Vegetation Measurement and Monitoring

13.3 Image Classification

Learning Guide

Image classification assigns pixels or groups of pixels in a remote-sensing image to one of a set of classes based on spectral (i.e., color), texture, or context information in the image. There are two types of classes:

  • Information classes are the categories that have meaning relative to the intended application of the classification, or meaning to the end user. Examples of information classes are: rock, water, sagebrush steppe, grassland. These are the categories that we wish to identify and map from the image.
  • Spectral classes are groups of pixels that have similar image values. The feature space plot below illustrates the relationship between spectral classes and information classes.

The goal of image classification is to relate the information classes (i.e., what you really want to map) to distinct spectral classes (i.e., what you can reliably discriminate in the image).

A feature space plot shows the value of individual pixels in two or three image bands. Pixels that naurally group together in the feature space plot are considered spectral classes. Pixels in the feature space plot can be labeled by their information class, and this helps determine whether an information class can be reliably mapped. Image from Hsiao and Chang (2016), https://www.mdpi.com/2072-4292/8/9/705/htm

 

Types of Image Classification

There are many different techniques for classifying images, but they basically fall into one of two categories:

Unsupervised Classification

In unsupervised classification, an image processing program or algorithm first classifies the image into a number of spectral classes, and then meaning or interpretation is applied to those classes (i.e., the spectral classes are assigned information class names). For unsupervised classification, you must tell the classification algorithm how many classes you want and it creates that number of classes in a way that maximizes the difference between the spectral classes.

 

Stitt et al. (2006) used unsupervised classification to map leafy spurge presence and distribution in Theodore Roosevelt National Park, North Dakota, from EO-1 Advanced Land Imager satellite data. Image from https://www.jstor.org/stable/pdf/3900010.pdf

 

Unsupervised classification is appealing because it is a simple approach that does not require a lot of training data or assumptions. It will produce a map with the desired number of classes where each pixel is assigned to a class. However, these classes may or may not correspond well to your desired information classes. Unsupervised classification often results in too many classes, and classes need to be combined to create a meaningful map. In other cases, the classification may result in spectral classes that contain multiple information classes, necessitating further splitting of the unsupervised classes. For this reason, unsupervised classifications are generally used only for data exploration or as a preliminary step for more sophisticated classifications.

Supervised Classification

Supervised classification starts with verified locations of the information classes and the classification uses that information as “training data” to describe the spectral “signature” for each information class. The algorithm then assigns each pixel in the image to the class it most likely belongs to (i.e., the class that it most closely resembles digitally). In supervised classification, the training areas must be homogeneous, representative samples of the information classes, and selection of appropriate training areas is critical. Many of the new machine learning or artificial intelligence (AI) techniques being used in remote sensing are instances of supervised classification.

 

The USGS National Land Cover Dataset uses supervised classification to map land cover types across the United States using Landsat satellite imagery. Image from https://www.mrlc.gov

 

Supervised classifications can be very effective and accurate in classifying satellite images and can be applied at the individual pixel level or to groupings of adjacent pixels (i.e., image objects). However, for the process to work effectively, the person supervising the classification must have good knowledge (field data, aerial imagery, first-hand experience) of where and what the information classes are and be able to identify them in the imagery.