Unformatted text preview: Classification Classification Unsupervised Methods Unsupervised Classification Unsupervised Classification
In an unsupervised classification, the identities of landcover In an unsupervised classification, the identities of landcover types to be specified as classes within aascene are not generally types to be specified as classes within scene are not generally known a priori because ground reference information is known a priori because ground reference information is lacking or surface features within the scene are not well lacking or surface features within the scene are not well defined. The computer is required to group pixels with similar defined. The computer is required to group pixels with similar spectral characteristics into unique clusters according to some spectral characteristics into unique clusters according to some statistically determined criteria. The analyst then relabels and statistically determined criteria. The analyst then relabels and combines the spectral clusters into information classes. combines the spectral clusters into information classes. Unsupervised Classification Unsupervised Classification
Unsupervised classification (commonly referred to as clustering) Unsupervised classification (commonly referred to as clustering) is an effective method of partitioning remote sensor image data is an effective method of partitioning remote sensor image data in multispectral feature space and extracting landcover in multispectral feature space and extracting landcover information. Compared to supervised classification, information. Compared to supervised classification, unsupervised classification normally requires only aaminimal unsupervised classification normally requires only minimal amount of initial input from the analyst. This is because amount of initial input from the analyst. This is because clustering does not normally require training data. clustering does not normally require training data. Unsupervised Classification Unsupervised Classification
Unsupervised classification is the process where numerical Unsupervised classification is the process where numerical operations are performed that search for natural groupings of operations are performed that search for natural groupings of the spectral properties of pixels, as examined in multispectral the spectral properties of pixels, as examined in multispectral feature space. The clustering process results in aaclassification feature space. The clustering process results in classification map consisting of m spectral classes. The analyst then attempts map consisting of m spectral classes. The analyst then attempts a posteriori (after the fact) to assign or transform the spectral a posteriori (after the fact) to assign or transform the spectral classes into thematic information classes of interest (e.g., classes into thematic information classes of interest (e.g., forest, agriculture). This may be difficult. Some spectral forest, agriculture). This may be difficult. Some spectral clusters may be meaningless because they represent mixed clusters may be meaningless because they represent mixed classes of Earth surface materials. The analyst must understand classes of Earth surface materials. The analyst must understand the spectral characteristics of the terrain well enough to be able the spectral characteristics of the terrain well enough to be able to label certain clusters as specific information classes. to label certain clusters as specific information classes. Kmeans clustering (also called cmeans clustering) 1. Select K points in the multispectral space as candidate clustering centers Let these points be Although m can be arbitrarily selected, it is suggested that they be selected evenly in the multispectral space. For example, they can be selected along the diagonal axis going through the origin of the multispectral space. 2. Assign each pixel x in=the image to the closest cluster center m
and return to step 2 to continue. Kmeans clustering (continued)
3. Generate a new set of cluster centers based on the processed result in 2. N is 4. If, number of pixels currently assigned to cluster i, and the summation of x the is the additive sum of their values. In other words, the cluster center is assigned to the mean (average value) of the (a small tolerance), the procedure isis the iteration number pixels assigned to it. n terminated. Otherwise let of step 2. (a small (a 4. If: small tolerance), the procedure is terminated. Otherwise let tolerance), the procedure is terminated. Otherwise let = and return to step 2 to continue. ISODATA Clustering ISODATA Clustering
The Iterative SelfOrganizing Data Analysis Technique ((ISODATA) The Iterative SelfOrganizing Data Analysis Technique ISODATA) represents aacomprehensive set of heuristic (rule of thumb) procedures represents comprehensive set of heuristic (rule of thumb) procedures that have been incorporated into an iterative classification algorithm. that have been incorporated into an iterative classification algorithm. Many of the steps incorporated into the algorithm are aaresult oof Many of the steps incorporated into the algorithm are result f experience gained through experimentation. experience gained through experimentation. The ISODATA algorithm is aamodification of the kkmeansclustering The ISODATA algorithm is modification of the means clustering algorithm, which includes a) merging clusters if their separation algorithm, which includes a) merging clusters if their separation distance in multispectral feature space is below aauserspecified distance in multispectral feature space is below userspecified threshold and b) rules for splitting aasingle cluster into two cclusters. threshold and b) rules for splitting single cluster into two lusters. ISODATA Clustering ISODATA Clustering
