ON BOARD IMAGE ANALYZER FOR PALAMEDE MICROSATELLITE

Transcription

1 O BOARD IMAGE AALYZER FOR PALAMEDE MICROAELLIE A. PROVERBIO 1, F. MALAI 1, A. FUMAGALLI 1, F. BERELLI-ZAZZERA 1 1 Dipartimento di Ingegneria Aerospaziale, Politecnico di Milano, Milano ABRAC he project Palamede, started in 1997 at Politecnico di Milano, gives students the opportunity to participate in the design, the development and the realization of a microsatellite. One of the payloads of Palamede is the Earth urface Imagining ystem. It is composed by a charge coupled device camera, that will take pictures of the Earth surface. A global positioning system receiver will determine the exact position of the satellite when a picture is taken. everal images will be taken from Palamede during its operative life, but, actually, not all pictures can be downloaded to the Earth. he telecommunication subsystem of Palamede is made by an Orbcomm modem that allows the satellite to communicate with the ground station by . his means that a picture needs to be translated in ACII code and divided into several text s, because only text is allowed into the s by the Orbcomm system. hese s must be downloaded to the ground station: the download of a single picture may take several days, due to the small amount of energy stored on Palamede. On the other hand the solid state mass memory on board is only 92 MB so only 13 pictures can be stored on board waiting to be downloaded. For these reasons a task of the on board oftware is dedicated to evaluate, by appropriate criteria, which picture can be downloaded and which must be deleted. A ranking of pictures is made in order to evaluate which picture has the highest quality and must be downloaded first. 1. IRODUCIO Palamede is a microsatellite built in the Department of Aerospace Engineering of the Politecnico di Milano [4]. he design philosophy for this project is to use non-space qualified components, cheaper and less reliable than the certified ones. his philosophy leads to certify the system by an extensive test campaign before launch. One of the aim of Palamede mission, in fact, is to verify the reliability of the satellite. he satellite size allows to use several launcher as against small changes at interface level and it allows to be lunched as piggyback as well.palamede's design is also in accordance with AA s last micro-satellite design philosophy that can be summarized in the phrase, a smaller, faster and cheaper mission. 1.1 he satellite Palamede is a satellite of about 20 Kg of mass. Its body has a cubic shape and a structure based on five honeycomb sandwich panels of aluminum layers. he baseplate is made with an aluminum panel milled, with a hole in the center, which allows the camera to take the pictures of the Earth. he on-board computer and the main boards are standard PC-104 technology. he choice to use this system allows to use reliable terrestrial technology which can also withstand to space environment. o reduce the power consumption, Palamede is equipped with a CPU (standard 80486), 64 Mbit of RAM memory and a 92 Mbytes olid tate Mass Memory (MM) technology hard disk,. in order to avoid any gyroscopic effect during the orbital mission phase [4].

