Lecture 10 Flashcards
What is supervised classification (definition)
A procedure for identifying spectrally similar areas on an image by identifying ‘training’ sites of known targets and then extrapolating those spectral signatures to other areas of unknown targets
Supervised classification needs training data - explain
based on locations where you know these locations are a particular landcover type - you know the spectral information* Fairly Homogeneous (the pixels don’t contain different landcover types - so spectral information all from same landcover type. ▪ Class Spectral Separability Indicates whether the training data from different classes overlaps. * Sufficiently large to capture the spectral variation of the land cover type - reflectance from grass for example varies throughout the image - different growth stages - some in shadow.
What are classification methods (supervised)
Apply difference ‘decision rules’ to classify the data
* Parallelepiped
* Minimum distance to means
* Maximum likelihood
What are the stages of supervised classification?
- User defined land cover classes
- Training site selection
- Generation of statistical parameters from the training sites
- Classification
- Accuracy assessment
How do you select training sites?
▪ Field visits ▪ High spatial resolution data (aerial imagery, Google Earth) ▪ Previous maps ▪ Investigator\expert knowledge ▪ Any/all of the above
What is the classification stage? (supervised)
▪ Need rule(s) to decide into which class we put a given pixel in
▪ Numerous mathematical approaches to spectral pattern recognition have been developed e.g. ▪ Examples include
▪ Parallelepiped classifier (BOX)
▪ Minimum distance to means (MDM)
▪ Maximum likelihood classifier (ML)
Describe the Parallelepiped classifier (BOX)
❑ Assign boundaries around the spread of a class in feature space i.e. take account of spectral variance
❑ All pixels in the image with values within the designated parallelepiped will be classified as that spectral class (water with water, urban with urban)
❑ Bounds (range of acceptable values in each band) are usually determined from training sets.
❑ The standard deviation (SD) in each band is determined. ▪ used to calculate the bounds of the cluster
There is some overlap of boxes – therefore can be misclassification.
Describe the Minimum distance to means (MDM)
like the unsupervised method - based on Euclidian distance. ▪ This is the same approach as the unsupervised classification clustering method
▪ Calculate of the mean spectral value for each training set in each band
▪ Put every unknown pixel into nearest class/cluster
▪ Compute the distance between the value of the unknown pixel and each of the category mean
▪ Define a limit beyond which a pixel remains unclassified
User can specify thresholds for SD, if a pixel is within 5 SD of mean – if it falls outside these, unclassified.
Describe Maximum likelihood classifier (ML
3.Maximum likelihood classifier (ML) Most common classification method For each pixel to be classified
❑ Assumes data in a class are (unimodal) Gaussian (normally) distributed
❑ The probability of classification is calculated for each class based upon the training data
❑ The pixel is classified as the class with the largest probability
❑ Theoretically the best classification
What are advantages of supervised classification
Advantages
▪ Analyst has control
▪ Processing is tied to specific areas of known identity
▪ Not faced with the problem of matching categories on the final map with field info
▪ Operator can detect errors in training data and often remedy them
Disadvantages of supervised classification
▪ Training classes based on field identification, not on spectral properties
▪ Training data selected by the analyst, may not be representative
▪ within class heterogeneity
▪ Unable to recognise and represent special or unique categories not represented in the training data
What are the goals of the accuracy assessment
*Assess how well the land cover map represents reality
*Understand how to interpret the usefulness of someone else’s classification Accuracy
What are the steps for the accuracy assessment?
1.Collect reference data: “ground truth”
*Determination of class types at specific locations
2. Compare reference data to classified map
*Does class type on classified map match the class type determined from reference data?
What are possible sources of ground truth
- High resolution satellite or airborne imagery
- Field visit with GPS – costly (time/money + impractical)
- GIS layers
How would you collect reference data
- Need to adequately sample the landscape
- Variety of spatial sampling schemes (e.g.) :
- Random * Stratified * Stratified random
- The greater the number of reference samples the better
- need to balance this with the cost (time/resources) required
then compare the reference data with the classified map - error matrix is produced.
