Detection of heat-emission sources using satellite imagery and morphological image processing Marcin Iwanowski Joint Research Center of the European Commision Institute of Environment and Sustainability T.P.262, via E.Fermi 1 21020 Ispra (VA) ITALY Institute of Control and Industrial Electronics Warsaw University of Technology ul.koszykowa 75 00-662 Warszawa POLAND e-mail: iwanowski@isep.pw.edu.pl Abstract In this paper the automatic detection of industrial heat sources is presented. The proposed method processes the thermal band images acquired by the TM sensor on board Landstat 5 satellite. It is based on a morphological image processing method which removes negligible heat sources such as sun-exposed materials, while preserving the significant ones. It can be used to estimate the energy losses by industry and as tool for large-scale energy conservation methods. I. INTRODUCTION Satellite imagery has become an important information source for environmental monitoring. The variety of image sources and the increase in computational power of modern computers encouraged the rapid growth of remote sensing applications. The ability of satellite sensors to acquire images in multiple wavelength bands (multispectral images) provides new image data that was previously unavailable. This new data has a great potential in many different areas. One of the most popular sources of remotely sensed data for environmental monitoring is the Landsat series of satellites. The Landsat program is a joint project of National Aeronautic and Space Agency (NASA) and the U.S. Geological Survey (USGS) [4]. The first satellite of this series, Landsat 1, was launched in 1972. Currently two Landsat satellites are operational: Landsat 5 and Landsat 7. The sensors on board the Landsat satellites are capable of measuring radiation in the 10.4 12.5µm wavelength range providing the temperature of the Earth (so-called at-satellite temperature). Consequently heat sources on the ground can be detected. Among the most important heat sources are industrial plants, e.g. power stations, steel works and the detection of heat emitters helps estimate of energy losses in various heavy industries. Such information can be used to develop an energyconservation plan on a large scale. This paper describes a method for detecting real heatsources based on the Landsat series of satellites. In the proposed method, the thermal band is processed using morphological image processing [3], which is used to filter the this band. This filtering is based on the morphological reconstruction operator which removes negligible heat sources while preserving the significant ones. The method is illustrated by analysing a Landsat 5 thermal band image acquired in 1990 covering the Silesia region of Poland and the Czech Republic. Locations of heavy industry and gas emitters are apparent in the optical bands. By analysing the thermal band one can detect the sources of high heat-emission. Morphological processing of the thermal band allows us to detect significant heat sources that are not a result of materials simply being exposed to direct sunlight, which are also characterised by a high at-satellite temperature. A. Satellite images II. BASICS The Earth observation from satellites for civilian purposes dates back to early 70 s when the Landsat satellite program was launched jointly by NASA and the USGS. It was the first civilian earth-observing satellite to meet the needs of Earth scientists [4]. The first satellite Landsat 1 launched on July 23, 1972 and was the first in a series. There are two currently operational satellites of this series, Landsat 5 (launched in 1984) and Landsat 7 (1999). The thematic mapper (TM) sensor on board Landsat 5 is capable of detecting reflected or emitted energy from the Earth. The TM sensor acquires images with a spatial resolution of 30 m for all optical bands, whereas the thermal infrared band has a resolution of 120 m The newer Landsat 7 is equipped with the Enhanced Thematic Tapper (ETM+) sensor with 7 bands including a thermal band with a spatial resolution of 60 m and 30 m spatial resolution of all visible and infrared bands. The images from these satellites are the most frequently used and easily accessible sources of information for researchers of various specialities studying the Earth surface. Due to the fact that thermal band imagery is also acquired, it is also possible to analyse the thermal radiance of the Earth s surface. From the point of view of the suitability of thermal images, the conditions of image acquisition are important. The choice of possible acquisition period is limited by multiple factors. The first is the image acquisition time,
