Skip to content
AI.info

Computer vision

Remote Sensing and Geospatial Vision

Learn multispectral imagery, spatial resolution, temporal stacks, georegistration, change detection, geospatial splits, and responsible earth-observation workflows.

By the end you can

Visual

Remote sensing observes the world along several resolutions

Improving one resolution trades against coverage, revisit time, noise, or cost. The trade is not abstract, and it is not free. Sentinel-2 buys 10 m pixels over a 290 km swath. Landsat 8 accepts 30 m over a 185 km swath on a 16-day cycle. Sentinel-1 gives up colour entirely to see through cloud. The case section below gives the published figures for all three.

FigureLayers · 5 layers
  1. 01

    Spatial resolution

    Ground area represented by a pixel or effective point-spread function.

  2. 02

    Spectral resolution

    Number, width, and placement of wavelength bands.

  3. 03

    Temporal resolution

    How often comparable observations revisit a location.

  4. 04

    Radiometric resolution

    Sensitivity and quantization of measured signal levels.

  5. 05

    Geographic coverage

    Area, orbit, swath, and acquisition availability.

A pixel may contain a mixture of land covers

At ten meters, one pixel can include roof, tree, road, and shadow; its spectral vector is an aggregate measurement, not a tiny photograph of one pure object.

The mixture is worse than the phrase "ten-meter pixel" suggests, because a scene is not acquired at one scale. Sentinel-2 “samples 13 spectral bands: four bands at 10 m, six bands at 20 m and three bands at 60 m spatial resolution”, according to SentiWiki, ESA's Copernicus mission documentation. A stacked ten-metre pixel therefore carries four measurements taken at its own footprint and nine resampled from the coarser 20 m and 60 m footprints. The vector is a mixture in space before any land cover mixes inside it.

Task labels should respect scale. Parcel classification, building footprints, canopy cover, and individual-tree detection require different imagery and annotation.

Ground sampling distance sets a lower bound on what can be claimed reliably.

Comparison

Optical, radar, and elevation sensors reveal different properties

Sensor fusion is useful because modalities fail differently. The difference is physics, and it has documented consequences. Sentinel-1 carries C-band synthetic aperture radar at 5.405 GHz and acquires “imagery regardless of the weather”, day and night, in SentiWiki's description. NASA Earthdata says the two satellites together provide “all-weather, day-and-night imagery of Earth’s surface”. An optical instrument has neither property. It measures reflected sunlight, and it cannot measure that through a cloud or at night. That is why a flood mapped at three in the morning under an overcast sky is a radar product, and why a crop-type map is not.

FigureComparison · 3 columns

Multispectral optical

Measures reflected sunlight in selected wavelength bands.

  • Rich color and vegetation cues
  • Affected by clouds and illumination
  • Season matters
  • Example: Sentinel-2

Synthetic aperture radar

Measures microwave backscatter and phase.

  • Works through clouds
  • Sensitive to geometry and moisture
  • Speckle requires care
  • Example: flood mapping

LiDAR or elevation

Measures distance or derived surface height.

  • Strong 3D structure
  • Coverage can be sparse or costly
  • Registration matters
  • Example: canopy height

Case

Resolution, swath and revisit cannot all be maximised at once

Sentinel-2, Landsat 8 and Sentinel-1 make the trade-offs concrete. Each mission publishes its own numbers.

Sentinel-2 is the fine-grained setting. It “samples 13 spectral bands: four bands at 10 m, six bands at 20 m and three bands at 60 m spatial resolution”, “the orbital swath width is 290 km”, and twin satellites “phased at 180°” deliver “a high revisit frequency of 5 days at the Equator”, records ESA's Copernicus mission documentation. ESA's own Sentinel-2 page states the same four figures: 13 spectral bands, 10 m resolution, a 290 km swath, a 5-day revisit from two identical satellites.

Landsat 8 works coarser and returns less often. The USGS mission page gives “15-meter panchromatic and 30-meter multi-spectral spatial resolutions”, “a 185 km (115 mi) swath” and “a 16-day repeat cycle with an equatorial crossing time of 10:12 a.m. +/- 5 minutes”. NASA lists the same resolutions alongside 100 m in the thermal infrared.

Sentinel-1 is the radar answer to cloud. The same ESA documentation gives C-band synthetic aperture radar at 5.405 GHz, all-weather day-and-night acquisition, a 12-day repeat for a single satellite and “a 6 day exact repeat cycle” for the constellation, with the Interferometric Wide mode covering a 250 km swath. Three missions, three settings of one dial: 10 m at 5 days over 290 km, 30 m at 16 days over 185 km, or no colour at all and no dependence on weather.

