09:35:59 >>ROBIN: Hello everybody.
09:35:59 My name is Robin.
09:36:05 I am a master student here the department of geography environments and a member of the Institute for space and earth exploration.
09:36:10 And I will be talking a bit about what is on my Masters research today galaxy mapping by machine learning classification.
09:36:20 So just to start off, have you ever looked up at the night sky, seeing the stars, maybe seen something like this?
09:36:28 And what are how our galaxy came to be and how it evolved, formed and ultimately how all matter formed.
09:36:39 More unfortunately we can't actually leave our galaxy to observe its full structure.
09:36:39 We can only see through the galaxy like this because we are inside the galaxy.
09:36:50 So therefore we need to actually look at other types of galaxies, look at galaxies in different stages of evolution, formation, and different components, and make inferences about our own galaxy.
09:36:57 So there's these three main types of galaxies.
09:37:03 Spiral in the top left here, irregular and elliptical.
09:37:07 So you can see they have different patterns of components.
09:37:09 Components, the main components being stars dust and gas.
09:37:18 If you noticed your spiral galaxy has kind of the most structure.
09:37:33 It has this variation of galaxy components, and you might notice it goes from about this reddish color to this bluish color, so it kind of exhibits the spatial and temporal components whereas these 2, you know, they don't really exhibit that is same patterns.
09:37:53 So just to talk about my research objectives briefly, we essentially want to evaluate how machine learning can be used to classify and map all galaxy components.
09:37:57 You want to really look in detail to galaxy and map all components, not to just a specific or not just a few ones but every single pixel.
09:38:06 We wanted to determine what input parameters are most useful for this as well as what machine learning algorithms are most useful.
09:38:16 Suggested to briefly talk about the data I used.
09:38:29 So we used Hubble space telescope data which as you might know, is a telescope that orbits Earth and takes the speed from high resolution images of galaxies and other special phenomenon, and you can send this galaxy here.
09:38:32 This is our study area.
09:38:32 You can see the really high resolution detail.
09:38:40 And you might think this galaxy is close, but it's actually very far away.
09:38:47 It would take you know it would take millions of years to get there or even trillions of years.
09:38:51 It's insane.
09:38:59 But the galaxy, you know, the telescope can take these beautiful images of it.
09:38:59 And so we really want to use the high-resolution data.
09:39:05 And we have three, blue visual and infrared.
09:39:07 I won't go into depth about these but they are essentially useful for galaxy component observation.
09:39:17 In this galaxy is extremely massive and very unusual galaxy.
09:39:23 So we kind of want to study it for this reason because it is the strange galaxy.
09:39:26 But also because it has a really wide variation of galaxy components.
09:39:30 So like I said, those patterns.
09:39:40 It exhibits the spatial and temporal patterns that are really useful for example GIS processing or remote methods, because on earth we have the spatial and temporal patterns.
09:39:42 So this is really compatible with those methods.
09:39:49 So just to briefly describe the kind of things that we produce from these images.
09:39:55 So here I just to show those three bands in color.
09:39:55 As an example.
09:40:01 And essentially we create these other theories text parameters from the images.
09:40:11 Euclidean distance is a commonly used GIS method at least in my projects I have used this a lot.
09:40:23 And this essentially just takes a future so for example in our case of the spiral arm, you can see here the spiral arms are kind of outlet year.
09:40:30 And it essentially calculates a for each pixel the distance from the spiral arms.
09:40:34 So it's a pretty simple thing to calculate.
09:40:36 You can just do it in arc map or something like that.
09:40:43 So this might be useful because of the spatial and temporal patterns that I have mentioned.
09:40:57 Hair like texture, hair-like is just the person who created this type texture.
09:40:57 The texture essentially tries to describe the tunnel variation in patterns.
09:40:57 It tries to imitate what we see as humans.
09:40:59 We recognize these patterns and it tries to imitate that.
09:41:12 What might be useful for using texture is that these galaxy components show different appearances.
09:41:13 They have different roughness or smoothness.
09:41:13 This might be particularly useful.
09:41:17 And finally, band ratios, we calculate these by performing band divisions.
09:41:30 So for example blue divided by infrared, blue divided by an visible light.
09:41:30 And this might show as different ratios of light coming from the galaxy.
09:41:30 So honestly different components within the galaxy would have different ratios of light.
09:41:39 So just a little simplified model of our flowchart of machine learning.
09:41:51 We preferred, machine learning is essentially when you take label data and you inputted into a machine learning model and it outputs, it automates whatever task you want to do.
09:42:07 So here on the left I just show examples of how we define the classes that we train with.
09:42:07 So there is just these six classes.
09:42:07 I want to go into detail about them but I kind of showed how we did that.
09:42:22 We draw polygons over parts of the spaces that represent different classes.
09:42:23 We input this training data as well as the different combinations, so use different combinations with T images, distance, ratios and texture parameters.
09:42:48 We input those into the machine running models, which you would use maximum Kosovar, random forest export rector machine, which I won't go into detail about but essentially maximum likelihood is more traditional model whereas random forest and support and vector or somewhat newer you could say or more powerful models is probably better.
09:42:54 Essentially we are comparing the more traditional model of the more powerful machine running models.
09:43:00 In this outputs a classified roster in our case because we want, we want this very detailed picture of all galaxy components.
09:43:07 So we really want to do that pixel classification.
09:43:09 So essentially every pixel image is just classified into one of these six classes.
09:43:21 So our results, and I included a map., For GIS, so machine learning is capable of classifying all galaxy components.
09:43:21 As you can see here.
09:43:27 So this is just one example.
09:43:27 But it does a pretty good job of classifying all those components.
09:43:44 We find that the more powerful she learning models, the random force and support vector are more useful than the traditional maximum likelihood model or classification and mapping of galaxy components.
09:43:52 And we find that actually a combination of different parameters, so here a combination of the original HST images, texture, and distance parameters results in highest accuracy.
09:43:56 With texture and distance being most important.
09:44:20 So this research is relevant in the age of James Webb space telescope, Romans space telescope in Euclid telescope, which are sent to be launched telescope's that will orbit the earth or orbit a little further from the earth and to take similar images.
09:44:26 So the stadium will become even more abundant, and you will see a process rapidly.
09:44:43 Augmenting the process of this data is very important.
09:44:43 So more than ever this research is relevant.
09:44:43 Also this research can be applied to other galaxies.
09:44:43 That doesn't just have to be applied to this galaxy.
09:44:55 Particularly spiral galaxies are probably the most compatible because they have those of spatial and floral patterns.
09:44:56 We would recommend further research on uses of textual analysis for galaxy classification.
09:45:03 Texture doesn't seem to be used as much and astronomy as it is and remote sensing.
09:45:09 So because we find it useful here, we really recommend more research on that.
09:45:20 And I just wanted to quickly thank my supervisors Doctor (name?) and Doctor Poling Barbie for their expertise and patient as well as thank Doctor Phyllis Duke for providing feedback on my research.
09:45:23 And thank you for listening to my presentation.