Big Universe, Big Data: Machine Learning and Image Analysis for Astronomy

TL;DR

This paper discusses how modern astronomy's large data volumes and unique challenges drive advancements in machine learning and image analysis techniques, highlighting their importance for real-time sky surveys and scientific discovery.

Contribution

It presents recent methodological advancements in machine learning tailored for astronomical data analysis and emphasizes astronomy as a key domain for pushing data science boundaries.

Findings

Development of scalable machine learning algorithms for astronomical data

Addressing bias and noise in astronomical datasets

Enhancement of real-time image analysis methods

Abstract

Astrophysics and cosmology are rich with data. The advent of wide-area digital cameras on large aperture telescopes has led to ever more ambitious surveys of the sky. Data volumes of entire surveys a decade ago can now be acquired in a single night and real-time analysis is often desired. Thus, modern astronomy requires big data know-how, in particular it demands highly efficient machine learning and image analysis algorithms. But scalability is not the only challenge: Astronomy applications touch several current machine learning research questions, such as learning from biased data and dealing with label and measurement noise. We argue that this makes astronomy a great domain for computer science research, as it pushes the boundaries of data analysis. In the following, we will present this exciting application area for data scientists. We will focus on exemplary results, discuss main challenges, and highlight some recent methodological advancements in machine learning and image analysis triggered by astronomical applications.

Figures

Equations

$$L(\beta,\widehat{\beta})=\sum_{x\in\mathcal{S}_{\mathrm{train}}}\big{(}\beta(x)-\widehat{\beta}(x)\big{)}^{2}p_{\mathrm{train}}(x)\,\mathrm{d}x\enspace,$$

Full text

Big Universe, Big Data: Machine Learning and Image Analysis for Astronomy

Jan Kremer

Department of Computer Science, University of Copenhagen, Universitetsparken 5, 2100 Copenhagen Ø, Denmark

Kristoffer Stensbo-Smidt

Department of Computer Science, University of Copenhagen, Universitetsparken 5, 2100 Copenhagen Ø, Denmark

Fabian Gieseke

Department of Computer Science, University of Copenhagen, Universitetsparken 5, 2100 Copenhagen Ø, Denmark

Kim Steenstrup Pedersen

Department of Computer Science, University of Copenhagen, Universitetsparken 5, 2100 Copenhagen Ø, Denmark

Christian Igel

Department of Computer Science, University of Copenhagen, Universitetsparken 5, 2100 Copenhagen Ø, Denmark

Astrophysics and cosmology are rich with data. The advent of wide-area digital cameras on large aperture telescopes has led to ever more ambitious surveys of the sky. Data volumes of entire surveys a decade ago can now be acquired in a single night and real-time analysis is often desired. Thus, modern astronomy requires big data know-how, in particular it demands highly efficient machine learning and image analysis algorithms. But scalability is not the only challenge: Astronomy applications touch several current machine learning research questions, such as learning from biased data and dealing with label and measurement noise. We argue that this makes astronomy a great domain for computer science research, as it pushes the boundaries of data analysis. In the following, we will present this exciting application area for data scientists. We will focus on exemplary results, discuss main challenges, and highlight some recent methodological advancements in machine learning and image analysis triggered by astronomical applications.

Ever-Larger Sky Surveys

One of the largest astronomical surveys to date is the Sloan Digital Sky Survey (SDSS, http://www.sdss.org). Each night, the SDSS telescope produces 200 GB of data, and to this day close to a million field images have been acquired, in which more than 200 million galaxies, and even more stars, have been detected.

Upcoming surveys will provide far greater data volumes.

Another promising future survey is the Large Synoptic Survey Telescope (LSST). It will deliver wide-field images of the sky, exposing galaxies that are too faint to be seen today. A main objective of LSST is to discover transients, objects that change brightness over time-scales of seconds to months. These changes are due to a plethora of reasons; some may be regarded as uninteresting while others will be extremely rare events, which cannot be missed. LSST is expected to see millions of transients per night, which need to be detected in real-time to allow for follow-up observations. With staggering 30 TB of images being produced per night, efficient and accurate detection will be a major challenge. Figure 1 shows how data rates have increased and will continue to increase as new surveys are initiated.

What do the data look like? Surveys usually make either spectroscopic or photometric observations, see Figure 2. Spectroscopy measures the photon count at thousands of wavelengths. The resulting spectrum allows for identifying chemical components of the observed object and thus enables determining many interesting properties. Photometry takes images using a CCD, typically acquired through only a handful of broad-band filters, making photometry much less informative than spectroscopy.

