Satellite image classification

Transcription

1 Satellite image classification EG2234 Earth Observation Image Classification Exercise 29 November & 6 December 2007 Introduction to the practical This practical, which runs over two weeks, is concerned with the automated classification of remotely sensed imagery. Working with Landsat ETM+ (AKA Landsat 7) satellite data for the city of Bristol and its environs, you will learn how image classification can be used to automatically establish the different land cover types present. This process enables the image analyst to generate a thematic map of the region of interest (ROI) which can be directly transferred to a GIS (or other) environment. Using IDRISI you will first pre-process the imagery to create a geometrically correct (Georeferenced) satellite scene before progressing to classify the different types of land cover present within the image scene using different classification rules. Image classification: a few pointers One of the most common procedures undertaken by any remote sensing scientist is a land cover assessment. Typically, this will involve the use of automated classification procedures to assign image pixels to a set of informational categories known as classes. This methodology takes advantage of the multispectral capabilities of remote sensing, especially satellite imagery. Remember, earth surface objects typically have wavelength dependent properties (i.e. spectral signatures) and therefore we need information from all available wavebands in order to maximise the information for processing and classifying land cover types from the image scene. There are numerous ways to classify multispectral remotely sensed imagery including both unsupervised and supervised techniques. Although you will briefly look at unsupervised methods, this practical primarily focuses on unsupervised methods. As their title suggests, unsupervised methods involve relatively little input from the image analyst, save the total number of classes required. Lillesand et al. (2004: 573) describe how this family of classifiers involves algorithms that examine the unknown pixels in an image and aggregate them into a number of classes based on the natural groupings or clusters present in the image values. Page 1 of 9 Image Classification EG2234

2 Campbell (2002: 324) identifies the advantages of such methods as: No extensive prior knowledge of the region is required The opportunity for human error is minimised Unique classes are recognised as distinct units Supervised image classification is an interactive process where the image analyst must first select the informational classes to be used in the analysis and then train the data by selecting training sites which characterise each of the chosen image classes. These training data are carefully selected to identify only pure or homogeneous pixels which represent the particular class category (Campbell 2002: 333). Only then are the data actually classified and image pixels assigned to the pre-determined classes. There are numerous decision rules available, but one of the most efficient is the Maximum Likelihood Classification (MLC) method. Historically, this has been a very widely used classification technique and is a very powerful classifier when it is based on good quality data and classes. It employs a probability function (a measure of the degree of membership) to allocate pixels to the most likely informational class (Curran 1985). To learn more about these and other image classification methods you are directed to any of the following textbooks: [1] Campbell, J.C. (2002) Introduction to remote sensing. Third edition. New York: Guilford Press. [2] Curran, P.J. (1985) Principles of remote sensing. Harlow: Longman. [3] Gibson, P.J. & Power, C.H. (2000) Introductory remote sensing: digital image processing and applications. London: Routledge. [4] Lillesand, T.M., Kiefer, R.W., & Chipman, J.W. (2004) Remote sensing and image interpretation. Fifth edition. New York: Wiley. [5] Mather, P.M. (2004) Computer processing of remotely-sensed images. Third edition. London: Wiley. Learning outcomes On completion of this extended practical you should be able to: Geometrically correct (georeference) a raw satellite image Select training data and sites to be used in the image classification process Employ the maximum likelihood classification decision rule to allocate image pixels to specific information classes Page 2 of 9 Image Classification EG2234

