Great video and thanks for the shoutout! I agree that obsessing over SNR can be a problem. A couple of random thoughts: 1) 8:42 You nailed it here! Getting better data is so very important! If you have really good single subs, the SNR calculators become much less important. 2) The SNR calculator I made gives raw SNR, not in decibels. dB are nice in that they try to scale things based on human perception (i.e., doubling dB should double the perceived quality), but... it's a log scale and you do need a reference point for dB to be really meaningful. That's why I think the pixinsight one is not that useful (it assumes too much). I think a raw SNR is more intuitive, which is why I left it in raw :) 2) I think doing SNR calculations are useful for experiments! Do you really want to know how well your filter is doing compared to no filter? Well, a good SNR measurement is actually a nice metric, especially if the experiment is based on limited imaging time. 3) I agree that SNR calculators aren't that useful for after you've stacked your images. But I think they're great for planning (if you're the kind of person who likes to plan of course). If I know that getting 30 raw SNR on the fainter parts of a nebula will make my image look good (the midtones and brighter areas will be much better), I can use the SNR calculator to get an estimate based on a single sub just on the fainter part. I can then plan my imaging time accordingly. 4) SNR is really, super important for deconvolution (someday, I'll make a video on it... I think!)
if you want scary good snr calculators there are ones made for the big science telescopes. they need a lot more info on ur setup but then theyre like magic
I started doing Astrophotography in 2020 and I'm sure like a lot of people I was a "stamp collector" trying to get a couple targets each night, spending maybe 2-4 hours on each before moving on. This year I switched to picking targets that I can spend multiple nights on, including a recent image of the Crescent Nebula that I spent 33 hours on over 11 nights. Now, I'm starting to earn Top Pick designations on AstroBin. It's a different kind of satisfying and I nice evolution on my journey.
Not that I know much about this but being an electronic engineer I did learn a little many years ago about electronic noise. It is basically generated randomly in any electronic circuit. It has to do with heat influencing electrons and from that we get the noise signal. The colder it is the less noise. At absolute zero Kelvin temperature there should be next to no noise if not zero noise. As noise is random and at least a sound signal is not it is possible to have a very weak input signal, amplify it through two amplifiers and then compare the output. This way it is possible to get a reasonably clear output from input signals well below the noise limit of a single amplifier. For digital pictures I would thing it should be possible to do the same. Take two pictures, compare the two and throw away everything that is not on both pictures. It should be possible to do the comparison electronically. I now come to a question that I have to every body. It is about the Cosmic Microwave Background Image. When scientists talk about it they often liken it to the noise signal on an analogue TV set. They even say that it is Cosmic Background signal. To my knowledge and understanding this signal in the TV is nothing other than random heat noise especially as there is no signal coming in to the Tv. I was fixing such TV sets many years ago and I know I could reduce this noise by fixing the TV.
As an Army Officer on tour I have had many different weapons pointed at at my head during negotiations overseas and kept my cool. Your threat to kick my tripod sent a shiver down my spine. You are a monster, Dylan. Particularly because I had an awful night last night. My tracking was all over and I was grabbing 5 minute subs (only managed to get 32 in the entire night . . . ouch). I still haven't figured out the problem. The hobby you love and hate. It does remind me of my years in RECCE where we took pride in our deprivation and pain.
"I will kick your tripod!" 😂 Love it! Anyway back to acting my age, personally I just do what you suggest Dylan, gather as much data as possible and hope there's enough there to play with in post. I've recently been getting data on C34 "Soul Neb'/Baby Neb'" and noticed it requires much more data than the circa 2 hours I initially got with my OSC camera. By contrast, it's neighbour IC1805 "Heart Neb'" with exact same mag' ( +6.5 ) really popped with the same amount of data. Data is king! Congratulations on reaching 50k+ subscribers Dylan! Without wanting to sound sycophantic Impo your channel deserves 1 million+ subscribers! 💥👍 Thanks as always Dylan! Wes.
Sweet! I are engineer... I care about the maths. I care about all the maths. This was fantastic! Although I wish I never had to know that astro-py existed. You and Seti are opening the literal Astro-ChatGPT-Pandora's box!!! I love it.
what if you used a cross reference from plate solving first so the program can cross refence known stars and remove them from the noise data to help distinguish the two for a better rating? not sure how to do it but just spit balling.
Hi Dylan, Great video and great tool. The argument I'd make for keeping this up and maybe a path to improvement is the following use case. When shooting mono, you don't necessarily know the amount of signal in each filter. That means I'm guessing blindly if it's woth doing 6 hours of O3 on a given target, only to discover that either there is no O3 in this target or that I should have shot 3 times more O3 than Ha. By evaluating a few shots of each filter, it could give me the ratio of Ha/O3/S2 to capture a reasonably balanced image.
Great video. I create a small preview of just the space and then use the PI SNR tool. You don't want to include the stars as these have a much higher SNR value. I am for a 42db SNR on a background preview and that is a pretty good rule of thumb. As soon as you run dbe, you lose the gradient and sky brightness so that reduces your overall SNR. I use SNR as a guide for where I want to shoot more. Generally speaking, a doubling of integration time will gain you 3 db of SNR and this is something you can detect. That goes back to a rule I learned early on. If you aren't happy with your image, double your integration time and try again.
@@DylanODonnell Thanks! I'm experiencing the effect of SNR first hand now that I've started shooting with a RASA. The larger aperture gives me a signficantly higher SNR per shot vs. my slower refractor and vastly reduces the amount of integration time I need on a target to get the result I want, especially on fainter targets. Always enjoy your videos and congrats on hitting 50k subscribers. Keep up the good work!
Maybe a database of popular targets with 'perfect' signal could be used to compare your own image with, at least a section of it. This could then be extrapolated to your whole image for a 'true' snr. Something like the Gaia database for SPCC and the upcoming MARS project for 'true' gradients. But then again, is knowing the Snr that necessary? In practice, I simply try to get enough data to make an image 'pleasant' to process. For my Edge11, this means no les than 20 hours on a target or the snr is too low. Some interesting ideas here! Cheers.
I used to bin RGB and shoot L full resolution. Now that i upgraded to an ASI294mm Pro, ive been considering dropping L and doing all exposure times RGB only. This tool can be useful for optimizing things since the weather has been so hazy this year, I just haven't had a lot of time for doing empirical tests with very long exposure stacks. Thanks for another great video.