While spectroscopy provides measurements of high precision, it has two drawbacks: First, it is not as sensitive as photometry, meaning that distant or otherwise faint objects cannot be measured. Second, only few objects can be captured at the same time, making it more expensive than photometry, which allows for acquiring images of thousands of objects in a single image. Photometry can capture objects that may be ten times fainter than what can be measured with spectroscopy. A faint galaxy is often more distant than a bright one—not just in space, but also in time. Discovering faint objects therefore offers the potential of looking further back into the history of the Universe, over time-scales of billions of years. Thus, photometric observations are invaluable to cosmologists, as they help understanding the early Universe.

Once raw observations have been acquired, a pipeline of algorithms needs to extract information from them. Much image-based astronomy currently relies to some extent on visual inspection. A wide range of measurements are still carried out by humans, but need to be addressed by automatic image analysis in light of growing data volumes. Examples are 3D orientation and chirality of galaxies, and detection of large-scale features, such as jets and streams. Challenges in these tasks include image artifacts, spurious effects, and discerning between merging galaxy pairs and galaxies that happen to overlap along the line of sight. Current survey pipelines often have trouble correctly identifying these types of problems, which then propagate into the databases.

A particular challenge is that cosmology relies on scientific analyses of long-exposure images. As such, the interest in image analysis techniques for preprocessing and de-noising is naturally great. This is particularly important for the detection of faint objects with very low signal-to-noise ratios. Automatic object detection is vital to any survey pipeline, with reliability and completeness being essential metrics. Completeness refers to the amount of detected objects, whereas reliability measures how many of the detections are actual objects. Maximizing these metrics requires advanced image analysis and machine learning techniques. Therefore, data science for astronomy is a quickly evolving field gaining more and more interest. In the following, we will highlight some of its success stories and open problems.

Large-scale Data Analysis in Astronomy

Machine learning methods are able to uncover the relation between input data (e.g., galaxy images) and outputs (e.g., physical properties of galaxies) based on input-output samples, and they have already proved successful in various astrophysical contexts. For example, Mortlock et al. 8 use Bayesian analysis to find the most distant quasar to date. These are extremely bright objects forming at the center of large galaxies and are very rare. Bayesian comparison has helped scientists to select a few most likely objects for re-observation from thousands of candidates.

In astronomy, distances from Earth to galaxies are measured by their redshifts, but accurate estimations need expensive spectroscopy. Getting accurate redshifts from photometry alone is an essential, but unsolved task, for which machine learning methods are widely applied.2 However, they are far from on a par with spectroscopy. Thus, better and faster algorithms are much desired.

Another application is the measurement of galaxy morphologies. Usually, one assigns a galaxy a class based on its appearance (see Figure 3), traditionally using visual inspection. Lately, this has been accelerated by the citizen science project Galaxy Zoo,7 which aims at involving the public in classifying galaxies. Volunteers have contributed more than 100 million classifications, which allow astrophysicists to look for links between the galaxies’ appearance (morphology) and internal and external properties. A number of discoveries have been made through the use of data from Galaxy Zoo, and the classifications have provided numerous hints to the correlations between various processes governing galaxy evolution. A galaxy’s morphology is difficult to quantize in a concise manner, and automated methods are high on the wish list of astrophysicists. There exists some work on reproducing the classifications using machine learning alone,3 but better systems will be necessary when dealing with the data products of next-generation telescopes.

A growing field in astrophysics is the search for planets outside our solar system (exoplanets). NASA’s Kepler spacecraft has been searching for exoplanets since 2009. Kepler is observing light curves of stars, that is, measuring a star’s brightness at regular intervals. The task is then to look for changes in the brightness indicating that a planet may have moved in front of it. If that happens with regular period, duration and decrease in brightness, the source is likely to be an exoplanet. While there is automated software detecting such changes in brightness, the citizen science project Planet Hunters has shown that the software does miss some exoplanets. Also, detecting Earth-sized planets, arguably the most interesting, is notoriously difficult, as the decrease in brightness can be close to the noise level. For next-generation space telescopes, such as Transiting Exoplanet Survey Satellite (TESS), scheduled for launch in 2017, algorithms for detecting exoplanets need to be significantly improved to more reliably detect Earth-sized exoplanet candidates for follow-up observations.

There are also problems that may directly affect our lives here on Earth, such as solar eruptions that, if headed towards Earth, can be dangerous to astronauts, damage satellites, affect airplanes and, if strong enough, cause severe damage to electrical grids. A number of spacecrafts monitor the Sun in real-time. While the ultimate goal is a better understanding of the Sun, the main reason for real-time monitoring is to be able to quickly detect and respond to solar eruptions. The continuous monitoring is done by automated software, but not all events are detected.13 Solar eruptions are known to be associated with sunspots, but the connection is not understood well enough that scientists can predict the onset or magnitude of an eruption. There may be a correlation with the complexity of the sunspots, and understanding this, as well as how the complexity develops over time, is crucial for future warning systems. While scientists are working towards a solution, for example through the citizen science project Sunspotter (https://www.sunspotter.org/), no automated method has yet been able to reliably and quantitatively measure the complexity.

