Home
Detecting clouds in a starlit sky
Contents
1. 0 0 1 0 2 0 3 0 4 0 5 0 6 0 7 0 8 0 9 Fractional visibility error a Fractional error 1 1 0 0 02 0 04 0 06 0 08 0 1 0 12 Visibility error b Absolute error Figure 4 6 Histograms of the error in the visibility due to random fluctuation of recorded flux values 26 For a further 18 of nights no information was recorded at all indicating some general failure of the camera In the remaining 13 the camera refused to operate due to weather conditions which might damage it A maximum wind speed was observed but this limit was never reached Relative humidity was used as an indi cator of incoming rain leading to insufficient images in 4 of nights For the final 9 the camera refused to operate as it could not obtain weather information or the weather information it obtained was greater than 5 minutes old Relative humidity is the only factor above that is directly related to cloud cover Since this is used as a predictor of rainfall it is likely to be cloudy when the camera does not operate for this reason This occurs for 10 of excluded nights which is similar to the percentage of nights I identified as entirely clouded
2. 1 1 06 0 9 en 1 04 Fa oF aH He TY 3 4 1 02 F tet MEE E ae lis E Be E gp 5 a 8 05r 9 oost Ht 2 i zu im gt educ 4 0 96 H T ER AA 0 3 E 0 94 F n l uw eu d os it E 0 92 J L ta E id p 0 1 p E i in 0 9 l 0 HO um m i 02 04 06 08 10 12 02 04 06 08 10 12 Time Hours Time Hours a Example clear night 24th March 2012 b Example variable night 26th March 2012 0 007 0 2 0 18 s 0 006 buds 0 16 0 005 E 0 14 d dis pk Z 0004H 4 ee ee d 2 2 0 1 i J a t B ri 0 003 at n opel i
3. 0 002 nH 4 0 06 ua p La 1 er ue sp a sm 0 04 les HH 0 001 Pu rar Je ae fuss sE EE T E H HEL oe 0 02 re H as T yo EE ay T rh T a ER pen ae ee mp p E eec 02 04 06 08 10 12 02 04 06 08 10 12 Time Hours Time Hours c Example cloudy night without moon 20th March 2012 No stars are visible for any of the images in this night d Example cloudy night with moon 9th March 2012 For all images after 05 00 no stars are visible Figure 3 2 Computed visibility for selected nights during March 2012 12 Visiblity v Range Description 0 0 v 0 03 Very cloudy almost always no visible stars Cloudy 0 03 v 0 15 Almost all of the image is covered by cloud but some stars are visible 0 15 v 0 3 Mostly cloudy a number of stars may be vis ible but dimmed considerably or in small patches of sky Mixed 0 3 lt v lt 0 6 On average sky is half covered by cloud 0 6 v 0 9 Predominately clear but may have some small Clear patches of cloud 0 9 v 1 2 Almost entirely clear Table 3 1 Categories of cloud cover that are easily distinguishable by eye and the corresponding ranges of visibility Images Cloudy Clear Non moonlit 23 66 All 28 51 Table 3 2 Percentages of cloudy and clear images including and excluding moonlit images There is similar variation in the visibility values for entirely c
4. 0 00 00 02 00 Sidereal time Figure 4 4 Visiblity for each image plotted against sidereal time as defined in section 2 3 13000 T r r r 12000 r 11000 10000 9000 r Flux estimate 8000 7000 6000 5000 L ll 1 ji 1 ll L 02 04 06 08 10 Time Hours a Beginning at horizon Flux estimate 40000 35000 30000 25000 20000 15000 10000 5000 m 02 04 06 08 10 12 Time Hours b Ending at horizon Figure 4 5 SExtractor s estimated flux values and errors for two example stars on a clear night 24th March 2012 see Figure 3 2a for calculated visibility of that night 25 Images Images 10000 9000 8000 7000 6000 5000 4000 3000 2000 1000 800 700 600 500 400 300 200 100
5. In Astronomy and Astrophysics Supplement 117 June 1996 pp 393 404 8 E Bertin SExtractor v2 13 User s Manual Institut d Astrophysique amp Observatorie de Paris URL https www astromatic net pubsvn software sextractor trunk doc sextractor pdf 29 Appendix A Visualising the Data A 1 Generating PNG images from FITS files A PNG representation was produced for each of the FITS images in this project These PNGs are used in some of the figures in this report To create these images a maximum pixel value much smaller than the satu ration value section 2 2 was chosen All pixel values exceeding this value were replaced by this new maximum value The pixel values were then scaled to fit within PNG s more limited range As a result of this flattening some moonlit images have entirely white areas which where stars would actually be distinguishable in the original FITS image These stars are also visible to SExtractor Pixels in the PNGs were rendered beginning in the top left corner This is different to the most common choice for FITS files which is to start in the bottom left corner As a result all the images in this report are mirrored top to bottom compared to the real view This change was introduced by accident but maintained to avoid losing the familiarity I had gained with the upside down images A 2 Data explorer interface I built a custom interface for exploring the data that I