2 Power supply is performed by 5 solar arrays body mounted and a battery which stores electrical energy during the sunlight and provide the electricity during peak phases, like the transmission one, and to all users during eclipse phase. he electric components on-board are all fed by a battery that is charged in the non-eclipsed part of the orbit. he devices working in DC need a constant voltage of 5V, while the components in AC work at 12V and are converted by a Converter. he battery is constantly charged by 5 solar panels produced by CEI, commissioned by EA and named CJH OP. he payload of the satellite is composed by: the top solar array, which is equipped with new kind of solar cells, the Charge Coupled Device (CCD) camera, connected to a Global Positioning ystem (GP) in order to evaluate the position when a picture is taken [5]. he satellite ADC (Attitude Determination and Control ystem) sensors and actuators are: six solar sensors, one for each face of the cube, a magnetometer and three magnetic as actuators [3]. his kind of architecture doesn t allow to have a three axis stabilization, but it allows to reach a good pointing with angular velocity below 0.2 deg/s for entire sunlight phase. [4] [9]. he telecommunication subsystem is based on Orbcomm Constellation ystem which allows to link the ground station to Palamede by a satellite constellation of 32 satellites placed on 6 orbital plans at 800 Km height. Palamede is equipped with a radio-modem Quake 1550 connected to the on-board computer, which is able to transmit text messages in the form of of 229 Bytes each one. In this way all data: telemetry, scientific data and picture, telecommand and any communications shall be performed by with only ACII text. 1.2 Mission tasks he use of non-space certified components has a very deep impact on the system, but it is the basis of the construction philosophy of this project and of the mission target. he reason of this choice lies in the fact that cost reduction has become a priority in many realities, specifically in low performance missions where a super satellite is not required. Unfortunately, this direction of work requires a very intensive testing of all the components both on the ground and during its operative life [4]. AR Acquisition Codificationm in ACII Analysis File splitting in packets Compression in JPEG Download toring ED Figure 1: General image processing Palamede has the simple aim of testing the components on-board. he two tasks of the mission are: o test, in terms of efficiency, endurance and performance, the space application of the solar panels. o test, in terms of performance, the CCD camera, in order to analyze the focus and the image quality that is sent on the ground. We must remember that Palamede is an university microsatellite, so overall, there is an academic task that constantly puts the students up against of the problem of design, building and programming a satellite [9].

3 2. IMAGE ACQUIIIO AD AALYI In Palamede there are many problems related to low cost, and consequentially, low performance of the components. his problem creates the need to use Orbcomm and to have a very reduced sending bit-rate. he image has to be acquired by the satellite, analyzed and stored. All these processes must work with 64 Mb of RAM and software, images and data shall be stored in 92 MB of mass memory [9]. Figure 2: Photo analysis general flow he CCD camera installed on-board is a model Ganz CMH-880. It is designed for taking pictures from a height of about 550km and to analyze Earth sectors of about 96x72km. It is able to acquire images of 753x582 pixels, but the software created and stores a 480 x 360 pixel images in Portable Pixel Map (PPM) format [5] [9]. In order to take good pictures the software allows the process only if the rotational speed remains under 1,9 deg/s. Each image taken uses B, the first 15 for the header and the others for the file body where each pixel is represented by a triple integer in the range 0 to 255, defining the intensity of red, green and blue. Due to Palamede telecommunication system, an image of such huge dimension, could not be downloaded in a short time. o a compression is necessary, but before this happens, it is also necessary to create a ranking of the pictures, in order to decide which is the best photo to be stored on-board. his task is due to the AOC system as mentioned before. An analysis is required and an ad hoc method has been studied. If the image is considered sufficiently good, it is then compressed in JPEG format, with a memory size reduction of about 1/20. After that, it is then necessary to convert it into ACII codification and, afterward, divide the text file into several packets, each one sent to ground by a single he software he onboard software is composed by several managers. Every manager coincide with a task to manage all the satellite life like: housekeeping generation, software management, AOC maintenance,picture generation, error evaluation, telecommunication and so on[9] [7]. One of this task is the Photo Analyzer. After several check for attitude data, power data and orbital position the software decide that a picture can be captured. o the CCD and GP are switched on and a picture, in.ppm format, is taken and linked with: the orbital position (latitude and longitude) and the onboard time (synchronized with GP time). All these data are

