Laserfiche WebLink
<br />. <br /> <br />. <br /> <br />. <br /> <br />'t <I' <br /> <br />Images: <br />Feature Space - Feature space images are a product of evaluating training samples for <br />image classification. For Irrigated Status and Crop Type classification there will be a <br />minimum of one Feature Space (FS) image for each Landsat image or Division (if the <br />Landsat images can be mosaiced for classification). Where only one date of imagery is <br />being used, the contractor will produce one FS image for the two band Principal <br />Components Analysis data and one for the Tasseled Cap data. The FS image will <br />show final signature ellipses used for the classification. <br />Threshold (or distance) - For each individual final classified image or for a mosaic (if an <br />area can be classified as a mosaic) a distance image file will be created to document <br />spectral distance from signature means. <br /> <br />Files: <br />Signature Separability - The signature separability files will be generated using two <br />layers per combination, the Transformed Divergence distance measure and the <br />CellArray output form with best average listing. <br />Signature File - The final signature set used for each image or mosaic classification will <br />be delivered in ERDAS .sig format. <br />Signature File - The initial signature set before merge, add and delete functions will be <br />delivered in ERDAS .sig format. <br />Histogram Files (if used) - Histogram files will be provided as .jpeg format. <br /> <br />Task 17 - Accuracy Assessment <br /> <br />When the classification has progressed to an acceptable state, accuracy assessment must be <br />performed. This process will be implemented for each division. Accuracy assessment involves <br />comparing samples that were collected in the field notes (ground truth) with samples taken from <br />the classification. Where the field category matches the classified category a correct <br />assignment is calculated. Accuracy figures are computed for each category and a table is <br />constructed so that the amount of confusion between any two categories is known. In addition <br />to computing overall accuracy the contractor will calculate the KHA T statistic. This figure is <br />used to determine what proportion of the sample points were correctly classified as a result of <br />random chance. After the initial classification the contractor will run a preliminary accuracy <br />assessment on the original ground truth polygons using a stratified random sampling. If this <br />assessment does not meet the accuracy goals the contractor will re-evaluate the signatures to <br />identify the problem signatures, set up a revised signature file and re-run the classification. If <br />this assessment meets or exceeds the accuracy goals, the contractor will proceed to classify <br />the entire scene for the final classification. The results will then be run through a second <br />accuracy assessment using the accuracy assessment points that were set aside earlier in the <br />process. <br /> <br />As specified by State, contractor will attain an overall accuracy of 95% for irrigated areas and <br />80% for crop type. For individual classes that are relatively small in size or that are not key to <br />the purpose of the study (the barren class for example), it is sometimes difficult to attain 95% <br />correspondence between the classification and ground reference data. The image will be <br />reclassified until an acceptable level of accuracy has been attained for every class, and the <br />overall accuracy goal is achieved. <br /> <br />Page 21 <br />