6. However the lack of logging information and failure to obtain recent weather information could also be related to cloud cover An example cause would be electrical outages which are more common during storms when cloud cover is high Without further data it is difficult to say whether there is much bias due to when the camera does and does not operate 4 2 Potential further work There were a number of approaches to improving my results that I had insufficient time to attempt To avoid bias in which objects are extracted subsection 4 1 6 all objects with fingerprints could be removed My visibility value is unable to distinguish between an image with thin cloud everywhere and one with a thick patch of cloud covering only part These could be differentiated by calculating the visibility for multiple small areas on the image and counting how many were cloud covered Visibility values for moonlit images could be improved by detecting the area covered by the moon and ignoring it My computed visibility values could be compared to historical meteo rological data for the site 27 Chapter 5 Conclusion Using this method a large number of stars were correctly identified across multiple images A smaller number were incorrectly identified A large im provement of this method would be to detect when such incorrect matches occur The relative visibility seem to indicate cloud cover well but with an error of the order of 10 Curr
7. visibility of their images 18 Chapter 4 Discussion 4 1 Sources of error 4 1 1 Assumptions when calculating the visibility In order to calculate the visibility it was assumed that stars have the same flux distribution Visual checking of sample star s distribution shows this is a good estimate However any deviations will result in an error in the visibility 4 1 2 Inclusion of moonlit images Throughout my results I have made various calculations both including and excluding moonlit images This is because the moon obscures the sky resulting in lower visibility values for clear moonlit images subsection 2 1 1 The real amount cloud cover for moonlit and non moonlit images is ex pected to be the same 4 In my example month s visibility values Figure 3 1 the differentia tion between cloudy and clear images is much more evident for the non moonlit plot In the moonlit plot this is obscured by the lower valued clear moonlit images Table 4 1 shows the ratio of several results for non moonlit and all images As expected individual non moonlit images are more likely to be considered clear and less likely to be cloudy than images in general Since there is no moon obscuration I expect the figures calculated without the moon to be more accurate However when comparing cloud cover across a night selecting only non moonlit images falsely decreases the effective length of the night A shorter night is more likely to be
8. 279 lt 0 3 d 0 3 lt 0 533 lt 0 6 e 0 6 lt 0 880 lt 0 9 f 0 9 lt 0 969 lt 1 2 Figure 3 3 Random selection of non moonlit images from each bin See sec tion A 1 for how these were generated 14 a 0 0 lt 0 020 lt 0 03 d 0 3 lt 0 494 lt 0 6 N 4 hd e 0 6 lt 0 634 0 9 0 9 lt 0 973 lt 1 2 Figure 3 4 Random selection of moonlit images from each bin The small black areas in a and c are due to the brightness of the moon causing pixel values to exceed my saturation filtering point section 2 2 No stars are visible in c this is a limitation of my rendering of these images section A 1 and a small number of stars can be seen in the raw file 15 Proportion of Images Proportion of Nights 0 4 0 35 0 3 0 25 0 2 0 1 0 05 0 5 0 45 0 4 0 35 0 3 0 25 0 2 0 15 0 1 0 05 All Images L Moonless Images 4 0 0 03 0 15 0 3 0 6 0 9 sv lt sv lt sv lt ENE sv lt sv lt 0 03 0 15 0 3 0 6 0 9 1 2 Visibility v Figure 3 5 Proportion of images in each visibility bin All Images 47s 22222222222 Moonless Images gt gt 9 L 2004 773 l 4 Ex al fF ee E CE z 0 0 lt p 0 375 0 375 lt p lt 0 625 0 625 lt p lt 1 0 1 0 Proportion of night clear of cloud p Figure 3 6 Proportion of n
9. 4 analysis is less than the sum of the number of images in each bin I assumed that the latter value was correct 17 Proportion of Nights Proportion of Nights 0 4 0 35 0 3 0 25 0 2 0 15 0 1 0 05 0 6 0 5 0 4 0 3 0 2 0 1 All Images L Moonless Images 3 a 0 0 03 0 15 0 3 0 6 0 9 sv lt sv lt sv lt sv lt sv lt sv lt 0 03 0 15 0 3 0 6 0 9 1 2 Visibility v a Mean visibility All Images Moonless Images 4 0 0 03 0 15 0 3 0 6 0 9 sv lt sv lt sv lt sv lt zv sv lt 0 03 0 15 0 3 0 6 0 9 1 2 Visibility v c Minimum visibility Proportion of Nights Proportion of Nights 0 4 All Images ae 0 35 Moonless Images o l 0 3 TEN 0 25 F 0 2 0 15 F A Oe tr teen ES 4 0 05 4 0 0 0 03 0 15 03 06 09 sv lt sv lt sv lt sv lt sv lt sv lt 0 03 0 15 03 06 09 12 Visibility v b Median visibility 0 6 All Images 3 e mE Moonless Images 4 l 0 5 Fo M 0 4 4 0 3 F eras 4 0 2 B 0 1 a em NI 0 0 03 0 15 03 O06 09 sv lt s lt v lt sv lt sv lt sv lt sv lt 0 03 0 15 0 3 06 0 9 1 2 Visibility v d Maximum visibility Figure 3 7 Proportion of nights in each bin for different aggregates on the
