Types of spectral analysis
.. and their application to the polar environment
Last updated
.. and their application to the polar environment
Last updated
Here we present a selection of spectral analysis methods which have been employed successfully to support projects focussing on extracting environmental information from imagery of the polar regions.
This set of methods of not exhaustive, and is set out to illustrate some examples of spectral analysis as a starting point to investigate further.
There are a number of unsupervised classifications that divide up an image into its most distinctive component parts. These can sometimes be use with other clustering algorithms such as K-means or Iso clustering which when combined can reduce the dimensionality of an image (i.e. the number of component parts). In some of these algorithms the user can choose the number of output classes, in others the algorithms decides.
Supervised classifications rely on user input to give the algorithm information (signatures) on what pixels should represent what surface feature.
An example algorithm we can use within GIS software is Multi-variate Maximum Likelihood Classification.
A typical workflow would involve identifying targets you wish to classify, and then running the algorithm with a specific number of classes to segregate the image into the constituent surface types. The two images below illustrate these two steps.
Supervised classifications have been successfully applied to very high resolution imagery of Emperor Penguin colonies in Antarctica, to provide population estimates.
Spectral indices are widely used within remote sensing analysis. They work by exploiting the different responses of earth surface characteristics within different spectral bands.
For example, vegetation has a much higher reflectance in the near-infrared than the red band. This rapid change in reflectance in vegetation is known as the 'Red-Edge'.
Vegetation can therefore be classified by taking the normalised difference between the reflectance of the red band and the near-infrared band (NIR), as described in the equation below:
NDVI will range between -1 and 1. Higher values will indicate likely presence of vegetation.
This method has been employed on the Antarctic Peninsula to characterise likely areas of vegetation.
Other normalized difference indices include:
NDVI – Vegetation
NDWI – Water
NDSI – Snow
RNDSI- Soil
NDMI- Moisture
GDVI- Generalized Vegetation Index
The Sentinel Playground allows users to experiment with applying a range of spectral indices to Sentinel-2 data.
Target finding algorithms include spectral angle mapping and are commonly applied to hyperspectral imagery processing.
These use the shape of the spectral profile irrespective of total brightness. In order to work, you only need a small number of pixels. The algorithm will return a ratio of “likeness”.
These methods use a mixture of spectral, shape and other conjectural information.
Methods include:
Object based image analysis: a type of computer vision that aims to identify features within images based on shape. Algorthims include edge detection and segmentation for example
Model based classifiers, ie. Random Forest and Support Vector Machines
CNNs and deep learning: these usually need large training and validation datasets to work effectively.
