Remote Sensing, Aerial Imagery, and Change Detection

Remote sensing is the collection of information about Earth’s surface from a distance, using sensors on satellites, aircraft and drones that detect reflected or emitted energy; aerial imagery is photography from the air; and change detection is the comparison of images from different dates to find what has changed. In brief: sensors measure electromagnetic radiation in bands such as visible light, near-infrared and thermal infrared, and different surfaces reflect these differently, so vegetation, water, soil and buildings can be identified. Key properties of an image are its spatial, spectral, temporal and radiometric resolution. The normalized difference vegetation index (NDVI) is a widely used measure of plant health: NDVI = (NIR − Red) ÷ (NIR + Red).

Satellite data allow scientists to monitor forests, cities, crops, ice, oceans and disasters over time and across the world. This guide explains how remote sensing works, the types of sensors and resolutions, how images are classified and compared, compares major satellite systems in a table, works through NDVI and change calculations, outlines the history and corrects common mistakes.

How does remote sensing work?

The Sun emits electromagnetic energy, which passes through the atmosphere, strikes the surface and is reflected or absorbed. A passive sensor records this reflected sunlight (and the heat emitted by the Earth), so it needs daylight for visible and near-infrared bands. An active sensor sends out its own signal and records the return: radar (such as Sentinel-1) uses microwaves that pass through clouds and work day and night, and lidar uses laser pulses to measure heights very precisely, for example to map terrain under a forest canopy. The spectral signature of a surface is the pattern of how much it reflects in each part of the spectrum. Healthy green vegetation absorbs red light for photosynthesis and strongly reflects near-infrared radiation, water absorbs most infrared, bare soil reflects more in the red and shortwave infrared, and snow reflects strongly in the visible and weakly in the shortwave infrared. These differences allow classification. The atmosphere affects the signal, so images often need atmospheric correction, and sensors are carried on platforms such as satellites, aircraft and increasingly drones (unmanned aerial vehicles), which give very detailed images of small areas.

What are the four types of resolution?

• Spatial resolution: the size of the ground area covered by one pixel. A pixel of 30 m (Landsat) can show fields and large buildings, a pixel of 10 m (Sentinel-2) shows more detail, and commercial satellites reach under 0.5 m, where cars are visible. Coarse images (250 m to 1 km, as from MODIS) cover large areas quickly.
• Spectral resolution: the number and width of the wavelength bands. Multispectral sensors have a few bands (such as 4 to 13), and hyperspectral sensors have hundreds of narrow bands that allow detailed identification of materials.
• Temporal resolution: how often the sensor revisits the same place, from daily (MODIS, weather satellites) to every 16 days (a single Landsat satellite). Frequent coverage is vital for monitoring changes, but clouds can block the view for optical sensors.
• Radiometric resolution: the number of brightness levels a sensor can record, for example 8-bit (256 levels) or 12-bit (4,096 levels), which determines how well it can detect subtle differences.

There are trade-offs: a high spatial resolution generally means a smaller area covered and fewer visits, and the best choice depends on the task.

How are images interpreted and classified?

Analysts use visual interpretation (tone, color, texture, shape, size, pattern and shadow) and computer methods. In false-color composites, the near-infrared band is shown in red, so healthy vegetation looks bright red. Spectral indices combine bands: the NDVI ranges from −1 to +1, with values of about 0.6 to 0.9 for dense healthy vegetation, 0.2 to 0.5 for sparse vegetation or crops, near 0 for bare soil, and negative for water, snow and clouds. Other indices are used for water (NDWI), burned areas (NBR) and built-up areas. Image classification assigns each pixel to a class such as forest, water, urban or cropland. In unsupervised classification, the computer groups pixels with similar values, and the analyst names the clusters afterward. In supervised classification, the analyst provides training areas of known classes, and the algorithm assigns the other pixels. Today, machine learning and deep learning are increasingly used. The accuracy of a classification is checked against reference data (ground truth) in a confusion matrix, from which the overall accuracy is calculated.