10. Detecting clouds in a starlit sky University of Cambridge Part III Physics Project Ben Webb Supervisor Dr D F Buscher June 27 2013 Abstract I describe a method for automatically determining cloud cover statis tics for images taken by an all sky CCD camera at the Magdalena Ridge Observatory Unlike previous techniques this method does not rely on properties of the lens or existing star catalogues Instead the same star was identified across multiple images by track ing its movement in small steps across composite images These composite images were produced by combining images taken at similar sidereal times The visibility an image was calculated as the ratio of the flux of all stars to the expected flux of all stars Errors of the order of 10 were found in this calculated visibility Various sources of error are discussed The appearance of the moon in images was found to have an adverse effect on the accuracy of the computed visibility Statistics were aggregated by night 24 of nights were found to be entirely clear of cloud whereas 9 were entirely cloudy The method used is expected to be applicable for images from other similar cameras This report is for the most part the same as the one submitted at the end of the project Some small alterations have been made for correctness The source code of the project can be found online https github com Bjwebb detecting clouds Contents 1 Introduction 2 Metho
11. ch pixellil BP hee IB e eM xu dE INSCR eI 21 Histogram of the proportional visibility difference for non moonlit images 4o aao Rcs 4X eee Pe 23 Example of stars being blocked by local camera conditions 23 Visiblity for each image plotted against sidereal time 25 Estimated flux values for two example stars on a clear night 25 Histograms of the error in the visibility due to random fluc tuation of recorded flux values 2 26 31
12. cloudy Instead I chose this threshold by eye However I can see by eye that some images are incorrectly classified which suggests my choice of this threshold is not the limiting factor 4 1 4 Falsely identifying objects Objects may be falsely identified as stars Figure 4 1 shows the density of extracted objects from the images Objects outside the sky are clearly visible before filtering whereas after filtering they are all removed thanks to my use of a mask The plot produced before filtering is also generally much noisier due to objects that are not stars Some of the single white pixels correspond to saturated pixels The other white patches are due to objects falsely identified as stars Such objects include pieces of cloud ghosting of the moonlight in the image and the lens flare False objects due to lens flare are the cause of the white rings The plot produced before filtering only uses objects that were suc cessfully matched to the composite images Many more non star objects would be seen without this step as it is coincidence that causes some to appear where a star could be In the plot produced after filtering most of this noise is gone The invalid points have been removed by a combination of filtering outliers and discounting objects that were not sufficiently matched section 2 7 One of the rings is still faintly visible after filtering This is due to the limit of my current filtering approach only objects wi
13. d 2 1 Images used 2 1 1 Moonlit images 2 2 Initial image filtering 2 3 Sidereal time grouping 2 4 Extracting objects 2 5 Extracted object filtering less 2 6 Matching extracted objects 0 0 2 7 Post matching filtering 2 8 Calculating image visibility 3 Results 3 1 Matching of stars across multiple images 3 2 Visibility values 3 3 Classifying visibility 3 4 Statistics by night 4 Discussion 4 1 Sourcesoferror 4 1 1 Assumptions when calculating the visibility 4 1 2 Inclusion of moonlit images 4 1 3 Measuring cloud cover by eye 4 1 4 Falsely identifying objects 4 1 5 Moonlit cloud 4 1 6 Error in identifying stars 4 1 7 Errors in matching 0 4 1 8 Variation in flux of individual stars 4 1 9 Camera operation bias 4 2 Potential further work 5 Conclusion 6 References A Visualising the Data A 1 Generating PNG images from FITS files A 2 Data explorer interface B List of Figures 10 10 10 13 17 19 19 19 19 20 20 20 22 24 24 24 27 28 29 30 30 30 31 Chapter 1 Introduction The Magdalena Ridge Observatory MRO is a recent telescope site home to a 2 4 meter fast tracking optical telescope A ten element op tical infrared interferometer is also under construction a
14. e This left 4 seconds unaccounted for The two images with a sidereal time in this range were discarded Any bias to my calculations due to losing these two images is much smaller than that due to other errors 1 AII dates and times given in UTC Figure 2 2 An example sidereal time bin composite image Stars should be in the same positions when the sidereal time is the same Thus in each bin the positions of stars will vary by 0 25 degrees Since the images show 180 degrees of sky in less than 640 pixels this corresponds to a variation of no more than 0 9 pixels Thus in each of the images in the bin the same stars will always be in the same position or adjacent positions Images containing the moon see section 2 1 were removed from these bins Then composite images were created for each bin by summing the pixel values from each image See Figure 2 2 for an example 2 4 Extracting objects SExtractor Source Extractor is a piece of software that detects and measures sources in astronomical images 7 In this project it was used to extract the positions of stars from the individual and composite images For individual images SExtractor s flux estimation was also obtained This flux is computed by summing the values of all pixels in an object and subtracting the estimated sum of pixel values if that object were not there As a result it does not have well defined units but comparisons of the flux between images will be
