p. 1 Lecture: Remote Sensing Sommersemester - Burckhardt
Transcription
p. 1 Lecture: Remote Sensing Sommersemester - Burckhardt
Lecture: Remote Sensing Sommersemester Prof. Dr. Christoph Kleinn Institut für Waldinventur und Waldwachstum Arbeitsbereich Fernerkundung und Waldinventur Sources of error in remotely sensed data Source: Lunetta, Congalton, Fenstermaker, Jensen, McGwire and Tinney 1981. Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Accuracy Assessment Slide No 1 Even in high resultion imagery it is not always clear where to define the boundary between forest and non-forest: It is a question of a clear definition. Four delineation options for forest boundary from aerial photographs: Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Accuracy Assessment Source: Slide No 2 Kenneweg. 2002. Quantifying classification accuracy Simple approach: - We select a sample of points in the image which we visit in the field (or interpret in another reference map that is supposed to be giving the “true” land cover): Ground control points. - Then we compare for each point the true classes with our classification result in the image. - The result can be presented in a so-called confusion matrix, which allows further analysis and quantification. Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Mod. 1 Accuracy Assessment Slide No 3 p. 1 Lecture: Remote Sensing Sommersemester Prof. Dr. Christoph Kleinn Institut für Waldinventur und Waldwachstum Arbeitsbereich Fernerkundung und Waldinventur Accuracy assessment Three basic components of an accuracy assessment are (after Stehman and Czaplewski 1999): 1. 2. 3. The sampling design to select the reference sample. The design how to measure truth on the sampling units selected. The analysis procedure to apply once the data have been collected. Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Accuracy Assessment Slide No 4 Guiding criteria for design desicions for accuracy assessment (from Stehman 1999) 1. 2. 3. 4. 5. 6. Probability sampling protocols. Simple to implement and analyse. Result in low variability for estimates requiring the highest accuracy levels. Allow for reliable variance estimation. Result in a well-distributed sample. Cost effective. Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Accuracy Assessment Slide No 5 Quantifying classification accuracy Per class: Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Mod. 1 Accuracy Assessment Slide No 6 Source: Congalton and Green 1999. p. 2 Lecture: Remote Sensing Sommersemester Prof. Dr. Christoph Kleinn Institut für Waldinventur und Waldwachstum Arbeitsbereich Fernerkundung und Waldinventur “Improving“ the accuracy measures - The more classes there are, the more are the confusion possibilities. - Separating only forest and non-forest will usually yield much higher accuracy values than the distinction of 10 further forest classes! - If we include large areas of easily identifyable features (the sea, for example), the overall accuracy is also considerably improved (because there are more observations in the diagonal). → These features must be taken into account when evaluating the accuracy assessment. Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Accuracy Assessment Slide No 7 Calculating statistics from an error matrix In statistics this type of matrix is called “contingency table”, giving the joint frequency distribution of two variables, and allowing for the test of numerous statistical hypothesis. For accuracy assessment the kappa statistic is frequently calculated, characterizing the degree of matching between reference data set and classication. In fact, it is a statistic that compares two matrices: here, it describes the difference between the agreement found in the actual matrix and the chance agreement given the same marginal frequencies: κˆ = Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen p0 − pc 1 − pc Accuracy Assessment Slide No 8 The Kappa statistic: A measure of agreement κˆ = p0 − pc 1 − pc ∑ pii = with p0 = sum of relative frequency in the diagonal of the actual error matrix, and ∑ pc = pi + p+ j = relative frequency of a random allocation of observations to the cells = chance agreement. The notation “i+” and “+j” is for the relative marginal frequencies. Weights wij may be assigned to each po and pc terms, if the accuracy of some classes is more important than for others. However, weights would be subjective! Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Mod. 1 Accuracy Assessment Slide No 9 p. 3 Lecture: Remote Sensing Sommersemester Prof. Dr. Christoph Kleinn Institut für Waldinventur und Waldwachstum Arbeitsbereich Fernerkundung und Waldinventur The Kappa statistic: Interpretation κˆ = p0 − pc 1 − pc κˆ = 0 indicates that obtained agreement equals chance agreement. κˆ > 0 indicates that obtained agreement is greater than chance agreement. κˆ < 0 indicates that obtained agreement is smaller than chance agreement. κˆ = 1 is perfect agreement (theoretical situation for p0=1). For kappa, a variance, and statistical significance can be given. Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Accuracy Assessment Slide No 10 Example calculation κˆ = p0 − pc 1 − pc p0 = 45 + 91 + 55 + 55 = 0.7321 336 pc = (85*55) (110*110) (75*72) (69*96) + + + 336 336 336 336 = 0.2551 336 κˆ = 0.7321 − 0.2551 = 0.64 1 − 0.2551 69 Source: Congalton and Green 1999. Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Accuracy Assessment Slide No 11 How to select “ground truth” data This is a sampling exercise! If we want to do further analysis with the ground truth data, i.e. making inferences that go beyond a mere description of what we found in those pixels only, then: principles of statistical sampling must be employed. In particular: the calculation of kappa and its variance must take place with the estimators that correspond to the sampling design applied! It has to be probability sampling – and not subjective sampling. - Stratification is frequently sensible, according to, for example - cover classes, or - assumed classification reliability. - Would it be a good idea to use the training sites (that had been observed for classification, anyway) also for accuracy assessment? Lecture Introduction to Remote Sensing Institut für Waldinventur und Waldwachstum, Georg-August-Universität Göttingen Mod. 1 Accuracy Assessment Slide No 12 p. 4