LANDSAT4 MSS And Thematic Mapper Data Quality And Information Content Analysis

222 IEEE TRANSACTIONSON GEOSCIENCEAND REMOTESENSING, VOL. GE-22, NO.3, MAY 1984TSRQPON L A N D S A T -4 MSS a n d T h e m a tic M apper D a ta Q u a lity a n d In fo r m a tio n C o n te n t A n a ly s is PAUL E. ANUTA, SENIOR MEMBER,IEEE, LUIS A. BARTOLUCCI, M. ELLEN DEAN, D. FABIAN LOZANO, ERICK MALARET, CLARE D. McGILLEM, FELLOW, IEEE, JOSE A. VALDES, AND CARLOS R. VALENZUELAWVUTSRQPONMLKJIHGFEDCBA A b s tr a c t-L a n d s a t-4 Thematic Mapper and Multispectral Scanner COMPARISON OF A- AND poTAPE DATA data were analyzed to obtain information on data quality and informa- To determine the effect of the geometric correction process tion content. Geometric evaluations were performed to test band-to- band registration accuracy. Thematic Mapper overall system resolution (resampling and gray-level interpolation) on the radiometry of was evaluated using scene objects which demonstrated sharp high the new resampled pixels when converting the data from an contrast edge responses. Radiometric evaluation included detector A-tape (30 m X 30 m pixels) to a P-tape (28.5 m X 28.5 m relative calibration, effects of resampling, and coherent noise effects. pixels) format, the means and standard deviations for homo- Information content evaluation was carried out using clustering, princi- geneous and heterogeneous areas on the ground were calcu- pal components, transformed divergence separability measure, and numerous supervised classifiers on data from Iowa and Illinois. A lated using the data from both the A and P tapes. A qualita- detailed spectral class analysis (multispectral classification) was carried tive comparison of both sets of statistics indicated that no out on data from the Des Moines, Iowa area to compare the informa- change (radiometric degradation) was caused by the nonzero tion content of the MSS and T M for a large number of scene classes. order (cubic convolution) interpolation process. Furthermore, to determine whether the cubic convolution INTRODUCTION interpolation has affected the structure of the data in feature space, a set of data from the A and P tapes for the same area L ANDSAT-4 Thematic Mapper and Multispectral Scanner data were analyzed in this study to obtain information on data quality and information content. Geometric evaluations on the ground was clustered into 16 cluster classes. Then the resulting 16 cluster classes from the A and P-tape data sets were merged together into a single 32-class file. Subsequently, were performed to test band-to-band registration accuracy. the pair-wise spectral separability for the 32 classes was com- Thematic Mapper overall system resolution was evaluated us- puted using a Transformed Divergence algorithm [1]. The ing scene objects which demonstrated sharp high contrast edge minimum transformed divergence between cluster classes responses. Radiometric evaluation included detector relative from the TM A tape and corresponding second cluster classes calibration, coherent noise analysis, and the effects of resarn- from the TM P tape is 22 and the maximum is 262. Since the pling. Information content evaluation was carried out using transformed divergence ( D CBA T ) measure ranges from 0- to 2000- clustering, principal components, and the transformed separa- where a value of 2000 indicates that the two classes in ques- bility measure on data from Iowa and Chicago, Illinois. A tion are completely separable (different) and any D T value detailed spectral class analysis (multispectral classification) between 0- and 500 indicates that the pair of classes is not was carried out on data from the Des Moines, Iowa area to separable (very similar )-it may be concluded tha t the 16 clus- compare the information content of the MSS and TM for a ter classes obtained from the A and P tapes are essentially large number of scene classes. This work is being carried out equivalent. Thus these results corroborate the previously stated as part of Landsat-4 data quality evaluation study. conclusion that the radiometry of the resampled pixels from TEST DATA SETS the P tape has not been significantly affected by the geometric correction process. Consequently, they meet GSFC specifica- The data sets used in this study were gathered over the Des tions for the Landsat-D system, i.e., resampling shall not Moines, IA area on Sept. 3, 1982 and the Chicago, IL area on introduce radiometric degradation of more than 1 quantum October 25, 1982. Both A tapes which are radiometrically level [2]. corrected and P tapes which are both radiometrically and geometrically corrected were acquired for these scenes. The REGISTRATION EVALUATION scene numbers were 40049-16264· for Iowa and 40101-16025 The first TM scene, which was of the Detroit area, contained for Illinois. Preliminary test scenes over Detroit and Arkansas several geometric and radiometric errors which did not appear were also available. in subsequent scenes. The overall visual quality of TM data is extremely good; detailed analysis is required to find subtle Manuscript received November 29, 1983; revised December 23, 1983. errors. Band-to-band registration was evaluated using a corre- This work was sponsored by the National Aeronautics and Space Ad- lation algorithm capable of estimating shifts to sub pixel accu- ministration under NASA Contract NAS5-26859. The authors are with the Laboratory for Applications of Remote racy. The algorithm implements the correlation coefficient Sensing, Purdue University, West Lafayette, IN 47906-1399. and uses cubic interpolation to estimate the peak to sub pixel 0196-2892/84/0500-0222$01.00 © 1984 IEEE ANUTA WVUTSRQPONMLKJIHGFEDCBA e t a l . : LANDSAT-4 MSS AND TM DATA QUALITY ANALYSIS ZYXWVUTSRQPONMLKJIHGFEDCBA 223 TABLE I CHANNEL-To-CHANNEL CALIBRATION EVALUATION DEVIATION OF DETECTOR MEAN FROM GRAND MEAN FOR WEBSTER COUNTY IOWA 'A' DATA; SCAN LINE SEQUENCE [)po te c to r 1 10 ihgfedcbaZYXWVUTSRQPONM II 12 13 14 15 16 B and -.08 - .4 .14 - .04 .29 .32 -.38 .44 .35 .07 -.1 -.1 .49 - .24 .2 - .29 .53 -.41 .35 -.47 .73 - .08 .06 .18 .2 .03 -.41 -.55 -.55 - .02 .42 .05 .45 -.13 .19 - .05 .1 '.77 TSRQPONMLKJIHGFEDCBA -.1 4 - .ZCBA ·.ll - .14 •. Z 3 .Z 1 ·.39 .04 .43 .76 .2 .4 .42 .18 1.0 .9 .64 •. 1 ·.17 ·.94 - .65 •. 59 -.38 •. 22 .08 -.17 .12 -.15 .ll - .48 .05 -.01 .49 - .25 ·.05 .04 ·.23 •. 44 _. C~ -.44 1.29 .JJ 6 - .29 0.0 .32 .04 .69 .25 .14 .36 •. 21 .61 .17 .02 ·.39 .18 .26 •. 35 accuracy. The means and standard deviations of shifts for 32 TABLE II sets of lines in TM imagery of the Iowa and Arkansas sites MAXIMUM DEVIATIONS OF MEAN AND VARIANCE OF ANY DETECTOR FOR MSS BANDS were computed. Results indicate that the primary focal plane bands (1, 2, 3, 4) are registered to 0.1 pixel, which is well W E B S TE R CO. C H IC A G O within the 0.2 pixel specification. Bands 5 and 7 are also well A D ata A D ata registered to within the 0.2 pixel specification. Problems were discovered, however, in the registration of the first band set (1, 2, 3, 4) to Bands 5 and 7. Approximately 0.6 B and M ean S td. A Dev. M ean S td. Dev. pixel misregistration was observed between these bands, which exceeds the specification of 0.3. The thermal IR band could not be numerically correlated due to the great differences in -.28 .