A Study of Emissivity Ratios and their Application In Determining Land Surface Temperature in IDL B. Todd Guest 25 March 2005 Dr. Hongjie Xie ES 6973, Image Processing
INTRODUCTION Remotely sensed images provide a great deal of scientific information that improves our knowledge of our environment s status and provides researchers with information that would otherwise be impractical or impossible to collect. Among the most revealing of these features is discerning land surface temperature. With it, the viewer may determine weather conditions and trends, heat islands, surface and terrain characteristics, and many other phenomena. Using the software program ENVI, as well as IDL, the language that it is based on, it is possible to determine the land surface temperature. An essential factor in determining the land surface temperature is to have knowledge of the emissivity of the surface that is being studied. Emissivity is a ratio between the radiant flux exiting a real-world selective radiating body and a blackbody at the same temperature (Jensen, 249). Put more simply, emissivity tells us the difference in energy emitted by an object (in remote sensing, this would be the land surface) compared to a theoretical blackbody that has the same temperature. Emissivity is an essential key in the land surface temperature algorithm used by IDL. The equation used to determine surface temperature is: Land Surface = Temperature Brightness Temperature 1+(0.0007991666*Brightness Temperature)* ln(emissivity) where the preset numbers and natural log are constants, the emissivity is a variable input ration (i.e., 0.988) and the brightness temperature is represented as: Brightness = 1282.71 Temperature ln((666.09/radiance)+1) where again the preset numbers are constants and the radiance is represented as: Radiance = 0.0370588 * High Gain Band 6 + 3.2 where the preset numbers are constants and the variable is each pixel of the band 6 high gain image. Just as the pixel value can change the outcome of the radiance, so too can the emissivity change the outcome of the land surface temperature. Although a number of methods exist for determining emissivity (Weng et al., 2004), a library is classified emissivities was constructed (Snyder, et al., 1998). The classes in this lookup table were derived from categories established by the International Geosphere-Biosphere Program (Weng et al., 2004), and were broken down by land use/land cover. Since every object has its own unique emissivity, any given remote sensing image may have hundreds, thousands, or tens of thousands of different emissivities, depending what is contained in the image. For example, in an image that is senescent woody savanna, the overarching emissivity may be 0.991 stemming from the specific emissivity of the trees (Snyder, et al., 1998), even though in a medium or high resolution image
other emissivities would most certainly by present due to the visibility of fallow soil or a copse of trees. In order to get an extremely accurate land surface temperature measurement of an image or to determine the land surface temperature of various features in a single image, it is necessary to determine the emissivity of a given feature and apply it along with other features emissivities to the land surface temperature algorithm. The purpose of this project is to do just that. The single umbrella emissivity-per-image, while simpler, does not provide the accuracy that is needed for many researchers. STUDY AREA AND DATA Images from the National Aeronautics and Space Administration s LANDSAT 7 Enahnced Thematic Mapper Plus (ETM+) instrument were used to acquire the study area image. I wanted to evaluate the variable emissivity land surface temperature formula on an area that would be familiar to the intended audience (ES 6973 students), would be easily acquired, and would provide a number of different land uses / land covers that fell within the lookup table, while still having a single overarching emissivity to compare the variable emissivity results to. An image of the east Texas / western Louisiana Great Piney Woods region met all of the requirements. The image (Reference Figure 1a-b.) was taken on 8 September 2002 on path 25 and row 38 of LANDSAT s orbital route. The pixel resolution was 30 meters, which provided a good sense of independent land use / land covers, while not being so detailed as to make the data set unwieldy. The projection was the Universal Transverse Mercator and the datum geoid was the World Geodetic System established in 1984, also known as WGS 84. Figure 1a-b. a. Landsat ETM+ RBG Image of East Texas Green Needle Forest b. ETM+ RBG Subset Image of East Texas Green Needle Forest. As already mentioned, the emissivities were determined from a look-up table. In conjunction with my LANDSAT image selection, I selected four emissivities that would
dovetail with the image. The primary emissivity to be studied was that of green needle forests, which were abundant in the image. In addition, I included green grass savanna, arid bare soil, and water. The arid bare soil, while minimal in area (it showed up only as individual pixels when classified) was spread throughout the image and made a good reference point to determine the accuracy of the classification. METHOD From this original, raw image a subset was created to make the image more manageable, and to prevent ENVI from locking up while processing the data. As with the full image, a subset was chosen that would have a good representation of land use / land covers that would provide for good experimentation. Initially, the next step was classification. However, during my oral presentation, it was brought up that I had not conducted an atmospheric correction of the image. I went back and ran a dark subtraction on the image. This program corrects for atmospheric scattering. I selected the minimum band subtraction method, as I thought that the results would be comparable to any that were user-determined. The corrected image visually appeared a few shades brighter, and the pixel value function in ENVI confirmed this. I next attempted to process the image through the relative atmospheric correction and reflectivity calculation algorithm in IDL using a sun distance of 1.0011 AUs and a sun elevation of 50.03. However, my image was too large and the memory requirements exceeded the capabilities of the computers I attempted this on. However, for the purpose of this project, I believe that any corrections would have been negligible. The next step accomplished was to classify the data according to the land use / land cover associated with the predetermined emissivities. I selected regions of interest using free-hand polygons in the Region of Interest Tool in ENVI. The regions were small but representative of the areas studied, and I feel yielded good results (Reference Figure 2a-e.). In addition to the four emissivity classes used, I also included a classification for cloud cover. I chose this as a control class because of its high contrast.