ISODATA is iterative because it makes aa large number of passes ISODATA is iterative because it makes large number of passes through the remote sensing dataset until specified results are obtained, through the remote sensing dataset until specified results are obtained, instead of just two passes. instead of just two passes. ISODATA does not allocate its initial mean vectors based on the ISODATA does not allocate its initial mean vectors based on the analysis of pixels in the first line of data the way the twopass algorithm analysis of pixels in the first line of data the way the twopass algorithm does. Rather, an initial arbitrary assignment of all Cmax clusters takes does. Rather, an initial arbitrary assignment of all Cmax clusters takes place along an nndimensional vector that runs between very specific place along an dimensional vector that runs between very specific points in feature space. The region in feature space is defined using the points in feature space. The region in feature space is defined using the mean, , ,and standard deviation, k, ,of each band in the analysis. This mean, kk and standard deviation, k of each band in the analysis. This method of automatically seeding the original Cmax vectors makes sure method of automatically seeding the original Cmax vectors makes sure that the first few lines of data do not bias the creation of clusters. that the first few lines of data do not bias the creation of clusters. ISODATA Clustering ISODATA Clustering
ISODATA is selforganizing because it requires relatively little human ISODATA is selforganizing because it requires relatively little human input. A sophisticated ISODATA algorithm normally requires the input. A sophisticated ISODATA algorithm normally requires the analyst to specify the following criteria: analyst to specify the following criteria: Cmax::the maximum number of clusters to be identified by the Cmax the maximum number of clusters to be identified by the algorithm (e.g., 20 clusters). However, it is not uncommon for ffewerto algorithm (e.g., 20 clusters). However, it is not uncommon for ewer to be found in the final classification map after splitting and merging take be found in the final classification map after splitting and merging take place. place. T: the maximum percentage of pixels whose class values are allowed T: the maximum percentage of pixels whose class values are allowed to be unchanged between iterations. When this number is reached, the to be unchanged between iterations. When this number is reached, the ISODATA algorithm terminates. Some datasets may never reach the ISODATA algorithm terminates. Some datasets may never reach the desired percentage unchanged. If this happens, it is necessary to desired percentage unchanged. If this happens, it is necessary to interrupt processing and edit the parameter. interrupt processing and edit the parameter. ISODATA Clustering ISODATA Clustering
M: the maximum number of times ISODATA is to classify pixels and M: the maximum number of times ISODATA is to classify pixels and recalculate cluster mean vectors. The ISODATA algorithm terminates recalculate cluster mean vectors. The ISODATA algorithm terminates when this number is reached. when this number is reached. Minimum members in aacluster (%): If aacluster contains less than the Minimum members in cluster (%): If cluster contains less than the minimum percentage of members, it is deleted and the members are minimum percentage of members, it is deleted and the members are assigned to an alternative cluster. This also affects whether aaclass is assigned to an alternative cluster. This also affects whether class is going to be split (see maximum standard deviation). The default going to be split (see maximum standard deviation). The default minimum percentage of members is often set to 0.01. minimum percentage of members is often set to 0.01. ISODATA Clustering ISODATA Clustering
Maximum standard deviation (max): When the standard deviation for aa Maximum standard deviation (max): When the standard deviation for cluster exceeds the specified maximum standard deviation and the cluster exceeds the specified maximum standard deviation and the number of members in the class is greater than twice the specified number of members in the class is greater than twice the specified minimum members in aaclass, the cluster is split into two clusters. The minimum members in class, the cluster is split into two clusters. The mean vectors for the two new clusters are the old class centers 1s. mean vectors for the two new clusters are the old class centers 1s. Maximum standard deviation values between 4.5 and 77are typical. Maximum standard deviation values between 4.5 and are typical. Split separation value: If this value is changed from 0.0, it takes the Split separation value: If this value is changed from 0.0, it takes the place of the standard deviation in determining the locations of the new place of the standard deviation in determining the locations of the new mean vectors plus and minus the split separation value. mean vectors plus and minus the split separation value. Minimum distance between cluster means (C): Clusters with aaweighted Minimum distance between cluster means (C): Clusters with weighted distance less than this value are merged. A default of 3.0 is often used. distance less than this value are merged. A default of 3.0 is often used. ISODATA Clustering ISODATA Clustering
