Changes between Version 34 and Version 35 of UserApp/Proba-V


Ignore:
Timestamp:
Dec 16, 2011, 1:26:59 PM (8 years ago)
Author:
Iliyankatsarski
Comment:

Legend:

Unmodified
Added
Removed
Modified
  • UserApp/Proba-V

    v34 v35  
    583583| 1.15 km || 250m (R, NIR), 500m (SWIR) || 300m (VNIR), 600m (SWIR)
    584584|}
     585= Generation of reference data =
     586
     587High resolution reference data were necessary for this study, in order to train and validate the classi-fication algorithms.
     588However, the only available European-wide data, was the JRC Forest Cover Map, compiled in 2006. Since our datasets were
     589collected in 2009, the Landsat TM data, acquired in 2009, were classified in order to produce a stocked/non-stocked classification
     590map, which would be used to evaluate the classification algorithms in the medium/low resolution datasets.
     591= Generation of reference data =
     592
     593High resolution reference data were necessary for this study, in order to train and validate the classi-fication algorithms.
     594However, the only available European-wide data, was the JRC Forest Cover Map, compiled in 2006. Since our datasets were collected
     595in 2009, the Landsat TM data, acquired in 2009, were classified in order to produce a stocked/non-stocked classification map,
     596which would be used to evaluate the classification algorithms in the medium/low resolution datasets.
     597
     598For the classification of Landsat TM in stocked and non-stocked areas, a set of 400 points was generated by stratified random
     599sampling, following the estimation of the extent of the two classes through an unsupervised classification. These points were
     600used as training samples, and were identified as stocked or non-stocked by using the JRC Forest Cover Map and visual interpretation
     601of Google Earth images. (Figure 2). A radius of two pixels was considered around each point in order to identify it as stocked
     602or non-stocked. A second set of 300 points was generated using the same method, to be used for the validation of the classification.
     603A Maximum Likelihood supervised classification was performed on three different band combina-tions: 6 bands (excluding the thermal band),
     6044 bands (Blue, Red, NIR and SWIR), and 4 bands + NDVI generated from the Landsat TM data, in order to find the best result. The same
     605training samples were used for the three band combinations. The four classes defined in these classifications were "stocked areas",
     606"non-stocked areas", “water” and "No data". The confusion matrices, for the three classification’s results, were performed using
     607the same 300 validation points.
     608= Forest cover reference maps creation =
     609
     610
     611The classification map with the higher accuracy was resampled to a 10 m resolution, using nearest neighbour,and the “no data”
     612and “water” classes were masked out, leaving the two main classes, “stocked” and “non-stocked” with values of 1 and 0, respectively.
     613
     614In order to create reference maps with the same resolution as the medium/low resolution images, the 10 m resolution map was
     615aggregated to 250, 300 and 1000 m resolution. The new value of each aggregated pixel was the percentage of the original “stocked”
     616pixels that were aggregated to the lower resolution pixel. According to the pixel values obtained after aggregation, each map was
     617reclassified to five classes according to the percentage of forest cover (0 – 10%, 11% - 30%, 31% - 50%, 51% - 75%, 76% - 100%),
     618see Figure 3. Those classes were chosen in order to ensure that the entire range of forest cover fractions would be
     619represented within the broad “stocked” and “non-stocked” classes.
     620
     621The Maximum Likelihood supervised classification on the three different combinations of Landsat TM bands produced the highest
     622overall accuracy (93.12%) and Kappa Coefficient (0.904) when all six bands (excluding the thermal) were used, as shown in Table 2.
     623The resulting stocked/non-stocked map was used as a reference map for the study.
     624
     625{| border="1"
     626|+ Results of the accuracy assessment of Landsat TM classification
     627! !! TM 6 bands !!TM 4 bands !! TM4 + NDVI
     628|-
     629! Overall accuracy
     630| 93.1214% || 92.9878% || 92.4135%
     631|-
     632! Kappa coefficient
     633| 0.9045 || 0.9027 || 0.8947
     634|}
     635
     636= Evaluation of classification algorithms =
     637
     638
     639The produced reference data were used to generate training samples for the classifications and the validation points
     640for the confusion matrices. Training samples were generated for the five clas-ses using stratified random sampling method and
     641subsequently the regions of interests were merged in order to create the two main classes, Non-stocked (0% - 10% and 11% - 30%)
     642and Stocked (31% - 50%, 51% - 75%, and 76% - 100%).These training samples were used to perform the Maximum Likelihood (ML),
     643Support Vector Machine (SVM) and Artificial Neural Network (ANN) supervised classifications on the two- and three-band MODIS,
     644VGT and PROBA-V datasets. Con-fusion matrices were generated for all classification results using the LandsatTM-derived 250m,
     645300m and 1000m reference maps , respectively for MODIS, PROBA-V and VGT classifications.