15. e cases when moonlit images were included and excluded The real amounts of cloud cover are expected to the same 4 In order to calculate results for only non moonlit images it was first necessary to determine which images were moonlit The moon leads to many more high valued pixels than all of the stars combined As a result the sum of an image s pixels is greater the more moonlight is visible However since the moon can illuminate the sky when it is just out of sight the amount of moonlight an image can have is continuous In order to definitely include all images that contain the moon I define images to be moonlit if X pixels in an image gt 2 x 108 2 2 Initial image filtering Saturated pixels were removed by replacing any values greater than a fixed threshold with 0 Some pixels lit by the moon also exceed this value but only if they are lit directly Since such pixels would otherwise contain the image of the moon there is no adverse effect on my calculation of cloud cover 2 3 Sidereal time grouping The following definition of sidereal time was used ts At mod L Where ts Sidereal time seconds At Seconds since 2011 01 01 00 00 00 L Length of mean sidereal day 86164 091 seconds 5 There is a fixed offset to other definitions of sidereal time This does not matter since I will not compare my sidereal times to any from other sources Images were binned ie grouped into minute long ranges of sidereal tim
16. e made less distinct by the cloud from other images If the number of stars missed varies across the sky then different areas of sky will be unfairly weighted in the calculation of visibility The most noticeable example of such variation is due to fingerprints These occur repeatedly in the same position so are quite visible in the composite images refer back to Figure 2 2 for an example Although some stars were successfully extracted within the prints the number is fewer than for other nearby areas of sky If the number of missed stars is greater for some composite images than for others then there will be a systematic error in the visibility that varies with sidereal time Figure 4 4 shows the visibility of images plotted against sidereal time There is a variation with sidereal time of up to 10 in the maximum visibility value This is likely to be a major cause of the error in the visibility for clear nights discussed in section 3 2 22 Images 3000 2500 r 2000 r 1500 1000 500 r 0 016 0 014 0 012 0 01 0 008 0 006 0 004 0 002 0 0 002 Proportional visibility difference Figure 4 2 Histogram of the proportional visibility difference UR Y for non moonlit images Figure 4 3 Example of stars being blocked by local camera conditions 23 4 1 7 Errors in matching The most serious error i
17. ently the most problematic error is due to the different behaviour of moonlit images Aggregating by night it is found that 24 are entirely clear whereas 9 as entirely cloudy It is evident that the Magdalena Ridge is a good choice for an optical telescope site This method could be used to track cloud cover in other locations provided a sufficiently large number of images from the same camera were available 28 Chapter 6 References 1 Daniel A Klinglesmith III et al Astronomical site monitoring system for the Magdalena Ridge Observatory In Astronom ical Telescopes and Instrumentation International Society for Optics and Photonics 2004 pp 1301 1309 2 TE Pickering The MMT all sky camera In Astronomical Telescopes and Instrumentation International Society for Op tics and Photonics 2006 62671A 62671A 3 Yin Jia et al Processing Method of Night Time Cloudiness for Astronomical Site Selection In Chinese Astronomy and Astrophysics 36 4 2012 pp 457 468 4 TJ Lauroesch JR Edinger Jr and JT Lauroesch Full moon and empty skies In International journal of climatology 16 1 1996 pp 113 117 5 US Nautical Almanac Office Astronomical Almanac 2013 US Nautical Almanac Office 2012 B9 6 P Kenneth Seidelmann Explanatory supplement to the astro nomical almanac University Science Books 2005 7 E Bertin and S Arnouts SExtractor Software for source extraction
18. free of cloud Selecting only these images Cloudy Clear Image 0 8 1 3 Median of night 0 8 1 3 Entire night 0 8 1 9 Table 4 1 Result using non moonlit images for the proportions of images nights that Result using all images are cloudy clear not clear ie cloudy or mixed 19 means that nights are 1 9 times as likely to be considered clear This is considerably larger than the 1 3 times increase for a single image There fore for whole night statistics the value computed using all images is likely to be more accurate The factor for cloudy nights is the same as that for cloudy images 0 8 I suspect this is because nights that are entirely cloudy no mixed images at all for part of the night tend to be entirely cloudy for the rest of the night 4 1 3 Measuring cloud cover by eye One of the large sources of error in the project was judging cloud cover by eye This is important because in general this is the limiting factor of my ability to judge the correctness of my visibility calculation Cloud cover in moonlit images was sometimes difficult to judge by eye as the moon illuminates the entire image This is especially true of my rendering to PNG images as included in this report since high values were flattened section A 1 This also has an impact on my statistics for cloudy night My visi bility calculation does not itself suggest what visibility ranges should be considered