(d) Fig. 1. Landsat 5 image Band 1, Band 2, Band 3, Band 4 (d). Image shows Bielsko-Biala city and its surroundings with Goczalkowice artificial lake (southern Poland). Fig. 2. Landsat 5 image cont.: Band 5, Band 6, Band 7. The same area as in Fig. 1 which equals approximately 10:00 a.m. local time (all Landsat satellites have a sun-synchronous orbit). Another is the repeat cycle, i.e. period between two successive scannings of the same area, which is equal to 16 days. Also the cloud-coverage around the area is an important factor which influences the availability of imagery of a given region of interests. A sample of Landsat 5 scene is shown in Fig. 1 and Fig. 2. Three bands reflecting the wavelengths of visible part of spectrum may be presented as RGB composite as shown in Fig. 3a. Although this composition is how humans would naturally observe the scene, for the analytical purposes it is not the optimal combination of bands. This is due to the fact that band 1 (blue) does not contain large amounts of information because the image is not well contrasted, is very sensitive to
T = ( ), (2) εc λ ln 1 πl λ λ + 1 5 where C 1 = 3.742 10 16 [Wm 2 ], C 2 = 0.0144 [mk] and ε is spectral emissivity. The last equation can be simplified to: C 2 14400 T = ( ) + 273.15, (3) 3.742 10 λ ln 7 πl 6λ + 1 5 where λ = 11.45 [µm] is the average wavelength of thermal band. The second term in the above subtraction refer to the shift from Kelvin to Celsius degrees. The L 6 variable describes spectral radiance in thermal band in [mw/cm 2 /µm/sr] and is calculated using the following equation ([7], [1]): Fig. 3. Color composites (RGB): bands 3,2,1, bands 4,3,2. L 6 = 0.005632 f(x,y) + 0.1238, (4) where f(x,y) is a value (digital number) of the pixel of the thermal band for which the temperature is computed. C. Morphological image processing Let f : {0,...,(x max 1)} {0,...,(y max 1)} {0,...,t max } be a single-band digital image of size x max y max, values of which can be between 0 and t max. The dilation of image f is defined as: atmospheric effects and has low SNR (see Fig. 1a). On the other hand, band 4 (near-infrared) contains more details and is much better contrasted (see Fig. 1d). Owing to this fact, another combination of bands is used for colour visualisation purposes - the false colour RGB composite of bands 4,3 and 2. An example of such a composite is shown in Fig. 3b 1. B. The thermal band The thermal band of the Landsat 5 Thematic Mapper (TM) image senses the emitted radiation from the ground of the earth it the wavelengths range between 10.4µm and 12.5µm. The value of the image pixel from the thermal band can be transformed into radiant temperature of surface bodies. This transformation can be done using the Planck s radiation equation for black-body for wavelengths equivalent to the bandwidth of thermal radiation [1]. The total spectral radiance in the thermal band is defined using the following equation: L λ = 12.5 10.4 ( )( 2πhc 2 λ 5 1 e hc λkt 1 ) dλ, (1) where: λ is the wavelength [m] in thermal band, T is the radiant at-satellite temperature in [K], h = 6.63 10 34 [Js] is the Planck constant, c = 3.0 10 8 [ms] is the speed of light, and K = 1.38 10 23 [J/K] is the Boltzmann constant. In order to get the temperature we transform the eq. 1 to: 1 Note, that with this representation the vegetated areas are in red. [δ B (f)](x,y) = max{f(x + p,y + q) : (p,q) B}, (5) where B stands for the structuring element describing the neighbourhood inside which the maximum value is detected. In the experiments described in the paper the 8-connected elementary structuring element is used. It is defined as the following set of two-element vector values: B = {( 1,0),(0, 1),(0,0),(0,1),(1,0),( 1, 1),( 1,1), (1, 1),(1,1)}. In the same way a dual operator of erosion (ε B (f)) is defined. In this case however, the max function in Eq. 5 is replaced by min. The dilation of a given size n is defined as: δ (n) B (f) = δ B(δ (n 1) B (f)),where δ (0) B (f) = f. (6) The geodesic dilation requires one additional image g, and is defined as: δ (n) B,g (f) = δ B(δ (n 1) B,g (f)) g,where δ (0) B,g (f) = f, (7) where stands for point-wise minimum of two images. While applying the above operation for increasing sizes (n) one can observe that at a certain level the image stops changing. This property is the basis of the definition of the morphological reconstruction transformation: R B,g (f) = δ (k) B,g (f) where k = min{n : δ(n) B,g = δ(n+1) B,g } (8)
