Abstract
Image differencing has been investigated as a technique for use with Landsatr digital data to delineate areas of load cover change in an irban environment. t Landsat data collected in April 1973 and April 1975 for Austin, Texas, were i_ geometrically corrected and precisely registered to United States Geological Survey (USGS) 7.5-minute quadrangle maps. At each pixel location reflectance ,> values for the corresponding bands were subtracted to produce four difference images. Areas of major reflectance differences are isolated by thresholding each of the difference images. The resulting images are combined to obtain an image data set of total change. These areas of reflectance differences were found, in general, to correspond to areas of land cover change. Informa tion on areas of land cover change was incorporated into a procedure to mr.sk out all Loncha_nge .eas and perform an unsupervised classification only for data in the change areas. This procedure identified three broad categories: Wr (1) areas of high reflectance (construction or extractive), (2) changes in agri cultural areas, and (3) areas of confusion between agricultural and other areas. Efforts are underway to include 1977 Landsat data and to rigorously evaluate the results of the image differencing technique and the unsupervised classifica tion. TABLE OF CONTENTS Section 1 - Background ................................ 1-1 Section 2 - Analysis Technique ........................... 2-1 Section 3 - Study Site ................................. 3-1 Section 4 - Data ..................................... 4-1 Section 5 - Results and Discussion ........................ 5-1 5.1 Difference Data Statistics ........................ 5-1 5.2 Difference Threshold ....................... . ... 5-4 5.3 Preliminary Evaluation ............ . .......... .. 5-7 5.4 Change Area Classification ....................... 5-11 Section 6 - Future Work 6-1 6.1 Scene Registration . . 0 .. 0 ..... 6-1 6.2 Ground Truth ................................. 6-2 6.3 Data Comparison ........... . .................. 6-2 Section 7 - Summary ................................. 7-1 References 4 tit LIST OF ILLUSTRATIONS Figure 4-1 Landeat Scenes 1237-16372 and 1974-16130 .............. 4-2 4-2 Enhanced Pictures of the Geometrically Corrected and Registered Portions of Landeat Scenes 1237-16372 and 1974-16130 ..................... 0 .. 0 0 0 0.00.0 4-3 5-1 Histograms of MSS 7 for 1973, 1975, and the Difference Data Set, With Mean Values Shown for Each Distribution ... 5-2 5-2 Histograms of the Four Difference Data Sets, With Mean Values Shown .. ........................0000... 5-3. 5-3 Daily Rainfall Data for Austin, Texas, During March 1973 and March 1975 ............................... 5-5 5-4 1975-1973 Difference Images by Band ................. 5-6 5-5 Areas of Reflectance Difference Greater Than Three Standard Deviations Above and Below Mean Values by Band ........ 5-8 5-6 Composite Images of Change Areas ................... 5-9 5-7 Total Change for the Scene ................... 0 .... 5-10 5-8 Total Change Displayed Against a Background of the 1975 Data .......................... 0 0.0.0... 5-12 5-9 Land Cover Classification of 1975 Data for Change Areas .... 5-14 Means and Standard Deviations of 1973, 1975, and Differenced Images ............................ 5-1 Threshold Values for Density Slice ................... 5-7 ^- r SECTION 1 - BACKGROUND In 1975 a joint research project between Goddard Space Flight Center (GSFC) and the Bureau of the Census began. The program is to examine the utility of Landsat remotely sensed data for monitoring land cover/land use activity around major urbanized areas in the United States. Early work in the project has shown that major land cover classes in the urban fringe (transition zone be tween urban and rural populations) can be identified from satellite data (Refer ence 1). Land cover change is a key element in the urbanized area boundary delineation program conducted by the Census Bureau. The frequent, synoptic nature of Landsat data makes it well suited for the detection of land cover change. In an extension of the earlier work on this project, a technique for the delineation of areas that have undergone major changes in spectral reflectance (indicating changes in land cover) has been developed. This document reports progress to date in the development of one approach for monitoring land cover change and also describes objectives of future research. 1-1 SECTION 2 - ANALYSIS TECHNIQUE The fundamental operation in automated change detection is the subtraction of one digital image from another. The differencing operation is carried out by subtracting the reflectance value for a point in one image from the reflectance value for the same location in the second image. Landsat I or II multispectral scanner (MSS) output consists of four bands or digital images. Thus, to differ ence all of the Landsat MSS data in a scene requires four operations per pixel location--one for each of the four MSS bands. The subtraction of one digital image from the other results in an image in which positive and negative values represent areas of change and zero values indicate areas that remained unchanged. When applied to Landsat data with a data range of 0 to 127, the potential range of difference values is -127 to +127. In the Small Interactive Image Processing System/Video Image Communication and Retrieval System (SMIPS/VICAR) system (Reference 2), all results are stored as positive values scaled from 0 to 255 by the addition of a constant. This oper ation can be expressed mathematically as OXijk = X(1)ijk - X(2)ijk + C (2-1) where AX = change image X(1) = image at time 1 X(2) = image at time 2 C = a constant i = 1... n number of lines j = 1... n number of columns k = 1... n number of bands When applied to Landsat data, this results in four different images that can be subsequently analyzed using standard image processing techniques. 