Where I’ve used the pixinsight Snr script and found it useful. Comparing the same image stacked multiple ways in particular 1x drizzle 0.9 with 1x drizzle 0.8 etc. I think there’s a point where drizzling works amazingly well based on the number and quality of subs below which. Images just get noisier. It also bothers me that there will inevitably be areas of an image (in most cases) where there’s simply very little signal. Perhaps when comparing images, it would be better to define a box size to sample and have two figures. One based on a background area and one for an area of high nebulosity which we are most concerned about. Great video by the way.
what if you feed a better external SNR image as a calibration to make it understand what a good SNR for a pixel. Might need some zooming and cropping. Long workflow
Having lived through about 55 years of the history of astrophotography so far, many times I've seen theory catch up with practice. That is, the photographers found out how to make things work better in some way, and then the theoreticians figured out why. Keep up the good work!
Isn't that the same with such a lot of things. They make a theory. They calculate and calculate and then tell us all how it should be. We then test it and find that it isn't quite as they said. Then they go over it again, change their theory a little and let us test again. Eventually we reach our capacity to test and they catch up. After that they claim that it was their theory that predicted the result and we confirmed it. I am an Engineer.
Decibels ARE ratios. Their most common usage is in sound volume (ratio of the peaks to the troughs if a sound wave) but they are applicable to any ratio.
Glad to see there are others using python for things like this. I wanted to know how much to expose as to not saturate a star .... took around 1000 lines of code, but the result was pretty good. With time you will love astropy more and more!
Very good that you talked about it! I was recently trying to compare different filters on the same target, but on different nights (different moon). And the results were just not fitting the clearly visible differences. So in the end I ended up using those SNR calculations just for comparing RELATIVE SNR for the same rig (including filters) and the same target. Don't spent your time, thinking about and calculating the SNR. Just go out and suck in some photons ;)
This is awesome, Dylan I have a question for you, I was trying to image the cresent nebula, I'm in a bortal 6-8 sky. I cant seem to get the middle portion of it. I was doing 2-3min exposures have about 100.Some people are able to get this cool blue vail around it. Any tips? Congrats on the 50k, you've inspired me a lot throughout the years.
So I'm not sure whether I do or do not want to see the ChatGPT transcript in that 4 day conversation. Ha! Great video...better equipment and more time...right answers! Congrats on the 50k!
Hah.. well I find them funny. It's mostly stuff like Me: "This is wrong, why?" ChatGPT: "It looks right to me" Me: "You forgot to multiply this number! that's why it's wrong!" ChatGPT : "You're right! here's the new code..."
What if you have Noise exterminator exterminator the hell out of the image. The result could be the theoretical best instead of the "perfect world" bit of the current equations? Would that then take into account all the other personal variables? Use that to find your exposure requirements?
@@DylanODonnell Python is the language most scientists use. Our PHD loved it. Anyways, I want to learn it too. I like the command line format. It is very similar to BASIC, in that respect. Cheers
Would it be correct, as a Siril user, to take a single frame pp_light.fit, open it in Siril, then save it as a 16-bit unsigned .fit for the purposes of uploading for the SNR calculation?
@@DylanODonnell Oddest thing my uncalibrated 10 second light is at 23dB; calibrated light is at 15dB. I tried calibrating a stack normally (flats, bias, darks) and also darks-only. Same result, reported SNR on the pp_ file is worse than on the light. Calibrations are at same 0 C as lights. 30 darks bult into a Master Dark; 50 bias into a Master Bias; and 30 Flats with histogram about 2/3s out. The calibrations are off a stock Siril script. And tbh, it's not like my stacked images are nasty. I'm just experimenting for the purpose of confirming my 10s procedure (vs going to 30s). Robin Glover expose led me to 10s in my Bortle 4 skies with 14" at f/5 into ASI2600MC Pro. I go 100-200 subs per run. I thought an SNR evaluation would maybe point to missing details that I'm not noticing by eye.
Great explanation!!! Thank you!!! I also read "The Deep-Sky Imaging Primer" by Charles Bracken. Which also gives a very deep/great explanation of SNR in the first few chapters.
Thanks for the down to earth analysis… I’ve reached the same conclusions… Many look at it as a noise problem, I switched my thinking awhile back and look at it as a signal problem only, it makes it easier… More signal = better image. At that point the relative noise becomes almost irrelevant…
Yes the real ratio is so small eventually it’s negligible.. and even if it isn’t a few little tweaks in post either with levels or denoising tool and boom .. nice image.
How about running StarXterminator first to remove the stars from your calculations? Harder to remove the nebula or galaxies from the general noise. But maybe mask them out
+1 for being the only astrophotography youtuber I wouldn't mind having a beer with. So I was going to complain about this SNR video because you didn't help me understand it any better than I already did... but you actually helped me untangle my mental block I think. Because this is what I want: I want to shoot a few subs in each color channel and then get a ballpark number as to what ratio of each I need to shoot at to get something usable. I always figured I wanted to compare the SNR of the different channels. But now I am thinking to forget SNR and instead is run each sub through star exterminator, so I'm only looking at noise + nebulosity, and then look at the intensity histogram. This should give me an idea of I need at least 3x more Oiii than Ha etc.... does that seem right?
Yeh that'l work .. but I tend to keep the exposures the same between HA and OIII so the stars don't bloat more than the other in one channel .. even though I'm removing them anyway for RGB .. But if you get enough data in OIII you can stretch it harder when you do your levels / histogram stretching.. I just stretch OIII harder then you can see all that oxygen over the Ha signal, even though you took the same exposure length.
An idea to help improve the graph for SNR is to have the SNR calculator calculate the SNR of different bits of raw data from various people at various locations and bortle scales. Try to get data (from various people or sources) from Bortle 1-2 zones, bortle 3-4 zones, until you reach 9-10 zones. The more data there is, the better. Utilizing all this data from the various telescopes, sensors, etc., will allow you to in a way, average out what the SNR should roughly look like, allowing the graph to give an additional line which I would call the "Mean SNR Line". This serves no other purpose than to show what the ideal SNR is. To further separate the data, dividing the sky based on the different constellations could further improve the SNR calculator, as the SNR is going to be different on a nebula compared to a galaxy (assuming all parameters are the same). However, this shouldn't be limited to specific targets, but rather all the targets within the constellation combined into 1 mean line, since people like astrofalls are constantly discovering new nebula seemingly every few months. No matter what, you are correct. No SNR calculator will ever be 100% accurate, and long exposure combined by taking 100's of images and only selecting the best ones will give the best results.