Figure 2a-e. Selected Regions of Interest. a. Green Needle Forest, b. Green Grass Savanna, c. Arid Bare Soil, d. Water, e. Cloud Cover. I used the parallelepiped supervised classification method for its ease of use, ability to designate/determine the classes, and its wide use among researchers. The first time I ran it, I used a single value with a 3.00 maximum standard deviation from the mean. At the recommendation of the class instructor I ran it again, but this time with a 2.00 standard deviation from the mean. The results (shown in Figure 3a-b) show a less distinct classification, with many more unclassified (black) pixels. In addition, the presence of striping is much more visible in the classification 2.00 maximum. I decided to keep the 3.00 maximum standard deviation from the mean. The difference is especially noticeable in both the cloud cover (color: thistle) and green needle forest (color: bright green) as shown in Figure 4a-b. Figure 3a-b. a) Parallelepiped Classification with a 3.00 Standard Deviation from the Mean. b) 2.00 Maximum Standard Deviation from the Mean.
Figure 4a-b. Parallelepiped Classification with a) 3.00 Standard Deviation from the Mean. b) 2.00 Maximum Standard Deviation from the Mean. The final products created by the parallelepiped classification method were as hoped. While there was a significant amount of unclassified pixels, there were no concerns because there were many other land use / land cover categories that were not included, as they were outside the scope of this project. Comparing the classifications with the RGB image produced a high confidence level in its accuracy (reference Figure 5a-c). Figure 5a-c. Parallelepiped Classification of a) Green Needle Forest [bright green] b) Green Grass Savanna [sea green] and c) Bare Arid Soil [red]. A classification accuracy assessment confirmed these findings. The method used was the confusion matrix, and it produced a 98.2574% overall accuracy with a kappa coefficient of 0.9863 (well above the 0.80 standard for strong agreement and good accuracy). On the match classes parameter window in ENVI, the ground truthing regions of interest (training sites) were matched against the regions in the classified image. These training
sites were in line with the results. There was a low omission error, as well as a very high user accuracy. In preparation for IDL, the image file (.img) was converted into a TiFF/GeoTiFF file (.tif). The next step was to write the IDL code to determine the land surface temperature with the emissivities derived from the classified image. This code was written by the course instructor for me (for which I am very grateful!). The desired method of the code was to take the band 6 image and first calculate the radiance of each pixel. The radiance would then be input into an equation to determine brightness temperature. Also included in this equation would be the emissivity of the different classes. It would be important to assign (or link) emissivity values to each class. For this project, the linkages are class one equaling 0.991 (green needle forest), class two equaling 0.986 (arid bare soil), class four not used, class four equaling 0.972 (water), and class five equaling 0.991 (green grass savanna). This way, IDL knows what each of the classes in the classified image represent, and are not just arbitrary. The brightness temperature is then put into a final land surface temperature equation. This algorithm also uses the emissivity classes mentioned above. The output would have been an image with each pixel value representing the land surface temperature measured in degrees Kelvin. The equation used in this project look like this: 1 pro ems 2 3 B6=READ_TIFF('C:\ES6973\MyWork\band6.tif', CHANNELS=[0,1], 4 geotiff=gtmodeltypegeokey) 5 ;read Band6 to B6, and channel 0 is the low gain, channel 1 is the high gain 6 c=read_tiff('c:\es6973\mywork\classify.tif', CHANNELS=[0], 7 geotiff=gtmodeltypegeokey) 8 ;read classification image to c, 9 10 B6h=temporary(B6(1,*,*)) ;assign the high gain (channel 1) of Band6 to B6h 11 L=temporary(0.0370588*B6h+3.2) ;calculate radiance 12 TB=temporary(1282.71/(alog((666.09/L)+1))) ;calculate brightness temperature 13 14 ;arrdim=size(c, /dimensions) ;get the dimensions of the ETM 15 cols = 3768 16 rows = 4062 17 ; c2 = bytarr(cols,rows) 18 for j = 0, cols-1 do begin 19 for i = 0, rows-1 do begin 20 if (c[j,i] EQ 1) then begin 21 c[j,i] = 0.991 22 endif else if (c[j,i] EQ 2) then begin 23 c[j,i] = 0.986 24 25 endif else if (c[j,i] EQ 4) then begin c[j,i] = 0.972 26 endif else if (c[j,i] EQ 5) then begin