~3 .2 .16 the nature of the image relative to any of the other bands, so a visual blink comparison was used on a digital display unit. A 2 .30 -.28 .2 .25 four 30 M pixel misregistration was observed in both Arkansas .82 .29 .13 .23 and Iowa thermal data. The MSS band-to-band registration was similarly evaluated .~ ~ .2 .11 - .12 and shifts were found to be less than the 0.2-pixel tolerance for the A data, but slightly greater than the tolerance for P data. The maximum shift of 0.27 was between bands 1 and 2. the visual inspection. The frequencies were all too high, how- ever, to be caused by jitter. A sampling, power supply, or DETECTOR-TO-DETECTOR CALIBRATION data processing error is suspected as the cause. Thematic Mapper detector-to-detector calibration was evalu- No visual noise effects were apparent in the thematic mapper ated by computing mean values of sets of scan lines spaced 16 data. However, Fourier analysis of low level Lake Michigan lines apart in A-tape data. Results are presented in Table I. data revealed two coherent noise frequencies in some of the Differences as large as 1.3 digital count were observed, but bands. Fig. 1 contains a plot of the power spectrum of data the typical deviations from the overall mean were less than 0.5 from band 4 detector 1 and two noise frequencies are ap- digital count. Those results indicate that striping should be parent. One is at a wavelength of 17 pixels and the other is at minimal in TM data. This is verified by visual inspection of 3.12 pixels. The power in these frequencies is 10 to 20 dB (10 the imagery. to 100 times) above the background noise level, but relative Similar tests were run on MSS data. The results are pre- to the power in the bright areas of the scene the noise power sented in Table II in terms of maximum deviation from overall is very low. The amplitude of the 3.12 pixel noise signal is means and variances. The values are all below one bin, indi- 24 dB and the low frequency power in the Sept. 23, 1984 cating that relative calibration is very good. Iowa TM data is in the 60-dB range, thus the noise is more than 30 dB less than the image power. This corresponds to a COHERENT NOISE ANALYSIS standard deviation of nominally 30 times less than the signal. Visual inspection of MSS imagery over Lake Michigan re- Band 4 standard deviation is 21 counts for typical sites in the vealed significant noise patterns in all bands. These wavelike Iowa data, thus the amplitude of the noise is nominally 0.6 patterns had a period of 14 to IS samples with higher fre- count. These noise frequencies exist in all detectors of band quency patterns superimposed. The magnitude of the noise is 4. This noise was observed in bands 1, 2, 3 also, but not in 4 to 5 counts peak-to-peak and it exists in all bands. It was bands 5, 6, and 7. first assumed that this noise was a jitter effect. Fourier anal- The MSS and TM noise frequencies are nominally the same ysis was performed on samples of the noise patterns and sev- suggesting a common cause, but no explanation is advanced eral significant frequencies were observed. The largest had here. These results are presented as evidence to aid researchers wavelengths of 3.31, 4.09, and 15.4 pixels which agreed with and data processing agencies in correction and refinement of 224 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. GE-22, NO.3, MAY 1984TSRQPO 3 5 .0 0 Lake Michigan, RQPONMLKJIHGFEDCBA A-tape use. In particular, agricultural field edges for high contrast fields with uniform texture were used. To determine the PSF, subpixel sampling is required since the sample rate of the sen- ihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA 2 8 .0 0 sor is nominally the width of the PSF. A procedure was de- r-r fined to allow subpixel sampling of edges. l:i 2 1 .0 0 Because of the orbital inclination of Landsat and the pro- LJ - pensity of man to arrange linear features, such as roads and ~'" •• 1 '1 .0 0 field boundaries, in the cardinal compass directions, it is gen- 0 N A erally found that there is a spatial displacement in the scene 7 .0 0 0 coordinates of the linear elements from one row or column to the next. This has the desirable effect of providing a fine o WVUTSRQPONMLKJIHGFEDCBA +----,---~--,...--__r-~ .0 0 0 grid of samples of the system response when values from .013 .016 0 .0 0 .0 0 3 .0 0 6 .0 1 0 adjacent rows or columns are combined after correction for Frequency [cycles/mJ Spectral Band y, detector 1 the spatial shift of the scene element. The procedure for Fig. 1. Power spectrum of low-level TM data showing noise peaks at combining the data is quite straightforward and can be illus- two frequencies: 3.12 pixels and 17 pixels per cycle. trated as follows for a north-south road. The coordinates of the peaks in the row data corresponding to the road are determined for a sequence of N rows. These the sensor and data processing stream to aid in assuring maxi- data are then fitted with a least-squares straightline, providing mum quality of Landsat-4 data. an analytical expression for the road coordinates. The x - RESOLUTION ESTIMATION IN TM DATA coordinates in each row are then modified by subtracting from them the least-squares estimate of the road location in that One geometric parameter of particular interest is the actual row. TIns converts the data to a coordinate system in which in-flight resolution of the sensor system which includes all en- zero is the road center. Because of the small angular differ- vironmental and data processing factors. The resolution is de- ence between the sensor coordinate system and the road di- termined by the point-spread function (PSF) of the system. rection, the change in road coordinates from row to row is This can be estimated in image data by observing the measured only a fraction of the pixel spacing and so represents a sam- response to scene elements of known shape. pled response to the scene element corresponding to a sub pixel The measured data can be expressed in the spatial domain translation. By combining the data from a number of rows, a as a convolution of the scene with an overall point-spread set of finely sampled data is found. These data can be gradu- functionCBA ated using splines or other smoothing functions to give an g (x ,y )= h (x ,y ) * [(x ,y ) average response function from which to estimate the system point-spread function. where [ ( x , y ) is the earth scene, h ( x , y ) is the overall point- The subpixel scheme was applied to field edges in the Iowa spread function of the sensor system, and g ( x , y ) is the result- data set. A dense set of samples was obtained for several edges ing image. in TM bands 4 and 5. The PSF was estimated in earlier work Given g ( x , y ) , we wish to determine h ( x , y ) . To do this, [3] using a finite sum of basis functions. The method used some deterministic element of the input [ ( x , y ) must be here is based on the following observations: known or assumed. Although the theory can take into ac- The scene element can be modeled as a step superimposed count the two-dimensional nature of the element, the initial on a constant background, i.e., experiments have been limited to the one-dimensional case. If the overall PSF is separable, i.e., if h ( x , y ) can be written [(x ) = A + B u (x ) as a product h e x ) h e y ) , then this approach provides a direct where A is the background and B is the amplitude change estimate of the two components. Otherwise it generates cross across the boundary. The output is then