Phase 1: ISODATA Cluster Building using many passes Phase 1: ISODATA Cluster Building using many passes through the dataset. through the dataset. Phase 2: Assignment of pixels to one of the Cmax clusters Phase 2: Assignment of pixels to one of the Cmax clusters using minimum distance to means classification logic. using minimum distance to means classification logic. a) ISODATA initial distribution of five a) ISODATA initial distribution of five hypothetical mean vectors using 1 hypothetical mean vectors using 1 standard deviations in both bands as standard deviations in both bands as beginning and ending points. b) In the first beginning and ending points. b) In the first iteration, each candidate pixel is compared iteration, each candidate pixel is compared to each cluster mean and assigned to the to each cluster mean and assigned to the cluster whose mean is closest in Euclidean cluster whose mean is closest in Euclidean distance. c) During the second iteration, aa distance. c) During the second iteration, new mean is calculated for each cluster new mean is calculated for each cluster based on the actual spectral locations of the based on the actual spectral locations of the pixels assigned to each cluster, instead of pixels assigned to each cluster, instead of the initial arbitrary calculation. This the initial arbitrary calculation. This involves analysis of several parameters to involves analysis of several parameters to merge or split clusters. After the new cluster merge or split clusters. After the new cluster mean vectors are selected, every pixel in the mean vectors are selected, every pixel in the scene is assigned to one of the new clusters. scene is assigned to one of the new clusters. d) This splitmergeassign process d) This splitmergeassign process continues until there is little change in class continues until there is little change in class assignment between iterations (the TT assignment between iterations (the threshold is reached) or the maximum threshold is reached) or the maximum number of iterations is reached (M). number of iterations is reached (M). a) Distribution of 20 ISODATA a) Distribution of 20 ISODATA mean vectors after just one mean vectors after just one iteration using Landsat TM iteration using Landsat TM band 33and 44data of Charleston, band and data of Charleston, SC. Notice that the initial mean SC. Notice that the initial mean vectors are distributed along aa vectors are distributed along diagonal in twodimensional diagonal in twodimensional feature space according to the feature space according to the 2 standard deviation logic 2 standard deviation logic discussed. b) Distribution of 20 discussed. b) Distribution of 20 ISODATA mean vectors after ISODATA mean vectors after 20 iterations. The bulk of the 20 iterations. The bulk of the important feature space (the important feature space (the gray background) is partitioned gray background) is partitioned rather well after just 20 rather well after just 20 iterations. iterations. Classification Classification Based on Based on ISODATA ISODATA Clustering Clustering ISODATA ISODATA Clustering Clustering Logic Logic Unsupervised Cluster Busting Unsupervised Cluster Busting It is common when performing unsupervised classification using the It is common when performing unsupervised classification using the chain algorithm or ISODATA to generate nnclusters (e.g., 100) and have chain algorithm or ISODATA to generate clusters (e.g., 100) and have no confidence in labeling qqof them to an appropriate information class no confidence in labeling of them to an appropriate information class (let us say 30 in this example). This is because (1) the terrain within the (let us say 30 in this example). This is because (1) the terrain within the IFOV of the sensor system contained at least two types of terrain, IFOV of the sensor system contained at least two types of terrain, causing the pixel to exhibit spectral characteristics unlike either of the causing the pixel to exhibit spectral characteristics unlike either of the two terrain components, or (2) the distribution of the mean vectors two terrain components, or (2) the distribution of the mean vectors generated during the unsupervised classification process was not good generated during the unsupervised classification process was not good enough to partition certain important portions of feature space. When enough to partition certain important portions of feature space. When this occurs, it may be possible to perform cluster busting if in fact there this occurs, it may be possible to perform cluster busting if in fact there is still some unextracted information of value in the dataset. is still some unextracted information of value in the dataset. Unsupervised Cluster Busting Unsupervised Cluster Busting
First, all the pixels associated with the qqclusters (30 in aahypothetical First, all the pixels associated with the clusters (30 in hypothetical example) that are difficult to label (e.g., mixed clusters 13, 222,45, 92, example) that are difficult to label (e.g., mixed clusters 13, 2, 45, 92, etc.) are all recoded to aavalue of 11and aabinary mask file is created. A etc.) are all recoded to value of and binary mask file is created. A mask program is then run using (1) the binary mask file and (2) the mask program is then run using (1) the binary mask file and (2) the original remote sensor data file. The output of the mask program is aa original remote sensor data file. The output of the mask program is new multiband image file consisting of only the pixels that could not be new multiband image file consisting of only the pixels that could not be adequately labeled during the initial unsupervised classification. The adequately labeled during the initial unsupervised classification. The analyst