27 c[j,i] = 0.991 28 endif else begin 29 c[j,i] = 0.0 30 endelse 31 32 endfor 33 endfor 34 35 RT=temporary(TB/(1+(0.0007991666*TB)*alog(c))) 36 ;supposing the same emissivity of 0.988 37 ;supposing the same emissivity of 0.988 38 WRITE_TIFF,'realT.tif', RT, geotiff=gtmodeltypegeokey, /FLOAT 39 ;write the temperature to an image called TMt.tif The first section (lines 3-8) directs IDL to use the appropriate image files for both the band 6 data and the classified data. The next section (lines 10-12) are the representations of the radiance and brightness temperature algorithms. The third section (specifically lines 19-30) are used to assign the emissivity values to the different classes. Fina lly, line 35-37 shows the land surface temperature algorithm and line 38 directs the resultin g image to a new file name. The desired end product should be a panchromatic image with the same unclassified black pixels and the lighter pixels representing the specific temperature of the land use / land cover for that time of day. RESULTS There were problems with the IDL code however, that prevented the creation of an image that displayed the surface temperature with the classified emissivities. After trying several variations, the end image only calculated the pixels in the first class, in this case green needle forest (reference figure 6). Figure 6. Land Surface Temperature Image after Processing through IDL Code.
The different shades of white and gray (note gray in bottom left hand corner of figure 6) do demonstrate that there is variance in the pixel values. Using the pixel value tool in ENVI, I confirmed that the data values were around the range of 290. Assuming this is Kelvin, a quick conversion to Fahrenheit shows that to be in the uppers 70 s, which is about what one would expect in East Texas at this time of day ion early September. CONCLUSION In producing a final, accurate image of an area s land surface temperature with classified emissivities, this project was unsuccessful. Given more time (and I fully acknowledge that I waited until too late in the semester to begin the code work), the bugs could have been worked out of the code to produce a code that was operable. The concept is there, and the goal is useful. Generalizations of an area s land surface temperature are useful only to a certain point. When applying a blanket emissivity to an entire image, often the result would be no more accurate than taking half a dozen in situ temperature readings at various locations and applying that to the entire area. This gives you a good idea of what the temperature is, but not a true idea. Scientific researchers depend on accurate, complete data in order to conduct their studies. By dividing the land use / land cover into classes, assigning each of those classes to a specific emissivity, and running a land surface temperature algorithm to each pixel, the user would be able to have a very accurate reading of an area s temperature. An interesting study would have been to run the statistics option in ENVI to find the mean and compare it to the mean of a land surface temperature of an algorithm using a single, generalized emissivity ration. The applications of this are numerous. For example, with high-resolution (< 12 inches) imagery, researchers studying heat islands would be able to determine the emissivities of roofing material, asphalt, paint, etc. (these emissivities are readily available). By running a classification and assigning emissivities, one would be able to get a very accurate idea of what the temperature of an urban area. With all of the variations in a city setting (everything from green parks to black tar paper roofs), specific assigned emissivities have the potential to dramatically change the overall temperature reading. More work needs to be done to determine where the IDL code is incorrect. When a solution is found, this will be an excellent tool for those who work with remote sensing. Works Cited Jensen, J., (2005). Introductory Digital Image Processing: A Remote Sensing Perspective, Pearson Prentice Hall, Upper Saddle River, NJ. Snyder, W.C., Wan, Z., Zhang, Y., & Feng, Y. Z. (1998). Classification-based emissivity for land surface temperature measurement from space. International Journal of Remote Sensing, 19, 2753-2774.
Weng, Q., Lu, Dengsheng., & Schubring., J. (2004). Estimation of land surface temperature vegetation abundance relationship for urban heat island studies. Remote Sensing of the Environment, 89, 467-483.