Analogy

Comparing calendars drawn with different pens

A diary tracks a landscape where some pages use watercolor, others pencil, and several are missing because of clouds. Change detection requires aligning dates, scales, and drawing styles before comparing marks.

Watercolor and pencil are choices of style, while sensors measure physical radiance or backscatter with calibrated geometry. Temporal gaps and modality mismatch are what the diary captures.

The missing pages are not always weather, and one of them has a date. “On May 31, 2003, the Scan Line Corrector (SLC), which compensated for the forward motion of the satellite, failed. Subsequent efforts to recover the SLC were not successful, and the failure was permanent”, records the U.S. Geological Survey on its Landsat 7 mission page. NASA Science, describing the same mission, says that “22% of each image was lost” and that “the SLC failure was irreversible”. From that date every ETM+ scene came down ruled through with wedge-shaped gaps that widen toward the scene edge. In the USGS wording, “78 percent of the scene's pixels still remain after the duplicated areas were removed”. Landsat 7 kept acquiring science data until 19 January 2024 and was decommissioned on 4 June 2025, so more than two decades of the archive carries that pattern. A difference image that reads those gaps as observations will report change where the instrument simply stopped writing.

A difference image mixes real change with sensor, season, atmosphere, and registration error.

Key idea

Nearby tiles are not independent examples

Adjacent patches share roads, fields, weather, geology, and acquisition artifacts; random tile splits can place nearly identical neighborhoods in training and test.

The size of that effect has been measured on one model and one dataset. A random forest was trained to map forest biomass from a central-African forest inventory of 11.8 million trees. Scored by conventional random cross-validation, it reported R2 = 0.53, an RMSPE of 56.5 Mg ha-1, about 19%. The same model, scored by a spatial 44-fold cross-validation that keeps training and test blocks apart on the ground, fell to R2 = 0.14 and an RMSPE of 77.5 Mg ha-1, about 26%. A null model that predicts the mean everywhere gave 82 Mg ha-1. “A standard nonspatial validation method suggests that the model predicts more than half of the forest biomass variation, while spatial validation methods accounting for SAC reveal quasi-null predictive power”, wrote Ploton and colleagues in Nature Communications in 2020.

Nothing about the model changed between 0.53 and 0.14. Only the split did. The finding travels beyond forests. Andy Stock, reporting a marine remote sensing case study in Frontiers in Remote Sensing, restates it: “Random splits suggested good predictive skill, but spatial cross-validation suggested no predictive skill, reflecting the known effects of data leakage when training and testing data are insufficiently separated”.

Split by geographic region, acquisition campaign, or future time according to deployment; report distance from training coverage and performance in novel regions.

Spatial autocorrelation can turn a map benchmark into a location-memory test.

Example

What can mimic environmental change

Change detection must control acquisition differences before declaring events. Two of the six below are not hypothetical. They are dated, documented steps in the archive that a naive time series will read as ground change.

  • Seasonality: Crops, snow, leaf cover, and water level change predictably.
  • Atmosphere: Haze, aerosol, and thin clouds alter optical reflectance.
  • Sun geometry: Shadow direction and brightness vary with date and latitude.
  • Registration error: A one-pixel shift creates false boundaries around every object.
  • Sensor transition: Band response and resolution differ across platforms, and an instrument can also change mid-mission. Landsat 7's Scan Line Corrector failed permanently on 31 May 2003. Every ETM+ scene after that date keeps about 78 percent of its pixels; the other 22 percent are wedge-shaped gaps that widen toward the scene edge.
  • Processing revision: New atmospheric correction or mosaicking changes pixel values. On 25 January 2022 Copernicus deployed Sentinel-2 Processing Baseline 04.00, adding a band-dependent constant radiometric offset (RADIO_ADD_OFFSET / BOA_ADD_OFFSET) so that negative reflectances over dark surfaces would not be truncated. The Google Earth Engine Data Catalog warns in its entry for COPERNICUS/S2_SR_HARMONIZED: “After 2022-01-25, Sentinel-2 scenes with PROCESSING_BASELINE '04.00' or above have their DN (value) range shifted by 1000. The HARMONIZED collection shifts data in newer scenes to be in the same range as in older scenes.” A differencing job spanning that date sees a step of 1000 across the scene, with no ground event to attach it to.

Steps

Evaluate a geospatial product at pixel, object, and area levels

Map quality depends on how predictions aggregate over land.

