Map of the Great Divide Basin, Wyoming, created using a neural network and used to find likely fossil beds
See: http://earthobservatory.nasa.gov/IOTD/view.php?id=79412&src=eoa-iotd
Using training data to classify digital imagery
How does supervised classification work?
• Associate areas on the image with informational classes from the field (training data!)
• Generate statistics to describe the spectral characteristics of each informational class in terms of the satellite bands and/or enhancements
• Assign unknown pixels to classes based on similarity to the statistical description of the training class.
Supervised vs. Unsupervised Classifications
Unsupervised
Choose Bands, enhancements, etc.
Supervised
Collect training data for your map area
Cluster pixels into spectral classes
Label clusters corresponding to informational classes
Evaluate result
Choose bands, enhancements, etc.
Assign pixels to most similar informational class
Evaluate result
Landsat image near Riverton, WY
A
B
C
A = sagebrush
B = water
C = agriculture
D = riparian vegetation
D
• Selection of training data is the most important part of a supervised classification!
• Training data must:
– Represent all of the classes that you want to map
– Represent the spectral variability within classes
• (Can split informational classes for classification purposes)
– Be carefully selected based on field work and examination of the image
– Be modified iteratively if necessary to improve your classification
Training Site Selection Resulting Classification
• There are many techniques for assigning pixels to training (informational) classes. Common methods include:
• Parallelpiped
• Minimum Distance
• Maximum Likelihood
• Determine the range of DNs for each class in each band
• Use these DN ranges to define multidimensional “boxes” in feature space
• If a pixel falls within a box, it is assumed to belong to that class
• If a pixel falls outside of all boxes, it is not classified
Parallelpiped “boxes” in 3D feature space
Advantages/Disadvantages of Parallelpiped
Classifier
• Does NOT assign every pixel to a class. Only the pixels that fall within ranges.
• Good for helping decide if you need additional classes (if there are unclassified pixels)
• Good for helping decide if you need additional predictor variables (spectral or ancillary)
• Problems when class ranges overlap—must develop rules to deal with overlap areas, or refine the training data.
• Assigns pixels to the class they are closest to in feature space in terms of Euclidean distance.
• Calculate the average DN for each class across all bands
(= the class centroid)
• Calculate the Euclidean distance from each pixel to each centroid in feature space
• Assign each pixel to the class with the closest centroid
Use Pythagorean theorem to calculate distance
(in terms of DNs) to each centroid. Assign unknown pixel to closest centroid.
Class 1 centroid
Unknown pixel Class 2 centroid
Class 3 centroid
Band X
Advantages/Disadvantages of Minimum
Distance Classifier
• Classifies every pixel in the image (regardless of probability that it is really in a class)
• Does not explicitly consider the variability
(variance) within classes
• Works well for some images and not as well for others
• Assigns unknown pixels to the class that it has the highest probability of belonging to
– Based on how many standard deviations the pixel is from the class centroid
• Should use with normally distributed data
(bell-shaped histograms) but we are often permissive about deviation from this
Advantages/Disadvantages of Maximum
Likelihood Classifier
• Classifies every pixel in the image
• Recognizes that some classes have lots of spectral variability and are more likely to include pixels that are “far” from the class centroid
• Image data are not always normally distributed
• Often, but not always, a better choice than minimum distance classifiers
• Supervised classification uses knowledge of the locations of informational classes to group pixels
• Requires close attention to development of training data
• Typically results in better maps than unsupervised classifications IF you have good training data.
• Requires more work (time/money) than unsupervised classifications
• Fuzzy classifiers
• Decision trees
• Classification and Regression Trees
(CART)
• Object-based image classification
• Many others – neural networks, expert systems, etc.
• Some of these are covered in Advanced
RS (BOT/GEOG 4211/5211)
• Recognize that in the real world, distinctions between classes are often not distinct
• Assigns the same place on the ground to all classes (with a membership probability)
• Similar to a dichotomous key that you might use to key out plants or animals
• Can combine many types of data to create classes (e.g., spectral data, elevation data, soil maps, etc.)
• Can be built “by hand” or using statistical techniques
• Includes CARTs that build trees based on statistics
• What is it that makes one feature of interest different from another?
• Can you capture that difference accurately with spatially distributed data?
• How can you best exploit the differences between features statistically or otherwise?
• Remember that we must often go beyond just using spectral data!