The morphological reconstruction is a very useful operation which allows one to remove some parts of an image while preserving the structure of the rest of it. For a given image f (called a marker image) the way of creating image g (called a mask image) controls the final result of operation. For example, when computing the marker as an erosion of a image g (f = ε (n) B (g)), the reconstruction produces a result of opening by reconstruction of the image g, which is a useful filter removing small objects from the image while preserving the structure of the rest of it (without any shape blurring). The computation of the morphological reconstruction from the definition, according to Eq. 8 is a rather time-consuming process. There are however fast and robust algorithms of reconstruction, e.g. described in [6]. The reconstruction operator can also be used to perform an image maxima detection. It can be performed as: MAX(f) = f R B,f (f 1) (9) Where f 1 stands for an image obtained by subtracting the value 1 from every pixel of f higher than 0. A. Input images III. DETECTING HEAT-SOURCES The principal motivation for using Landsat images in the current study was their easy access via Internet. Unfortunately not all available images could be used to detect the high-temperature areas. The main factor that influences the measurement of the temperature is the date and hour when the image was taken. During the summer, on sunny days the surface of the ground is heated by direct sunlight. Higher temperatures are perceived over areas without vegetation and in urban and industrial regions with large surfaces covered e.g. by concrete, asphalt (roads, parking lots) or aluminium (roofs of big buildings, shopping centres, factories etc.). In order to properly detect the sources of heat which are not a result of direct sun-exposure, the influence of the sun should be either minimised by choosing the appropriate acquisition period and/or filtered-out using image processing. For current study the most appropriate winter images were not available and therefore late-summer images were chosen. Images presented in the paper are parts of a Landsat 5 scene acquired on the 31st of August 1990. The whole scene (Landsat 5 path 189, row 25) covers the large part of South- Western Poland including the Silesia industrial area as well as a part of the Czech republic. In order to extract the required information morphological image processing tools were used. The hot spots referring to industrial heat sources on the ground are characterised by the high pixel values in the thermal band. An obstacle exists however, which makes this detection task not a trivial one. The images were acquired in late summer on a sunny day, late morning. Due to that fact, there exist relatively large areas of high temperature caused by sun exposure. They are visible in Fig. 4: in the RGB composite from bands 3,2,1 is shown, in the content of the thermal band. The image in 4b contains Fig. 4. Colour composite of bands 1,2,3, thermal (5th) band, filtered thermal band. the areas which refer to sun-exposed pieces of the ground, marked by the yellow colour and warmer spots referring to the industrial heat sources (marked in red). Colours on the pictures showing the detection results refer to the temperature ranges listed in Table I. The conversion between pixel values (digital numbers) and Celsius degrees has been performed using Eq. 3. B. The detection algorithm The aim of the detection algorithm is to remove from the image the sun-exposed areas while preserving those indicating industrial heat sources. On the thermal-band image every area is represented as a region located around a regional image maximum. The distinction between both types of areas is
TABLE I TEMPERATURE RANGES pixel value temperature marking range [DN] range [C] colour 160-166 34-36,5 yellow 167-169 36,5-37,8 orange 170-174 37,8-39,7 red 175-180 39,7-42,4 magenta 181 > 42,4 brown performed according to the measurement of the height of the maximum associated with a particular area. Due to the fact that thermal images do not contain any significant noise, there was no need to apply any pre-filtering before preceeding with the proper detection. The algorithm can be summarised in the following points: 1) Computation of image maxima At first, the maxima are extracted from the input thermal image. Every maximum is associated with a high temperature region. The value of a maximum refers to the highest temperature detected inside the region. The resulting image with detected maxima produced using the Eq. 9 is a twovalued image where pixel of value 1 belongs to a maximum, while the other pixels of value 0 do not. 2) Thresholding of detected maxima The maxima are now checked in order to remove those which represent lower temperatures that can be caused by the sunexposure. This is done by a thresholding with a threshold set-up manually by the user: m θ = T θ (f (MAX(f) t max )) (10) where m θ is an image containing the maxima of pixel values higher than θ, T θ stands for the thresholding operator which is sets to 0 for all pixels of value lower than θ. The argument of the thresholding operator is an image which, for every maximum contains values of this maximum in the original image. 