given by sections through the two-dimensional PSF along the x- and j-axes. g ( x ) = [A + B u ( x ) ] * h e x ). Three scene elements that would be useful for this type of Taking the derivative of both sides of this equation gives analysis are: g '( x ) = (A + B u (x ) * hex) 1) An impulse represented by a narrow-width discontinuity along a row or column of the data. = B 8 (x ) * hex) 2) A step function represented by an abrupt change in gray = B h (x ) level along a row or column of the data. 3) A rectangular pulse represented by a sequence of two h e x ) = g '( x ) / B . steps in opposite directions along a row or column of the The PSF is therefore the derivative of the measured step func- data. tion. The measured subpixel sampled step data were smoothed It was observed by experimentation with the three types of to produce a well-behaved derivative. If this had not been scene elements that the step edge was the most desirable to done, the derivative would have been excessively noisy. Two AN UTAWVUTSRQPONMLKJIHGFEDCBA e t a l . : LANDSAT-4 MSS AND TM DATA QUALITY ANALYSIS 225 1.00'0' RQPONMLKJIHGFEDCBA TABLE III RESULTS OF T M RESOLUTION ESTIMATION USING STEP DIFFERENTIATION ZYXWVUTSRQPO .(..J o e ihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA TRIGONOMETRIC SMOOTHING SPLINE SMOOTHING • (.!I O D W (P ix.) C arr. W (M ) K nots W (P ix.) C arr. W (M ) '" -c .. OJ . ~::lOO o, B and 4 1. 60 1.114 43.2 1.13 1.31 39.3 E <I .l:.Z O O B and 5 1. 62 1.46 43.8 1.29 1.49 44.7 0 .0 0 0 TSRQPONMLKJIHGFEDCBA due to scene variation and both linear and multiple regressions -.D O D -------r-.----r-~-...., -]0 .0 - < '. 0 0 -C.DD E .O O '.0 0 10.0 failed to produce significant equations for all bands (Table IV), d but a general decline in response does seem to exist for some Fig. 2. TM band 4 point spread function estimate using trigonometric if not all of the bands, as indicated by a negative slope ob- smoothing with 29 coefficients. tained for all regressions from both reverse and forward scans for all bands. No significant difference between the forward methods of smoothing were used: 1) Trigonometric poly- and reverse scans for any band was found for the Iowa TM A nomials, and 2) cubic spline functions [4]_ Tape using an alpha level = 0.05. A simulation experiment was conducted using an ideal edge DIMENSIONALITY OF TM VERSUS MSS with Gaussian noise to test the smoothing methods. A test PSF was convolved with the noisy edge and the result was The spectral dimensionality of the TM sensor is of great in- smoothed and the PSF width was determined by differentia- terest since it potentially represents nearly double the infor- tion. It was observed that the spline smoothing tended to mation over the MSS. The dimensionality was investigated by underestimate the width by 10 percent and the trigonometric applying the principal components transformation to test smoothing overestimated by 15 percent. The results were used data. as correction factors when the method was applied to real One of the major advantages of using orthogonal linear trans- data. formations, such as the Principal Component or Karhunen- Field-edge data for two cases were extracted from the Iowa Loeve transformation, is their ability to compress overall scene and the derivative PSF estimation scheme was applied. data variance in a multifeature space onto a relatively few The resulting estimated point spread function for band 4 is transformed orthogonal axes. This essentially separates non- shown in Fig. 2. The results for the two types of smoothing random variance (information content) from random variance are listed in Table III. The width measure is the half ampli- (noise) while, concurrently, uncorrelating the transformed tude of the estimated PSF. These results indicate that the axes in such a way that any information redundancy (due to actual overall resolution is less than the optical system resolu- interband correlation) is eliminated. Such transformations tion. This would be due to atmospheric and data processing therefore allow a compression of the data into a fewer num- effects or to the characteristics of the field edge. The effect ber of dimensions while retaining a maximum amount of sig- of smoothing may also be significant but the simulation ex- nificant information content and removing much of the ran- periment provided a correction which is believed to compen- dom variance or noise from the data. sate for this. These results arc based on a limited sample Principal components were generated for both the MSS and under one atmospheric state and further estimation research TM P-tape data sets from the Chicago O'Hare test site. Statis- must be carried out to determine the variability of the resolu- tics used in calculating the principal components were gen- tion estimates. erated from data samples of the original MSS and TM data sets from every fifth line and fifth column and all bands. Since SCAN ANG LE EI·TECTS the sum of the eigenvalues is equal to the trace of the original The effect of scan-look angle on the recorded digital re- covariance matrix, i.e., the total variance, the importance or sponse value was studied using the TM A tape from Iowa percent of total variance explained by eigenvectorCBA 0 ' 1 ) is (Scene 10:40049-1(264). Average digital values calculated given by: from 3 blocks each of dimensions 16 lines (detectors) by 10 x, columns were obtained for both the forward and reverse scans; tr S each set of 3 blocks were selected at intervals of 100 columns for each of the ref1ective bands. Fig. 3 shows a plot of the where average digital response values for the forward and reverse AI = eigenvalue i or characteristic root of the ith component scans for band I. This figure shows a general decline in aver- tr S is the trace of covariance matrix S or total variance. age response across-track, with a total drop of about 4 to 5 digital counts. Some of the other channels displayed similar, Therefore, each of the eigenvalues of the ordered components although somewhat lower and less distinct drops in digital divided by the sum of the eigenvalues represents the amount counts across-track. Much of the variance across-track was of total data variance or information content accounted for by  226 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. GE·22, NO.3, MAY 1984ihgfedcba S C A N A N G LE RESPONSE (FO R W A R D & REVERSE SCANS) [0 W A flHl rAPE f RUN. 82000202 & 820002031 CHflNNELol 600 l200 l800 2.t(J0 3000 3600 42.00 4800 5.t(J0 COWHN RQPONMLKJIHGFEDCBA lEENO. rRACK ••......•••.... F U R t.lR R O g.{]-El REVERSE Fig. 3. Mean scan angle response for band 1 showing 5 digital count variation over width of frame. TABLE IV TABLE V LINEAR REGRESSION RESULTS OF THE FORWARD AND REVERSE SCANS TABLE OF EIGENVECTORS (PRINCIPAL COMPONENTS) AND EIGENVALUES CBA IOWA ' A ' TAPE (SCENE ID: 40049-16264) OF THE MSS-'P' TAPE DATA FOR THE CHICAGO O'HARE TEST SITE (SCENE ID: 40101-10625) ~ Total drop in digital ZYXWVUTSRQPONMLKJIHGFEDCBA Channel Scan R Slope _ c o u n t across-scan Matrix of Eigenvector's -------- Forward 0.413 -0.00087 5.22 Princ:tpal Component (Eigenvector) Reverse 0.261 -0.00067 4.02 W ave- length 2 Forward 0.153 -0.0003 1.8 Band 1 2 __ 3_ 4 1. 32 --- Reverse 0.041 -0.00022 0.39291 -0.41975 0.36717 0.73117 Forward 0.041 -0.00022 1. 