Classification Overview
Introduction
Computer aided classification can help provide data for landscape-level planning and assessment. Image classification can also help with tracking and modeling changes in the environment over time. There are several approaches to image classification and this module will focus on computer based classification techniques. The main types of image classification are:
- Manual Classification
- Pixel Based Classification
- Unsupervised Classification
- Supervised Classification
- Object or Feature Based Classification
Manual Classification
Some people consider manual classification the most rudimentary form of image classification. Manual or visual classification refers to the interpretation and classification of imagery by the human eye. Before computers, this was the only way classification could be done. In modern geospatial analysis, this is now achieved through "heads-up" digitizing, which was covered in previous modules. Manual digitizing can be useful and appropriate in many scenarios. It tends to work well for small contiguous areas but it may not be ideal for large, non-contiguous areas of study. Digitizing can also be time consuming and repetitive, but when done well can produce reliable and consistent results.
Manual classification or digitizing might work well to accurately digitize the glacier in the image on the left, but it would be time consuming and not very efficient to digitize all of the areas with recent timber harvesting in the image on the right.Pixel Based Classification
Pixel Based classification methods use the pixel values in an image to assign every image pixel to a class. Each pixel is assigned to a class based on its spectral characteristic, this is known as Spectral Pattern Recognition. The objective of pixel based classification is to assign all pixels in the image to particular classes or themes (e.g. water, coniferous forest, deciduous forest, agriculture). The number and type of classes are decided by the analyst. The two types of pixel based classification are:
- Unsupervised Classification
- Supervised Classification
Feature Space
Pixel based classification methods use what's known as feature space to classify the pixels. Feature space is essentially a scatter plot of the spectral values for two bands for all the pixels in the image. When we look at a satellite image the pixels are organized in a grid so that objects such as lakes, rivers, and forests appear in a way similar to the way they would in a photograph or on a map. We can also view image data using a feature space reference system. In a feature space plot, the axes represent the range of possible values for a specific feature. When used in an image classification algorithm the feature space can have several dimensions, one dimension for each band in the image. In other words if you are processing a 7 band image, the classification algorithm will be working with a 7-dimensional feature space. Obviously it is difficult to view more than three dimensions at a time, so for practical reasons all possible combinations of two bands are plotted separately. For example, a Landsat image with 7 bands would have 21 different feature space plots (all possible combinations of two bands). Most software packages allow you to use feature space to preview the classes you have created to see how the classes are distributed.
The above figure shows a feature space plot (left) for the NIR and red bands of a Landsat image (right). The areas shown in red represent the distribution of water pixels. The pixels that represent water generally have low red values and NIR values close to zero. This makes sense as we known water reflects minimal in the NIR and SWIR.Mixed Pixel Problem
A pixel is the smallest spatial area for classification and most classification algorithms assume that a pixel covers a homogeneous land cover type. In reality this is not often the case. For example, Landsat 8 has a spatial resolution of 30m and within that 900 m2 ground area there could be a variety of different land features. The digital numbers or pixel values of these pixels represent the average of several spectral classes within the area that it covers on the ground. These pixels are known as Mixed Pixels. Mixed pixels are common in data with relatively coarse spatial resolution and along the edges of features. When a pixel whose ground area is composed of two or more areas that differ greatly in terms of spectral brightness, then the "averaged" spectral value that represents the pixel may not accurately represent either of the categories present. This can lead to the frequent misclassification of mixed pixels. These pixels are especially common in urban areas where there may be a variety of features (paved, lawns, trees) within small areas.
Various techniques have been developed in attempts to ‘unmix’ the spectral information. Some of the common techniques are Sub-pixel classification and Fuzzy classification. Sub-pixel classification attempts to identify the proportions of the different land cover types in a pixel. For example in the Landsat pixel highlighted in the image below, a sub-pixel classification could indicate 70% forest and 30% bare ground within the pixel. Fuzzy classification attempts to address the mixed pixel problem by using the concept that a single pixel may belong in more than one category or class. Fuzzy classification allows pixels to belong to more than one class with a degree of membership in each class.
The two images above cover the same area, but as you can see in the NAIP image within the red square these is both forest and a road. In the Landsat image this isn't apparent and the pixel color and spectral data are a "mix" of these two cover types.