Digital Image ProcggSensing PeFW5LectuAccuracy Assessment aAccuracy Assessment aAccuracy assessment is a general tegeographical data that are assumedgeographical data that are assumedaccuracy of the classification proceUsually theassumedtrue data are dUsually,theassumed-true data are dIt is usually not practical to groundclassified image. cessing: A Remote gerspective5560ure 18and Change Detectionand Change Detectionerm for comparing the classification to d to be true, in order to determine thed to be true, in order to determine the ess. derived from ground truth dataderived from ground truth data.d truth or otherwise test every pixel of aError or Accuracy Assessment MThe error matrix simply compares thepoints in ac×cmatrix wherecis thpoints in a c ×c matrix, where c is th0).Matrixe reference points to the classified e number of classes (including classe number of classes (including classOverall accuracy: (# pixels correctly c= (911+343+176+27) / 1687 = 0.864 = Average (Class) Accuracy: The averag(sum of producer class accuracies) / num+ 58 7) / 4 0 754 75 3%+ 58.7) / 4 = 0.754 = 75.3%lassified) / (total # of pixels)86.4%ge of the individual class accuracy. mber of classes) = (97.7 + 75.1 + 69.8Producers accuracy: Producer accurcorrectly classified as hardwood) / (#correctly classified as hardwood) / (# hardwood) = 343/457 = 0.751 = 75.1%Oii EEldi ilOmission Error:Excluding a pixel tclass (i.e., omission error = 1 -producracy (hardwood) = (# of pixels ground reference pixels inground reference pixels in %hhldh b ilddihthat should have been included in the cers accuracy = 1 – 0.751)User’s Accuracy: (hardwood) = # of pi(total # of pixels classified as hardwoodCommission Error: Including a pixel iexcluded (i.e., commission error=1-coexcluded (i.e., commission error 1 coixels correctly classified as hardwood) / d) = 343/425 = 0.807 = 80.7%n a class when it should have been onsumer's accuracy=1–0.807).onsumer s accuracy 1 0.807).Overall accuracy: (# pixels correctly c= (911+343+176+27) / 1687 = 0.864 = Producers accuracy: Producer accuracclassified as hardwood) / (# ground refeclassified as hardwood) / (# ground refe0.751 = 75.1%lassified) / (total # of pixels)86.4%cy (hardwood) = (# of pixels correctly erence pixels in hardwood)=343/457=erence pixels in hardwood) 343/457Overall accuracy: (# pixels correctly c= (911+343+176+27) / 1687 = 0.864 = Producers accuracy: Producer accuracclassified as hardwood) / (# ground refeclassified as hardwood) / (# ground refe0.751 = 75.1%lassified) / (total # of pixels)86.4%cy (hardwood) = (# of pixels correctly erence pixels in hardwood)=343/457=erence pixels in hardwood) 343/457Change Detection Time PeriodSometimes the time period selected oidi h lChange Detection Time Periodmonitored is too short or too long to The analyst must be careful to identifhe analyst must be ca eful to identifperiod(s). This selection is dictated btransportation studies might require afew seconds or minutesfew seconds or minutes.Images obtained monthly or seasonathe greening of a continent.Careful selection of the change detecCareful selection of the change detecresource analysis funds are not wasteover which change is to be hif i ficapture the information of interest. fy the optimal change detection time fy the optimal change detection timeby the nature of the problem. Traffic a change detection period of just a ally might be sufficient to monitor ction time period can ensure thatction time period can ensure that ed.Select an Appropriate Land Cover/LIt is wise to use an established, standclassification system for change deteclassification system for change deteUse of standardized classification sysb d ith th t dibe compared with other studies.Also classification must be the same Cross walking is a common procedurLand Use Classification Schemedardized land-cover/land-use ctionction.stems allows change information to to compare various dates/timesreHard and Fuzzy Change Detection LMost change detection studies havegmultiple-date hard land-cover classiThltithtifhdTheresultisthecreation ofahardinformation about the change indforest, agriculture). Still the most wiHowever we now recognize that itifuzzychangesinthelandscapefuzzychangesinthelandscape.Logicebeen basedon the comparison ofpfications of remotely sensed data.hdt tiitifchangedetection map consisting ofdiscrete categories (e.g., change inidely employed.is ideal to capture both discrete andPerpixel or Objectoriented ChangeMost digital image change detectionPer-pixel or Object-oriented Change gggn and Date n+1classification mapscommonly referred to as per pixel chObject-oriented change detection invtwo or more scenes consisting of maimage objects (patches or segments)relatively homogeneous image objecsubjected to various change detectiosubjected to various change detectioDetectionn is based on processing Date Detectionpgs pixel by pixel. This is hange detection.volves the comparison of any relatively homogenous . The smaller number of cts in the two scenes are then n techniques.n techniques.Remote Sensing System ConsideratioSuccessful remote sensing change deid• remote sensor system considera• environmental characteristics. Failure to understand the impact of thdetection process can lead to inaccurIdeally, the remotely sensed data usedacquired by a remote sensor system tlil(dlkconstant: temporal, spatial (and look onsetection requires careful attention to: idations, and he various parameters on the change ate results.d to perform change detection is that holds the following resolutions l ) l d di iangle), spectral, and radiometric.SPTlRlTwo temporal resolutions should be System Parameter: Temporal Resolutdetection, if possible.First, use a sensor system that acquirFirst, use a sensor system that acquirtime of day. For example, Landsat TMfor most of the conterminous U. S. Teffects that can ca se anomalo s diffeffects that can cause anomalous diffof the remote sensor data.Second, acquire remote sensor data o2005, and Feb 1, 2007. Anniversary of seasonal Sun-angle and plantphenof seasonal Sunangle and plant phennegatively impact a change detectionheld constant during change tionres data at approximately thesameres data at approximately the same M data are acquired before 9:45 a.m. This eliminates diurnal Sun angle fferences in the reflectance propertiesfferences in the reflectance properties on anniversary dates, e.g., Feb 1, date imagery minimizes the influence nologicaldifferences that