4 stored in MM. ext the Analyzer Manager is called to perform its task. It opens the file, reads the image and progressively counts the number of pixels in the important levels and the statistics variables of interest. hose procedures require an expensive program that has four for cycles inside. After that, a rank index (BIPBIP) is computed, and thanks to this, the photo is inserted into the right position in the preview structure. he image is inserted into the position and all of the photos of lesser quality are moved down a position. If the photo is of good quality and is not the last image, it is inserted and converted into JPEG format. Otherwise, the photo will be deleted if the position exceeds the maximum structure dimension. In this way a sort of ranking is made. Only 13 photos are allowed, so when the fourteen photo is taken, if it is reach a BIPBIP better than the bottom one, the new photo is stored on MM and the last one of the ranking is deleted. When the satellite has an image with a BIPBIP higher than a threshold or the acquiring period of 20 days is [7] finished, the W Manager calls the Encoder Manager which encodes the image in an ACII codification and splits it in packets of 229B, the dimension required by Orbcomm for the transmission. Orbcomm downloads the packets corresponding to an image in a time of about 3 days, transmitting the data only in a predefined part of orbit, and the Orbcomm service sends the Palamede G an containing the photo that should be recomposed with a code called Decoder Manager. 3. PICURE AALYI ALGORIHM he main problem with the on-board picture management is selecting the most interesting picture that should be downloaded to the Ground tation. Many satellites have the task of analyzing images onboard, but they have different features than Palamede. hey are generally high-definition, multi-spectral cameras and computationally able to analyze the photo, matching the different photos of the same subject onboard with a high cost (in terms of flops and memory). One available technique is checking the gradient of tonality inside the image and detecting boundaries and objects [1] [2] [8]. Unfortunately, our satellite has neither a high-definition camera nor a multi-spectral camera, so other less accurate techniques must be adopted. Palamede is designed for orbiting in a sun-synchronous orbit and for taking all the pictures during the sunlight of the orbit. ix sensor allows the photo to be captured only when the camera is pointed at the Earth s surface [5]. Unfortunately, it can sometimes happen that a part of the image represents deep space or is very bright as the result of a direct sun ray exposure. econdly, it is important not to download a sector of earth covered by haze or fog. Figure 3: Boundary color for overexposed photos hirdly, we want to identify interesting sectors or portions of the Earth that are not really useful for the mission s task. We had the goal of downloading a portion of Earth s surface that represents flat areas, and not cloud coverage areas or sea areas. Actually, we are interested in computing an index of quality for any analyzed picture: the Brightness Percentage of Best Picture (BIPBIP)

5 3.1 Over or Under-exposed photo identification In a photo captured from the space, a part of it could be very dark because of the representation of deep space. We want to avoid the onboard storage of a photo representing the deep space. 3 Consider the color of each pixel of the image as a vector, in R, where each component represents the value of the single intensity of color (red, green and blue). In this way, it is possible to identify two spheres, the first representing the dark pixels, the second representing the bright ones. he dark pixels sphere has its center in the point (0 0 0), that means the total absence of color, and a radius of 30. If a pixel it is identified as being very dark, it is probably an image of deep space. he software passes over all the pixels of the photo with a for cycle, and if the number of pixels with this characteristics exceeds a threshold of 30% of the total amount, the photo is considered to have captured a dark place, or a deep space image but not a subject of interest like a night vision of the Earth s surface. Figure 4: Boundary color for underexposed photos Exactly the same process has been applied for determining the overexposed images. It has been defined a sphere with the center in ( ) the brighter color in the pixel map, and it has a radius of 20. Just as previously stated, if bright pixels are more than the 30% of the total amount, the image is considered overexposed. Figure 5: Image division in sector 3.2 Hazy photo identification Another kind of image that we absolutely do not want to download is a photo that is too homogeneous. his kind of images generally represents the Earth s surface covered by fog. hese are less interesting to our purpose being