32 Reverse * 2 0.51763 -0.57144 -0.54223 -0.33392 Forward 0.06 -0.00136 8.16 0.64029 0.31652 0.54675 -0.43693 Reverse 0.148 -0.00217 13.02 0.40951 0.63015 -0.52176 0.40371 Forward 0.127 -0.00127 7. 62 Reverse 0.275 -0.00172 10.32 Percent CumuIa t t ve Forward 0.058 -0.00044 2.64 Eigenvalue Variance Percentage Reverse 0.077 -0.00046 68.83 65.30% 65.30% • No significant regression could be found. 2 33.18 31.48% 96.78% 1. 98 1.88% 98.66% each eigenvector. Tables V and VI list the eigenvalues and the 1. 42 1.34% 100.00% corresponding amount of data variance that is accounted for by their respective eigenvectors for both the MSS and the TM data sets. As shown in these tables, the first two principal components of the MSS Pvtape account for almost 97 percent images of the TM principal components also show that the of the total data variance. This is also evident from the images first four components contain a significant amount of scene of the ordered MSS principal components (Fig. 4). The first contrast with a sequential decrease in the lower-ordered com- two components contain a significant amount of scene con- ponents (Fig. 5). trast and data structure, while the last two MSS components From both the listed eigenvalues of the MSS and TM data contain mostly random noise. sets, it is possible to conclude that the first two components In the TM data set, the first two ordered components to- of the MSS and the first four components of the TM data gether account for 90 percent of the data variance, the first sets contain a maximum amount of significant information. three account for 97.45 percent, and the first four cornpo- It could be possible, therefore, to conclude that for this data nents account for almost 99 percent of the data variance. The set of the Chicago O'Hare area, the TM data contain more sig- ANUTAWVUTSRQPONMLKJIHGFEDCBA e t a l . : LANDSAT-4 MSS AND TM DATA QUALITY ANALYSIS 227 TABLE VI TABLE OF EIGENVECTORS (PRINCIPLE COMPONENTS AND EIGENVALUES OF THE TM 'P' TAPE DATA FOR THE CHICAGO O'HARE TEST SITE (SCENE ID: 40101-10625) Principal Component (E1 genvector)~ ihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA W aveband 1 2 _3_ 5 6 7 " 0.389 -0. "8" 0.335 -0.1011 -0.213 -<).666 -0.057 0.258 -D .21111 -0.189 -0.009 -0.020 0.361 0.8", 0.376 -0.357 0.173 -0.065 -0.051 0.639 -0.535 0.231 0.599 0.728 0.070 O.221l -0.013 -0.0"8 0.639 o. "56 -0. "36 -0.157 -O.1l10 -0.025 0.032 0.421 -0.103 -0.321 0.25" 0.792 -0.133 -0.005 O.OIJ7 -0.021 O. OlD 0.9"" -0.325 0.007 -0.025 Eige~~ctor: E1genval~,,: Pet of Var Cum P.!:..!:. --- "02.90 71.8"% 71.8"% 102.29 18.2" 90.07 u 1. 37 7.38 97. "5 5.90 1. 05 98.50 ".96 0.88 99.39 2.20 0.39 99.78 1.211 0.22 100.00 nificant information, i.e., a higher significant dimensionality, addition, Fig. 8 graphically shows the percent of total data than the corresponding MSS data set. The significant dimen- variance accounted for by each eigenvector (principal compo- sionality as defined by a principal component analysis may be nent) of the Chicago O'Hare and Des Moines, IA TM data sets. substantially different for other data sets and other ground It is interesting to compare these two data sets in that they cover types. However, these results do imply that for the same demonstrate the importance of different bands for different area, the TM may contain more significant information than data sets of varying cover type. The first eigenvector or com- the corresponding MSS. ponent of the O'Hare data set shows relatively equal weighting of all of the original reflective bands (Fig. 6). By comparison, PRINCIPAL COMPONENT COMPARISON the greater amount of vegetation (primarily agricultural crops C h i c a g o O 'H a r e a n d D e s M o i n e s , I o w a T M D a ta and forest areas) associated with the Des Moines TM data set In order to investigate both the significant dimensionality has resulted in a significant amount of the total variance (ac- and the importance of the individual TM bands for different counted for by PC1, Fig. 8) of the data to be contained on data sets, a principal component analysis was performed on band 4 and, to a lesser extent, band 5. The second component the Des Moines, Iowa TM data (Scene ID :40049-16264) and of the Des Moines TM data set contains more information compared with the principal components analysis of the from the rest of the bands. It is interesting to note that com- Chicago O'Hare TM data set discussed above. ponent four of both the Chicago O'Hare and Des Moines TM Since the coefficients of a particular principal component data sets has the greatest contribution from band 6 (thermal correspond to the direction cosines of the original bands to IR) (Figs. 6 and 7), which is clearly evident from a comparison the new component [5], it is possible to describe the relative of the two images, i.e., component 4 and the original band 6 direction of the new component axis in the feature space with (Fig. 5). This may imply that although the thermal IR band respect to the original bands. The greater the coefficient of a may not contain a significant amount of total data variance particular band to a specific component, the larger the cosine (the fourth component of both data sets describe only 1-2 and thus the smaller the angle subtended between the original percent of the total data variance), the information or variance feature axis (band) and the new component axis, i.e., the more may be distinctly unique from the rest of the bands. In addi- the new component axis lies in the direction of the original tion, although most of the significant variance was incorpo- feature axis (band). In this way it is possible to describe the rated on different bands for these two data sets (as described relative influence or "pull" of the original bands on each of by the first eigenvector for each set), both sets indicate a sig- the new component axes. nificant dimensionality of between 3 and 4 dimensions (Fig. 8). Using the coefficients of the higher-ordered principal compo- nents, i.e., those components which incorporate most of the SPECTRAL CLASS ANALYSIS IN TM AND MSS DATA total data variance, it is possible to describe which of the origi- A detailed spectral analysis was conducted of thematic map- nal bands contain most of the significant variance or informa- per and MSS data for an area near Des Moines, Iowa from the tion content for a particular da ta set. Figs. 6 and 7 graphically September 3, 1982 data set. Data were utilized from 7 blocks show the loadings or coefficients of the Chicago O'Hare air- distribu ted throughou t the area which included agricultural, port and the Des Moines, IA TM data sets, respectively. In forest, suburban, urban, and water scene types. The blocks 228 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. GE-22, NO.3, MAY 1984ZYXWVUTSRQ MSS Band 1 (0.5 - 0.6~m) MSS Band 2 (0.6 - 0.7pm) MSS Band 3 (0.7 -ihgfedcbaZYXWVUTSRQPONML a.8I1m ) MSS Band 4 (0.8 - l lum) v MSS PCl MSS PC2 MSS PC3 MSS PC4 Fig. 4. Images of bands 1-4 and principal components 1-4 of Landsat-a MSS of the Chicago O'Hare test site. were processed using a clustering algorithm to produce up to half of the TM result. This result is considered to be a very 18 cluster groupings for each block. Each cluster class was significant indicator of the dimensionality of TM relative to then identified with a ground-cover class using aerial photog- MSS. The MSS class occurrences are indicated in Table VII. raphy and maps of the area. The cluster classes from each of The maximum divergence values of anyone class with respect the 7 blocks were inspected with regard to separability, means, to all others also were much less for the MSS classes relative variances, and number of points in each class were either de- to TM. Table VIII contains the maximum and average trans- leted or pooled with spectrally similar clusters [19]. The formed divergence values for the 42 spectral classes and for separability measure used in the transformed divergence func- the best su bsets of TM spectral bands. It should be noted that tion or processor [6] measures the statistical distance between the best spectral band for any com bina tion of bands 1-7 is classes based on class means and covariance matrices. The the first middle IR band (1.55-1.75f.1m). The next best band measure has a maximum value of 2000 and the minimum of O. is the near IR (0.76-0.90CBA f . 