     646= Results =
     647
     648
     649RESULTS
     650The results of the confusion matrices are presented in Figure 4. It appeared that the two-band 250 m MODIS
     651dataset produced more accurate classifications with all classification methods, in comparison with the
     652three-band dataset, which employed the SWIR data resampled from 500 to 250 metres.
     653Comparison between the classification methods for all datasets shows that the ANN classification gives a slightly higher
     654overall accuracy than the other methods for the two- and three-band MODIS data, as well as the simulated PROBA-V, while
     655SVM gives the highest overall accuracy for VGT. However the differences between the different classifiers for each dataset
     656are very small and could be considered to be of low significance.
     657According to these preliminary results, the best accuracy and kappa coefficients were achieved by the ANN classification on
     658MODIS image with two bands at 250 m spatial resolution and the lowest overall accuracy was performed by ML classification on
     659simulated PROBA-V image .
     660
     661In addition to the classification accuracy the percentage of stocked area distribution in each map was also calculated .
     662For MODIS and VGT data, the estimated forest cover was overestimated significantly, from 27% which was the estimated
     663forest cover using the Landsat TM data, to 41 to 48%. On the contrary forest cover estimation with PROBA-V, using the
     664small subset of the scene, for which the simulated PROBA-V data were available, showed significant differ-ences between
     665the classifiers. The Landsat TM data produced an approximate 70% forest cover, while Maximum Likelihood underestimated the
     666forest cover (57.55%) and MODIS and VGT over-estimated the forest cover (80.72 and 79.39% respectively).
     667
     668{| align="center"
     669|+'''Figure 4: Overall accuracy and Kappa Coefficients or ML, ANN and SVM classification result on MODIS, VGT and PROBA-V (B: bands)'''
     670|-
     671|[wiki:File:Table.png File:Table.png]
     672|}
     673
     674
     675{| align="center"
     676|+'''Figure 5: Class distribution of the different classification results comparing to Landsat TM refer-ence (B: bands).'''
     677|-
     678|[wiki:File:Table2.png File:Table2.png]
     679|}
     680
     681
     682{| align="center"
     683|+'''Figure 6: Classes distribution of simulated PROBA-V classification compared to the subset of the Landsat reference map.'''
     684|-
     685|[wiki:File:Table3.png File:Table3.png]
     686|}
     687= Discussion and Conclusions =
     688
     689
     690The error matrix is the most common method to evaluate classification accuracy (4), and it is the starting point
     691for many analysis techniques (4,5). In addition to the overall accuracy, the entire confusion matrix was considered
     692through the use of the kappa statistic. These analyses were used to allow a comparison between classification results.
     693
     694Forest cover in the study area showed significant fragmentation, which tends to pose problems when classifying satellite
     695images (6). This fact proved to be detrimental to the performance of cer-tain datasets. The main aim of this study was to
     696compare the potential performance of PROBA-V in estimating forest cover, with that of VEGETATION (VGT). Because of the small
     697extent of the simulated PROBA-V scene, 250 m MODIS data were also evaluated as an alternative sensor with comparable spatial
     698resolution, which could deliver more statistically significant results. The lack of 250 m SWIR data (the native 250 m product
     699contains data only in the Red and NIR channels) forced the resampling of the 500 m SWIR data to 250 m. Comparison between the
     700classification accuracies between the three-band (with the resampled SWIR data) and the two-band (without the SWIR data)
     701MODIS data showed that the latter achieves more accurate classifications. Spectral variation of the SWIR signal within the
     702500 m pixel could not be retrieved with the resampling pro-cess and, as a result, the combination of accurate 250 m Red and
     703NIR data with “false” SWIR data led to considerable misclassifications between stocked and non-stocked areas On the contra-ry,
     704the use of just the Red and NIR data produced more accurate classifications.
     705
     706Comparison between classifications of the two-band MODIS data and the VGT data, revealed that the former again produced
     707more accurate classifications. The availability of SWIR data in the VGT dataset could not assist the classification
     708sufficiently in order to make up for the increased uncer-tainty brought about by the 1 km spatial resolution of the data.
     709As expected, extraction of forest cover information in a fragmented landscape using low resolution data, proves to
     710be quite prob-lematic at a local to regional scale (4).
