C153918 Contract - Laserfiche WebLink

Home Browse Search

. . . . . ( information classes) is then refined through appropriate merger, deletion and/or augmentation. The resulting composite spectral sub-class statistics are then used as the basis for a maximum-likelihood classification of the entire image area to be classified. The classification results from the various sub-classes are then collapsed into the original target information classes. Clearly, guided clustering involves elements of both supervised and unsupervised classification, but avoids many of the problems associated with using either individually. For example, initial training areas that are not homogeneous are "automatically" separated into multiple spectral sub-classes by the techniques (by virtue of the clustering operation). In this manner, sub-classes that are not initially apparent to the analyst can be readily discriminated. Similarly, spurious pixels (such as inclusions of non-target cover types, or edges between types) can often be detected and edited from the data (Lillesand, 1996) The training sample locations acquired from the water commissioner's will be extracted from the imagery for each crop type or irrigation status (information class) from the appropriate PCA or Tasseled Cap image. The training sample polygon will be converted to GRID and buffered internally to avoid field boundary edge pixels. The pixels identified for each information class will be processed using the ISODA T A unsupervised classifier. Minimum of 15 sub-classes (signatures) will be generated for each class. These sub-classes will be reviewed in signature editor for separability. It is expected that signatures within an information class will be spectrally very similar and would have extensive overlap as plotted in feature space. Where sub-class signatures are separable, it indicates a different crop type or irrigation status than the information class being studied. These will be eliminated from the sub-class signature set. For review purposes sub-class signatures will be merged into their information class for comparison in Feature Space (2 Standard Deviations) and Signature Separability. If two signatures, which represent different categories, are statistically similar, one will be deleted or changed to the appropriate category. All of the signatures will be compared to one another before the first classification is performed. Thresholds will be determined for each information class. Classification thresholds will be determined by examining training sample signatures in spectral space (ERDAS Feature Space). The Greenness - Wetness and PCA components 1 and 2 scatter plots will be linked to the imagery so that the spectral and geographic locations of the data can be reviewed simultaneously. When an acceptable signature set has been created, it is applied to the satellite imagery to create a classification map. The maximum likelihood algorithm will be used for assigning each pixel in the image to one of the land cover categories. A threshold image will also be generated for the classification image. The first classification produced always requires evaluation and reclassification. The process continues using visual interpretation and local knowledge until an acceptable product is produced. At this point it is often helpful for the image analyst to draw on expertise from the field staff in making recommendations for changes or confirming accuracy. This need can be fulfilled by sending image plots back and forth via ftp or by working side by side for a few days. Supervised classification is generally more labor intensive than the unsupervised approach, but can also produce a better product when adequate ground reference data are available -- this is especially true for guided clustering. Deliverables to State: A task memorandum detailing the steps involved in each subtask Page 20