Skip to content
Tutorials Center
My learning

Topics

Lesson 36 of 46 ยท Detection, Segmentation and Video

Satellite and Remote Sensing Imagery

Learn remote sensing basics for computer vision: four kinds of resolution, spectral bands, NDVI and NDWI indices, land-cover clustering and change detection, with NumPy code.

A satellite or aircraft image is not just a big photograph. Sensors record many wavelength bands, including some the eye cannot see; the pixels are tied to places on Earth; and the pictures are taken again and again, so time becomes a dimension too. That changes how computer vision is applied: instead of asking "what object is this?" you ask "how much forest is left?", "where did the flood go?" or "which field needs water?".

This lesson introduces the vocabulary of remote sensing, a few of its most useful pixel computations (spectral indices), and the standard analysis tasks (land-cover classification and change detection), all as small NumPy exercises. Basic array handling comes from Images as Arrays and clustering from Clustering and Region Segmentation.

Four kinds of resolution

Remote sensing talks about four different resolutions, and a sensor is always a compromise between them.

Resolution Question it answers Examples
Spatial How big is one pixel on the ground? 30 m (Landsat), 10 m (Sentinel-2), 0.3 m (commercial)
Spectral How many bands, how narrow? 3 (RGB), 4 to 13 (multispectral), hundreds (hyperspectral)
Radiometric How many brightness levels per band? 8 bit = 256 levels, 12 bit = 4096, 16 bit = 65,536
Temporal How often is the same place imaged? Every 16 days (Landsat), every 5 days (Sentinel-2), daily (small-satellite fleets)

Finer spatial detail usually means a narrower strip and more data, fewer bands or longer revisit time; the trade-off is physical (light-gathering, bandwidth, orbit).

Python Radiometric resolution: bits per pixel

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

In this dark, low-contrast patch a 4-bit sensor collapses all 400 pixels to a single value, 6 bits keep two, and even 8 bits keep only six distinct levels, so the smooth terrain turns into steps. This is why remote-sensing sensors record 11 to 16 bits per band even though a monitor shows only 8: subtle differences in water, vegetation and shadow need the extra precision.

Spectral bands and composites

A multispectral sensor stores each wavelength range as its own grayscale image, a band. A typical set (Sentinel-2, simplified): blue, green, red, near-infrared (NIR) and short-wave infrared (SWIR). A pixel is then a short vector of numbers, its spectral signature. Different surfaces reflect differently:

  • Healthy vegetation absorbs red light (chlorophyll) but reflects NIR strongly.
  • Water reflects little in NIR and SWIR (very dark in those bands).
  • Bare soil and built-up areas reflect fairly evenly, rising slowly with wavelength.
  • Snow and cloud are bright in visible bands.

You display three bands at a time as red, green and blue. A true-colour composite uses (red, green, blue). A false-colour composite uses, for example, (NIR, red, green), which makes healthy vegetation bright red and water black, so it jumps out.

Spectral indices

An index combines two or more bands per pixel into one number that highlights a property. The classic NDVI (Normalized Difference Vegetation Index):

NDVI = (NIR - Red) / (NIR + Red)

ranging from -1 to +1: values near or below 0 mean water, snow or bare surfaces, small positive values (0.1 to 0.3) mean sparse vegetation or soil and values above 0.5 mean dense, healthy vegetation. Dividing by the sum cancels much of the effect of lighting differences, so the index is more stable than raw bands. A water index, NDWI = (Green - NIR) / (Green + NIR), is positive over water. Other indices cover burned area (NBR), built-up area (NDBI) and snow (NDSI).

Python NDVI and NDWI on a tiny scene

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

Forest and crops have high NDVI, soil and urban sit near zero, and water is negative. NDWI is positive only for water. Two simple threshold rules turn two bands of raw numbers into land-cover masks. Always check thresholds against ground truth: they differ by sensor, season and region.

Land-cover classification

Assigning every pixel to a class (water, forest, crops, urban...) is land-cover classification.

  • Unsupervised: cluster the spectral vectors (k-means) and name the clusters afterwards.
  • Supervised: label some training areas, then train a classifier on their pixel vectors (maximum likelihood, random forest, SVM) or a CNN or U-Net that also uses neighbouring pixels (see Segmentation Metrics).

