Midterm/Final Exam Reviewer

• Image classification and Analysis
• Bands of a single image


  Used to identify and separate spectral signatures of landscape features
• Ordination and other statistical techniques


  Used to "cluster" pixels of similar spectral signatures in a theoretical space
• Maximum likelihood classifier


  Most often used
• Image classification


  1. Assign pixels to categories

  2. Apply categories to image to create classified image

  3. Resulting classified image can be used and interpreted as a map
• The resulting classified image will have errors! Accuracy assessment is critical. Maps created by image classification should report an estimate of accuracy.
• Tasseled Cap Transformation


  • Transforms image bands into new bands that represent physical properties of the land surface
• Spectral signatures


  Unique reflectance patterns of different land cover types
• Supervised classification


  Classifier which requires a training sample for each class, then classifies based on how close each pixel is to the training samples
• Unsupervised classification


  Classifier that groups pixels into clusters based on their spectral similarity, without using training samples
• Aim of image classification is to assign all pixels in the image to particular classes or themes
• Types of classes


  • Information classes: categories of interest (water, forest)
  • Spectral classes: group of uniform pixels (brightness)
• Objective of image classification is to match the spectral classes to the information classes
• Classification procedures


  • Supervised classification
  • Unsupervised classification
• Digital image classification


  Process of assigning pixels to classes, usually treating each pixel as an individual unit composed of values in several spectral bands
• Informational classes


  Categories of interest (e.g. water, forest)
• Spectral classes


  Groups of uniform pixels (based on brightness)
• Spectral subclasses arise from variations in illumination, species composition, and density within an informational class
• Supervised classification


  1. Collect empirical data from training areas on spectral response patterns

  2. Compare each unknown pixel to spectral patterns and assign to most similar category

  3. Present results as tables, digital data files
• Training samples


  Pixels with known identity used to train the classifier
• Training samples are used to define the relationship between classes and feature vectors (spectral values)
• Guidelines for selecting training areas include: homogeneity, large uniform areas, easy to locate, and minimum size
• Feature space


  Multidimensional space where each pixel is represented by a vector of its spectral values
• Pixels belonging to the same class form clusters in the feature space
• Collecting class statistics


  1. Calculate means and standard deviations of training samples in the feature space

  2. Use to define boundaries between classes
• Overlap between class clusters in the feature space leads to classification errors
• Supervised classification requires training samples, while unsupervised classification groups pixels into clusters based on spectral similarity without training samples
• Classes


  • A
  • B
  • CB
• Overlap between the classes

  Clan B
• UNEP-ITC RS/GIS for Monitoring and Assessment of Iraqi Marshland

  1. 10 Feb 2005
• Distances and Clusters in Feature Space

  • band y (units of 5 DN)
  • Max y
  • (0,0)
  • band x (units of 5 DN)
  • Min y
  • Euclidian distance
  • ++
  • *
  • (0,0)
  • Min x
  • Max x
  • Cluster
• Supervised Classification


  1. Analyst selects training areas where he/she knows what is on the ground and then digitizes a polygon within that area

  2. Computer creates mean spectral signatures
• Known areas


  • Conifer
  • Water
  • Deciduous
• The computer then creates mean spectral signatures
• The result is a land cover map
• Unsupervised Classification


  • Classifier which does not compare pixels to be classified with training data
  • Examines a large number of unknown data vectors and divides them into classes based on properties inherent to the data themselves
• UNEP Unsupervised Approach


  • Based on spectral groupings
  • Considers only spectral distance measures
  • Minimum user interaction
  • Requires interpretation after classification
• Unsupervised Classification Process


  1. Questions: number of classes, number of bands, spectral distance or radius in spectral space, spectral space distance parameters when merging clusters

  2. Spectral plot of the whole image

  3. Initiate the class centers

  4. Start separating (e.g. merging)

  5. Iteration stage

  6. Define the statistics of the classes

  7. Finalizing the decision boundary

  8. Creating the classification map
• Unsupervised Classification Input Parameters

  • Number of clusters
  • Size of cluster
  • Distance between the clusters
  • Cluster elimination value (convergence threshold)
• Unsupervised Classification Output


  • Distance or divergence measure between clusters
  • Cluster mean vector plots
  • Cluster histogram / feature space plots
  • Cluster variance-covariance matrices