19. generated This made use of the PNG files described above Plots of the visibility and the flux for each star are accompanied by the relevant images These images can be animated through tracking the corresponding time on the plot This allowed me to very quickly verify whether the visibility changed due to the cloud that I could see by eye For individual stars these animations were annotated with a box in dicating the tracked object This allowed me to quickly verify by eye that stars were being tracked correctly 30 Appendix B List of Figures 2 1 2 2 2 3 2 4 3 1 3 2 3 3 3 4 3 5 3 6 3 7 4 1 4 2 4 3 4 4 4 5 4 6 Overview of the method used 4 An example sidereal time bin composite image 6 The mask applied to objects after extraction 7 Lines drawn for the tracks of a random selection of the objects extracted from a composite image 8 Computed visibility for images taken during March 2012 11 Computed visibility for selected nights during March 2012 12 Random selection of non moonlit images from each bin 14 Random selection of moonlit images from each bin 15 Proportion of images in each visibility bin 16 Proportion of nights with varying proportions of cloud cover 16 Proportion of nights in each bin for different aggregates on the visibility of their images les 18 Density plot of the objects extracted and matched at ea
20. ights with varying proportions of cloud cover Val ues of all images and non moonlit images are shown The bins are chosen to correspond roughly to entirely cloudy 25 50 and 75 and entirely clear These are the bins used by Klinglesmith s 2004 analysis of all sky images from the same telescope site 1 Values from this are also shown 16 Entirely Median Images Cloudy Clear Cloudy Clear All 9 24 30 49 Non moonlit 7 46 24 66 Table 3 3 Percentages of nights that are entirely clear and cloudy and where the median image is clear and cloudy 3 4 Statistics by night Statistics were also calculated for each night using the same bins Fig ure 3 7 Nights with 10 or fewer valid images were excluded to avoid bias from a small number of unrepresentative images The median visibility is more likely to be extremal than the mean This is probably due to nights which are mostly entirely cloudy or clear but then conditions change Entirely cloudy clear nights can easily be identified as those where the maximum minimum visibility is above below the relevant thresholds These results are summarised in Table 3 3 Nights were also grouped by the proportion of time they were clear Figure 3 6 The same bins were used as for the 2004 analysis of the all sky camera 1 The results for this analysis are similar to the values I computed using all images T he stated total number of images in the 200
21. imate used for the error bars on Figure 4 5 is computed based on the background flux 8 which does not scale In dimmer images the estimated error is comparable to the actual variation However for the brighter stars the lack of relation is clear In addition the same variation can be also be observed by manually inspecting the pixel values of a star across a series of images The variation in flux is much larger than the variation in visibility This suggests that stars do not all vary in the same way between images so this variation behaves as a random error in the flux These qualities suggest that the variation is a result of the camera setup itself A potential source of the error is that starlight sometimes illuminate the gaps between the pixels instead of the pixels themselves Assuming the errors in flux measurements are independent and that Aflux z flux the fractional error in the visibility is Av _ es _ gt median flux v V QS flux V5 median flux Figure 4 6 shows the range of values of this error The modal fractional error is about 0 1 meaning that this could be the major contributor to the visiblity error 4 1 9 Camera operation bias My results have necessarily only been calculated with images that have been taken by the camera For the date range considered only 6996 of nights had the 10 images or greater required to be included in my results 24 Visibility R Iaa Ast Lad a AES
22. is calculated relative visibility simply as the visibility Chapter 3 Results 3 1 Matching of stars across multiple im ages SExtractor successfully extracts the supermajority of stars visible in im ages Using the method described above 3667 distinct stars were found These corresponds to 53 percent of objects extracted by SExtractor which do not fail the most basic tests of not being a star such as being outside the sky or being overly large On average there were 503 correctly matched stars in each image Inspecting a sample of such objects by eye shows that the majority are matched correctly The most common problem is for a single star to have been falsely identified as being two or more distinct stars ie the tracking of that star s progress across the sky in the composite images has been broken note the broken lines in Figure 2 4 There were also a small number of stars where two different stars were being tracked These problems are discussed further in subsection 4 1 7 3 2 Visibility values My computed visibility values Figure 3 1 for a sample month were seen to be a reasonable indicator of cloud cover when compared by eye to the cloud cover at each point For nights where the amount of cloud cover varies e g Figure 3 2b the change in cloud cover in the images is strongly related to the change in the computed visibility This can be seen clearly on an animation through the images that tracks the pos
