GE 113 REMOTE SENSING

Transcription

1 GE 113 REMOTE SENSING Topic 8. Image Classification and Accuracy Assessment Lecturer: Engr. Jojene R. Santillan Division of Geodetic Engineering College of Engineering and Information Technology Caraga State University

2 Outline Part 1. Image Classification Concepts What is image classification? Supervised vs. Unsupervised Classification The Different Stages of Supervised Image Classification Part 2. Different Types of Supervised and Unsupervised Image Classifiers Part 3. Image Classification Accuracy Assessment 2

3 Expected Outcomes The students would be able to: Discuss the principles of image classification Identify the different methods of image classification Identify the procedures involved in image classification and accuracy assessment Perform image classification and accuracy assessment using a remote sensing image analysis software Discuss the importance of yielding high accuracies in image classification 3

4 PART 1. IMAGE CLASSIFICATION CONCEPTS 4

5 Image Classification: 5

6 Image Classification: An operation that replace visual analysis of the image data with quantitative techniques for automating the identification of features in a scene Involves: Analysis of multispectral image data Application of statistically-based decision rules for determining the identity of each pixel in an image 6

7 What is the intent of image classification? To categorize all pixels in a digital image into one of several classes or themes Or to assign all pixels in the image to particular classes or themes (e.g. water, forest, barren land, built-up, etc.). Note: A class or theme could be anything like landcover classes, type of vegetation, tree species, etc. 7

8 What is the basic output of an image classification? The resulting classified image is comprised of: a mosaic of pixels, each of which belong to a particular class or theme, and is essentially a thematic "map" of the original image. 8

9 Image Classification Example Input Image Classified Image CSU Phil-LiDAR 2 9

10 Image Classification Concepts (1) Image classification uses the concept of Spectral Pattern Recognition 10

11 (In general) 11

12 Image Classification Concepts (1) Image classification uses the concept of Spectral Pattern Recognition 12

13 Image Classification Concepts (1) Image classification uses the concept of Spectral Pattern Recognition Spectral Patterns 13

14 Image Classification Concepts (1) Image classification uses the concept of Spectral Pattern Recognition spectral information represented by the Digital Numbers or Radiance or Reflectance in one or more spectral bands is used by a classifier, and attempts to classify each individual pixel based on this spectral information. Spectral Patterns 14

15 Basic Illustration: 15

16 What are the common image classification procedures? Common classification procedures can be broken down into two broad subdivisions based on the method used: Supervised classification Unsupervised classification 16

17 What are the main differences between the two? Supervised The image analyst supervises the pixel categorization process by specifying, to the computer algorithm (or classifier ), numerical descriptors of the various land-cover types present in an image Involves a training step followed by a classification step Utilizes training data Unsupervised The image analyst lets the computer algorithm to determine which pixels are related and groups them into classes The land-cover identity of these groups are determined by comparing with ground reference data Involves a classification step followed by a training step Do not utilize training data as basis during the start of the classification 17

18 Basic Steps involved in a Supervised Classification Step 1: TRAINING STAGE The analyst identifies in the imagery homogeneous representative samples of the different surface cover types (information classes) of interest. These samples are referred to as training areas. The selection of appropriate training areas is based on the analyst's familiarity with the geographical area and their knowledge of the actual surface cover types present in the image

19 What are the uses of training areas? Training areas are used to develop a numerical description of the spectral attributes of each land-cover type of interest in an image The numerical information in all spectral bands for the pixels comprising these areas are used to "train" a computer to recognize spectrally similar areas for each class. The computer uses a special program or algorithm ( classifier ), to determine the numerical "signatures" for each training class. 19

20 Basic Steps involved in a Supervised Classification Step 2: CLASSIFICATION STAGE Once the computer has determined the signatures for each class, each pixel in the image is compared to these signatures and labeled as the class it most closely "resembles" digitally

21 Basic Illustration: In supervised image classification, the multidimensional image matrix is used to develop a matrix of interpreted land cover category types. 21