then performs aanew unsupervised classification on this file, analyst then performs new unsupervised classification on this file, perhaps requesting an additional 25 clusters. The analyst displays these perhaps requesting an additional 25 clusters. The analyst displays these clusters using standard techniques and keeps as many of these new clusters using standard techniques and keeps as many of these new clusters as possible (e.g., 15). Usually, there are still some cclustersthat clusters as possible (e.g., 15). Usually, there are still some lusters that contain mixed pixels, but the proportion definitely goes down. The contain mixed pixels, but the proportion definitely goes down. The analyst may want to iterate the process one more time to see if an analyst may want to iterate the process one more time to see if an additional unsupervised classification breaks out additional clusters. additional unsupervised classification breaks out additional clusters. Perhaps five good clusters are extracted during the final iteration. Perhaps five good clusters are extracted during the final iteration. Unsupervised Cluster Busting Unsupervised Cluster Busting
In this hypothetical example, the final cluster map would be composed In this hypothetical example, the final cluster map would be composed of :: of 70 good clusters from the initial classification, 70 good clusters from the initial classification, 15 good clusters from the first clusterbusting pass (recoded as values 15 good clusters from the first clusterbusting pass (recoded as values 71 to 85), and 71 to 85), and 55good clusters from the second clusterbusting pass (recoded as good clusters from the second clusterbusting pass (recoded as values 86 to 90). values 86 to 90). The final cluster map file may be put together using aasimple GIS The final cluster map file may be put together using simple GIS maximum dominate function. The final cluster map is then recoded to maximum dominate function. The final cluster map is then recoded to create the final classification map. create the final classification map. Clustering Using the Chain Method Clustering Using the Chain Method
The Chain Method clustering algorithm operates in aatwopass mode (i.e., it The Chain Method clustering algorithm operates in twopass mode (i.e., it passes through the multispectral dataset two times). passes through the multispectral dataset two times). Pass #1: The program reads through the dataset and sequentially builds Pass #1: The program reads through the dataset and sequentially builds clusters (groups of points in spectral space). A mean vector is then clusters (groups of points in spectral space). A mean vector is then associated with each cluster. associated with each cluster. Pass #2: A minimum distance to means classification algorithm is applied Pass #2: A minimum distance to means classification algorithm is applied to the whole dataset on aapixelbypixel basis whereby each pixel is to the whole dataset on pixelbypixel basis whereby each pixel is assigned to one of the mean vectors created in pass 11.The first pass, assigned to one of the mean vectors created in pass . The first pass, therefore, automatically creates the cluster signatures (class mean vectors) therefore, automatically creates the cluster signatures (class mean vectors) to be used by the minimum distance to means classifier. to be used by the minimum distance to means classifier. Clustering Using the Chain Method Clustering Using the Chain Method
Phase 1: Cluster Building Phase 1: Cluster Building R, aaradius distance in spectral space used to determine when R, radius distance in spectral space used to determine when aanew cluster should be formed (e.g., when raw remote sensor new cluster should be formed (e.g., when raw remote sensor data are used, it might be set at 15 brightness value units). data are used, it might be set at 15 brightness value units). C, aaspectral space distance parameter used when merging C, spectral space distance parameter used when merging clusters (e.g., 30 units ))when N is reached. clusters (e.g., 30 units when N is reached. N, the number of pixels to be evaluated between each major N, the number of pixels to be evaluated between each major merging of the clusters (e.g., 2000 pixels). merging of the clusters (e.g., 2000 pixels). Cmax,,the maximum number of clusters to be identified by the Cmax the maximum number of clusters to be identified by the clustering algorithm (e.g., 20 clusters). clustering algorithm (e.g., 20 clusters). Phase 2: Assignment of pixels to one of the Cmax clusters Phase 2: Assignment of pixels to one of the Cmax clusters using minimum distance to means classification logic using minimum distance to means classification logic Clustering Using the Chain Method Clustering Using the Chain Method
Starting at the origin of the multispectral dataset (i.e., line 1, Starting at the origin of the multispectral dataset (i.e., line 1, column 1), pixels are evaluated sequentially from left to right column 1), pixels are evaluated sequentially from left to right as if in aachain. After one line is processed, the next line of as if in chain. After one line is processed, the next line of data is considered. We will analyze the clustering of only the data is considered. We will analyze the clustering of only the first three pixels in aahypothetical image and label them pixels first three pixels in hypothetical image and label them pixels 1, 2, and 3. 