This glimpse of success stories and open problems of big data analysis in astronomy is by no means exhaustive. An overview of machine learning in astronomy can be found in the survey by Ball and Brunner.1

Astronomy Driving Data Science

In the following, we present three examples from our own work showing how astronomical data analysis can trigger methodological advancements in machine learning and image analysis.

Describing the Shape of a Galaxy

Image analysis does not only allow for automatic classification, but can also inspire new ways to look at morphology.9; 11 For instance, we examined how well one of the most fundamental measures of galaxy evolution, the star-formation rate, could be predicted from the shape index. The shape index measures the local structure around a pixel going from dark blobs over valley-, saddle point- and ridge-like structures to white blobs. It can thus be used as a measure of the local morphology on a per-pixel scale, see Figure 4. The study showed that the shape index does indeed capture some fundamental information about galaxies, which is missed by traditional methods. Adding shape index features resulted in a $12\%$ decrease in root-mean-square error (RMSE).

Dealing with Sample Selection Bias

In supervised machine learning, models are constructed based on labeled examples, that is, observations (e.g., images, photometric features) together with their outputs (also referred to as labels, e.g., the corresponding redshift or galaxy type). Most machine learning algorithms are built on the assumption that training and future test data follow the same distribution. This allows for generalization, enabling the model built from labeled examples in the training set to accurately predict target variables in an unlabeled test set. In real-life applications this assumption is often violated—we refer to this as sample selection bias. Certain examples are more likely to be labeled than others due to factors like availability or acquisition cost regardless of their representation in the population. Sample selection bias can be very pronounced in astronomical data,12 and machine learning methods have to address this bias to achieve good generalization. Often only training data sets from old surveys are initially available, while upcoming missions will probe never-before-seen regions in the astrophysical parameter space.

To correct the sample selection bias, we can resort to a technique called importance-weighting. The idea is to give more weight to examples in the training sample which lie in regions of the feature space that are underrepresented in the test sample and, likewise, give less weights to examples whose location in the feature space is overrepresented in the test set. If these weights are estimated correctly, the model we learn from the training data is an unbiased estimate of the model we would learn from a sample that follows the population’s distribution. The challenge lies in estimating these weights reliably and efficiently. Given a sufficiently large sample, a simple strategy can be followed: Using a nearest neighbor-based approach, we can count the number of test examples that fall within a hypersphere whose radius is defined by the distance to the $K$th neighbor of a training example. The weight is then the ratio of the number of these test examples over $K$. This flexibly handles regions which are sparse in the training sample. In the case of redshift estimation, we could alleviate a selection bias by utilizing a large sample of photometric observations to determine the weights for the spectroscopically confirmed training set. 6

To measure how well we approximated the true weight we used the squared difference between true and estimated weight, that is,

[TABLE]

where $\mathcal{S}_{\mathrm{train}}$ is the training sample, $\beta$ and $\widehat{\beta}$ are true and estimated weight, respectively, and $p_{\mathrm{train}}$ is the training density. The nearest neighbor estimator achieved similar or lower error compared to other methods. At the same time the estimator’s running time is three orders of magnitude lower than the best competitor for lower sample sizes. Furthermore, it is able to scale up to millions of examples (code is available at https://github.com/kremerj/nnratio).

Scaling-up Nearest Neighbor Search

Nearest neighbor methods are not only useful for addressing sample selection bias, they also provide excellent prediction results in astrophysics and cosmology. For example, they are used to generate candidates for quasars at high redshift.10 Such methods work particularly well when the number of training examples is high and the input space is low-dimensional. This makes them a good choice for analyzing large sky surveys where objects are described by photometric features (e.g., the five broad-band filters shown in Figure 2). However, searching for nearest neighbors becomes a computational bottleneck in such big data settings.

To compute nearest neighbors for a given query, search structures such as k-d trees are an established way to accelerate the search. If input space dimensionality is moderate (say, below 30), runtime can often be reduced by several orders of magnitude.

While approximate schemes are valuable alternatives, one is usually interested in exact nearest neighbor search for astronomical data. In this context, massively-parallel devices, such as graphics processing units (GPUs), show great promise. Unfortunately, nearest neighbor search based on spatial data structures cannot be parallelized in an obvious way for these devices. To this end, we developed a new tree structure that is more amenable to massively-parallel traversals via GPUs, see Figure 5.4 The framework can achieve a significant runtime reduction at a much lower cost compared to traditional parallel architectures (code available on http://bufferkdtree.readthedocs.io). We expect such scalable approaches to be crucial for upcoming data-intensive analyses in astronomy.