2-1 and these results combined to obtain an image representing total change. Change areas for each band of the difference image were determined by identify ing those pixels with intensity values above or below calculated threshold levels (i. e. , density slicing the image). Pixels in the difference image with reflec tances above the high threshold represent areas that have increased in reflec tance from time 1 to time 2 (high reflectance change), whereas pixels with values below the lower threshold represent areas that have decreased in reflec tance from time 1 to time 2 (low reflectance change). Pixels with reflectances that fall between the threshold values are considered to have remained un changed. This procedure results in a series of images depicting areas of change for each band which can then be combined to obtain an image showing all areas of change. This process makes two assumptions: (1) changes in land cover/land use will result in corresponding changes in reflectance values, and (2) these reflectance changes will be of sufficient magnitude to distinguish them from the normal var iation of features within the images. 2-2 The test site for this study is an area of approximately 1200 square kilometers encompassing Austin, Texas, and its environs. This area was chosen because the Census Bureau conducted pretests there for the 1980 census at the start of this work. The Bureau intended to compare pretest results with satellite data products. As part of that earlier work, 1973 and 1975 images were registered to United States Geological Survey (USGS) topographic maps. The availability of registered images and the interest of the Bureau of the Census led to the use of Austin as the site for image differencing/change detection experimentation. To date, however, results of the census pretest have not been made available. Austin is located in central Texas between the Coastal Plain and the Edwards Plateau. Its terrain varies from level, open farmland and fields in the east to tree-covered hills in the west. The Austin central business district (CBD) is surrounded by commercial, industrial, and high-density residential develop ment. Further from the center city--primarily to the north, east, and south-lie low-density residential development and agricultural areas. In addition, a limited number of estates and trailer parks are mixed into this diverse land scape. 3-1 SECTION 4 - DATA Subsets of Landsat data collected on March 17, 1973 (Figure 4-1(A)), and March 24, 1975 (Figure 4-1(B)), were used as the principal input for this study. i Near-anniversary data sets were selected to minimize phenological and sea sonal varlatiors. Each subimage was geometrically corrected and precisely registered to USGS 7-1/2-minute quadrangle maps and resampled to 60-meter - { square picture elements (Figures 4-2(A) and (B)) using the Digital Image Recti A fication System (DIRS) implemented at GSFC (Reference 3). The SMIPS/VICAR i system available at GSFC was used for all subsequent processing and analysis. Supporting data for the study included USGS topographic maps (photo-revised in 1973), low-altitude black and white photography taken in 1973 and 1974, a few frames of low-altitude black and white photography taken in 1975, and 1977 high altitude black and white imagery. 4-1 Q c H a^ k c^u .^ k .M..^ fA O N I G N ^ c^ M :n N ^ N ^ cz c> ^ 3 ^ ^ 'fl 1 ^ N ^J M cd N ^ to a c CA y U . d rn S. d' I a (A 4-2 Q7 M d Q 1 cc V 1 r OF POOR QUALITY ORIGINAL F:. r Ln • an r N I m w L r) a^ c co U o N M c 'v U c1 tz C7 ti I Cu .+ C7N y V ^E C C) W (n V M In cn cl c a 0 ^-~ ^ m vsr u c^ a ^' N I ^I i )^^lP►VrY'n'^^^yrf PA E r r ^^n ORIGINAL PAGE j' B^K A "'0 WH;TE PHOTOGRAPH ORIGNAt. PAGE 1 OF POOR QUALITY SECTION 5 - RESULTS AND DISCUSSION For this study a difference image was produced by subtracting 1973 pixel values • from the corresponding elements in the 1975 image and adding a constant of 128. Thus, Equation (2-1) becomes Axiik = X(1975)iik - X(1973)iik + 128 The four resulting difference images are the basis of all further analysis. 5.1 DIFFERENCE DATA STATISTICS Differencing procedure relsuits are illustrated by histograms of MSS 7 for the 1973, 1975, and difference data Lets (Figure 5-1). The histogram of the differ enced data rFigure 5-1(C)) is centered near 228 with a much tighter standard deviation than the histograms of the raw data. The histograms of all four dif ference images (Figure 5-2) are tightly clustered near 128. The means and standard deviations for each of these images and their raw data counterparts are presented in Table 5-1. Mean values near 129 are Rxpected for the differ ence) data sets because most areas have remained unchanged during the 2-year period. Table 5-1. Means and Standard Deviations of 1973, 1975, and Differenced Images MIS BAND 1673 147E DIFFERENCE MEAN STANDARD MEAN STANDARD WAN ETANDARO DEVIATION DEVIATION DEVIATION "A 11.3 61.7 11.1 126.6 6.7 E 61.4 15.3 $3.1 1E.2 1274 6.3 6 71.4 16.2 ME 10.4 132.2 10.1 7 741.6 17.E 75.7 16.7 130.7 11.6 5-1 5-2 f i W V W ^ V In WWW w A a EI d A W m w yy IA A t w° 3 Fa R .r w ^ W ri V Q1 .