3 Determine the accuracy of the MLC supervised classification Understand the relative merits and limitations associated with standard pixel-based classifiers Data In this practical class you will be working with Landsat 7 satellite imagery for the area of Bristol. All of the data are in IDRISI format so you can open them directly within the image processing software. Raw, uncorrected satellite imagery (band 4) for Bristol and its environs Georeferenced multispectral imagery (bands 1-5, and 7) for Bristol and its environs All of the imagery has been placed on the GAIA server (the J:\ drive) in E409 and can be found in the EG2234 directory folder called classification and its sub folders: raw and classify. J:\EG2234\classification\ Copy the entire folder to your own personal drive space (please note that all of the raw data sets alone before you have undertaken any image processing take up approximately 5MB of storage space). Once you have copied the data over successfully start up the IDRISI image processing software package from the desktop. When the software opens the first thing you will need to do is set up your working environment. Make sure that you identify the working directory to be the classification folder on your own drive space. You should also add two resource folders raw and classify before proceeding with the practical. Image preprocessing Using the DISPLAY LAUNCHER open the Near Infrared (NIR band 4) waveband image. You should ensure that you use the greyscale display function. This imagery covers the city of Bristol and the surrounding countryside. Take a few minutes to familiarise yourself with the imagery there are Ordnance Survey (OS) maps available to help you do this. Using the NIR band you will see that the area is pretty mixed in terms of land cover, varying from semi-rural to highly urban. The satellite scene is currently in raw (geometrically uncorrected) format and before undertaking any further image processing, or using it as a base map layer (Lillesand et al. 2004) it is necessary to geometrically correct the image and resample the data to OS British National Grid map co-ordinates. This procedure corrects for known distortions present within airborne and satellite imagery such as earth curvature and relief displacement (see Lillesand et al. 2004: 495). To perform this correction you need to collect a series of ground control points, known as GCPs. These are points Page 3 of 9 Image Classification EG2234

4 which are of known (map) location and are easily recognised within the image scene. It is common practice to use easily detected features within the imagery in this process, such as cultural features (e.g. road intersections, edges of built-up areas, airport runways etc.) The selected points should also cover the image area as widely as possible, and not cluster in any one single part of the image scene this will help to create a more accurate rectification. The GCP data are then used to resample or register the data to map co-ordinates. This involves calculating the positional error of the selected points and the most accurately identified points will be used in the registration process. The results will enable the image analyst to use the image as a base map layer. The geometric correction process requires that the geographic (map) locations of the GCPs are recorded as well as their image co-ordinates. Select a series of control points from the image, noting their on-screen x and y co-ordinates as well as the Eastings and Northings (12 figure grid references required) from the OS maps or the online maps available at if you find this easier. List the GCP coordinates in the table below. Point Feature Identified On-screen co-ordinates OS Grid co-ordinates X Y Easting Northing Page 4 of 9 Image Classification EG2234

5 When you have collected enough GCPs you can open the Resample tool. Reformat Resample You should be presented with a dialogue box like that shown in Figure 1 below. Figure 1. The resample tool On the left-hand side of the dialogue box shown in Figure 1 you can select your input image (b4_nir.rst) and then provide an appropriate name for the output image (e.g. b4_corr), making sure that you save the output file somewhere you can easily access on your I:\ drive. You should then click on the Output Reference Parameters button and input the following details: Number of columns: 900 Number of rows: 750 Minimum X co-ordinate: Maximum X co-ordinate: Minimum Y co-ordinate: Maximum Y co-ordinate: Reference system: Plane Reference units: Meters Unit distance: 1.0 After putting in the above parameters click the OK button. You can now start to input the GCPs you collected these should be placed in the middle section of the Resample tool menu in the following format: Image X co-ordinate, Image Y co-ordinate, Map X co-ordinate, Map Y co-ordinate Task 1 Run the geometric correction on band 4 of the Bristol satellite data. What is the RMS accuracy of the corrected imagery? How many GCPs did you retain in the procedure? Were there any common issues or characteristics regarding the GCPs you discarded from the process? Page 5 of 9 Image Classification EG2234

6 Image classification In this part of the practical you will learn how to implement both supervised and unsupervised image classification. First, you will look at the use of unsupervised methods and explore some of the advantages and disadvantages of this method before going on to implement the supervised MLC decision rule described earlier. In the previous part of this exercise you geometrically corrected the satellite image for Bristol. Now for this part of the practical you will employ an already corrected satellite image scene for the city and its environs (this is to ensure that all imagery used in the classification stages is accurate to within 1 pixel i.e. RMS < 1). The corrected image data may now be employed within a GIS framework, however, it would be preferable to reduce the data set to more readily usable information classes (categories). The most obvious step is to classify the imagery using ancillary (mapbased) information, but before you proceed to this stage you will first explore an unsupervised classification known as ISOCLUST. Unsupervised classification Run ISOCLUST using all the available wavebands for the Bristol imagery (located in the classify sub-directory). This technique identifies clusters in spectral space and tries to assign pixels to these natural spectral groupings. All that it requires from the image analyst is the input bands (use all corrected wavebands in this exercise), the number of iterations (3 for this exercise) and the expected number of clusters (5 for this exercise). However, you cannot predetermine the type of classes it identifies the computer does this arbitrarily. Only afterwards can you begin to identify informational classes from the resultant classified image. Image Processing Hard Classifiers ISOCLUST Keep the results of this classification you will need to create a map compilation as well as compare the results of this classification with the supervised methods used in the next part of the practical. Supervised classification You will now perform a supervised classification of the Landsat 7 data using the maximum likelihood classification algorithm. This supervised classification process involves 3 stages: 1. Training site selection 2. Classification of imagery 3. Testing of classification accuracy Page 6 of 9 Image Classification EG2234

