Using Deep Neural Networks to Measure Photometric Redshifts of Galaxies 🌌

Understanding the distance of galaxies is a fundamental aspect of cosmology. It helps scientists determine key physical parameters such as mass, luminosity, star formation rate, and metallicity. One of the most accurate ways to measure these distances is through redshift—how much a galaxy’s light shifts due to the expansion of the universe.

The Challenge of Measuring Redshift 🔭

Redshift can be determined using spectroscopic or photometric methods:

• Spectroscopic redshift is highly accurate but requires expensive equipment and can only measure a limited number of bright galaxies.
• Photometric redshift is less precise but allows scientists to analyze a much larger dataset using images from different filters.

With the Sloan Digital Sky Survey (SDSS) DR10 dataset, which contains 250 times more photometric than spectroscopic data, leveraging machine learning to improve photometric redshift measurement becomes an exciting opportunity.

Preparing the Dataset

To train a Deep Neural Network (DNN), a well-structured dataset was required. The steps to prepare the dataset included:

1. Querying the SDSS DR10 Database

   • Galaxy images and redshift values were retrieved from the SDSS DR10 dataset using SQL-based queries.
   • Selection criteria:
     • Diameter less than 30 arcseconds
     • Valid magnitude values in the g, r, i, and z filters
     • Spectroscopic redshift below 2, with an error margin under 0.1

2. Random Sampling

   • To optimize computational efficiency, 59,861 galaxies were randomly selected from the dataset.
   • This selection ensured a balanced distribution of redshifts across different ranges.

3. Downloading Multi-Band Images

   • Each selected galaxy’s images in the g, r, i, and z filters were downloaded using Python’s `astroquery` library.

4. Preprocessing the Images

   • Image Registration: Aligning the images from different filters to correct positional mismatches.
   • Flux Conversion: Converting pixel values to asinh magnitudes, which handle low signal-to-noise data better.
   • K-Correction (Optional): Adjusting brightness for redshift effects, though this was later found to reduce accuracy.

5. Dataset Formatting

   • Each galaxy image was converted into a standardized format suitable for training the DNN.
   • Labels were assigned based on their spectroscopic redshift values.

The Role of Deep Learning 🧠

To enhance photometric redshift measurement, this research explores the use of Deep Neural Networks (DNNs)—specifically DenseNet, a convolutional neural network architecture. Unlike traditional regression methods, DNNs can detect complex patterns in multi-band galaxy images, leading to better estimations.

Key Findings 📊

1. Larger Training Data Improves Accuracy

   • The study used 59,861 galaxy images from SDSS DR10 for training and validation.
   • Models trained on 100% of the dataset performed significantly better than those trained on only 50%.

2. Early Stopping Model vs. Cross-Validation Model

   • The Early Stopping Model (which saves the best performing model during training) produced more accurate results.
   • The Cross-Validation Model (which uses k-fold validation) also performed well but was slightly less effective.

3. The K-Correction Factor Dilemma

   • Applying K-correction (which adjusts for galaxy brightness variations due to redshift) unexpectedly reduced model accuracy.
   • Future research is needed to train models with K-corrected data from the start.

4. Deep Neural Networks Outperform Traditional Methods

   • Compared to linear and non-linear regression techniques, DNNs achieved higher precision in photometric redshift estimation.
   • The DenseNet architecture was particularly effective in extracting relevant features from galaxy images.

5. Computational Challenges 🖥️

   • Training DNNs requires significant computing power, making optimization techniques like dropout layers and batch normalization essential.
   • The model was trained using Python and machine learning libraries on a multi-core Xeon processor with NVIDIA Quadro GPU.

6. Potential for Future AI-powered Astronomy

   • AI models like this could help process massive datasets in future space surveys, such as those from the Vera C. Rubin Observatory or the James Webb Space Telescope (JWST).
   • Further improvements—such as adding more filter bands or using generative models—could push accuracy even further.

Future Prospects 🔮

The study highlights that Deep Neural Networks have great potential in astrophysics, especially in handling massive datasets like SDSS. However, further optimizations—such as data augmentation and additional filter bands—could improve performance even more.

This approach paves the way for AI-powered astronomical research, helping scientists analyze the vast universe more efficiently. 🚀✨