Living Landscapes: Remote Sensing &
Climate Science

  Click image to expand

In supervised classification, you show the computer how to group pixels into categories. You, the analyst, decide what landcover categories exist in your image. After you draw training samples for each category, the computer lumps all pixels into your groups.

Some trial and error is often involved (1) creating categories your software can tell apart from reflectance, and (2) drawing training areas that represent each category.

With unsupervised classification, you, the analyst, decide how many groups exist in your image; your software then lumps similar pixels into that many classes. Your computer takes the first turn at categorizing pixels, based on statistical grouping of reflectance.

Your computer does not know what the groups are, it just knows the pixels are similar (groups are labeled 1, 2, 3 rather than water, forest, pasture, for example). The image analyst is then responsible for identifying the classes and lumping classes into useful groups (e.g. combining deep water & shallow water). Trial and error may lead you to begin with more or less groups; starting high and then lumping is often useful.

  Click image to expand

Thematic accuracy measures how well pixels are classified into the correct group (theme). The true “reference” values are collected by ground truthing—checking classifications in the field. Higher resolution imagery may also be useful when ground truthing is not possible.

Overall accuracy shows how many pixels (%) were classified correctly. In the above table, correctly classified pixels were 21+31+22 = 74. Divided by 95 total pixels = 77.9% overall accuracy.

Errors of Omission are reference pixels that were left out of the correct class—go down reference columns to calculate. Water reference cells were incorrectly classified 5+7 = 12 times out of 33 reference cells. Omission error for water is 12/33 = 36%

Producer’s Accuracy is from the map maker’s perspective: how often are real features on the ground correctly shown on the map. It is a compliment of Omission: producers accuracy = 100% - Error of Omission.
Water = 100% - 36% = 64%

Errors of Commission look across classification rows for incorrect classifiations. Water incorrectly classified pixels are 6+0 = 6 out of 27 classified pixels. Commission error for water is 6/27 = 22%

User’s Accuracy is from the map users perspective: how often the classification on the map will actually be present on the ground. It is a compliment of Commission: user’s accuracy = 100% - Error of Commission. Water = 100% - 22% = 78%

  Click image to expand

  Click image to expand

Band combinations can visually highlight features in an image, showing different wavelength bands as Red, Green, & Blue (RGB) on your display. The following indices & ratios are ways to quantify that spectral reflectance.

Simple band ratios divide one band by another (e.g. NIR / red). Band ratios help remove effects from illumination angle, shadows, and terrain—revealing information about actual earth surfaces.

Band ratios can also be displayed in RGB composites (in place of regular band reflectance values). Sultan et al (1987) provided groundbreaking insights on remote sensing for geology and used a mineral-revealing composite of Landsat TM band ratios: band-5/1, band-5/7, and band-5/4 x 3/4, composite-displayed as Red, Green, and Blue.

Index values are calculated from spectral reflectance for each pixel; they allow us to take biophysical measurements of earth surfaces, with relative consistency, compared to visual analysis.

Indices have been developed for a very wide range of applications, including vegetation, atmosphere (temperature, ozone, NOx, fire smoke), water (temperature, sediment, algae), snow & ice, wildland fire burn severity, etc.

All sorts of human activities are affected by vegetation, and vice versa. The ability to measure vegetation type and abundance is useful for predicting crop yields, tracking tropical deforestation, describing wildlife habitat, estimating biomass, leaf area index (LAI), and carbon cycling.

Normalized Difference Vegetation Index (NDVI) was developed soon after the launch of Landsat 1 in 1972. Rouse et al (1973) found a ratio of NIR & red reflectance, used to monitor spring green-up in the Great Plains, normalized much of the difference in sun illumination between lower and more northern latitudes. NDVI is perhaps the most-used index in remote sensing.

Enhanced Vegetation Index (EVI) was developed to reduce errors from atmospheric aerosols and provide better measurements in densely vegetated areas (e.g. rainforest) where NDVI tends to saturate.

NDVI can range from -1 to +1. Water heavily absorbs near infrared (NIR) light, resulting in often negative NDVI values. NDVI for other surfaces:

• Barren areas of rock, sand, or snow = very low values of NDVI (0.1 and below)
  
• Shrub and grassland = moderate NDVI (0.2 to 0.3)
  
• Temperate and tropical rainforests = high NDVI (0.6 to 0.8)
  *Threshold values from: John Weier and David Herring, NASA. August 30, 2000. https://earthobservatory.nasa.gov/features/MeasuringVegetation
  
• Now calculate NDVI for both surfaces with the equation:
  
  
  
  
• Your computer will perform this calculation for every pixel, when you create NDVI in GIS
  
• Write down your NDVI for both willow and concrete.
• Then check your work by clicking here for the answers.
  

  Click image to expand

• Identify imagery to use. Late summer images were used to show the minimum snow & ice extent. Landsat was chosen because (1) decades of imagery, (2) detects wavelengths needed for snow & ice index, (3) spatial resolution of 30 meters is small enough to detect most ice fields.
  
• Masking removed image pixels below summer ice elevation. This removed error from water appearing as ice, and reduced other “noise”.
  
• Calculate NDSII. Index values range from -1 to +1 (or scaled from -100 to +100), but what is the threshold for ice vs. non-ice (e.g. bare ground)?
  
• Threshold value for snow & ice. High-resolution 1 meter aerial imagery was used to identify ice vs other land cover. True “ground truthing” not possible in grizzly bear conservation area. Compared to Landsat NDSII from same date. Scaled NDSII > 30 considered snow/ice.
  
• Extract snow & ice pixels (NDSII > 30) for each image over time, calculate area for each year. Do any trends emerge in ice extent over time?

The tutorials at the end of this course will show you how use ArcGIS Image Analyst software. GIS software is powerful, adaptive, and in demand for careers ranging from natural resource, historic preservation & archeology, urban & emergency planning, etc. Typical GIS analysis requires a fair amount of training and expensive software licensing.

Cloud-based computer processing is opening the door for geospatial and image analysis, performed with relative ease without the need for installing software or even data on personal computers. Of course, varying levels of background knowledge (which you are developing) are required for varying confidence of results. However the usefulness of cloud-based image processing for educational purposes is, at the very least, a small revolution. Time required to download & radiometrically correct a time series of imagery is drastically reduced.

Next you will get a quick introduction to remote sensing image analysis with ClimateEngine.org, a web app that provides access to Google’s EarthEngine (which otherwise requires knowledge of the scripting languages Python or Java).