^ N w w r, w ^O d QO , Ia 0o O 'DM(;INAL RAGE &3 OF POOR QUALI7'y w • — E!—efy 1if > —c" N o ti N ..rm O 60 m w x a^ w 0 Q ^wwwa^ G A go ca 3 O 0s r O I I^ a a 5-3 Mean values and standard deviations for MSS-6 and MSS-7 difference data are higher and broader, respectively, than the means or standard deviations in MSS-4 and MSS-5 difference images. This suggests that an overall shift in ra diance has occurred for these spectral bands. Indeed, rainfall data for 1 arch 1973 and March 1975 (Figure 5-3) show a cumulative rainfall of 1.2 inches for the 2 days prior to the March 17, 1974, image, whereas no appreciable rainfall was recorded for the 10 days prior to the March 24, 1975, image. Such a dif ference in ground moisture would contribute to the reflectance difference be tween the images, particularly for MSS 6 and MSS 7, because both bare soil and healthy vegetation decrease their reflectance of infrared radiation as their moisture content increases. Impervious areas, like the CBD, would not exhibit such changes. Given the difference in rainfall, responses for MSS 6 and MSS 7 for 1973 would be expected to be more varied and on the average lower than the values for 1975. Indeed, the mean values and standard deviations for MSS 6 and MSS 7 of the difference image (1975-1973; sire higher and broader than MSS 4 and MSS 5 of the same product. 5.2 DIFFERENCE THRESHOLD A direct display of the difference data produces a flat, low-contrast picture thht is difficult to interpret (Figure 5-4). The intensity of an area in the image indi cates the magnitude of the reflectance difference between the two dates for that location. Areas that have experienced major change appear light (increased re flectance) and dark (decreased reflectance) in the image. These light and dark areas correspond to the data located in the tails of the difference data histo grams. The pixel locations corresponding to these differences can be selec tively displayed by density slicing each of the difference images. In this experiment the threshold values for the density slice were set at three standard deviations above and below the mean difference value (Table 5-2). Initially, threshold values of plus and minus two standard deviations were se lected, but this appeared to overestimate the probable amount of change. 5-4 I QUP'UV of poOR INCHES OF RAIN d0 5 10 15 20 25 30 17 DAY OF MONTH MARCH 1973 INCHES OF RAIN 14 1.2 1.0 08 0.6 04 0.2 0 5 10 15 20 25 30 24 DAY OF MONTH MARCH 1975 Figure 5-3. Daily Rainfall Data for Austin, Texas, During March 1973 and ]March 1975 5-5 MSS 4 41 IMSS 5 ORIGINAL PAGE BLACK AND WHITE PHOTOGR M Figure 5-4. 1975-1973 Difference Images by Rand 5-6 ORIGiiNA- PAGE POOR Table 5-2. Threshold Values for Density Slice | | | | | | | | -3 STANDARD | | | *3 STANDARD | | | | | | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | | | DIFFERE14CED | | 3 STANDARD | | | | DEVIATIONS | | | DEVIATIONS | | | | | | | MISS BAND | | DEVIATIONS | | | | | | | | | | | A | | | | | | | | (LOW THRESHOLD) | | | (HIGH THRESHOLD) | | | | | | | | | 4 | | 20.1 | | | 108 | | | 149 | | | | | | | | 5 | | 272 | | | 101 | | | 157 | | | | | | | | 8 | | 30.3 | | | 101 | | | 182 | | | | | | | | 7 | | 36.7 | | | W | | | 188 | | | | | Picture elements with differences less than, or greater than, the threshold value are displayed in Figure 5-5. The top row of images (Figures 5(A) through (D)) shows those pixels that have decreased in reflectance from 1973 to 1975, for each of the MSS ban,is. The bottom row of images (Figures 5-5(E) through (li)) displays those pixels that have increased in reflectance. The areas delin eated in these images correspond to all locations that had a variation in meas ured reflectance greater than the noise threshold of three standard deviations. These images were combined to generate a series of data sets that depict areas of decreased reflectance (low change), increased reflectance (high change), and decreased and increased reflectance (total change) (Figure 5-6). 5.3 PRELIMINARY EVALUATION The areas delineated in this way may provide useful information for identifying certain types of land cover change. Preliminary comparisons indicate that areas which have decreased in reflectance (low reflectance change) are primar ily vegetated or agricultural areas. A reas of increased reflectance, however, seem to include both agricultural areas and areas of construction activity. A comparison of the low-reflectance change image and high-reflectance change image shows that 407 pixels increased in reflectance in one or more bands and also decreased in reflectance in one or more of the remaining bands. A color display of the total change image (Figure 5-7) highlights these areas as white, 5-7 ^, o U fy o ^ 6m co (n Cn ul y q U L^ S. m 4) cz Cd C Cd F c. a^ a^ cd b Cd c. b Q a^ cd 0 c m U W
Showing the abstract — retrieve the full paper via the Exa API.