The S/N ratio can be liken to watering down beer. Too much water (noise) makes the beer (signal) taste weak. In a perfect world there is only pure beer.
Another good explainer mate👌 and remember people "Comparison is the thief of Joy". We are not all imaging under the same skies with the same gear. Damo
I wonder how PixInsight calculates the SNR if it doesn't know what the pure signal is. I've done some SNR stuff before for dental x-ray calibration but it was always relative to a long reference exposure. EDIT: yeah should have watched more of the video lol
Here's an idea for next calculator - have a reference db of pictures classified by humans as high/mid/low SNR to be used as additional input for your script. Congrats on 50k!
The problen with "get more data" is that it's difficult to guess when to stop. As a beginner, I always run in to a dilemma where I want to shoot many targets as there is so much to shoot but also I don't want to spend 50 hours on one target just to get better SNR.
I loved this video, especially the reference to the reality that SNR calculators have no real idea what they are looking at. It's so easy to get meaningless results and, like all statistical measures, the information is never absolute and must be interpreted. TBH, I never bother to look at SNR. The image is done when the data shows the detail I am after.
While it was great fun to watch someone else get all wrapped around the axle on something mostly useless, I think you nailed it when you told us to go to High Point and buy better gear. Gonna do that now. And maybe, just maybe, this will be the year I get off my ass and haul my (new) gear out to the desert. Thanks.
Any chance you could post a video talking about OSC imaging but using narrowband filters? I’ve seen some amazing images with people using OSC with narrowband filters but I can’t find any info on how they edit, and no one responds to comments LOL
@@DylanODonnell I WANT TO! But I wanna ace OSC first! Only got my telescope November last year, and then got my Astro camera like 2-3 months ago. Don’t tempt me just yet 🤤🤤
All of this video 👏 Just shooting until you see diminishing returns under your conditions with your setup. Who really cares beyond that :D Btw, I watched til the very end, congrats on the 50K. Kick my tripod though and it'll kick you back twice as hard.
welcome to astropy...matplotlib numpy etc anaconda makes it pretty easy to jump in... that was alot of code for what you did... i think. What'd you use for the web stuff
Yeh there’s some archiving stuff for reviewing submissions and a lot of bombproofing to handle mono and OSC .. server is just CPanel with python3 .. straight html .. I don’t know what I’m doing really.
Oh heck yeah. I have never trusted any SNR calculator (because my images with clouds or haze in them tell me they have better SNR than those without mwahaha) and I definitely don't trust WBPP SNR weights (number of stars, baby!) Because I don't trust SNR measures, I'm not sure what to do when people ask me to do quantitative, objective image comparison using objective measures like SNR. No, thank you very much but no. I'll keep to my subjective comparisons of just looking at and comparing the darn images, and how easy/difficult they are to process... Ok small rant, thanks for coming to my TED talk!
0:32 *chucks sacred scroll of truth into the seas* Athlete should be aware of what sorts of conditions will lead to their blood-oxygen levels being optimal. Idea: You could have it ask for a picture of a blank black piece of cardboard (darkness/no light captured) from your camera, so perhaps it can know what baseline noise the sensor produces, to compare against an image?
Congrats on the 50K. SNR? Not sure how much practical value there. As an analytical chemist, I looked at SNR to determine if a given substance was detected looking at a signal of 3x the standard deviation of the blank (noise). In astrophotography, I like to just zoom into an image and see how much detail is apparent. The ratio is generally sky high in astrophotography.
You accurately described all my ChatGPT-assisted coding experiences. It makes up functions and features of a language or API which don't exist, then you have to convince it that it's making shit up. LLM's are cancer.
4:39 And this is why ChatGPT is a waste of time. It's far quicker and less error prone just to do a task oneself rather than go through and manually verify that the algorithm and code are correct (including all edge cases).
OMG! Hilarious (just so you know, I purchase from Highpoint so you don't have to buy a plane ticket to New Jersey, USA). I haven't seen a comment noting that these calculators are all web-based! So you go out, do your best to take photos all night but only find out if your SNR is good or crap after an all night session!
I see people getting far too obsessed with this and they don't learn to use their eyes which I think is much more important. Kudos to you for speaking the truth.
Noodling out SNR is not unlike calculating your error bars for an experiment. For hard science it's required. Otherwise; experience or your gut tells you how good something is or isn't (this is not a hard rule as bias will lead one down a primrose path of delusion but I digress). In any event; assumptions can be thorny.
Noise is a value that can be quoted for an image, but SNR is a value that has to be tied to a specific object with a specific signal level... is that the source of the confusion?
I'm a simple man so like to think of things in simple terms. To me, signal to noise is intuitively obvious to comprehend. The anwser is in the name. Noise is random, signal is not. Lets represent a pixel with a 1 or 0 value. '1' being signal and a '0' being no signal and noise being a random combination of both with respect to time. Eg, no star over 4 subs could be 0,1,0,1 then star 1,1,1,1 then no star 1,0,0,1. Get the average of each pixel (0+1+0+1)/4 =0.5. (1+1+1+1)/4=1 and lastly (1+0+0+1)/4=0.5. Hopefully over time we get more 0's in the noise than 1's thus reducing the average noise with respect to the constant signal that would always remain very close to 1. This ratio is then normally expressed in terms of dBs.