1 m ) , followed by the red band and Spectrally, very close classes will typically have values as low then the thermal IR. The best combination of four bands in- . as 50 to 500. cludes one from each of the four regions of the spectrum For the TM data, initially 94 classes were defined and the (visible, near IR, middle IR, and thermal IR). Table IX con- pooling and deleting process reduced these to 42 final spec- tains the minimum and average transformed divergence values trally separable classes. Table VII lists these classes. The MSS for the best combination of MSS bands. data were then analyzed using the same clustering and merging The high average divergence indicates that the 21 spectral sequence. The number of separable classes in the MSS is 21, classes found in the MSS were about as separable as the TM ANUTAWVUTSRQPONMLKJIHGFEDCBA e ta l.: LANDSAT-4 MSS AND TM DATA QUALITY ANALYSISZYXWVUTSRQPONMLKJIHGFEDCBA 229 T/>1 Rand 1 (l).lo5 - ihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA n.Slum' TM B and 2 (0.52 - O .60um ) 11-1 Band 3 (0.63 - O.69urn) 1M B and ~ (0.76 - O .90um ) TH Band 5 (1.55 - 1.75um) T M B and 7 (2.08 - 2.3S um ) TN Band 6 (10.40 - 12.50\1m) TM pel HI pe2 TM PC3 TM PC4 TM pes TM PC6 11-1 PC7 Fig. 5. Images of bands 1-7 and principal components 1-7 of Landsat-4 TM of the Chicago O'Hare test site. 230 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. GE-22, NO_ 3, MAY 1984 TM-PCl H1-PCl B ands Bands -1 -1 rx-scz T M -pe2 eo ihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA . WVUTSRQPONMLKJIHGFEDCBA .~ -o s B ands -1 B ands -1 TM -P C 3 TM -P C 3 B ands -1 -1 B ands B ands B ands -1 -1 rs-scs T!'I-PCS B ands B ands -1 -1 TM -P C 6 TM -P C 6 B ands -1 -1 B ands TM -P C 7 TM-PCl ec . 5 "o -' B ands B ands -1 -1 Fig. 6. The loadings or coefficients of the original wavelength bands Fig. 7. The loadings of coefficients of the original wavelength bands on on each of the ordered TM principal components of the Chicago each of the ordered TM principal components of the Des Moines, O'Hare test site. Iowa test site. ANUTA WVUTSRQPONMLKJIHGFEDCBA e t a l . : LANDSAT·4 MSS AND TM DATA QUALITY ANALYSIS 231 TM-Principal Components TABLE VII SPECTRALLY SEPARABLE CLASSES IN TM AND DATA OF (Des Moines. Iowa) DES MOINES, IOWA AREA S4.44 CLASS TM CLASS EXISTS CLASS T/1 CLASS EXISTS NUMBER NAME in MSS NAME NAME in MSS ZYXWVUTSRQPONML Forestl 22 Substation 2 Forest2 x 23 Quarry Cornl x 24 Concrete x Corn2 25 Sludge x L02 0.71 0.20 Soyl ihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA x 26 Industrial1 0.11 Soy2 x 27 Industrial2 PCl PC2 PC3 PC4 PCS PC6 PC7 Urban/Hiway Soy3 x 28 8 Soy4 x 29 Soil/Hiway Soy5 x 30 Residentiall 7L84 TM-Principal Components 10 Soy6 x 31 • Residentia12 x (Chicago 0 'Hare) Beachl 11 Wheat R esi.due x 32 12 Grassl x 33 Beach2 13 Grass2 34 Beach3 14 Grass3 35 Soilwetl 15 Soil/Vegl 36 Soi.1wet2 16 Soil/Veg2 37 M a rsh 17 Soil/Veg3 38 Waterl x LOS 0.88 0.39 0.22 18 Farm/Grass 39 Water2 x 19 Road/Farm 40 Water3 x PCl PC2 PC3 PC4 PCS PC6 PC7 20 Baresoill x 41 Water4 x Fig. 8. Percent of total variance accounted for by each of the ordered TM principal components of the Des Moines, Iowa and Chicago 21 Baresoi12 42 Water5 O'Hare test data sets. classes. However, there was twice the number of equally sepa- TABLE VIII SEPARABILITY (TRANSFORMED DIVERGENCE) FOR 42 CLASSES IN THEMATIC rable TM classes. This is considered to be the most significant MAPPER DATA result of the spectral analysis. CHANNEL COM- DIVERGENCE BEST A final test using a preliminary set of test data was carried BINATIONS MIN. AVE. CHANNELS out using the small amount of ground truth available. A set ---- ----~- 1574 of 5615 TM and 1376 MSS pixels containing forest, corn, 2 210 1880 4 5 soybean, soil, water, and urban classes was extracted from the TM and MSS data where the cover classes were known or 522 1949 3 4 could be inferred from aerial photography. The overall cor- 1090 1973 3 4 5 6 rect recognition was 95.7 percent for the TM classification 1356 1979 3 4 567 using all 7 bands, 92.6 percent for the TM using a subset of 1405 1983 2 3 4 567 the best four bands, and 67.4 percent for the MSS classifica- tion using all 4 bands and a per-point, Gaussian Maximum 1553 1986 1 234 567 Likelihood (GML) classifier. The results are listed in Table X -------------- along with the amount of CPU time required to classify each data set. the fact that since cover classes are more likely to occur in An additional classification using all 7 TM bands was per- homogeneous areas larger than one pixel in size (i.e., larger formed on the test data using a contextual algorithm, SECHO than 30 m), adjacent pixels are highly correlated, with the de- (Supervised Extraction and Classification of Homogeneous gree of correlation diminishing with an increasing distance Objects), in order to demonstrate the effectiveness of such between the pixels [7]. Thus SECHO assigns an analyst- contextual algorithms over per-point algorithms for use with specified threshold value, below which adjacent pixels will the 30-m resolution TM data. The contextual or per-field be grouped into a homogeneous field. Statistics for these algorithm SECHO first divides the scene to be classified into fields are calculated and compared to the original cover class homogeneous fields and then classifies these fields using an statistics and a "homogeneous field" is classified as a unit into extension of the GML algorithm [7]. SECHO incorporates that class which it most closely resembles. 232 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. GE-22, NO.3, MAY 1984 TABLE IX TEMPERATURE MAPPING OF A COOLING POND SEPARABILITY FOR 21 CLASSES IN MSS DATA AND THERMAL PLUME FROM A NUCLEAR CHANNEL DIVERGENCE POWER PLANT COMBINATIONS MINIMUM AVERAGE ---- 32 1842 The use of thermal infrared (IR) data in conjunction with reo 2 3 730 1957 flective multispectral data, obtained from aircraft altitudes, 1968 has been proven to be an effective means for increasing the 2 3 4 1032 classification accuracy of earth surface features [9] , [10]. 