22 Basic Steps involved in a Supervised Classification Step 2: CLASSIFICATION STAGE Once the computer has determined the signatures for each class, each pixel in the image is compared to these signatures and labeled as the class it most closely "resembles" digitally. What if the pixel is insufficiently similar to any training data? 22 22

23 Basic Steps involved in a Supervised Classification Step 3: THE OUTPUT STAGE After the entire data set has been categorized, the results are presented in the output stage Three typical forms of output products: Thematic maps Tables of land-cover statistics Digital data files that can be used in a GIS The classification output becomes a GIS input 23 23

24 Examples of Supervised Classification Outputs 24

25 Examples of Supervised Classification Outputs 25

26 PART 2A. SUPERVISED IMAGE CLASSIFIERS 26

27 Examples of Supervised Image Classifiers Minimum-Distance-to-Means Classifier (also known as Minimum Distance ) Parallelepiped Classifier Maximum Likelihood Classifier How does each classifier work? 27

28 Consider the following Landsat image overlaid with the Training Areas corresponding to 6 Land-cover classes: Water (sea) Water (river) Urban Grassland Forest Barren 28

29 Consider the following Landsat image overlaid with the Training Areas corresponding to 6 Land-cover classes: Water (sea) Water (river) Urban Grassland Forest Barren 29

30 The band values for each pixel for each landcover class of the training areas can be extracted: 30

31 Band 4 Digital Numbers From the extracted values, a 2D scatter plot or scatter diagram can be generated: Forest Grassland Barren Built-up 20 Water (Sea) 0 Water (River) Band 3 Digital Numbers 31

32 Or even a n-dimensional visualization (where n = no. of bands used) 32

33 The clouds of points represent multidimensional descriptions of the spectral response patterns of each category of land-cover type to be interpreted. These training set descriptions are used by the supervised image classifiers as interpretation keys by which pixels of unidentified cover type are categorized into their appropriate classes. 33

34 To understand how the supervised image classifier works, let s consider Band 3 and Band 4 of the Landsat image as input for classification. 34

35 Band 4 Digital Numbers What would be the land-cover type of Pixels X, Y and Z? Answer: It depends on what classifier to use! Forest X Grassland Barren Y Built-up 20 Water (Sea) 0 Z Water (River) Band 3 Digital Numbers 35

36 The Minimum-Distance-to-Means Classifier The mean, or average, spectral value in each band for each category is determined from the training areas. These values comprise the mean vector for each category A pixel of unknown identity maybe classified by computing the distance between the value of unknown pixel and each of the category means. The unknown pixel is assigned to the closest class. Note: An image analyst can define a maximum allowable distance between the mean to the unknown pixel. If a pixel is farther than the defined distance, it would be classified as unknown. 36

37 Band 4 Digital Numbers Using the Minimum Distance classifier, what would be the landcover type of Pixel X? FOREST! Forest + + Grassland 80 X Barren + Y + Built-up 20 Water (Sea) 0 + Z + Water (River) Band 3 Digital Numbers 37

38 Band 4 Digital Numbers Using the Minimum Distance classifier, what would be the landcover type of Pixel Y? (Seatwork) Forest + + Grassland 80 X Barren + Y + Built-up 20 Water (Sea) 0 + Z + Water (River) Band 3 Digital Numbers 38

39 Band 4 Digital Numbers Using the Minimum Distance classifier, what would be the landcover type of Pixel Z? (Seatwork) Forest + + Grassland 80 X Barren + Y + Built-up 20 Water (Sea) 0 + Z + Water (River) Band 3 Digital Numbers 39

40 Advantages and Disadvantages of the Minimum Distance Classifier Advantage Mathematically simple Computationally efficient Disadvantage Insensitive to different degrees of variance in the spectral response data Because of this, it is not widely used in applications where spectral classes are close to one another in the measurement space and have high variance. 40

41 Band 4 Digital Numbers Forest Grassland High variance in Built-up Barren Built-up 20 Water (Sea) 0 Water (River) Band 3 Digital Numbers 41