And... that's exactly where you are wrong. Signal is also random. If you had a perfect detector, you make 100 exposures of the same ultra stable light source, with the same conditions, and if you look at the value of given pixel of your image, it won't be always the same pixel value. You will be able to take the 100 values obtained on that pixel, and calculate an average, and therefore a variation around this average value (root mean square). The number of photons in a given time interval varies around a value, such that the RMS value is proportional to the square root of that value (it's called Poisson's statistics, and has nothing to do with fishes (fish is poisson in french), but it's the name of the mathematician who discovered this type of statistical distribution, hence the uppercase P in Poisson). See en.wikipedia.org/wiki/Poisson_distribution if you like to understand (or not). If the average count is 100 (meaning for example 100 photons detected), the variation around this number will be 10 (square root of 100) and you will be able to know the value of that signal to only 10%. If the value is around 10000 photons, the variation will be 100 photons (square root of 10000), leading to the knowledge of the "real value" of that pixel to 1% (100/10000=1%). In fact, even if you have your detector in the dark, and make short exposures (bias frame), and read the successive values of a given pixel, you get a statistical distribution around an average. If you make long exposures in the dark, it gets worse, because on top of that "average" value, you will have a signal (with it own noise) caused by the thermal emission of the pixels. That's why it's important to make many bias and dark frames and average them (in fact there are better ways to get an average than just averaging, because some pixels may have received cosmic rays, i.e. median, kappa sigma, etc...) because the signal of that parasitic source (bias and thermal signal) has then less noise/less variation around the average, to it. Another way to proceed, because on average :), people are interested in detecting the faintest stars on their image is to measure the signal variation over many pixels. if you take a square window of 100x100 pixels, you get 10000 pixels on which you can calculate the average value and variation around this value, therefore the SNR on a single frame. The SNR of the sky background on the final (preprocessed, stacked, etc...) image is what's interesting. Except on an image like that of M8, there is basically no sky background because it's full of nebula. Then there is endless discussion about the usefullness of making 100s of hours of exposure, and hundreds of calibration exposures (since on top of that the conditions during which the calibration exposures have been taken may have varied with time), etc... In the end, it's indeed a big mess and no wonder different software packages give different values. Again in a very "rough" manner, if your pixel sees one square arc second on the sky, and if the sky brightness is magnitude 21 per square arc second, if you have a magnitude 21 star on that pixel, it will double the value of the pixel, compared to the pixel next to it where there is no star. Then if there is a magnitude 26 star on that pixel (5 magnitudes less, leading to 100 times less signal) you will need to measure that sky brightness to better than 1% in order to see that there may be a magnitude 26 star on that pixel. Because the star value will be a very small value compared to the value of the sky background. Then if your sky brightness is magnitude 16 per square arc second, you should stop shooting images during the day, or change the location of your observatory :).
@@alainmaury5941 That’s an uber epic reply to which I’m eternally grateful for but I think you missed my point. I was trying to explain things in simple terms with 0,1 as a min/max value. The basic concept of signal to noise for the lay person, Its why my comment started with saying I’m just a simple man. Sometimes we just need to stick to KISS, in the way we explain things to give early learners a head start, learning to crawl before we learn to walk. My reply is not wrong. You’re just being more specific and expanding on it to a much higher degree of detail.
fun fact, if you want to be annoyingly correct: the longer your exposures, the more noise is in your images people often get that mixed up with a higher SNR.
Great video and thanks for the shoutout! I agree that obsessing over SNR can be a problem. A couple of random thoughts:
1) 8:42 You nailed it here! Getting better data is so very important! If you have really good single subs, the SNR calculators become much less important.
2) The SNR calculator I made gives raw SNR, not in decibels. dB are nice in that they try to scale things based on human perception (i.e., doubling dB should double the perceived quality), but... it's a log scale and you do need a reference point for dB to be really meaningful. That's why I think the pixinsight one is not that useful (it assumes too much). I think a raw SNR is more intuitive, which is why I left it in raw :)
2) I think doing SNR calculations are useful for experiments! Do you really want to know how well your filter is doing compared to no filter? Well, a good SNR measurement is actually a nice metric, especially if the experiment is based on limited imaging time.
3) I agree that SNR calculators aren't that useful for after you've stacked your images. But I think they're great for planning (if you're the kind of person who likes to plan of course). If I know that getting 30 raw SNR on the fainter parts of a nebula will make my image look good (the midtones and brighter areas will be much better), I can use the SNR calculator to get an estimate based on a single sub just on the fainter part. I can then plan my imaging time accordingly.
4) SNR is really, super important for deconvolution (someday, I'll make a video on it... I think!)
if you want scary good snr calculators there are ones made for the big science telescopes. they need a lot more info on ur setup but then theyre like magic
@@markc7899 do you have any links? :)
"I will kick your tripod"! That's gold! 😂😂😂
Don’t try me ! 😆
@@DylanODonnellmy tripod has two counterweights hanging in it’s sack 🤣
@@AstroIsland so do I
😂😂😂😂😂😂😂
@@DylanODonnell 😆😆😂😂🤣🤣
I saw the thumbnail and thought, really? Who didn't notice that! Then I saw it was Dylan and it all made sense
Def not here in the comments to say something similar
I started doing Astrophotography in 2020 and I'm sure like a lot of people I was a "stamp collector" trying to get a couple targets each night, spending maybe 2-4 hours on each before moving on. This year I switched to picking targets that I can spend multiple nights on, including a recent image of the Crescent Nebula that I spent 33 hours on over 11 nights. Now, I'm starting to earn Top Pick designations on AstroBin. It's a different kind of satisfying and I nice evolution on my journey.
Congrats on the 50K. Clearly you have been stacking lots and lots of subscribers so that should give you a pretty good subscriber to noise ratio.
Haha thx Loges !
@@LogansAstro good one! 😆
Congratulations on 50k subcribers. Have a cold beer to celebrate to this exellent astro channel. Clear skies stellar mate.
Thanks Mr Knight !
How do you only have 50K? (congrats). Channel deserves far more than that!
Thanks !
Happy journey my friend, 50k and growing.
"In PHP" I knew that wasn't gonna last long 😂
Not that I know much about this but being an electronic engineer I did learn a little many years ago about electronic noise. It is basically generated randomly in any electronic circuit. It has to do with heat influencing electrons and from that we get the noise signal. The colder it is the less noise. At absolute zero Kelvin temperature there should be next to no noise if not zero noise.
As noise is random and at least a sound signal is not it is possible to have a very weak input signal, amplify it through two amplifiers and then compare the output. This way it is possible to get a reasonably clear output from input signals well below the noise limit of a single amplifier. For digital pictures I would thing it should be possible to do the same. Take two pictures, compare the two and throw away everything that is not on both pictures. It should be possible to do the comparison electronically.