1 2 3 4 1112 1973 Also it has been demonstrated that radiometrically cali- brated thermal IR data obtained from airborne platforms can be used to produce accurate temperature maps of water bodies TABLE X CLASSIFICATION ACCURACY AND CPU TIME COMPARISON 0 TEST [11]-[14]. DATA IN THE DES MOINES IOWA AREA (SCENE 1 0 : 40049-16264) ihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA Accurate temperature maps of water bodies have been ob- TH TM TM M SS tained also from calibrated thermal JR data gathered in 1973 GMt Per-Poi nt GMt Per-Pol nt SECHO G M t P e r-P o in t C'l as s t r t et- C'l as s t r t er- Classifier Classtfier by the Skylab SL-2 S 192 scanner system [15]. However, ( All 7 B ands ) (B e st 4 B ands ) ( All 7 B ands) ( All 4 B ands) %Corre~ %Correct %Correct ~rec~_ for almost ten years the remote sensing data users have not CLASS --~~ ---- Forest 99.0 97.1 100. a 91.2 had available high spatial resolution thermal JR data (in con- C orn 92.0 76.8 97.7 30.8 trast to the coarse spatial resolution data gathered by meteor- S oybeans 100.0 99.8 100.0 99.3 ological satellites) acquired from space borne platforms, until Bare Soil 99.7 99.0 100.0 55.6 the TM thermal data were recen tly collected. Grass 96.8 87.6 98.1 1.9 In order to determine the radiometric quality of the TM W a te r 100.0 96.8 100.0 98.9 thermal data for temperature mapping of surface water, a test U rb a n 91.7 99.9 95.8 50.2 site was selected within the area covered by the TM scene Overall 95.7 92.6 97. 9 67.4 (Scene ID: 40101-16025) gathered over Illinois. This site was chosen because it includes a surface water body with a large range of temperatures, i.e., a cooling pond for the Dresden N o. of No. of pt xe l s N o. of nuclear power plant and the junction of two rivers: the Kanka- Classes Classi....!:..i~ bands C P U T im e (h o u rs ) kee and the Des Plaines Rivers which once merged form the M SS 21 250,000 0.4 Illinois River. TH 42 1,000,000 2.7 The Dresden power generating station is located approxi- TM 42 ',000,000 7.9 mately 50 miles southwest of Chicago near Morris, Illinois. The 'Station houses three nuclear reactors which together have a maximum nameplate generating capacity of 1 656630 kW. These "test" fields are limited in number of pixels and so The water which is taken from the Kankakee River and is used are not really evauating how representative the final spectral to cool the steam condensers is channeled into a man-made classes are of the entire scene but rather how separable the lake which is designed to dissipate heat. This l300-acre lake has classes are. Deletion of certain spectral classes (e.g., Corn 2), a dike in its center which directs the water flow up, around, due to low separability, resulted in much confusion of corn and down the lake in an 8-mile loop that takes two-and-a-half with trees in the MSS but not the TM for both the best four days to complete. During its course the water is cooled by and for all seven TM bands. Also, the resolution of TM natural evaporation and is either recycled back through the actually allowed "purer" cluster classes to be defined since station or discharged into the Illinois River. A schematic dia- smaller areas (e.g., beaches, roads) were distinct. These re- gram of the power plant, channels, cooling lake, and the three sults along with the listed CPU time for classification indi- rivers is shown in Fig. 9. cate that a subset of the best four TM bands incorporates the Because the radiant temperature of the various cover types advantages of the higher spatial resolution (e.g., "purer" clus- in the test site is a function not only of the kinetic tempera- ter classes) over the equivalent MSS data without simulta- ture of the materials, but also a function of the intrinsic emis- neously incurring as substantial an increase in computer time sive properties (emissivity) of the objects, and to avoid differ- required for classification as with all 7 bands. In addition, ential emissivity complications, it was decided to perform the these test results support the use of contextual algorithms, calibration of the TM thermal IR data corresponding to only such as SECHO, over per-point algorithms for use with TM those pixels representing water, which has an emissivity close data. An earlier study using simulated TM data [8] demon- to that of a perfect radiator or blackbody. strated that higher classification performances could be ex- To accomplish the thermal calibration of water only, a pected with the SECHO classifier for the 30-m TM data, es- heirarchical classification of the TM data had to be performed, pecially in those areas where the cover class field sizes are i.e., one of the middle IR (1.55-1.75WVUTSRQPONMLKJIHGFEDCBA p m ) bands was utilized relatively large, relative to the scanner FOV, e.g., most agri- to discriminate the water pixels from all the other cover types cultural areas. using the LARSYS layered classifier [12], [16], [17]. Once ANUTA WVUTSRQPONMLKJIHGFEDCBA e ta /.: L A N D S A T -4 M SS AND TM DATA Q U A L IT Y A N A L Y S IS 233 ILLINOIS RIVERZYXWVUTSRQPONMLKJIHGFEDCBA DRESDEN COOLING LAKEzyxwvutsrqponmlkjihgfedcbaZYXWVUTSRQPONMLK Fig. 9. Schematic diagram of the Dresden Nuclear Power Plant, channels, cooling lake, and the Kankakee, Des Plaines, and Illinois River.RQPONMLKJIHGFEDCBA WATER TEMPERATURE MAPPING LAYERED C l..A sS IF 1 C A T leJ o I T>£JlllTlC ~. MTA VEG L VEG 2, VEG 3, SOILS. WATER M AX, L IK E L IH O O D TM BANO 5 C L A S S IF IC A T IO N VEG L VEG 2, VEG 3. SOILS E£e LEVEL TM BAND 6 SLICING & CAL I BRAT 1 ON TEM P .20 Fig. 10. Decision tree utilized to carry out the hierarchical classification and calibration of the water bodies present in the Dresden Nuclear Power Plan t test site. all the water pixels were separated from everything else, their plotted versus the corresponding temperatures, as illustrated thermal IR responses (relative digital counts) were converted in Fig. 13. Note in this figure that if one assumes a linear rela- to radiant temperatures. Fig. 10 shows the decision tree tionship between the emitted energy from a blackbody and its utilized to carry out the hierarchical or layered classification temperature for a range of temperatures in excess of 10 K (the and calibration of the water bodies present in the test site. range between the cold and hot internal calibration sources of Figs. II and 12 show the images of the test site corresponding the TM system is 60 K), errors on the order of 4 K ( 4 ° C ) to the 1.55-1.75 and 10.4-12.5 pm bands, respectively. would be introduced. The conversion of the relative thermal IR responses into The nonlinear calibration function is given by (I) below radian t temperature measurements was accom plished using a non linear calibration function derived specifically for the TM TeC ) = - 12.5809 + 0.2917D - 0 .0 0 0 2 3 3 D 2 (1) thermal IR band and the range of temperatures of the two in- ternal calibration reference blackbodies. where T is the temperature in degre.es Celsius and D is the The nonlinear calibration function was derived through the relative digital counts for the TM thermal IR band. integration of Planck's equation for a spectral band covering If a linear interpolation was performed, the linear calibra- the 10.4-12.5 pm spectral range, and for a range of tempera- tion function would be represented by (2) tures between 260 and 320 K [18] at increments of 0.1 K. The resulting in-band radiances (in Watts/cm ' -sr) were then TeC ) = -13 + 0 .2 3 5 2 9 D . (2) 234 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. GE-22, NO. 3, MAY 1984 Fig. 11. Image of band 5 (1.55-1.75 urn) of Landsat-4 TM of the Fig. 12. Image of band 6CBA ( 1 0 .4 - 1 2 .5 - llm ) of Landsat-4 TM of the Dresden Power Plant test site. Dresden Power Plant test site.ihgfedcbaZYX INTEGRATION OF PLANCK'S EQUATION WVUTSRQPONMLKJIHGFEDCBA II 0' 11 8' '0 ' + " BAND: 10,4 - 12,5 UM .. .. + '0 TEM PERATURE RANG E: 250 - 320K o RQPONMLKJIHGFEDCBA 2 2 4 -+ + " e:; 0 20Q-+ ..... + 36 N~ ~ 0 TSRQPONMLKJIHGFEDCBA tq 3 -+ :,. 