6 useless for the test campaign: they do not allow to study the capability of the camera to focus on a subject. Just as in the last paragraph, we consider the color of each pixel as if it is were a vector in a tri-dimensional space. he software uses the same for cycle and checks to see if the there is more than the 30% of the total amount of pixel inside a sphere of radius 5 and centers the mean color. Figure 6: APER versus number of variable ince fog results as a very homogeneous kind of cloud, all the pixels of the image have more or less the same color. In accordance with this observation, we suppose that fog covers the entire image, and so the medium color of the photo is the medium color of the cloud. he radius has been valued by empirical consideration based on tests on many images. 3.3 ector identification Each photo has a dimension, (360x480) in terms of pixels, that allows them to be divided into 36 sectors. his division is necessary because, in our working condition, the image represents a surface of about 300 km2. o, statistically speaking, there is a low probability that the same kind of terrain is featured in all the photos. 36 is a good compromise between the necessity to split up the image and the question of maintaining a statistically meaningful number of pixels in each sample. A sector is a portion of 60 x 80 pixels, so it is composed of 4800 elements. his great dimension allows the application of many statistical tests. In the first instance, we have computed the mean, the variance, the skewness and the kurtosis for each color level [8]. After that we have counted the number of pixels for each tonality that falls in a given range. First we have divided the 256 levels in groups of 10, and then we have plotted the APER over a variation of the range of the interval [6]. he results, shown in figure 7, demonstrate that, over a span of 40 pixels per interval, the error becomes constant if we do not enlarge the range too much. We have chosen to split the range into intervals of 90 levels: [0,89] [90,179] [180,255]. he decision of using 90 as range is due to the need to have a reduced number of intervals and variables to be analyzed and to have the greatest possible similar dimensions among the intervals.

7 Figure 7: APER versus pixel interval dimension We have decided to not use luminance or chrominance values because these are all linear combinations of the color levels. he supervised analysis accomplished is a linear discriminant analysis, where the starting number of variable is 21: 3 values of mean, 3 of variance, 3 of skewness, 3 of kurtosis and 3 for each color for the level of intensity of color [6]. A sample of 302 pictures has been used with images representing desert, from the ahara, exas and California, Australia and Mongolia areas; forest in Brazil and Colombia; mountains from the Rocky Mountains, Andes, Alps, Pyrenees, Morocco; and zones of flat terrain taken in Pennsylvania, Virginia, Argentina, France, Italy, Australia, Ivory Coast and outh Africa. his distribution has been taken so that Palamede will capture photos only in a part of the Earth s surface that is from a latitude of 40E to 40E. Besides this kind of sector, other typologies have been added like areas with light (less than 3/8) cloud coverage, medium (4/8 to 6/8) and heavy coverage (more than 7/8) and coastal areas. With this picture we have had been able to analyze and create a first discrimination model with all these variables, but to implement the onboard software, we have decided to reduce the less meaningful coefficients in the analysis and we have progressively reduced the number of coefficients to7 with a progressively increased error expected [6]. his reduction has produced an increase in the APER as shown in figure 6. At the end of the process we have finally create a method using only: mean of the red and green variance of red, green and blue number of pixel with the red intensity falling in the [90 179] interval number of pixel with the green intensity falling in the [90 179] interval number of pixel with the blue intensity falling in the [ ] interval. 3.4 Quality index computing he goal of the analysis is to associate an index of quality to each photo. First of all, we would not waste time downloading images that do not represent a part of the Earth s surface or that have the phenomena of overexposure.

8 We have decided that in the technical specifics of Palamede, the index will be called BIPBIP and will have values in the range from 0 to 100. he worst possible image has an index of 0, the best, 100 [9]. We have decided to mark as not good photos that have captured a part of deep space or are overexposed, and to let their index start from 50. he BIPBIP associated to these images, after being cataloged as over or underexposed, will be penalized in proportion to the percentage of pixels close to black or white. BIPBIP is computed as: CB BIPBIP = 50 1 B C W W (1) Where is the total number of pixels in the whole photo, B the number of dark pixels, W of bright pixels and CB and CW are empirical coefficients set equal to 1. As self evident, this formula does not allow to have a BIPBIP grater than 50 and is used only if one of the following conditions is verify: B W (2) Even if the image is interesting (not over or underexposed) the index of quality should consider both the sector analysis and the homogeneity of the picture. he index is computed as: K BIPBIP = H H K C 1 CH K C 2 CM K (3) Where, with the same notation as before, H is the number of pixels considered homogeneous at the mean color and KH is the corresponding coefficient set equal to his formula is used if is verify the following condition. H 0.3 (4) If it is not, the formula to apply is a variation of the (3): K BIPBIP = C 1 CH K C 2 CM K (5) Here, the ratios are based on the number of sectors and not on the pixels. As soon as the software has identified the number of sectors for each typology, the ratio computed between the number CH, of the sectors totally covered by clouds and the total number of sectors (set as 36) is multiplied for a coefficient KC1, empirically defined as 0.9. It has a high value because it considers very penalizing for a picture to represent a cloudy vision of the Earth surface. he same process is applied to partially covered sectors, where C2 is the number of them and KC2 is equal to 0.5. At the end the sectors representing the sea are processed. hose are equal in number to and the coefficient K is equal to 0.7. After this process, the value of the BIPBIP is the final value that describes the quality of the image and catalogues the pictures. 4. COCLUIO he method implemented here can easily detect images that are overexposed or do not represent the Earth s surface. All the images tested showing space have been valued with a low BIPBIP by the software, and not one photo showing terrain has been mismatched. his first result is very important because sets a very deep threshold