I now come to a question that I have to every body. It is about the Cosmic Microwave Background Image. When scientists talk about it they often liken it to the noise signal on an analogue TV set. They even say that it is Cosmic Background signal. To my knowledge and understanding this signal in the TV is nothing other than random heat noise especially as there is no signal coming in to the Tv. I was fixing such TV sets many years ago and I know I could reduce this noise by fixing the TV.
Congrats on hitting 50k!
As an Army Officer on tour I have had many different weapons pointed at at my head during negotiations overseas and kept my cool.
Your threat to kick my tripod sent a shiver down my spine. You are a monster, Dylan.
Particularly because I had an awful night last night. My tracking was all over and I was grabbing 5 minute subs (only managed to get 32 in the entire night . . . ouch). I still haven't figured out the problem.
The hobby you love and hate. It does remind me of my years in RECCE where we took pride in our deprivation and pain.
Haha!
"I will kick your tripod!" 😂 Love it! Anyway back to acting my age, personally I just do what you suggest Dylan, gather as much data as possible and hope there's enough there to play with in post. I've recently been getting data on C34 "Soul Neb'/Baby Neb'" and noticed it requires much more data than the circa 2 hours I initially got with my OSC camera. By contrast, it's neighbour IC1805 "Heart Neb'" with exact same mag' ( +6.5 ) really popped with the same amount of data. Data is king!
Congratulations on reaching 50k+ subscribers Dylan! Without wanting to sound sycophantic Impo your channel deserves 1 million+ subscribers! 💥👍
Thanks as always Dylan! Wes.
Thanks for your support Wes! Don’t think I don’t notice it :) if I had a million subs I’d be doing this 5 days a week instead of 1!
@@DylanODonnell Ahh thanks Dylan! And hey maybe one day, you never know! 👌👍
You are crushing it on your channel Dillon! Keep it up!! Thank you!
50k! Congrats dylan:)
Next step: 100k. (If you want that its better to keep creating shorts videos like about useful astro stuff)
Thanks for the tip :)
I watched the Highpoint Scientific ad as well! Now please don't kick my tripod! 🤣
Good good.
Genius, Dylan! Keep up the great work!
Dylan saying “I will kick your tripod” needs to be a meme immediately. 😂
grats for 50k!
Sweet! I are engineer... I care about the maths. I care about all the maths. This was fantastic! Although I wish I never had to know that astro-py existed.
You and Seti are opening the literal Astro-ChatGPT-Pandora's box!!! I love it.
I get it !
what if you used a cross reference from plate solving first so the program can cross refence known stars and remove them from the noise data to help distinguish the two for a better rating? not sure how to do it but just spit balling.
Hi Dylan,
Great video and great tool.
The argument I'd make for keeping this up and maybe a path to improvement is the following use case. When shooting mono, you don't necessarily know the amount of signal in each filter. That means I'm guessing blindly if it's woth doing 6 hours of O3 on a given target, only to discover that either there is no O3 in this target or that I should have shot 3 times more O3 than Ha. By evaluating a few shots of each filter, it could give me the ratio of Ha/O3/S2 to capture a reasonably balanced image.
Hey Bestie! Congrats on the 50k! 👏
Thanks BFF
Great video. I create a small preview of just the space and then use the PI SNR tool. You don't want to include the stars as these have a much higher SNR value. I am for a 42db SNR on a background preview and that is a pretty good rule of thumb. As soon as you run dbe, you lose the gradient and sky brightness so that reduces your overall SNR. I use SNR as a guide for where I want to shoot more. Generally speaking, a doubling of integration time will gain you 3 db of SNR and this is something you can detect. That goes back to a rule I learned early on. If you aren't happy with your image, double your integration time and try again.
Great tips !
@@DylanODonnell Thanks! I'm experiencing the effect of SNR first hand now that I've started shooting with a RASA. The larger aperture gives me a signficantly higher SNR per shot vs. my slower refractor and vastly reduces the amount of integration time I need on a target to get the result I want, especially on fainter targets. Always enjoy your videos and congrats on hitting 50k subscribers. Keep up the good work!
Maybe a database of popular targets with 'perfect' signal could be used to compare your own image with, at least a section of it. This could then be extrapolated to your whole image for a 'true' snr. Something like the Gaia database for SPCC and the upcoming MARS project for 'true' gradients.
But then again, is knowing the Snr that necessary? In practice, I simply try to get enough data to make an image 'pleasant' to process. For my Edge11, this means no les than 20 hours on a target or the snr is too low. Some interesting ideas here! Cheers.
I used to bin RGB and shoot L full resolution. Now that i upgraded to an ASI294mm Pro, ive been considering dropping L and doing all exposure times RGB only. This tool can be useful for optimizing things since the weather has been so hazy this year, I just haven't had a lot of time for doing empirical tests with very long exposure stacks. Thanks for another great video.
Where I’ve used the pixinsight Snr script and found it useful.
Comparing the same image stacked multiple ways in particular 1x drizzle 0.9 with 1x drizzle 0.8 etc. I think there’s a point where drizzling works amazingly well based on the number and quality of subs below which. Images just get noisier.
It also bothers me that there will inevitably be areas of an image (in most cases) where there’s simply very little signal. Perhaps when comparing images, it would be better to define a box size to sample and have two figures. One based on a background area and one for an area of high nebulosity which we are most concerned about.
Great video by the way.
As always, great video! Thank you for all your hard work!
Thx! This one was more work than usual hah
Very well worth it!!!
what if you feed a better external SNR image as a calibration to make it understand what a good SNR for a pixel. Might need some zooming and cropping. Long workflow
So many great nuggets in here, love that thumbnail, haha. Super video! Butterfingers would be proud.. 🙂
Thanks Marcel :)
Having lived through about 55 years of the history of astrophotography so far, many times I've seen theory catch up with practice. That is, the photographers found out how to make things work better in some way, and then the theoreticians figured out why. Keep up the good work!
Isn't that the same with such a lot of things. They make a theory. They calculate and calculate and then tell us all how it should be. We then test it and find that it isn't quite as they said. Then they go over it again, change their theory a little and let us test again. Eventually we reach our capacity to test and they catch up. After that they claim that it was their theory that predicted the result and we confirmed it.