42 Parallelepiped Classifier Uses the range of values in each category training set to define decision regions that are then used to classify an unknown pixel. The range is defined by highest and lowest pixel values* in each band The decision region appears as: Rectangle in a 2D space Parallelepiped in multidimensional space An unknown pixel is classified according to the category range, or decision region, in which it lies. It remains unknown if it lies outside all regions. It if it lies in the overlap, can be arbitrarily placed in one (or both) of the two overlapping classes. *pixel values can be DN, radiance or reflectance 42

43 Parallelepiped Classifier 43

44 Band 4 Digital Numbers Using the Parallelepiped classifier, what would be the land-cover type of Pixels X, Y and Z? (Seatwork) Forest Grassland 80 X Barren Y Built-up 20 Z Water (River) Water (Sea) Band 3 Digital Numbers 44

45 Advantages and Disadvantages of Parallelepiped Advantages: Mathematically simple Computationally efficient Sensitive to category variance by considering the range of values in each category training set Major Disadvantage: Difficulties in classification can be encountered when category ranges overlap. 45

46 Problem (Seatwork) Given below are the statistics of 2 land-cover types based on training areas extracted from a Landsat 7 ETM+ image: Landcover: URBAN Min DN Max DN Mean DN Band Band Band Landcover: FOREST Min DN Max DN Mean DN Band Band Band Using the Minimum Distance and Parallelepiped Classifiers, classify the following unknown Pixels, with only Band 2 and Band 3 as input. DN Values Pixel A Pixel B Pixel C Band Band Band

47 The Maximum Likelihood Classifier Also called MLC The most common supervised classification method used with remote sensing data The MLC quantitatively evaluates the mean, variance and covariance of the category spectral response patterns (training data) when classifying an unknown pixel. 47

48 The Maximum Likelihood Classifier Classification is based on probability Through the use of probability distribution/probability density functions Probability is based on the mean, variance, covariance computed for each land-cover type 48

49 Variance? Co-variance? The variance refers to the spread of the data set how far apart the numbers are in relation to the mean, for instance. A covariance refers to the measure of how two variables will change together and is used to calculate the correlation between variables. In MLC, these variables are the band values (e.g., Band 1, Band 2) of a training dataset for a land-cover type 49

50 Band 4 Digital Numbers Forest Grassland Barren Built-up y = 0.253x R² = Water (Sea) 0 Water (River) Band 3 Digital Numbers 50

51 The Maximum Likelihood Classifier Using sufficient training data for each class, the mean, variance, covariance is computed for each land-cover type Given these parameters, the statistical probability of a given pixel being a member of a particular land-cover type is computed (also called probability distribution ) This probability distribution describes the chance or likelihood of finding a pixel from a certain cover type/class The probability distribution is computed for each landcover type for each unknown pixel After evaluating the probability for each category, the unknown pixel would be assigned to the most likely class (highest probability value). 51

52 Example Probability Distribution for an Unknown Pixel 52

53 Questions or clarifications? 53

54 PART 2B. UNSUPERVISED IMAGE CLASSIFIERS 54

55 Unsupervised Classifiers These classifiers identify the distinct spectral classes present in the image data Many of these classes might not be initially apparent/obvious to the analyst applying a supervised classifier The spectral classes maybe so numerous that it would be difficult to train on all of them In the unsupervised approach, they are found automatically 55

56 Commonly used Unsupervised Classifiers Also called clustering algorithms: K-Means ISODATA (Iterative Self-Organizing Data Analysis Techniques A) 56

57 The K-Means Classifier Used to determine the natural spectral groupings present in a dataset The number of groupings or clusters is defined by the analyst 57

58 The K-Means Classifier How does it work? Given the number of clusters to determine, the algorithm seeds or locates the number of cluster centers ( mean vectors or centroids ) in the multidimensional measurement space Each pixel in the image is then assigned to the cluster whose arbitrarily mean vector is closest (similar to Minimum Distance) After all pixels have been classified in this manner, revised mean vectors for each cluster are computed The revised means are then used as basis to reclassify the image data The procedure continues until there is no significant change in the location of class mean vectors between successive iterations of the algorithm. Once this point is reached, the analyst determines the land cover identify of each spectral class using reference data (e.g., from field surveys) 58