How does change detection work?

Change detection compares images of the same area at two or more dates, to find changes in land cover, such as deforestation, urban growth, flooding, glacier retreat or crop condition. The images must be aligned precisely (co-registered) and use similar conditions, such as season, sun angle and atmosphere, so differences are due to real change. Common methods are image differencing (subtracting the pixel values of one date from the other, or the NDVI values, so that large differences stand out), post-classification comparison (classifying each image and comparing the maps, to produce a “from-to” change matrix), time-series analysis (looking at changes through many images, as with MODIS or Landsat archives) and radar methods that detect flooding or ground movement. Results are validated with field data or high-resolution images. Applications include Global Forest Watch (annual forest loss maps), monitoring of sea ice and glaciers, damage assessment after earthquakes and wildfires, and urban sprawl.

Satellite systems compared

Worked example: NDVI, pixels and accuracy

1. NDVI of vegetation: a pixel of dense forest has near-infrared reflectance of 0.50 and red reflectance of 0.08. NDVI = (0.50 − 0.08) ÷ (0.50 + 0.08) = 0.42 ÷ 0.58 = 0.72, a value for healthy vegetation.
2. NDVI after change: after clearing, the same pixel has NIR = 0.30 and red = 0.15. NDVI = (0.30 − 0.15) ÷ (0.30 + 0.15) = 0.15 ÷ 0.45 = 0.33. The change is 0.33 − 0.72 = −0.39, a large decrease that indicates a loss of vegetation.
3. Area of changed pixels: a Landsat pixel is 30 m × 30 m = 900 m². If 2,500 pixels show forest loss, the area is 2,500 × 900 = 2,250,000 m², which is 2.25 km² or 225 hectares. If the forest area at the first date was 20 km², the loss is 2.25 ÷ 20 = 11.25%.
4. Pixels per area: a square kilometer (1,000,000 m²) holds 1,000,000 ÷ 900 = about 1,111 pixels at 30 m resolution, and 1,000,000 ÷ 100 = 10,000 pixels at 10 m resolution, nine times as many.
5. Classification accuracy: 100 reference points are checked, and 85 of them are classified correctly. The overall accuracy is 85 ÷ 100 = 85%. If forest was correct in 28 of 32 reference forest points, the producer’s accuracy for forest is 28 ÷ 32 = 87.5%.
6. Revisit time and clouds: a satellite with a 5-day revisit offers 365 ÷ 5 = 73 images a year. If 40% are too cloudy, 0.60 × 73 = about 44 usable images.

How did remote sensing develop?

Aerial photography began in the 19th century with cameras on balloons and kites, and aircraft were used extensively in the First and Second World Wars, which advanced interpretation techniques. The launch of Sputnik (1957) and then weather satellites such as TIROS-1 (1960) opened the space age for observation. The first Earth Resources Technology Satellite, later renamed Landsat 1, was launched in 1972, and the Landsat program has provided more than 50 years of continuous images, which became free to the public in 2008. Radar satellites began in the 1970s and 1990s, and the Copernicus program of the European Union has launched Sentinel satellites since 2014 with free and open data. Commercial high-resolution satellites appeared from 1999, small satellites (“cubesats”) in the 2010s have created daily global coverage, and cloud computing, such as Google Earth Engine, has made it possible to analyze archives of images at planetary scale.

Which mistakes are common?

Mistake 1: Treating a satellite image as a photograph that shows everything

Sensors measure bands of energy, and the information depends on resolution, wavelength and processing. Many features are too small or too similar to distinguish.

Mistake 2: Comparing images from different seasons without care

Changes in vegetation or snow cover between seasons can look like land cover change. Use images from similar times of year and conditions.

Mistake 3: Believing that higher spatial resolution is always better

Fine images cover small areas, are costly and produce huge files. A coarser image with frequent coverage can be better for large-scale monitoring.

Mistake 4: Skipping validation

Classifications and change maps contain errors. Their accuracy should be checked against reference data before the results are used for decisions.