I am an Engineer.
ONLY 50K 😢 Wah! You deserve way more. Keep it up Dylan.
Hey thanks Phil :)
Decibels ARE ratios. Their most common usage is in sound volume (ratio of the peaks to the troughs if a sound wave) but they are applicable to any ratio.
Glad to see there are others using python for things like this. I wanted to know how much to expose as to not saturate a star .... took around 1000 lines of code, but the result was pretty good. With time you will love astropy more and more!
Thx! I’m intrigued!
Very good that you talked about it! I was recently trying to compare different filters on the same target, but on different nights (different moon). And the results were just not fitting the clearly visible differences.
So in the end I ended up using those SNR calculations just for comparing RELATIVE SNR for the same rig (including filters) and the same target.
Don't spent your time, thinking about and calculating the SNR. Just go out and suck in some photons ;)
Exactly !
This is awesome, Dylan I have a question for you, I was trying to image the cresent nebula, I'm in a bortal 6-8 sky. I cant seem to get the middle portion of it. I was doing 2-3min exposures have about 100.Some people are able to get this cool blue vail around it. Any tips? Congrats on the 50k, you've inspired me a lot throughout the years.
I'm not familiar with the target but Oxygen signal is always much much lower. When I combine Ha and OIII I always stretch the OIII much harder
So I'm not sure whether I do or do not want to see the ChatGPT transcript in that 4 day conversation. Ha! Great video...better equipment and more time...right answers! Congrats on the 50k!
Hah.. well I find them funny. It's mostly stuff like Me: "This is wrong, why?" ChatGPT: "It looks right to me" Me: "You forgot to multiply this number! that's why it's wrong!" ChatGPT : "You're right! here's the new code..."
@@DylanODonnell 🤣 same experience...
What if you have Noise exterminator exterminator the hell out of the image. The result could be the theoretical best instead of the "perfect world" bit of the current equations? Would that then take into account all the other personal variables? Use that to find your exposure requirements?
Youre a genius Dylan, your videos are always very pleasant to watch , keep being a rebel ✊
Thx Charles :)
Congratts on the 50 Dylan,, well deserved ,,, !!!
Thanks buddy :)
I love the technical stuff. Is the code available with your new SNR tool, Dylan?
It’s my first attempt at python I’m not sure it has value to anyone but I’ll keep playing with it :)
@@DylanODonnell Python is the language most scientists use. Our PHD loved it. Anyways, I want to learn it too. I like the command line format. It is very similar to BASIC, in that respect. Cheers
Would it be correct, as a Siril user, to take a single frame pp_light.fit, open it in Siril, then save it as a 16-bit unsigned .fit for the purposes of uploading for the SNR calculation?
Yeh that should work !
@@DylanODonnell Oddest thing my uncalibrated 10 second light is at 23dB; calibrated light is at 15dB. I tried calibrating a stack normally (flats, bias, darks) and also darks-only. Same result, reported SNR on the pp_ file is worse than on the light. Calibrations are at same 0 C as lights. 30 darks bult into a Master Dark; 50 bias into a Master Bias; and 30 Flats with histogram about 2/3s out. The calibrations are off a stock Siril script. And tbh, it's not like my stacked images are nasty. I'm just experimenting for the purpose of confirming my 10s procedure (vs going to 30s). Robin Glover expose led me to 10s in my Bortle 4 skies with 14" at f/5 into ASI2600MC Pro. I go 100-200 subs per run. I thought an SNR evaluation would maybe point to missing details that I'm not noticing by eye.
Congratulations on 50k.
Thanks !
Great explanation!!! Thank you!!! I also read "The Deep-Sky Imaging Primer" by Charles Bracken. Which also gives a very deep/great explanation of SNR in the first few chapters.
Thanks for the down to earth analysis… I’ve reached the same conclusions… Many look at it as a noise problem, I switched my thinking awhile back and look at it as a signal problem only, it makes it easier… More signal = better image. At that point the relative noise becomes almost irrelevant…
Yes the real ratio is so small eventually it’s negligible.. and even if it isn’t a few little tweaks in post either with levels or denoising tool and boom .. nice image.
How about running StarXterminator first to remove the stars from your calculations? Harder to remove the nebula or galaxies from the general noise. But maybe mask them out
Great video! I tried to figure out SNR in the beginning and ended up figuring it was all BS. I just try to stack more.
+1 for being the only astrophotography youtuber I wouldn't mind having a beer with.
So I was going to complain about this SNR video because you didn't help me understand it any better than I already did... but you actually helped me untangle my mental block I think. Because this is what I want: I want to shoot a few subs in each color channel and then get a ballpark number as to what ratio of each I need to shoot at to get something usable. I always figured I wanted to compare the SNR of the different channels. But now I am thinking to forget SNR and instead is run each sub through star exterminator, so I'm only looking at noise + nebulosity, and then look at the intensity histogram. This should give me an idea of I need at least 3x more Oiii than Ha etc.... does that seem right?
Yeh that'l work .. but I tend to keep the exposures the same between HA and OIII so the stars don't bloat more than the other in one channel .. even though I'm removing them anyway for RGB .. But if you get enough data in OIII you can stretch it harder when you do your levels / histogram stretching.. I just stretch OIII harder then you can see all that oxygen over the Ha signal, even though you took the same exposure length.
An idea to help improve the graph for SNR is to have the SNR calculator calculate the SNR of different bits of raw data from various people at various locations and bortle scales.
Try to get data (from various people or sources) from Bortle 1-2 zones, bortle 3-4 zones, until you reach 9-10 zones. The more data there is, the better. Utilizing all this data from the various telescopes, sensors, etc., will allow you to in a way, average out what the SNR should roughly look like, allowing the graph to give an additional line which I would call the "Mean SNR Line". This serves no other purpose than to show what the ideal SNR is.
To further separate the data, dividing the sky based on the different constellations could further improve the SNR calculator, as the SNR is going to be different on a nebula compared to a galaxy (assuming all parameters are the same). However, this shouldn't be limited to specific targets, but rather all the targets within the constellation combined into 1 mean line, since people like astrofalls are constantly discovering new nebula seemingly every few months.