Physical Models vs. Machine Learning Models

A big concern data scientists meet when bringing forward data-driven machine learning models in astrophysics and cosmology is lack of interpretability. There are two different approaches to predictive modeling in astronomy: physical modeling and data-driven modeling. Building physical models, which can incorporate all necessary astrophysical background knowledge, is the traditional approach. These models can be used for prediction, for example, by running Monte Carlo simulations. Ideally, this approach ensures that the predictions are physically plausible. In contrast, extrapolations by purely data-driven machine learning models may violate physical laws. Another decisive feature of physical models is that they allow for understanding and explaining observations. This interpretability of predictions is typically not provided when using a machine learning approach.

Physical models have the drawbacks that they are difficult to construct and that inference may take a long time (e.g., in the case of Monte Carlo simulations). Most importantly, the quality of the predictions depends on the quality of the physical model, which is typically limited by necessary simplifications and incomplete scientific knowledge. In our experience, data-driven models typically outperform physical models in terms of prediction accuracy. For example, a simple $k$ nearest neighbors model can reduce the RMSE by $22\%$ when estimating star formation rates.14; 15 Thus, we strongly advocate data-driven models when accurate predictions are the main objective. And this is indeed often the case, for example, if we want to estimate properties of objects in the sky for quickly identifying observations worth a follow-up investigation or for conducting large-scale statistical analyses.

Generic machine learning methods are not meant to replace physical modeling, because they typically do not provide scientific insights beyond the predicted values. Still, we argue that if prediction accuracy is what matters, one should favor the more accurate model, whether it is interpretable or not. While the black-and-white portrayal of the two approaches may help to illustrate common misunderstandings between data scientists and physicists, it is of course shortsighted. Physical and machine learning modeling are not mutually exclusive: Physical models can inform machine learning algorithms, and machine learning can support physical modeling. A simple example of the latter is using machine learning to estimate error residuals of a physical model.9

Dealing with uncertainties is a major issue in astronomical data analysis. Data scientists are asked to provide error bars for their predictions and have to think about how to deal with input noise. In astronomy, both input and output data have (non-Gaussian) errors attached to them. Often these measurement errors have been quantified (e.g., by incorporating weather conditions during observation), and it is desirable to consider these errors in the prediction. Bayesian modeling and Monte Carlo methods simulating physical models offer solutions, however, often they do not scale for big data. Alternatively, one can modify machine learning methods to process error bars, as attempted for nearest neighbor regression by modifying the distance function.10

Getting Started on Astronomy and Big Data

Most astronomical surveys make their entire data collection, including derived parameters, available online in the form of large databases. These provide entry points for the computer scientist wanting to get engaged in astronomical research. In the following, we highlight three resources for getting started on tackling some of the open problems mentioned earlier.

The Galaxy Zoo website (https://www.galaxyzoo.org) provides data with classifications of about one million galaxies. It is an excellent resource for developing and testing image analysis and computer vision algorithms for automatic classifications of galaxies.

Much of the Kepler data for exoplanet discovery is publicly available through Mikulski Archive for Space Telescopes (http://archive.stsci.edu/kepler). These include light curves for confirmed exoplanets and false positives, making it a valuable dataset for testing detection algorithms.

Having being monitored continuously for years, there is an incredible amount of imaging data for the Sun, from archival data to near real-time images. One place to find such is Debrecen Sunspot Data archive (http://fenyi.solarobs.unideb.hu/ESA/HMIDD.html). These images allow for the development and testing of new complexity measures for image data or solar eruption warning systems.

A Peek Into the Future

Within the next few years, image analysis and machine learning systems that can process terabytes of data in near real-time with high accuracy will be essential.

There are great opportunities for making novel discoveries, even in databases that have been available for decades. The volunteers of Galaxy Zoo have demonstrated this multiple times by discovering structures in SDSS images that have later been confirmed to be new types of objects. These volunteers are not trained scientists, yet they make new scientific discoveries.

Even today, only a fraction of the images of SDSS have been inspected by humans. Without doubt, the data still hold many surprises, and upcoming surveys, such as LSST, are bound to image previously unknown objects. It will not be possible to manually inspect all images produced by these surveys, making advanced image analysis and machine learning algorithms of vital importance.

One may use such systems to answer questions like how many types of galaxies there are, what distinguishes the different classes, whether the current classification scheme is good enough, and whether there are important sub-classes or undiscovered classes. These questions require data science knowledge rather than astrophysical knowledge, yet the discoveries will still help astrophysics tremendously.

In this new data-rich era, astronomy and computer science can benefit greatly from each other. There are new problems to be tackled, novel discoveries to be made, and above all, new knowledge to be gained in both fields.

Notes

This text has been edited and published as

J. Kremer, K. Stensbo-Smidt, F. Gieseke, K. Steenstrup Pedersen, and C. Igel. Big universe, big data: Machine learning and image analysis for astronomy. IEEE Intelligent Systems 32:16–22, 2017.

The IEEE Intelligent Systems magazine restricts the number of references to 15.