31 '"~ ~ 0 1 7 8 -. + 2. c ~ 0 lo 3 - ~ - - - - - - - - - - - - :7 lr - :~ ' + 21 ... ... o:l i:. 0 1 4 1 -. + '6ZYXWVUTSRQPONMLKJIHG O. 1 3 2 - .0 - + II o 1 1 7 -. o 101- , . + ,. ! , , . + .. ! ...., .• of .+ , , ze 000 27 200 :2 8 4 0 0 2q 000 3 0 .2 0 0 30. BOO 31 400 32 000 TEM PERATURE (K) -2 IIE R T IC ••.. L SCALE FACTOR. 10 H O R T tlO N T A L SCALE F"A C TO A • 10 ~B E A P O IN T S "LO TTED " bO I R E S O L U T IO N J8. ONE U N IT ON '( - A X IS " O .3 0 1 E - 0 4 ONE U N tT ON 1-"'U9. 0 bO O E+O O P L O T T IN O ERROR IB PLUS OR " 'IN U S 1 /2 SCALE U N IT Fig. 13. Integration of Planck's equation for a spectral band covering the 1 0 . 4 - 1 2 . 5 - l l m spectral range, and for a range of temperatures between 260 and 320 K at 0.1 K increments. The pixels of water at the intake (Kankakee River) have a ence of approximately 1°C with respect to the radiant tem- relative thermal IR response equal to 95 digital counts, and perature derived from the TM thermal IR data. Fig. 14 shows those corresponding to the water at the entrance of the cool- a temperature map of the Dresden power plant test site. ing pond equal to 135 digital counts. If these two relative Although neither the temperature of the water recorded at thermal responses (digital counts) are substituted in (1), one the power plant nor the precise temperature of the internal obtains the following radiant temperatures: calibration reference blackbodies in the TM system was reli- able, the results obtained during this preliminary assessment D ig ita l R a d ia n t of the TM radiometric data quality are very encouraging. C ount T e m p e r a tu r e SUMMARY AND CONCLUSIONS 95 13.l °c Landsat-4 Thematic Mapper and Multispectral Scanner data 135 22.6°C quality was evaluated using data from several frames over Since at the time the TM data were recorded (October 25, midwestern United States. Four categories of evaluation were presented: 1982 at 16: 02 GMT) reference (ground truth) data were not being collected, it was not possible to assess reliably the 1) Basic radiometric and geometric characteristics. accuracy of the radiant temperatures derived from the TM 2) Overall TM system resolution. thermal IR data. However, the personnel from the Dresden 3) Spectral dimensionality. nuclear power plant keep a record of the temperature of the 4) Spectral class separability. water at the entrance to the cooling lake. The temperature 5) Use of TM thermal IR data for temperature mapping of recorded for that date was 71°F or 21.7°C, which is a differ- water bodies. AN UTAWVUTSRQPONMLKJIHGFEDCBA e t a l . : LANDSAT-4 MSS AND TM DATA QUALITY ANALYSISihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA 235 ACKNOWLEDGMENT The author's sincere appreciation is extended to Ms. K. Hunt, systems analyst at LARS, who carried out the data formatting and data-set preparation tasks necessary 'before any of the analysis work could be done. REFERENCES [1 ]L. A. Bartolucci, M. E. Dean, P. E. Anuta, "Evaluation of the radiometric quality of the TM using clustering and multispectral distance measures," in P r o c . L a n d s a t - 4 S c i e n t i f i c C h a r a c t e r i z a - t i o n E a r l y R e s u l t s S y m p . , NASA Goddard Space Flight Center, Greenbelt. MD. Feb. 22-24.1983. [2] Goddard Space Flight Center, Greenbelt, MD 20771, S p e c i f i c a - , RQPONMLKJIHGFEDCBA .,..> ;::. TSRQPONMLKJIHGFEDCBA t i o n s f o r t h e L a n d s a t - D S y s t e m , Revision C, Mar. 1981, pp. 3-88 . [3] C. D. McGillem, P. E. Anuta, E. Malaret, and K. B. Yu, "Estima- • ~..... #"!"!- ....• ,. -Q. 18.8-19.3~C tion of a remote sensing system point-spread function from mea- sured imagery," in P r o c . 9 t h I n t l . S y m p . o n M a c h i n e P r o c e s s i n g o f R e m o te ly S e n s e d D a t a , LARS/Purdue University, West Lafay- t t t t t t t t t t t t t t ette, IN, pp. 62-68, June 21-23, 1983. 2 3 3 '1 5 5 [4] C. D. McGillem, E. Malaret, and P. E. Anuta, "Resolution estima- B 2 6 '1 2 6 0 0 0 0 0 0 tion for the Landsat-4 Thematic Mapper," in P r o c . C a n t T e c h - n iq u e s fo r E x tr a c tio n of In fo r m a tio n fr o m R e m o te ly Sensed 1 M IL E Society of Photographic Scientists and Engineers, Spring- Im a g e s, field, MD, to be published. Fig. 14. Temperature map of the Dresden Nuclear Power Plant test site [5] D. F. Morrison, M u l t i v a r i a t e S t a t i s t i c a l M e t h o d s . New York: derived from the Landsat-4 TM data. McGraw-Hill,1967. [6] P. H. Swain, T. V. Robertson, and A. G. Wacker, "Comparison of the divergence and B-distance in feature selection," LARS Infor- The overall quality of the T M data is very good with only a mation Note 020871, LARS/Purdue University, 1971. few error conditions, which in the case of band-to-band mis- [7] R. L. Kettig and D. A. Landgrebe, "Classification of multispectral registration can be corrected by ground processing. image data by extraction and classification of homogeneous ob- jects," LARS Information Note 062375, LARS/Purdue Univer- Detector-to-detector calibration was very good, such that sity, 1975. striping should be minimal in TM. MSS demonstrated more [8] R. S. Latty, "Computer-based forest cover classification using noise problems, primarily due to high frequency wave pat- multispectral Scanner data of different spatial resolutions," LARS Tech. Rep. 052081, LARS/Purdue University, 1981. terns. In general, the MSS appeared somewhat noisier than [9] M. E. Bauer, P. H. Swain, R. P. Mroczynski, P. E. Anuta, and previous Landsat MSS data. R. B. MacDonald, "Detection of southern corn leaf blight by The resolution estimation experiments resulted in relatively remote sensing techniques," in P r o c . S e v e n t h I n t . S y m p . o n R e m o t e S e n s i n g o f t h e E n v i r o n . , vol. 1, pp. 693-704, May 17-21, wide half-amplitude widths for the point-spread function of 1971. the TM. The reason for this is not understood; it is likely due [10] R. Kumar and L. F. Silva, "Emission and reflectance from healthy to a combination of system and atmospheric effects plus the and stressed natural targets with computer analysis of spectro- radiometric and Multispectral Scanner data," LARS Information uncertainty in the actual shape of the step edges being used Note 072473, LARS/Purdue University, July 1973. as a test input. [11] B. H. Atwell, R. B. MacDonald, and L. A. Bartolucci, "Thermal Dimensionality was investigated using principal components mapping of streams from airborne radiometric scanning," Water Resources Bull., J . A m e r . W a t e r R e s o u r c e s A s s . , vol. 7, no. 2, analysis. The results indicate two significant dimensions in the Apr. 1971. MSS and up to four in the TM. The spectral analysis was the [i2] L. A. Bartolucci, R. M. Hoffer, and T. R. West, "Computer-aided most revealing by producing twice as many separable classes in processing of remotely-sensed data for temperature mapping of surface water from aircraft altitudes," LARS Information Note T M than in MSS. The result was due both to the greater num- 042373, LARS/Purdue University, ApI. 1973. ber of spectral bands and to greater spatial resolution of the [13] L. A. Bartolucci, R. M. Hoffer, and J. R. Gammon, "Effects of TM. More classes, such as beach, roads, field edges, and