No matter what, you are correct. No SNR calculator will ever be 100% accurate, and long exposure combined by taking 100's of images and only selecting the best ones will give the best results.
Great ideas!
The S/N ratio can be liken to watering down beer. Too much water (noise) makes the beer (signal) taste weak. In a perfect world there is only pure beer.
Perfectly phrased.
Congrats ok the 50K mate I guess you have been behaving & not pissing people off lately lol, look forward to seeing more content.
hah biting my tongue :)
Nice shot of the mach diamonds under the Starship Booster. Epic machine. ..... and great video - about something pointless!
Thx Phil! That launch photo is an epic shot !
Another good explainer mate👌
and remember people "Comparison is the thief of Joy". We are not all imaging under the same skies with the same gear.
Damo
Exactly
I wonder how PixInsight calculates the SNR if it doesn't know what the pure signal is. I've done some SNR stuff before for dental x-ray calibration but it was always relative to a long reference exposure. EDIT: yeah should have watched more of the video lol
I woke up this morning and my tripod was knocked over. Never skipping the ad again :(
Lesson learned.
Congrats on 50K. Remember, the bigger the number the more betterer.
MATH!
50k Well done Sir
Isn't the ultimate measure of quality how satisfied "you" are with the result? Congrats on the 50K!
Exactly !
Congrats with 500 K Dylan next week LOL nice Is this SNR not possible for a script in Pixinsight ? THANKS
I don’t know how to do that. I’ll see if I can work it out.
@@DylanODonnell Ask for help from Frank Seti Astro, nice guy
Congrats on 50K, you're right...i don't care about the math, and I must say your hair is looking really good. The best ever.
Haha cheers man. Ppl will say I need a haircut soon.
Here's an idea for next calculator - have a reference db of pictures classified by humans as high/mid/low SNR to be used as additional input for your script.
Congrats on 50k!
I like the idea of an anchor to the expected signal of a given target, thx!
@@DylanODonnellnice, so no kicking my tripod!
The problen with "get more data" is that it's difficult to guess when to stop. As a beginner, I always run in to a dilemma where I want to shoot many targets as there is so much to shoot but also I don't want to spend 50 hours on one target just to get better SNR.
Congrats on the 50k! I'll downvote the video so it doesn't get to your head 👍
Haha cheers Lasser
I loved this video, especially the reference to the reality that SNR calculators have no real idea what they are looking at. It's so easy to get meaningless results and, like all statistical measures, the information is never absolute and must be interpreted. TBH, I never bother to look at SNR. The image is done when the data shows the detail I am after.
Yeh I think most people don’t really *use* SNR for anything and still win awards :)
The other issue that's never addressed is that the SNR doesn't differentiate between the high and low signal portions of the image.
While it was great fun to watch someone else get all wrapped around the axle on something mostly useless, I think you nailed it when you told us to go to High Point and buy better gear. Gonna do that now. And maybe, just maybe, this will be the year I get off my ass and haul my (new) gear out to the desert. Thanks.
tool website down?
Any chance you could post a video talking about OSC imaging but using narrowband filters? I’ve seen some amazing images with people using OSC with narrowband filters but I can’t find any info on how they edit, and no one responds to comments LOL
I would .. but mono is better .. you should go mono :)
@@DylanODonnell I WANT TO! But I wanna ace OSC first! Only got my telescope November last year, and then got my Astro camera like 2-3 months ago. Don’t tempt me just yet 🤤🤤
Nice Video as always. Enjoyed looking it 😅. Congrats to 50k !🎉 ist that a ratio btw?😅
My best friend won't kick my tripod, but may fall on it in a stupor while doing drunken astrophotography.
I skipped the ad because I like living on the edge (and to see you come to Belgium just to kick my pier 😅 )
All of this video 👏 Just shooting until you see diminishing returns under your conditions with your setup. Who really cares beyond that :D Btw, I watched til the very end, congrats on the 50K. Kick my tripod though and it'll kick you back twice as hard.
Haha thanks mate!
welcome to astropy...matplotlib numpy etc anaconda makes it pretty easy to jump in... that was alot of code for what you did... i think. What'd you use for the web stuff
Yeh there’s some archiving stuff for reviewing submissions and a lot of bombproofing to handle mono and OSC .. server is just CPanel with python3 .. straight html .. I don’t know what I’m doing really.
Oh heck yeah. I have never trusted any SNR calculator (because my images with clouds or haze in them tell me they have better SNR than those without mwahaha) and I definitely don't trust WBPP SNR weights (number of stars, baby!)
Because I don't trust SNR measures, I'm not sure what to do when people ask me to do quantitative, objective image comparison using objective measures like SNR. No, thank you very much but no. I'll keep to my subjective comparisons of just looking at and comparing the darn images, and how easy/difficult they are to process...
Ok small rant, thanks for coming to my TED talk!
Agree!
Infinite humor... did he do that??? yup! Congrats on 50k!
Hehe
0:32 *chucks sacred scroll of truth into the seas*
Athlete should be aware of what sorts of conditions will lead to their blood-oxygen levels being optimal.
Idea: You could have it ask for a picture of a blank black piece of cardboard (darkness/no light captured) from your camera, so perhaps it can know what baseline noise the sensor produces, to compare against an image?
Congrats on the 50K. SNR? Not sure how much practical value there. As an analytical chemist, I looked at SNR to determine if a given substance was detected looking at a signal of 3x the standard deviation of the blank (noise). In astrophotography, I like to just zoom into an image and see how much detail is apparent. The ratio is generally sky high in astrophotography.
You accurately described all my ChatGPT-assisted coding experiences. It makes up functions and features of a language or API which don't exist, then you have to convince it that it's making shit up. LLM's are cancer.
Haha yeh like it saved me the initial time suck of setting up working code .. then I had to make it real lol
4:39 And this is why ChatGPT is a waste of time. It's far quicker and less error prone just to do a task oneself rather than go through and manually verify that the algorithm and code are correct (including all edge cases).
It’s bad enough when I do it Dylan, now I’m worried about you popping up in the dead of night kicking my tripod! 😂😂😂
“… and I will kick your tripod…” lol
My SNR is myself. I am just an equipment operator after all. 😃
OMG! Hilarious (just so you know, I purchase from Highpoint so you don't have to buy a plane ticket to New Jersey, USA). I haven't seen a comment noting that these calculators are all web-based! So you go out, do your best to take photos all night but only find out if your SNR is good or crap after an all night session!
