authors warn that a reduced mission risks delivering less and lower quality data than the current satellites. In May 2026 NASA and the USGS chose the single 26-band observatory described above.
• Data volume. The USGS expects Landsat 10 to add about 3.3 TB a day, with Level-1 scenes of about 3.7 GB. Downloading scenes one at a time does not scale to that.

What scale changes

• Units people decide on. Zonal statistics summarize the pixels inside each tract, parcel, field, or building. Policy and pricing work uses values for those units.
• Joins across sources. Surface temperature, building footprints, census estimates, and tree canopy sit on different grids and geometries. The analysis joins them where they overlap.
• Time at scale. Fifty years of 8- to 16-day observations over every polygon in a country is a distributed join, the workload Apache Sedona was designed to run. The Dallas query scales the same way, with more scenes, polygons, and joins in one distributed job.

Conclusions

The missions, Collection 2, and the Dallas test show what Landsat is in practice. It is a 50-year, consistently processed record. Its values need masking and a chosen unit before comparison.

1. Landsat is the source for dates before 2015 and for surface temperature. For detail under 30 m or revisits under 8 days, use Sentinel-2, or the Harmonized Landsat and Sentinel-2 product for one denser 30 m series.
2. Landsat Level-2 values are scaled integers, so put the scale factors in the query. Surface temperature in Kelvin is the pixel value times 0.00341802 plus 149.0. A tract mean pixel value of 52,837 is 56.4 °C.
3. A Landsat scene with low reported cloud still needs a QA_PIXEL mask. Scene cloud cover is only a first filter. Three of the nine Dallas scenes report under 5% cloud.
4. A Landsat result depends on the unit and the weighting, so decide both before reading it. The median Dallas tract averaged 52.4 °C, and the pixel-weighted county mean was 50.8 °C. The five coolest tracts, 37.9 to 43.5 °C, each hold lake water. When the question is about people, drop water pixels or clip tracts to land first.
5. The Landsat archive keeps growing, so read scenes in place. Landsat 10 is expected to add about 3.3 TB a day. Distributed joins over Cloud Optimized GeoTIFFs with STAC replace per-scene downloads.
6. Landsat thermal pixels are 100 m, resampled to 30 m. A tract mean averages many thermal footprints, and values within one footprint are not independent. Because of that, use Landsat surface temperature for tract and neighborhood comparisons, and a finer sensor for single buildings.

Related terms

• Sentinel-2: the 10 m European mission most often paired with Landsat
• NDVI: the vegetation index built on Landsat's red and near-infrared bands
• Change detection: the analysis the 50-year Landsat archive is built for
• Remote sensing: the field Landsat helped define
• GeoTIFF: the file format Landsat bands are delivered in
• Land cover classification: mapping surface types from Landsat bands
• Satellite imagery analysis: the wider set of methods for working with scenes
• Raster vs vector data: why joining Landsat pixels to tracts needs both

Read more from Wherobots

• Raster spatial joins at scale: Google Earth Engine and BigQuery vs Apache Sedona and Wherobots
• Introducing Sedona 1.5: raster and vector data processing at scale
• Introduction to the Data Hub
• Zonal statistics: BigQuery + Earth Engine comparison
• Bring your own rasters with RasterFlow

Query Landsat surface temperature and reflectance in SQL with WherobotsDB with a Wherobots free trial at cloud.wherobots.com.