23. ition on the flux graph Nights that are entirely clear are expected to have a constant visibility since all images are entirely free of cloud However there is variation in my computed visibility values The amount of variation is similar for all such nights By inspecting the magnitude of the variation for a typical clear night Figure 3 2a an error estimate for of 0 1 was obtained for the high visibility limit 1A custom interface was built to allow such animations to be quickly viewed for any visibility or flux plot See section A 2 for more details 10 ur rat pit wee uso MEL L wor Visibility Visibility 0 8 0 6 amp 0 4 0 2 1 2 0 8 0 6 0 4 0 2 Time Day of month a All Images POP i is I M LEA Dod 1 Po d T 2 TEE Nc E Vag tg 518 ROC O8 n n te A 4 m pez l i E a 1 F f T n 5 ee y i i i 1 i 1 1 1 Lt i lam i 1 i 02 04 06 08 10 12 14 16 18 20 26 28 30 01 Wr RES rydid EE i i d o4 Pe dp o eeocut S ig 02 04 06 08 10 12 14 16 18 20 Time Day of month b Non moonlit images only Figure 3 1 Computed visibility for images taken during March 2012 11 26 28 30 01 1 08
24. loudy images where no real stars are visible Figures 3 2c and 3 2d show typical no star cases for moonless and moonlit nights Ideally the visibility should be zero here so the error can be estimated to be the largest value that the visibility commonly obtains Thus error estimates of 0 005 for non moonlit images and 0 05 for moonlit images can be made for the low visibility limit 3 3 Classifying visibility My visibility parameter is non linear To aid comparison over my whole data set I chose a set of bins such that each covered a category of cloud cover that I could distinguish by eye These categories and the corre sponding bounds of the visibility bins are shown in Table 3 1 I also grouped each of them into broader categories of cloudy mixed and clear Figures 3 3 and 3 4 show examples of these images from each of the bins for non moonlit and moonlit nights respectively Random selections of images such as these show that most of the images match their bin description well The most common error is for images to be placed in the bin higher or lower than the one they would be most suited to Figure 3 5 shows the proportion of images in each bin A summary of the percentages of cloudy and clear images is shown in Table 3 2 Images are almost twice as likely to be cloudy Non moonlit images were almost three times as likely to be cloudy 13 a 0 0 lt 0 001 lt 0 03 b 0 03 lt 0 034 lt 0 15 c 0 15 lt 0
25. mallest sidereal time difference were chosen Of those the object at the shortest distance was used By this method the line a star takes is slowly tracked across the sky A graphical representation of several such tracks can be seen in Figure 2 4 Objects extracted from individual images were then matched against those from the corresponding composite image Those pairs of objects with positions closer than 3 pixels were considered to match In the case of conflicts the nearest objects were chosen It can now be inferred which objects match between different individ ual images ie those which are the same star as they will have both been matched to composite objects which have been matched to each other 2 7 Post matching filtering Any suspected stars where less than 200 objects were matched in the composite images were discounted A typical star will appear in 600 or more composite images Objects appearing in less than 200 are likely to not be stars or to be stars that are not extracted or matched consistently see subsection 4 1 7 for more discussion of this Objects with negative fluxes or fluxes anomalously high for their line were also discarded Negative fluxes arise when the flux of an object is smaller than that of the surrounding background These occur due to patches of sky sur rounded by illuminated cloud Such objects may be safely discarded Some objects that are not stars are falsely matched to lines These objec
26. n matching is falsely identifying two objects as the same star The likelihood of this is increased by the possibility of SExtractor did not extract the correct star However the number of instances of this in practice are quite small The incidence of this error has been reduced by my filtering Bad matching generally results in a small number of objects being matched These cases were eliminated by removing objects with a small number of matches section 2 7 However this removal also exaggerates the effect of stars missed during extraction The broken tracks are more likely to be removed from the data This means small obstructions can exclude the star across areas much larger than themselves For example the broken paths between the two fingerprints in Fig ure 2 2 were removed This meant that cloud between the two fingerprints had a lesser effect on the visibility than other similar areas of sky 4 1 8 Variation in flux of individual stars For a visibly clear night it is expected that the flux of stars will be close to constant with some dimming due to atmospheric extinction at the horizon Figure 4 5 shows the flux of two stars on such a night Near the horizon the stars show the expected dimming However when the stars are away from the horizon there is a large random variation in their estimated flux This scales with the flux of the star the maximum flux is always approximately twice the minimum flux SExtractor s error est