59 From: lva1-app6892/95/k-meanclustering-algorithm jpg?cb=

60 K-means: How it works (2D Example) Note: X and Y can be any band of the image 60

61 The ISODATA Classifier A variant of K-means Same procedure except that: It permits the number of clusters to change from on iteration to the next, by merging, splitting, and deleting clusters 61

62 The ISODATA Classifier The basis for merging, splitting and/or deleting clusters is defined by the analyst Merging: if the distance between the mean points of two clusters is less than some predefined minimum distance, the two clusters are merged together Splitting: a single cluster is split if it has a standard deviation (in any one dimension) that is greater than a predefined maximum standard deviation value Deleting: clusters with fewer than the specified minimum number of pixels are deleted 62

63 ISODATA Illustration From: 63

64 The ISODATA Classifier Requires from the analyst A range of number of clusters (not a single value like that of K-means) Minimum Distance between clusters Minimum number of pixels for each cluster Maximum standard deviation for each cluster 64

65 Notes on Unsupervised Classification The goal of unsupervised classification is simply the identification of spectrally distinct classes in image data The analyst must still use reference data to associate the spectral classes with the cover types of interest 65

66 Questions or clarifications? 66

67 PART 3. IMAGE CLASSIFICATION ACCURACY ASSESSMENT 67

68 Image Classification Accuracy Assessment A classification is not complete until its accuracy is assessed. 68

69 How do we assessed the accuracy of image classification? Use of reference data (also called ground truth data) E.g.: from site visits/field surveys, from aerial photographs, or from high resolution satellite images Accuracy of a classified pixel can be assessed by comparing its actual class based on the reference data 69

70 How do we assessed the accuracy of image classification? (2) Two ways to assess: Compare reference data to the classified image If a pixel is classified as forest on the ground, is it also forest on the classified image? Compare the classified image to the reference data If a pixel is classified as forest classified image, is it also forest on the ground? Accuracy assessment requires sufficient number of reference data for each land-cover class If reference data is many, some classified pixels maybe correct and some maybe not In this case, how can we determine the accuracy of the classified image? 70

71 Example: Pixel No. Reference/Ground Truth Data Classified Result 1 Forest Grassland 2 Forest Forest 3 Forest Forest 4 Forest Water 5 Urban Urban 6 Urban Urban 7 Urban Urban 8 Grassland Grassland 9 Grassland Grassland 10 Grassland Forest 11 Grassland Urban 12 Water Water 13 Water Forest 14 Water Grassland 15 Water Water 71

72 Example: Pixel No. Reference/Ground Truth Data Classified Result 1 Forest Grassland 2 Forest Forest 3 Forest Forest 4 Forest Water 5 Urban Urban 6 Urban Urban HOW DO WE 7 Urban Urban DETERMINE 8 Grassland Grassland 9 Grassland Grassland ACCURACY? 10 Grassland Forest 11 Grassland Urban 12 Water Water 13 Water Forest 14 Water Grassland 15 Water Water 72

73 The Use of Classification Error Matrix One of the most common means of expressing classification accuracy is the preparation of an error matrix Also called: Confusion matrix Contingency table An error matrix compares, on a category-bycategory basis, the relationship between known reference data and the corresponding results of the classification The matrix is square, with the number of rows and columns equal to the number of categories whose classification accuracy is being assessed. 73

74 A Simple Example of an Error Matrix 74

75 OVERALL ACCURACY AS OBTAINED FROM THE ERROR MATRIX: The major diagonal (running from upper left to lower right) represent pixels that are classified into their proper category (i.e., classified correctly) Overall accuracy (OA): computed by dividing the total number of correctly classified pixels (sum of diagonal) by the total number of reference pixels In the given example: OA = ( )/( ) = 63/100 = 0.63 or 63% 75

