ALI assignment – see amended instructions Webpage – see also article and website outlines TIF 1 2 3 4 5 6 7 8 9 10
Classification review Unsupervised classification: clustering into classes identification of classes Supervised classification: training areas to ‘train’ the classification, check the statistics of the classes created check resulting coverage for errors
Classification display Classes are numbered 1 to 'n', so the channel 'DNs' are low. Colours are assigned arbitrarily (can be changed by editing the PCT)
Unsupervised – how it works Algorithm starts with statistical seed points Assigns each pixel to the closest seed Calculates group mean Re-assigns pixels to the closest group mean Re-calculates group mean Iterates (10?) until relatively little change and fixes groupings
Iterations in unsupervised classification Final step .. Assigning names to clusters (and merge some)
Unsupervised classification Input bands selected – minimum 3 or 4; more acceptable
Post-classification steps Checking the display Merging and adding classes Sieving (and filtering) Accuracy assessment Conversion of results to vectors – GIS layers Applying results to other areas ? Advanced classification options
Merging and adding classes Classes can be merged if they overlap spatially or are not distinguishable spectrally. Merging Unsupervised – clusters are not really separate features Supervised – when its not possible to separate two types Splitting / adding Unsupervised: one cluster covers too much area – run again with more clusters Can generate many clusters, and then group merge Supervised: Class covers too much area – create new training class or delete training areas Areas unclassed – create new training class
Sieving (and filtering) Classification ALWAYS produces a 'salt and pepper' effect with isolated pixels This is a result of a. the local variations in DNs and b. using ‘per-pixel’ classifiers
Mt. Edziza – classification and sieve (150 hectares minimum) - recognises connectivity of adjacent pixels in the same class - special classes e.g. wetlands can be specified and preserved - removes small sub-areas; does not ‘blur’ edges like filtering
Accuracy assessment This requires knowing what is reality at some pixels (ground truthing), and how they were classified. This generates a ‘confusion matrix’ The diagonal represents pixels correctly classified An off diagonal column element = an error of omission An off diagonal row element = error of commission http://www.gisdevelopment.net/application/nrm/overview/mma09_Mustapha.htm
Supervised classification methods a. Minimum distance b Supervised classification methods a. Minimum distance b. Parallelepiped c. Maximum likelihood
Calculation of mapping accuracies http://rst.gsfc.nasa.gov/Sect13/Sect13_3.html
Measuring accuracy The overall yardstick of 85% accuracy is held up as a (rarely achieved) ideal. User's accuracy: the % of a type of pixels that are correctly classified Producer's accuracy: the % within a classified class that are really that type Kappa coefficient provides a more unbiased estimate of overall agreement κ Interpretation < 0 No agreement 0.0 — 0.20 Slight agreement 0.21 — 0.40 Fair agreement 0.41 — 0.60 Moderate agreement 0.61 — 0.80 Substantial agreement 0.81 — 1.00 Almost perfect agreement
Vector conversion After classes are finalised, the adjacent same class pixels can be transformed into vector polygons, using polygon growing or raster to vector options (see feature extraction and lab 7)
Applying classification results to other areas (or other dates) spectral signatures are not usually transferable due to: environmental, local, atmospheric, and topographic variables - also temporal variables - time of year and advance of the seasons
Input channels for classification Input channels for classification .. not just bands how many and which ones? Landsat TM bands Channels created from Bands Ratios Indices Greenness DEMs Use of masks
Advanced classification algorithms These incorporate other interpretation elements beyond DN alone (‘colour/tone’): - Texture (example - Ruiz et al, 2004) PCI - TEX - Contextual classifiers PCI - CONTEXT context table - fuzzy classifiers PCI - FUZCLUS Object oriented programs (eCognition) Project example Neural Networks PCI: NNCLASS
Classification summary There are many articles on classification approaches: Input channel combinations Best algorithms New approaches There is no single best solution http://www.youtube.com/watch?v=ydbbd-4oEds