27. ntly small that double peaks will not occur for one object so deblending is always optimal Thus the minimum contrast was set to 0 2 5 Extracted object filtering SExtractor extracts several types of objects that are not stars These include objects outside of the sky clouds illuminated by the moon and planets In this set of images none of the stars are greater than 20 pixels wide or high Therefore all such objects are ignored A simple mask was created Figure 2 3 that covers only those parts of the image where the sky is visible Objects outside this mask are ignored Unfortunately some small areas of sky near the horizon are not contained in the mask due to the difficultly of including these areas whilst excluding non sky areas 2 6 Matching extracted objects Two objects in different composite images are identified as being the same star if they have similar positions and the composite images are for sim ilar sidereal times This equivalence is transitive I limited matching to objects within a 3 pixel distance of each other and where the sidereal time difference is no greater than 4 minutes Objects are matched across Figure 2 4 Lines drawn for the tracks of a random selection of the objects extracted from a composite image several bins in case the star is missing from the adjacent bin discussed in subsection 4 1 6 Sometimes multiple matches were found for an object In this case the objects with the s
28. nts Visibility 82 4 82 5 82 6 Cloud cover was estimated by comparing the flux of stars in an image to the flux of the same stars in other images section 2 8 This requires identifying the stars in an image section 2 4 and then matching the same star across multiple images Since the images do not offer a continuous view of a star for example due to cloud cover and daytime it is not easy to track a star across them Instead composite images were produced for a series of sidereal time ranges section 2 3 These were used to follow the positions of stars The same star in multiple individual images was then identified using the information from the composite images section 2 6 Various filtering steps sections 2 2 2 5 and 2 7 were applied to ensure that only real stars were used 82 7 82 8 2 1 Images used Images taken by the all sky camera between 25th March 2011 and 30th November 2012 were used During this time period images were recorded with a 15 second exposure every 5 minutes when possible subsection 4 1 9 discusses this in more detail Earlier images for 2011 were available but were unsuitable due to differing exposure times and image resolutions 2 1 1 Moonlit images Images containing the moon present several challenges The brightness of the moon obscures stars leading to a lower calculated visibility than for the same conditions without the moon For these reasons I calculated re sults for both th
29. t the same site Optical telescopes require clear sky to function well In order to measure cloud cover an all sky camera was installed at the site A 2004 analysis of images from this camera found that 23 of nights were completely clear of clouds 1 This number was obtained by measuring the cloud cover for each night by eye This becomes tedious as the number of images increases In this report I present a method of automatically determining the cloud cover I apply it to images taken by same camera during 2012 and 2013 Such automatic measurements have been made for other all sky cam eras This has been achieved by using information about projection of the lens to compare the stars in images to those in existing star cata logues 2 3 In contrast my method does not rely on such properties of the camera setup or require the use of existing star catalogues Instead dimming of stars is identified by comparing each stars across multiple images 2 1 2 2 2 3 Chapter 2 Method Figure 2 1 Overview of the method used Remove Select ee L Short mages Lines Y Filter Filter Sr Remove Saturated gt SExtractor gt Extracted o Outlying i Sidereal Pixels Objects Fluxes J Points C N Be t TN Group by Filter Connect Calculate Sidereal gt SExtractor Extracted 4 Average Image Time Objects Poi