76 CLASSIFICATION ERRORS AS OBTAINED FROM THE ERROR MATRIX: Classification Errors: Ommision (Exclusion) Errors Commission (Inclusion) Errors All non-diagonal elements of the matrix represent errors of omission or commission Omission and commission errors are determined according to category 76

77 OMMISSION ERRORS Omission errors correspond to nondiagonal column elements Example: 14 pixels + 1 pixel that should have been classified as Water were omitted from that category, and falsely classified as Forest (14 pixels) and Urban (1 pixel) Omission Error for a Category = number of omitted pixels/total number of reference pixels In the above error matrix, the omission errors can be computed as follows: For Forest: Omission Error = (1+1)/30 = 2/30 = 0.07 or 7% For Water: Omission Error = (14+1)/30 = 15/30 = 50% For Urban: Omission Error = 50% 77

78 COMMISSION ERRORS Commission errors correspond to nondiagonal row elements Example: 1 pixel (Water)+ 5 pixels (Urban) were falsely included/classified as Water Commission Error for a Category = number of committed or included pixels/total number of classified pixels In the above error matrix, the commission errors can be computed as follows: For Forest: Commission Error = (14+15)/57= 29/57 = 0.51 or 51% For Water: Commission Error = (1+5)/21 = 6/21 = 29% For Urban: Commission Error = 2/22=9% 78

79 79

80 ACCURACIES OF INDIVIDUAL CATEGORIES AS OBTAINED FROM THE ERROR MATRIX: PRODUCER S ACCURACY (PA) Computed by dividing the number of correctly classified pixels in a category (on the major diagonal) by the number of reference pixels for that category (column total) The Producer s accuracy of category would tell us how many of the reference pixels for that category were correctly classified as such. E.g., If 30 pixels of Forest were found on the ground, how many of it were correctly classified as forest in the classified image? It is a measure how good the classified image is in correctly classifying a particular category. In the above error matrix, the Producer s Accuracy for Forest can be computed as follows: For Forest: PA = 28/30 = 0.93 or 93% Alternatively: PA =100% Omission Error For Forest, PA = 100% - 7% =93% 80

81 ACCURACIES OF INDIVIDUAL CATEGORIES AS OBTAINED FROM THE ERROR MATRIX: USER S ACCURACY (UA) Computed by dividing the number of correctly classified pixels in a category (on the major diagonal) by the number of pixels that were classified for that category (row total) The User s accuracy of category would tell us how many of the classified pixels for that category were correctly classified as such. E.g., If 57 pixels of Forest were classified, how many of it were are actually forest? It indicates the probability that a pixel classified into a given category actually represents that category on the ground. In the above error matrix, the User s Accuracy for Forest can be computed as follows: For Forest: UA = 28/57 = 0.49 or 49% Alternatively: UA =100% Commission Error For Forest, UA = 100% - 51% =49% 81

82 Some notes to remember Producer s accuracy: Compare reference data to classified data E.g., visit forest locations on the ground, and compare with the classified image. How many of it were correctly classified as forest? User s accuracy Compare classified data to reference data E.g., get pixels classified as forest in the classified image. Then, visit these locations on the ground. How many of it were actually forest? 82

83 Example: Build a confusion matrix based on this table, and compute the different measures of errors and accuracies. Pixel No. Reference/Ground Truth Data Classified Result 1 Forest Grassland 2 Forest Forest 3 Forest Forest 4 Forest Water 5 Urban Urban 6 Urban Urban 7 Urban Urban 8 Grassland Grassland 9 Grassland Grassland 10 Grassland Forest 11 Grassland Urban 12 Water Water 13 Water Forest 14 Water Grassland 15 Water Water 83

84 Solution: Overall accuracy: 9/15 = 60% 84

85 References/Further Reading Lillesand, T. M., Kiefer, R. W., & Chipman, J. W. (2008). Remote Sensing and Image Interpretation 6th Edition. United States of America: John Wiley & Sons, Inc. 85