Haha true !
Ffs, I'd just polar aligned perfectly and then I skipped the ad for HighPoint Scientific.....please Dylan, don't kick my tripod! 😂
SMACK ! Enjoy your 10 are seconds of guiding error!
Hell yea 50k, now you’re on the path to that internet big money! Bling Bling…
Now this is one of your better videos.
Thanks man .. definitely more work and technical depth in this one :)
Another fun video but please don't kick my tripod. I've got that covered.
No need for you to kick my tripod, I usually do that myself.😂
I see people getting far too obsessed with this and they don't learn to use their eyes which I think is much more important. Kudos to you for speaking the truth.
Thanks !
I see what you did with the thumbnail 😲🤣
what did he do?
He linked SNR to the imaginary number, "i"... Either that or it's a red rocket 🚀 one is easily confused these days 😉
You tube is testing rotating thumbnails so you have a 1/3 chance of seeing the thumbnail in question ;)
@@DylanODonnell Hahaha! That's gold. I'm glad I scored that one
Congrats on fifty thousand subscribers! Do your kids now acknowledge your rizz?
No cap
Thanks Dylan.
Noodling out SNR is not unlike calculating your error bars for an experiment. For hard science it's required. Otherwise; experience or your gut tells you how good something is or isn't (this is not a hard rule as bias will lead one down a primrose path of delusion but I digress). In any event; assumptions can be thorny.
Noise is a value that can be quoted for an image, but SNR is a value that has to be tied to a specific object with a specific signal level... is that the source of the confusion?
Don't care about noise in the software which is in decibel, much easier to use the volum-knob at the end of your SC - telescope.
Cranking to 11 !
jfc the raygun... you're truly a master antagonist.
She’s a hero.
@@DylanODonnell reminds me of the wow signal from seti. Mistaken for a microwave.
I'm a simple man so like to think of things in simple terms. To me, signal to noise is intuitively obvious to comprehend. The anwser is in the name. Noise is random, signal is not. Lets represent a pixel with a 1 or 0 value. '1' being signal and a '0' being no signal and noise being a random combination of both with respect to time. Eg, no star over 4 subs could be 0,1,0,1 then star 1,1,1,1 then no star 1,0,0,1. Get the average of each pixel (0+1+0+1)/4 =0.5. (1+1+1+1)/4=1 and lastly (1+0+0+1)/4=0.5. Hopefully over time we get more 0's in the noise than 1's thus reducing the average noise with respect to the constant signal that would always remain very close to 1. This ratio is then normally expressed in terms of dBs.
Yep! It’s intuitively obvious. :)
And... that's exactly where you are wrong. Signal is also random. If you had a perfect detector, you make 100 exposures of the same ultra stable light source, with the same conditions, and if you look at the value of given pixel of your image, it won't be always the same pixel value. You will be able to take the 100 values obtained on that pixel, and calculate an average, and therefore a variation around this average value (root mean square). The number of photons in a given time interval varies around a value, such that the RMS value is proportional to the square root of that value (it's called Poisson's statistics, and has nothing to do with fishes (fish is poisson in french), but it's the name of the mathematician who discovered this type of statistical distribution, hence the uppercase P in Poisson). See en.wikipedia.org/wiki/Poisson_distribution if you like to understand (or not). If the average count is 100 (meaning for example 100 photons detected), the variation around this number will be 10 (square root of 100) and you will be able to know the value of that signal to only 10%. If the value is around 10000 photons, the variation will be 100 photons (square root of 10000), leading to the knowledge of the "real value" of that pixel to 1% (100/10000=1%). In fact, even if you have your detector in the dark, and make short exposures (bias frame), and read the successive values of a given pixel, you get a statistical distribution around an average. If you make long exposures in the dark, it gets worse, because on top of that "average" value, you will have a signal (with it own noise) caused by the thermal emission of the pixels. That's why it's important to make many bias and dark frames and average them (in fact there are better ways to get an average than just averaging, because some pixels may have received cosmic rays, i.e. median, kappa sigma, etc...) because the signal of that parasitic source (bias and thermal signal) has then less noise/less variation around the average, to it. Another way to proceed, because on average :), people are interested in detecting the faintest stars on their image is to measure the signal variation over many pixels. if you take a square window of 100x100 pixels, you get 10000 pixels on which you can calculate the average value and variation around this value, therefore the SNR on a single frame. The SNR of the sky background on the final (preprocessed, stacked, etc...) image is what's interesting. Except on an image like that of M8, there is basically no sky background because it's full of nebula. Then there is endless discussion about the usefullness of making 100s of hours of exposure, and hundreds of calibration exposures (since on top of that the conditions during which the calibration exposures have been taken may have varied with time), etc... In the end, it's indeed a big mess and no wonder different software packages give different values. Again in a very "rough" manner, if your pixel sees one square arc second on the sky, and if the sky brightness is magnitude 21 per square arc second, if you have a magnitude 21 star on that pixel, it will double the value of the pixel, compared to the pixel next to it where there is no star. Then if there is a magnitude 26 star on that pixel (5 magnitudes less, leading to 100 times less signal) you will need to measure that sky brightness to better than 1% in order to see that there may be a magnitude 26 star on that pixel. Because the star value will be a very small value compared to the value of the sky background. Then if your sky brightness is magnitude 16 per square arc second, you should stop shooting images during the day, or change the location of your observatory :).
@@alainmaury5941 That’s an uber epic reply to which I’m eternally grateful for but I think you missed my point. I was trying to explain things in simple terms with 0,1 as a min/max value. The basic concept of signal to noise for the lay person, Its why my comment started with saying I’m just a simple man. Sometimes we just need to stick to KISS, in the way we explain things to give early learners a head start, learning to crawl before we learn to walk. My reply is not wrong. You’re just being more specific and expanding on it to a much higher degree of detail.
Wonderfull conclusion : ”And remember ..etc..” 🤣
hehe yes!
fun fact, if you want to be annoyingly correct: the longer your exposures, the more noise is in your images
people often get that mixed up with a higher SNR.
wild thumbnail man 😂
Heheh