     711
     712The simulated PROBA-V data covered only a 20 „e 20 km2 area and consisted of approximately 3600 pixels. From a statistical
     713point of view it is very difficult to extract statistically significant re-sults from such a small sample size, which is the
     714reason MODIS data were also used in this study. Nevertheless, the classification accuracy of PROBA-V data was very similar to
     715the one produced by VEGETATION. On the other hand, the low kappa values of the classifications are caused by the amplification
     716of the errors in the confusion matrix, brought about the small sample size.
     717
     718The accuracies of the classifications produced by the different classification algorithms on the same dataset showed that
     719there is effectively no difference in their performance on all three da-tasets. The only exception was Maximum Likelihood
     720classifier on the simulated PROBA-V data, which was by 5% less accurate in comparison with ANN and SVM, a fact probably
     721attributed to the quality of the simulated PROBA-V data The Artificial Neural Network classification provides the most
     722accurate classifications and appears to deal best with the mixture of different cover types within each coarse resolution
     723pixel (7) and the fragmentation of the landscape (8).
     724
     725Non-thematic errors may be one major problem in the use of confusion matrix and associated ac-curacy analysis (9). This is
     726particularly true for errors of image misregistration, since the images were registered to an image of a higher spatial
     727resolution. The methodology used of aggregation and forest coverage percentage class was used to minimize these errors.
     728
     729In conclusion, comparison between classification algorithms on MODIS and VGT data, showed no significant differences between
     730them, in terms of classification accuracy. The higher accuracy achieved by those classifiers when using the MODIS data, which
     731have comparable spatial resolu-tion with PROBA-V, compared to that achieved with the VGT data, suggest that the increased spa-tial
     732resolution will provide more accurate forest cover mapping, particularly in areas with fragment-ed forest coverage. The usefulness
     733of SWIR data could not be evaluated, but it could prove useful in discriminating between forest and agricultural land, and assist
     734even further the accurate map-ping of forest cover. More research on this topic is recommended in order to quantify the useful-ness
     735of SWIR data at a 300 m resolution.
     736
    585737= User Segment =
    586738
     
    727879The mission is planned to have a lifetime of around two years, following a launch in 2015-2016. The satellites will be launched
    728880together in a stack configuration, with the larger ‘coronagraph’ spacecraft on the bottom to provide control and the smaller
    729 ‘occulter’ spacecraft adopting a more passive role. 
    730  
    731 Following launcher separation, the stack will perform a series of perigee (or bottom of orbit) burns to raise its apogee (or top of orbit)
    732 to 60,524 km. This will be followed by a single apogee burn to raise its perigee to 800 km. After a short preparatory period the
    733 two satellites will be separated from each other and injected into a safe tandem orbit with no chance of collision.
    734 
    735 An approximately three-month commissioning period will follow. One important test will be to demonstrate that the mission’s Collision
    736 Avoidance Manoeuvre (CAM) functionality works. The automated CAM is activated in the event that the two satellites grow dangerously
    737 close, so they can be left safely in a relative orbit.
    738  
    739  
    740 Nominal operations will include both formation flying manoeuvres and coronagraph observation. The cost in fuel of maintaining formation
    741 throughout the orbit would be too great so each orbit will be divided between six hours of formation flying manoeuvres and the two hours
    742 closest to Earth in ‘perigee passage’, based on passive drifting with mid-course corrections.
    743 
    744 Formation flying experiments will routinely take place during the week, with coronagraph experiments being performed during the
    745 weekend. The coronagraph experiment will fill the entire 16 gigabit mass memory of its satellite. This data will then be progressively
    746 downlinked to the Redu ground station during 20 hours of passes the following week.
    747 == Platforms ==
    748 
    749 Both the larger coronagraph and smaller occulter satellites are derived from the Proba standard platform.
    750  
    751 The 475 kg coronagraph satellite has a volume of 1010 x 1010 x 1410 mm. It hosts the coronagraph instrument which
    752 is pointed directly at the occulter satellite and observes only the corona of the Sun. It has a doubled H-shaped
    753 main structure with a set of external panels that make a box for its supporting structure on the port side of the
    754 satellite to hold its deployable solar array. There are also further solar cells located on the support structure itself.
    755 
    756 The satellite’s front panel allows for the opening of the coronagraph instrument, optical metrology sensors and a
    757 Sun sensor but no solar cells as this area will be kept in shadow by the occulter satellite during science operations.