other altitude and wavelength band selection on remote measurements of water temperature," in P r o c . F i r s t P a n A m e r i c a n S y m p . o n small features, were separated out in T M where they are mix- R e m o t e S e n s i n g , Panama City, Panama, May 1973. tures in MSS and in many cases not spectrally distinct. These [14] R. M. Hoffer and L. A. Bartolucci, "Remote sensing techniques results further verify the high quality of the T M data and for measurements of water temperature," LARS Information Note 111671, LARS/Purdue University, 1971. promise significant increase in usefulness of these data in [15] L. A. Bartolucci, "Hydrologic features survey," ch, 4, LARS earth resources applications. Information Note 121275, LARS/Purdue University, 1975. The radiometric quality of the T M thermal IR band was [16] L. A. Bartolucci, P. H. Swain, and C. L. Wu, "Selective radiant temperature mapping using a layeted classifier," I E E E T r a n s . studied using both a linear calibration function and a non G e o s c i . E l e c t r o n . , vol. GE-14, no. 2, Apr. 1976, also LARS Infor- linear calibration function derived from the integration of mationNote 111175, 1975. [17] C. L. Wu, D. A. Landgrebe, and P. H. Swain, "The decision tree Plank's equation for a spectral band covering theCBA 1 0 A - 1 2 .5 - llm approach to classification," LARS Information Note 090174, spectral range, and for a range of temperatures between 260 LARS/Purdue University, Sept. 1974. and 320 K at increments of 0.1 K. These preliminary results [18] T. L. Barker, "TM sensor description and status," presented at First European Landsat-4 Workshop, ESA/EARTHNET Program, support the use of the nonlinear calibration function with the Frascati, Italy, Nov. 1982. thermal IR band for effective use of the thermal band for [19] S. M. Davis, "Polk County (Iowa) case study," LARS Tech. Rep. temperature mapping of water bodies.ZYXWVUTSRQPONMLKJIHGFEDCBA 073183, LARS/Purdue University, Aug. 1983. 236 IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. GE-22, NO.3, MAY 1984 Paul E. Anuta (S'55-M'56-M'61-SM'77) re- Erick Malaret received the B.S. degree in phys- ceived the B.S. degree in electrical engineering ics from the University of Puerto Rico in 1980 from Purdue University in 1957, the M.S.E.E. and the M.S.E.E. degree from Purdue University degree from the University of Connecticut in in 1983. He is currently working toward the 1962, and the M.S. degree in computer science Ph.D. degree in electrical engineering at Purdue from Purdue University in 1967. University. He joined the Laboratory for Applications He is a graduate research assistant at the of Remote Sensing (LARS), Purdue University Laboratory for Applications of Remote Sensing in 1967 and has researched data handling sys- (LARS), Purdue University. tems, interferometer spectrometer, and other sensors. He is currently Associate Program Leader for Data Handling Research at LARS. Key data research han- dling areas are: image registration, geometric correction, and resolution * enhancement of satellite multispectral imagery. His current interests Clare D. McGilIem (M'52-SM'73-F'75) received are in the area of multitype data integration and preprocessing and the B.S.E.E. degree from the University of analysis methods. Mr. Anuta is a member of Tau Beta Pi, Eta Kappa Michigan in 1947 and the M.S.E. and Ph.D. de- Nu, and the American Society of Photogrammetry.ihgfedcbaZYXWVUTSRQPONMLKJIHGFEDCBA grees from Purdue University in 1949 and 1955, respectively. * After a number of years in government and industry working in the area of aerospace elec- Luis A. Bartolucci received the B.S., M.S., and tronics, he joined the faculty of Purdue Univer- Ph.D. degrees in geophysics from Purdue sity in 1963. From 1968 to 1972 he served as University. Associate Dean of Engineering and Director of He has been involved in Remote Sensing the Engineering Experiment Station at Purdue. Research since 1969. He has played an active He is currently Professor of Electrical Engineering at Purdue and is a role in the development of remote sensing tech- member of the technical staff of LARS. He is active in research and nology for applications in the area of water teaching in communication theory, radar, and signal processing. resources and has also made outstanding con- tribu tions in the field of thermal infrared radia- tion for remote sensing applications. In addi- * tion, he has served as consultant to the U.S. Information Agency, the U.S. Agency for International Development, and the InterAmerican Development Bank and to several Latin Ameri- Jose A. Valdes studied agronomy at the Instituto can development agencies. He has been Principal Investigator and Tecnologico y de Estudios Superiores de Mon- Project Director of several domestic and international research and terrey, campus Queretaro, where he presented a training programs involving computer-aided processing and analysis Bachelor's thesis on the applications of remote of remotely sensed data for earth resources inventories. sensing to agronomic surveys. He is currently pursuing the M.S. degree in agronomical remote sensing at Purdue University. * He is currently a graduate research assistant at the Laboratory for Applications of Remote M . Ellen Dean received the B.S. degree in for- Sensing (LARS), Purdue University. He had estry from Virginia Polytechnic Institute and been a staff member of the Remote Sensing State University in 1979. She is currently pur- Department of the Direction General de Geografia del Territorio Na- suing the M.S. degree in remote sensing through cional in Mexico. the Department of Forestry and Natural Re- Mr. Valdes is a member of the Sociedad Mexicana de Fotogrametria, sources at Purdue University. Fotointerpretacion y Geodesia. She is currently a research associate at the Laboratory for Applications in Remote Sensing (LARS), Purdue University and has been pri- marily involved with the evaluation of analysis * techniques for use with simulated and, cur- rently, for actual Thematic Mapper data. Carlos R. Valenzuela received the B.S. degree from the Agricultural State College, Deventer, The Netherlands, the M.S. degree from the * International Institute for Aerial Survey and Earth Sciences (ITC), Enschede, The Nether- D. Fabian Lozano received the B.S. degree in lands and is currently pursuing the Ph.D. degree biological sciences, with emphasis on Remote at Purdue University. Sensing of vegetation in arid regions, from He has been Soils Investigator at the ERTS/ Universidad Nacional Autonoma de Mexico in GEOBOL Bolivian Remote Sensing Program 1977. He is currenly pursuing the M.S. degree from 1977 to 1980, in charge of developing in the Department of Forestry and Natural soil survey methodologies using LANDSAT Resources at Purdue University. MSS data. He has participated as a Visiting Scientists at the Laboratory He is currently a graduate research assistant for Applications of Remote Sensing (LARS) in the development of a at the Laboratory for Applications of Remote Geographic Information System for Bolivia. He has also taught a Sensing (LARS), Purdue University. He was a Remote Sensing course at the University of Ball State at graduate and research staff member of the Instituto Nacional undergraduate levels. He is currently a Research Instructor LARS, de Investigaciones sobreRecursos Bioticos until 1981. Purdue University.