The last step has a published standard. Olofsson and colleagues set out eight good-practice recommendations in Remote Sensing of Environment in 2014. Among them: “implement a probability sampling design”; “summarize the accuracy assessment by reporting the estimated error matrix in terms of proportion of area and estimates of overall accuracy, user’s accuracy (or commission error), and producer’s accuracy (or omission error)”; “estimate area of classes … based on the reference classification of the sample units”; “quantify uncertainty by reporting confidence intervals for accuracy and area parameters”; and “document deviations from good practice that may substantially affect the results”.

What that discipline produces is unglamorous and honest. ESA's WorldCover 2021 v200 maps 11 land-cover classes at 10 m from Sentinel-1 and Sentinel-2, and it is about as good as a global product gets. “This new WorldCover map for 2021 was released on 28 October 2022 and resulted in a global overall accuracy of 76.7%”, states the project's About page. The FAO data catalogue confirms it independently: “The WorldCover 2021 v200 reaches an overall accuracy of 76.7%”. Digital Earth Africa's data specification supplies the intervals the good-practice paper asks for: 76.7% ±0.5 globally and 76.5% ±1.3 over Africa, against 74.4% ±0.1 and 73.6% ±0.2 for the 2020 version. Note what the numbers do rather than what they promise. They put a bound on the map. They show the accuracy moving between versions. They show the bound is looser over Africa than it is globally. A map without those numbers is a picture.

FigureProcess · 5 steps

  1. 1. Verify geolocation

    Measure registration error and coordinate-system consistency.

  2. 2. Score local predictions

    Use pixel, object, boundary, or parcel metrics appropriate to the task.

  3. 3. Estimate area bias

    Compare total mapped area and stratified uncertainty.

  4. 4. Test transfer

    Hold out regions, seasons, sensors, and extreme events.

  5. 5. Validate in the field

    Sample sites using a plan that supports defensible population estimates.

Key idea

Maps can create consequences far from the model interface

A land-cover map may influence subsidies, enforcement, insurance, conservation, or disaster aid; errors are not distributed evenly when imagery quality and ground validation differ by region.

The clearest instance is a map that became a reference number for deforestation policy. “In this study, Earth observation satellite data were used to map global forest loss (2.3 million square kilometers) and gain (0.8 million square kilometers) from 2000 to 2012 at a spatial resolution of 30 meters”, opens the abstract of a 2013 paper in Science by Hansen and colleagues. The same work put tropical forest loss on a trend rising by 2,101 km2 per year. Every one of those quantities is the output of a classifier over Landsat pixels, aggregated across the planet, and it carries whatever error the classifier carries. Published as global totals, it reads instead like a measurement of the Earth.

Publish uncertainty, coverage gaps, update dates, and intended use. Give people on the ground a way to correct the map and to contest it.

A geospatial prediction can become an administrative fact unless its uncertainty remains visible.

Example

Practice: design a flood mapping service

The service combines radar and optical imagery after storms and must support emergency teams within hours. One documented activation shows the shape of the deliverable. On 29 October 2024 the Community of Valencia flooded. NASA Earth Observatory reports that “On October 29, 2024, more than 300 millimeters (12 inches) of rain fell in parts of the province”, and that “In the town of Chiva, nearly 500 millimeters (20 inches) fell in 8 hours”. The Copernicus Emergency Management Service was activated as EMSR773. Its delineation for the situation on 31 October 2024 at 18:02 UTC was reported days later: “The map reveals that more than 53,000 hectares were affected by the floods, with more than 190,000 people and 3,200 km of roads potentially affected”, Copernicus wrote on 3 November 2024. Landsat 8's OLI imaged the flooding on 30 October, with data from the U.S. Geological Survey. Design the service to produce that: an activation code, a situation timestamp, a mapped area, a population and a length of road, each of which somebody will act on.

  • Define the map unit, latency target, and minimum detectable flooded area. EMSR773 answered in hectares, people and kilometres of road, pinned to a situation time of 31 October 2024 at 18:02 UTC.
  • Create a split that holds out entire river basins and future storms. The collapse from R2 = 0.53 to R2 = 0.14 in the biomass study is exactly what a random split conceals.
  • Specify registration, cloud, and sensor-quality checks, including the Processing Baseline 04.00 offset for any Sentinel-2 series crossing 25 January 2022 and the ETM+ gaps in any Landsat 7 scene after 31 May 2003.
  • Measure pixel overlap, missed settlements, area bias, and time-to-map, and report confidence intervals on the area estimate as Olofsson and colleagues require.
  • Design a correction workflow for local responders with newer evidence.

Key takeaways