9 between images that could be download, and the ones that absolutely must not be considered good. It is important for the strength of the method. For the homogeneous images, the result changes a little. he software does not identify every hazy image because the phenomena in observation is very different in each situation, but the software could penalize a photo in which homogeneity does not permit it to be used in the test campaign. he results for the sector analysis does not appear as good, but let us analyze them. CM CMF CMM CH F D Fo M CoF CoFo CM CMF CMM CH F D Fo M CoF CoFo able 1: Results of the analysis of the sectors he APER is 27,45%. he index is not very low, but it is interesting to observe that the error of misclassification is generally addressed to penalize the sector of interest. If we consider that Palamede has the capability of storing 13 photos, an error like this is not very important because if a good photo has a low BIPBIP, another one has a better quality index. In addition, we have to realize that for many misclassified sectors, like confusing sea with flat terrain, a human operator may also not be able to distinguish between them. In other kinds of sectors, like coast and flat terrain, their classification could be influenced by the amount of plain and sea in the photo. All this consideration helps us to conclude that the method has good performance and it is able to give good results with the advantage of an easy and light implementation. Better results can be achieved with more computational and memory resources which are not available on Palamede. REFERECE [1] J. hmetz, P. Pili,. jemkes, D. Just, An Introduction to Meteosat second generation, AM journal, Vol. 83 Issue 7, 2002, pp [2] G. eiz, M. Baltsavias, A. Gruen, High-resolution cloud motion analysis with Meteosat-6 rapid scans, MIR and AER. EUMEA Meteorological atellite Conference, Weimar, 28 eptember - 3 October 2003 [3] F. Bernelli-Zazzera, M. Molina, M. Vanotti, M. Vasile, ew algorithm for attitude determination using GP signals, AIDAA XVI congress Palermo eptember 2001 [4] F. Bernelli-Zazzera, G. Chessa, A.E.Finzi, M. Molina, Preliminary design and operations of Palamede microsatellite, AIDAA XVI congress Palermo eptember 2001

10 [5] F. Bernelli-Zazzera, C. Daeder, M. Molina, Low cost commercial camera use on small satellites for Earth observation, AIDAA XVI congress Palermo eptember 2001 [6] R. A. Johnson, D. W. Wichern, Applied Multivariate tatistical Analysis, 5 th ed., Pearson Education, ovember 2001 [7] D. Meloni, oftware Architectural design of Palamede microsatellite, Mc. hesis (in Italian, title: Progetto dell architettura software del microsatellite Palamede), Politecnico di Milano, [8] M. Massari, I. Ferrario, Autonomous navigation of a rover for space exploration, Mc hesis (in Italian, title: avigazione autonoma di un veicolo per esplorazione spaziale), Politecnico di Milano, 2000 [9] F. Malnati, D. Meloni, G. Giardini, Palamede oftware Requirements, Internal document,politecnico di Milano, 2003