7 Select one of the geometrically corrected image bands for display you will use this to locate and identify your training sites for image classification. You may find band 4 (near infrared) the most visually expressive display for this purpose. For the purposes of this practical class choose 5 land cover categories to use in the automated image classification. You may find that you will have to choose some classes that are spectrally similar do not worry about this. OS maps are on hand to help you decide upon appropriate class types. Training The training sites selected need to represent pure areas of each land cover class. You should steer clear of any pixels or areas within the image that you think may show heterogeneous characteristics. The training sites need to be defined using the on-screen digitising function within IDRISI. This is located using the cross-hair button from the toolbar. When using this function make sure that for each class you define polygon features. You should also only create one vector file called TRAIN just make sure that you use a separate ID number for each set of training polygons created within the different classes, e.g. urban ID = 1, rural ID = 2. When you have completed digitising each set of class training areas you should click on the Arrow button (stop and save digitising) next to the cross-hair button. Signature generation in IDRISI Once you have generated all of the training areas you will need to generate the statistics for the classification process. IDRISI requires you to generate signature files for each of the different proposed classes. Select the MAKESIG command from: Image Processing Signature Development MAKESIG You will need to create signatures separately for each class type. When you employ the MAKESIG function you will be provided with a dialogue box where you will need to identify the classes for signature generation you will need to select a signature filename for each training site ID (Figure 2). Classification The second step in the supervised classification process is to actually assign the pixels to the predefined classes using the Maximum Likelihood Classification process: Image Processing Hard Classifiers MAXLIKE Page 7 of 9 Image Classification EG2234

8 Select 5 signature files as inputs into the classification process making sure you keep class probabilities equal (20% for each category, based upon 5 information classes). Select an appropriate output filename and then ensure that all available LANDSAT 7 bands are used in the classification process. Run the classification process. You will be presented with a newly classified image. How does it look? Is it what you expected? Does it seem accurate? In order to find out we will need to produce a formal statement of classification accuracy. Accuracy assessment To determine the accuracy of the classification procedure you will need to generate a confusion (error) matrix. A table is provided below by way of example. To create this matrix you will need to select approximately 40 sample points and then determine whether these have been accurately assigned to the different information (land cover category) classes. IDRISI allows you to create a random set of sample points to enable error check between map and original imagery with the newly classified imagery. Image Processing Accuracy Assessment SAMPLE For the reference image use the Max Like image as the input and select Stratified Random as the sampling strategy. Select 40 points. Call the output vector file SAMPLE. This vector file can then layered on top of your classified image and one of the raw image bands to enable you to check for accuracy. Use the cursor inquiry key to determine how the point has been classified and determine its accuracy using the raw image data and the available OS maps. Using the results you can start to complete the following table: Reference class Assigned class Total Total Page 8 of 9 Image Classification EG2234

9 Task 2 Create a map compilation including a legend, scale bar and north arrow of the ISOCLUST and MLC classified images. Can you clearly interpret the results of the ISOCLUST classification? Do the clusters relate to any identifiable informational classes? Can you see any potential problems with using such a technique? How accurate was the MLC classification? Which classes were less accurately classified using this procedure? Can you think of any reason for this? Acknowledgements The datasets: The source for the Bristol imagery was the Global Land Cover Facility (GLCF) at the University of Maryland, USA. URL: The Landsat data are distributed by the Land Processes Distributed Active Archive Center (LP DAAC), located at the USGS Center for Earth Resource Observation and Science (EROS) in Souix Falls, South Dakota PLEASE NOTE THAT THE DATASETS ARE ALL COPYRIGHT RESTRICTED AND SHOULD ONLY BE USED FOR THE PURPOSES OF THIS CLASS Page 9 of 9 Image Classification EG2234