1, 2, and 3. Original brightness values of pixels 1, 2, and Original brightness values of pixels 1, 2, and 3 as measured in Bands 4 and 5 of the 3 as measured in Bands 4 and 5 of the hypothetical remote sensed data. hypothetical remote sensed data. The distance ((D)in 22dimensionalspectral space The distance D) in dimensional spectral space between pixel 11(cluster 1) and pixel 22(potential cluster between pixel (cluster 1) and pixel (potential cluster 2) in the first iteration is computed and tested against the 2) in the first iteration is computed and tested against the value of R=15, the minimum acceptable radius. In this value of R=15, the minimum acceptable radius. In this case, D does not exceed R. Therefore, we merge clusters case, D does not exceed R. Therefore, we merge clusters 11and 22as shown in the next illustration. and as shown in the next illustration. Pixels 11and 22now represent cluster #1. Note that the location Pixels and now represent cluster #1. Note that the location of cluster 11has migrated from 10,10 to 15,15 after the first of cluster has migrated from 10,10 to 15,15 after the first iteration. Now, pixel 33distance (D=15.81) is computed to see iteration. Now, pixel distance (D=15.81) is computed to see if it is greater than the minimum threshold, R=15. It is, so if it is greater than the minimum threshold, R=15. It is, so pixel location 33becomes cluster #2. This process continues pixel location becomes cluster #2. This process continues until all 20 clusters are identified. Then the 20 clusters are until all 20 clusters are identified. Then the 20 clusters are evaluated using aadistance measure, C (not shown), to merge evaluated using distance measure, C (not shown), to merge the clusters that are closest to one another. the clusters that are closest to one another. How clusters migrate during the several iterations of aa How clusters migrate during the several iterations of clustering algorithm. The final ending point represents clustering algorithm. The final ending point represents the mean vector that would be used in phase 22of the the mean vector that would be used in phase of the clustering process when the minimum distance clustering process when the minimum distance classification is performed. classification is performed. Note: As more points are added to aa cluster, the mean Note: As more points are added to cluster, the mean shifts less dramatically since the new computed mean is shifts less dramatically since the new computed mean is weighted by the number of pixels currently in aa cluster. weighted by the number of pixels currently in cluster. The ending point is the spectral location of the final The ending point is the spectral location of the final mean vector that is used as aa signature in the minimum mean vector that is used as signature in the minimum distance classifier applied in pass 2. distance classifier applied in pass 2. Some clustering algorithms allow the analyst to Some clustering algorithms allow the analyst to initially seed the mean vector for several of the initially seed the mean vector for several of the important classes. The seed data are usually obtained in important classes. The seed data are usually obtained in aasupervised fashion, as discussed previously. Others supervised fashion, as discussed previously. Others allow the analyst to use a priori information to direct the allow the analyst to use a priori information to direct the clustering process. clustering process. Pass 2: Assignment of Pixels to One of the Cmax Clusters Using Pass 2: Assignment of Pixels to One of the Cmax Clusters Using Minimum Distance Classification Logic Minimum Distance Classification Logic The final cluster mean data vectors are used in aa The final cluster mean data vectors are used in minimum distance to means classification algorithm to minimum distance to means classification algorithm to classify all the pixels in the image into one of the Cmax classify all the pixels in the image into one of the Cmax clusters. The analyst usually produces aacospectral plot clusters. The analyst usually produces cospectral plot display to document where the clusters reside in threedisplay to document where the clusters reside in threedimensional feature space. It is then necessary to dimensional feature space. It is then necessary to evaluate the location of the clusters in the image, label evaluate the location of the clusters in the image, label them if possible, and see if any should be combined. It is them if possible, and see if any should be combined. It is usually necessary to combine some clusters. This is usually necessary to combine some clusters. This is where an intimate knowledge of the terrain is critical. where an intimate knowledge of the terrain is critical. The mean vectors of the 20 clusters displayed using only The mean vectors of the 20 clusters displayed using only bands 22and 3. The mean vector values are summarized bands and 3. The mean vector values are summarized in Table 9911.Notice the substantial amount of overlap in Table 11. Notice the substantial amount of overlap among clusters 11through 55and 19. among clusters through and 19. The mean vectors of the 20 clusters displayed using only The mean vectors of the 20 clusters displayed using only bands 33and 44data. The mean vector values are bands and data. The mean vector values are summarized in Table 9911. summarized in Table 11. Grouping (labeling) of the original 20 spectral clusters Grouping (labeling) of the original 20 spectral clusters into information classes. The labeling was performed by into information classes. The labeling was performed by analyzing the mean vector locations in bands 33and 4. analyzing the mean vector locations in bands and 4. ...
View
Full
Document
This note was uploaded on 09/08/2010 for the course GEOG 182 at San Jose State.
 '10
 Pereira,Gary

Click to edit the document details