The k-means version is a few lines and works surprisingly well on well-separated spectra.

Python Unsupervised land-cover clustering with k-means

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The clusters land on the true class centres, so a quick unsupervised map is possible with no labelled data. Real scenes are harder: mixed pixels (half forest, half field), shadows, clouds, seasons and sensor differences, which is why trained deep models and good labels matter.

Change detection

Because satellites revisit, you can compare two dates. The simplest change detection is image differencing: subtract the earlier image from the later one (after making sure they are registered, pixel for pixel on the same ground), and threshold the difference. A more targeted version differences an index: a large drop in NDVI marks deforestation or crop loss; a new region with positive NDWI marks flooding.

Python Detecting deforestation by differencing NDVI

Press Run (or Ctrl + Enter) to see the result here.

Program input ยท what you'd type while the program runs

Type the answers before you press Run, one per line. Each time the program reads a line (with input()), it gets the next line from this box.

The noise stays within a few hundredths, far from the -0.3 threshold, so only the real clearing is flagged. In practice, differences are also caused by seasons, clouds, haze, sun angle and misregistration, so analysts compare the same time of year, mask clouds, and apply atmospheric correction (converting the sensor's top-of-atmosphere values to surface reflectance).

Preprocessing and working with huge images

  • Georeferencing and orthorectification. Each pixel needs coordinates (a coordinate reference system and an affine transform). Terrain and the viewing angle distort images, and orthorectification removes it. Formats such as GeoTIFF store the coordinates with the pixels.
  • Radiometric and atmospheric correction. Convert digital numbers to radiance or reflectance, correct for haze and sun angle.
  • Cloud and shadow masking. Remove or flag pixels you cannot trust.
  • Tiling. A scene can be 10,000 by 10,000 pixels or more: cut it into overlapping tiles (for example 256 by 256), run the model per tile and stitch the predictions, blending the overlaps to avoid seams.
  • Pan-sharpening. Fuse a sharp panchromatic band with coarser colour bands to get a sharp colour image.
  • Class imbalance and scarce labels. Buildings or ships cover a tiny fraction of the pixels; use appropriate losses (see Segmentation Metrics) and pretrained backbones.

Applications

Agriculture (crop health, yield forecasts, irrigation), forestry and deforestation monitoring, urban planning and building footprint extraction, disaster response (floods, fires, earthquakes), water and coastline monitoring, mining and oil-and-gas inspection, climate science (glaciers, sea ice), defence and maritime surveillance. See Pitfalls, Ethics and Deployment for the privacy and misuse concerns of high-resolution imagery.

Check your understanding

Quick check ยท 5 questions

of 5 answered You scored / 5 ยท perfect!

  1. 1. What does spectral resolution describe?

    Spatial resolution is pixel size on the ground, radiometric is bit depth, temporal is revisit time, and spectral is the band structure.
  2. 2. Why does NDVI highlight healthy vegetation?

    Chlorophyll absorbs red light while leaf structure reflects NIR, so the normalised difference is high for dense, healthy vegetation.
  3. 3. What is a false-colour composite?

    Mapping NIR to red, for example, makes vegetation appear bright red so it is easy to see.
  4. 4. What must be true of two images before you subtract them for change detection?

    Misregistration creates false changes along every edge, so alignment (and similar radiometry) is essential.
  5. 5. Why are large satellite scenes processed in tiles?

    Tiling (with overlap, stitched afterwards) fits the data to GPU memory and the model's input size.

Summary

  • Remote sensing images have four resolutions (spatial, spectral, radiometric, temporal), and every sensor trades them off.
  • Multispectral bands let you separate surfaces by their spectral signatures; composites and indices such as NDVI and NDWI turn bands into meaningful maps.
  • Land-cover classification can be unsupervised (clustering) or supervised (classifiers, CNNs).
  • Change detection differences registered images or indices from two dates, after correcting for clouds and atmosphere.
  • Georeferencing, atmospheric correction, cloud masking and tiling are standard preprocessing steps for large scenes.

Next: Evaluating Vision Models.