3) Reconstruction Once the important maxima has been preserved and the others removed, the original thermal image can be modified. The modification is performed using the morphological reconstruction (Eq. 8). As a marker image, the image with filtered maxima is used, whereas the mask is the input thermal image: f = R B,f (m θ ) (11) The resulting image shows all the areas associated with a maxima of height lower than a given threshold that was removed. All the other maxima were preserved without any local contour changes. Due to the latter, the distribution of temperatures around the heat sources were also preserved. The result of filtering and detection is shown in Fig. 4c. Comparing this image with the previous one (Fig. 4b) the Fig. 5. Czestochowa city and Czestochowa steelworks. Fig. 6. Katowice steelworks. filtered image does not contain the areas which refer to sunexposed pieces of the ground. Due to their temperature, which was below the threshold defined by the user, those pixels were removed during the maxima thresholding and consequently not reconstructed by the morphological reconstruction operator. C. Results The only parameter of the algorithm which has to be set manually is the threshold θ. By analysing the content of the image, for the current study it was set to θ = 170 which, according to Eq. 3 is equivalent to an at-satellite temperature of approx. 38 degrees Celsius. Considering the date and time of acquisition as well as weather conditions (approx. 10 a.m., end of August, sunny day) it can be equivalent to the temperature above the ambient temperature of sun-exposed areas. The algorithm correctly detected all the significant sources of high temperature on the input thermal image. It also removed the sun-exposed areas as shown in Fig. 4. This area shows the region around Zdzieszowice in Lower Silesia. The peak temperatures are found within a coking plant. Other important industrial areas were also detected in the thermal band. As an example the following are shown: 1) in fig. 5 - Czestochowa city with the steelworks (larger spot), 2) in fig. 6 - Katowice steelworks close to Dabrowa Gornicza city,
Fig. 7. Ostrawa city (Czech Republic). IV. CONCLUSIONS In this paper the automatic detection of industrial heat sources was presented. The proposed method is based on the processing of thermal band images from the Landsat series of satellites. Due to morphological image processing the method is able to distinguish between sun-exposed areas and real heat-sources. Thus winter images are not necessary in order to identify point heat sources. The method may be used to estimate the energy losses in industry and therefore be a tool for the creation of large-scale energy conservation methods. It can also be applied to detect other high-temperature sources such as forest fires. An example shown in the paper is based on the Landsat 5 image. The same processing can be applied to Landsat 7 images, easily accessible in e.g. the Image2000 database (available at http://image2000.jrc.it). The automatic identification of point heat sources can be supported by also considering other information sources, such as the CORINE Land Cover map [2] - a pan-european digital map describing land use. ACKNOWLEDGEMENTS Author would like to thank Conrad Bielski, Jacek Kozak and Pierre Soille for stimulating discussions and help in preparing the paper. (d) Fig. 8. Various heat sources: natural - forest fire, mining dump site, industrial zone in Tychy and Trzyniec, Czech Republic (d) REFERENCES [1] T.R. Martha, A.Bhattacharya, and K. Vinod Kumar, Coal-fire detection and monitoring in Raniganj coalfield, India - A remote sensing approach, Current Science, vol.88, No.1, Jan.2005. [2] Nunes de Lima, V., Image2000 and CLC2000 - Products and Methods, EC JRC Report, 2005. [3] P. Soille, Morphological Image Analysis, Principles and Applications, Springer 2003. [4] USGS Fact Sheet 023-03 Landsat: A Global Land-Observing Program, March 2003, U.S. Department of Interior, U.S.Geological Survey [5] J.A. Richards, Remote Sensing Digital Image Analysis, An Introduction, Springer Verlag, 2000. [6] L.Vincent, Morphological grayscale reconstruction in image analysis: applications and efficient algorithms, IEEE Trans. on Image Processing, 2(2), April 1993. [7] B.L.Markham and J.L.Barker, EOSAT Landsat Technical Notes 1, Earth Observation Satellite Co, Lnaham, Maryland, August 1986. 3) in fig. 7 - Ostrawa city in the Czech Republic with its steelworks. The high-temperature areas detected in the thermal band were marked according to the values from Table I and superimposed on the RGB colour composite. Apart from the industrial heat sources some natural sources were also detected. They refer either to forest fires or fires started on mining dump sites. The characteristic feature of this second type of detected heat source is the relatively the high gradient on the boundary of high-temperature region. This is visible in Fig. 8 where items and show the forest and mining dump site fire respectively, while and (d) present the industrial heat sources.