Lecture 19-20

Image classification; Unsupervised classification

What is image classification, and what is the significance

• Process of assigning raw data to information category by assigning multispectral pixels (each with a digital number) to an information category

• Important tool for examination of digital images, image analysis, pattern recognition

• Grouping similar pixels into classes to solve problems by enhancing information

What is a classified image

• Image has been processed to put each pixel into a category

• Produces vegetation maps, LULC maps, or maps with other grouped features

Spectral pattern recognition

• When the decision rules used for classification are derived based on the spectral radiances observed in the image data

• It is often done pixel-by-pixel

Spatial pattern recognition

• When the decision rules are derived based on the geometric shapes, sizes, patterns, and spatial relationships of pixels with neighbors in the image data

• This is more complicated than spectral pattern recognition

What is a classifier

loosely refers to a computer program that is used to implement a specific procedure for image classification, specifically refers to a classification strategy of image

What are spectral classes

• Groups of pixels that have similar DN or brightness values in several channels

• Can have subclasses (i.e young, mature, old growth forest)

What are information classes

• Categories of interest to the users of the data. They are objects such as different land use and land cover, different forests, etc.

• Conveys information to planners, managers, administrators, and scientists

• Internal variability is minimized, and distinctiveness is maximized

Why can information classes have multiple spectral classes

Because of variation caused by different environmental, climate, and illumination conditions, and human activities

What allowed digital image classification to become effective?

With the advent of digital multispectral remote sensing data and allows the application of statistical logistics

Disadvantages of digital image classification

Lacks spatial dimension

Pixels might not meet needs of client

7 steps of image classification; short description for each

• Selection of data (RS, availability, size of study area, ancillary data)

• Selection of classification system (spatial, temporal, radiometric, and spectral resolution)

• Sampling design (how many? Simple random sampling, systematic, or stratified random sampling

• Data preprocessing (geometric, radiometric, atmospheric correction)

• Feature extraction and selection (suitable variables, principle component analysis, graphic analysis and spectral feature (spectral signature, vegetation indices)

• Classification methods (unsupervised/supervised/hybrid classification)

• Accuracy assessment (Visual and spatial comparisons, Error matrix)

4 errors to consider with the error matrix portion of an accuracy assessment for image classification

• Error of commission

• Error of omission

• Overall accuracy

• Kappa coefficient

Some applications for image classification

o Land use/land cover change

o Natural resources management

o Urban studies

o Forestry

o Hydrology

o Agriculture

Supervised vs unsupervised classification

• Supervised: We impose our perceptions on spectral data

• Unsupervised: spectral data imposes constraints on our interpretation

Describe unsupervised classification

• Identification of natural groups, or structures, within multispectral data and groupings of spectrally homogenous classes

• Transformation of spectral classes into thematic info

• no extensive prior knowledge of study area needed

Describe distance measures for unsupervised classification

• Distances to other pixels can always be used to define group membership

• Many distance calculation techniques for determining similarities for the many pixels and groups

  • Calculate spectral distance

Unsupervised classification process

• extract spectrally similar clusters

• pixels from clusters should be classified via spectral classes

• group the clusters into meaningful classes for information

What is ISODATA

• Iterative Self-Organizing Data Analysis Technique

• Special case of minimum distance clustering

• Parameters you must enter include:

  • N - the maximum number of clusters that you want

  • T - a convergence threshold and

  • M - the maximum number of iterations to be performed. Enter the number of maximum times that the ISODATA utility should recluster the data

What is the Convergence Threshold

• The maximum percentage of pixels whose cluster assignments can go unchanged between iterations

• i.e a convergence threshold of .95 means when 95% or more of pixels stay in the same cluster between iterations, the utility should stop processing

ISODATA Procedure

• Arbitrary cluster means are established

• The image is classified using a minimum distance classifier

• A new mean for each cluster is calculated

• The image is classified again using the new cluster means

• Another new mean for each cluster is calculated

• The image is classified again...

• After each iteration, the algorithm calculates the percentage

of pixels that remained in the same cluster between iterations

• When this percentage exceeds T (convergence threshold), the

program stops

Pros and cons of ISODATA

• Non-parametric (no assumptions made and data doesn’t need to be normally distributed)

• helpful at finding true clusters within data if enough iterations allowed

• Signatures from ISODATA easily incorporated and manipulated with supervised spectral signatures

• Slowest of the clustering procedures

Advantages (5) and Disadvantages (3) of unsupervised classification

Advantages

• No extensive prior knowledge of the study area is needed

• Opportunities for human error is minimized

• Unique classes are recognized as distinct units

• Takes maximum advantage of spectral variability in an image

• Reliance on ‘natural groupings’. What if both forest land and agricultural land are in the same group?

Disadvantages

• Grouping of classes that are of non-interest to the analyst

• Spectral values changes over time

• The maximally-separable clusters in spectral space may not match our perception of the important classes on the landscape