    758 On the opposite side are another Sun sensor and star trackers. 
    759  
    760 Its guidance, navigation and control (GNC) system consists of four reaction wheels, six gyroscopes, one star tracker
    761 with three heads, six Sun sensors, and two GPS systems and antennas. The majority of the formation flying system is
    762 located within the coronagraph satellite as it performs most of the manoeuvres.
    763 
    764 The 245 kg occulter satellite has a volume of 900 x 1100 x 900 mm. For scientific operations its task is to block
    765 the Sun, leaving only the solar corona visible to the coronagraph instrument.
    766  
    767  
    768 It is basically box-shaped, a H-shaped primary structure with all avionics and instrument equipment mounted on the
    769 inner panels, with an 1520-mm diameter occulting disk on the face pointing away from the Sun. The opposite side
    770 accommodates the satellite’s solar array. Its GNC system consists of four reaction wheels, six gyroscopes, one star
    771 tracker with three heads, six Sun sensors and two GPS systems and antennas.
    772  
    773 The Launch date of Proba 3 is somewhere between 2015-2016 ,expect more info [http://www.esa.int/esaMI/Proba/SEMG2R4PVFG_0.html here]
    774 = ESA and RTEMS validation and tools =
    775 
    776  
    777 Saab Space AB performed a validation of the real-time operating system RTEMS. Since it is available for many different targets and includes a multitude of functionality, ranging from I/O drivers to file-systems and beyond, it was agreed to only focus on the parts that were applicable for European space community applications. This implied that only the ERC32 target and a limited sub-set of the configurable RTEMS managers had to be considered.
    778 
    779 Subsequently, Edisoft has reached an agreement with OAR to implement an RTEMS maintenance centre (see related link) in Europe. Edisoft has complemented the validation and the toolset associated with RTEMS for the specific needs of the European space industry. Gaisler Research also provides services based around RTEMS on ERC32 and Leon.
    780 
    781 RTEMS has already been used in several space applications, in particular FedSat (a scientific Research and Development microsatellite), the Surrey's Solid State Data Recorder (a component used in the Disaster Monitoring Constellation), ChipSat (a System-on-Chip architecture), the Electra UHF antenna of the Mars Reconnaissance Orbiter and in the Galileo GIOVE-A and Herschel-Planck satellites.
    782 
    783 
    784 =  See also =
    785 
    786 
    787  * [http://www.esa.int/esaMI/Proba/index.html ESA Proba Missions 400px]
    788  * [http://probav-iuc.org/assets/investigations/Aitkenhead.pdf Aitkenhead 400px]
    789  * [http://probav-iuc.org/assets/201102/PROBA-V_PDF.pdf PROBA-V 400px]
    790  * [http://probav-iuc.org/assets/investigations/DiBella.pdf Investigation DiBella 400px]
    791  * [http://www.esa.int/esaMI/Proba/SEMJJ5ZVNUF_0.html Official Page 400px]
    792  * [http://www.esa.int/esaMI/Proba_web_site/index.html ESA Proba1 400px]
    793  * [http://ilrs.gsfc.nasa.gov/docs/ESA4S_06_11d.pdf PROBA-2 MISSION AND NEW TECHNOLOGIES OVERVIEW 400px]
    794  * [http://ilrs.gsfc.nasa.gov/docs/AIAA_ASC_087086.pdf Autonomous and Precise Navigation of the PROBA-2 Spacecraft 400px]
    795  * [http://esamultimedia.esa.int/docs/Proba/Proba-2_Factsheet_8oct.pdf Factsheet 400px]
    796  * [http://events.eoportal.org/presentations/7111/10001905.html PROBA-V (Project for On-Board Autonomy - Vegetation) 400px]
    797  * [http://www.esa.int/esaCP/SEMNOJRTJRG_index_0.html NEWS about Proba2 400px]
    798  * [http://www.esa.int/esaCP/index.html ESA 400px]
    799  * [http://ilrs.gsfc.nasa.gov/docs/PROBA2_DLR_TN_0020.pdf Technical Information for PROBA-2 Laser Tracking Support  400px]
    800  * [[http://ilrs.gsfc.nasa.gov/satellite_missions/list_of_satellites/prb2_stadata.html Station Data]]
    801  * [[http://inforum.org.pt/INForum2009/docs/full/paper_88.pdf RTEMS Improvement – Space Qualification of RTEMS]]
    802  * [[http://news.bbc.co.uk/2/hi/science/nature/8481321.stm Esa's Proba-2 demonstration satellite views eclipse ]]
    803  * [[http://www.esa.int/TEC/Software_engineering_and_standardisation/TECLUMKNUQE_2.html ESA Operating Systems]]