30. th much higher fluxes than the star they are misidentified as can be removed Reducing my filtering limit would begin removing legitimate instances of stars and would skew the median flux of that star 4 1 5 Moonlit cloud Moonlit cloud is the most common object misidentified as a star This probably causes the tenfold increase in error for moonlit cloudy images 20 b After Filtering Figure 4 1 Density plot of the objects extracted and matched at each pixel summed over all images Black indicates no stars lighter colours indicate large numbers of stars 21 compared to equivalent non moonlit images It is also possible that including moonlit images for computing medians would affect the visibility of non moonlit images Erroneous high flux objects from moonlit images increase the median flux of a star This decreases the visibility of other images However due to the filtering of outlying fluxes section 2 7 this effect should be minimal Figure 4 2 shows the distribution of the proportional difference be tween visibilities calculated with and without moonlit images in the me dian The visibility as expected almost always decreases with moonlit images in the median However this decrease is always smaller than 1 6 much smaller than the other causes of error 4 1 6 Error in identifying stars Also visible in Figure 4 1 are a number of darker areas where less stars were correctly extracted One dark patch is around
31. the horizon where stars appear dimmer due to atmospheric extinction Atmospheric ex tinction is the dimming of stars due to light passing through a greater volume of air This is true for stars near the horizon due to the geometry of detection The section of the sky most near the horizon is black since it is not included in my mask The other dark patches are due to fingerprints on the lens in some images blocking stars from being detected in those images The bias against objects on the horizon may be acceptable as areas near the horizon are less useful for astronomical observation Fingerprints meanwhile are an example of objects near the camera which are misinterpreted as cloud cover Other such objects include rain drops frost and a spider Figure 4 3 This effect could be mitigated by filtering out such images In general such problems are difficult to detect automatically Also most such conditions affect only a small portion of the sky so the effect on the cloud cover estimate for an image is minimal In general SExtractor is not able to extract all stars from an image which leads to an error in the visibility This error may be systematic if there is some bias to which stars are missed Enough missed stars also leads to broken tracks where a single star is incorrectly identified as multiple different stars when it is in different places This is especially problematic for my composite images where stars from clear nights ar
32. ts can have wildly different fluxes which will skew the calculated visibility Visual inspection shows that the flux of almost all valid stars does not exceed three times their median flux Therefore all points lying outside this range were discarded 2 8 Calculating image visibility With ideal data the absolute visibility of an image could be defined 9 Flux stars in image gt D Maximum flux stars expected Visibility The expected stars for an image are those in the corresponding composite image The maximum flux is calculated across all images However the real maximum flux of a star is difficult to determine since other objects with greater flux are misidentified as it Assuming all stars have the same flux distribution replacing the me dian gives a scaled version of the real visibility This allows us to calculate the relative visibility 9 Flux stars in image bE Median flux stars expected Visibility The median flux is calculated across all images Whereas the absolute visibility has a maximum value of 1 the relative visibility could take arbitrarily high values In practice however stars are detected more often when they are close to their maximum flux so the median is close to the real maximum value As a result the relative visibility for clear images is close to 1 No images were found with a relative visibility any greater than 1 2 Throughout the rest of this report I will refer to th
33. valid if they are taken using the same A mean sidereal day is actually defined as the time for one rotation of the Earth relative to the vernal equinox Since this precesses a mean sidereal day is actually 0 0084 seconds shorter than the actual period of the rotation relative to the fixed stars 6 This amounts to just over 5 seconds difference from my first image to my last image Since this is much smaller than my bin size it will not have an important effect 3360 number of bins 360 39190 0 25 Figure 2 3 The mask applied to objects after extraction The black area indi cates objects that were included It consists of two semicircles fitted by eye to cover as much sky as possible whilst only including sky setup In order to extract the maximum number of valid stars several of the default SExtractor configuration parameters were changed Minimum Detection Area Set to 1 pixel since some stars in the im ages only fill 1 pixel Background Mesh Size Reduced considerably since the background of these images is much more variable due to the possibility of cloud cover and the edge of the sky than the clean images of star fields SExtractor is often used for Minimum Contrast for Deblending Deblending refers to treating flux maxima separated by a non background minimum as separate ob jects The minimum contrast controls how deep this minimum must be for deblending to occur The stars in these images are sufficie
Download Pdf Manuals
Related Search
Related Contents
Kenroy Home 91557CH Installation Guide Descarga ficha técnica Ref. 4301 Ref. 4302 Ref. 4304 Ref. 4381* Ref. 4308 Ref. 4309 Ref Draft JCONSTRUCCIÓN Y FUNCIONAMIENTO 1. Panel de control del Copyright © All rights reserved.
Failed to retrieve file