AstroSpect: Predicting Redshift of Celestial Objects from Multi-Band Galaxy Images

A lightweight Convolutional Neural Network (CNN) that predicts photometric redshift (photo-z) from multi-band galaxy images, trained on data from the Sloan Digital Sky Survey (SDSS DR12). This approach offers a fast, scalable alternative to traditional spectroscopic redshift estimation.

---

Overview

Understanding galaxy redshift is crucial in cosmology — it tells us how far away and how old a galaxy is. But measuring redshift using spectroscopy requires expensive equipment and takes a lot of time.

This project presents a faster, image-based alternative using deep learning. We built a CNN that takes 32x32x5 galaxy image cubes — basically, small images of galaxies captured in 5 different color filters: u, g, r, i, z — and learns to predict how redshifted (i.e., distant) the galaxy is.

---

Model Architecture

We designed a CNN (Convolutional Neural Network) that works like a smart pattern detector. It looks at the galaxy images and learns to estimate the redshift without needing any manual calculations or domain-specific formulas.

Architecture Breakdown

Here’s what happens step-by-step:

1. Input Layer: A 32x32 image with 5 channels (one for each photometric band).
2. Conv2D Layer 1: Learns 32 patterns using 5x5 filters, followed by ReLU activation.
3. MaxPooling2D: Shrinks image size to retain only important info (2x2 pooling).
4. Conv2D Layer 2: Learns 64 more complex patterns, again with 5x5 filters + ReLU.
5. MaxPooling2D: Again reduces size to focus on most meaningful parts.
6. Flatten: Converts the data into a 1D array for the next layers.
7. Dense Layer 1: Fully connected layer with 220 neurons + ReLU.
8. Dropout: Randomly drops 50% of the connections to prevent overfitting.
9. Dense Layer 2: Another fully connected layer with 64 neurons + ReLU.
10. Dropout: Another dropout layer to regularize.
11. Output Layer: Single neuron with Sigmoid activation to predict a value between 0 and 1 (normalized redshift).

---

Data Pipeline

To predict how far away a galaxy is (its redshift), we first prepare the data so our neural network model can understand it.

Input Images

• Each galaxy is converted into a small image of size 32x32 pixels.
• Instead of a regular color image (Red, Green, Blue), we use five filters:
  • u (ultraviolet)
  • g (green)
  • r (red)
  • i (infrared)
  • z (deep infrared)
• This gives us a 5-channel image or tensor: 32x32x5.

Sample Input Image Cube

Below is an example of a galaxy shown in the five SDSS photometric bands:

Preprocessing

• All image values are scaled (normalized) between 0 and 1.
• Redshift values are also normalized so the model can learn efficiently.

What the CNN Does

1. It takes the 5-band galaxy image as input.
2. Learns patterns related to shape, brightness, and band combinations.
3. Outputs a redshift prediction (a number between 0 and 1).
4. Later, this is scaled back to get the real redshift.

This whole process removes the need for hand-crafted features or traditional redshift fitting.

---

Redshift Prediction Results

We tested our model on two versions of the dataset:

1. Full Dataset: Includes all galaxies, even those with extreme redshift values.
2. Clipped Dataset: Removes galaxies with very high redshift (e.g., z > 0.4) to reduce outliers.

Key Findings

Metric	Full Dataset	Clipped Dataset
Mean Absolute Error (MAE)	0.0304	0.0556
Mean Squared Error (MSE)	0.0017	0.0056
R-squared Score	0.9041	0.8826
Precision	0.0269	0.0427
Catastrophic Outliers (	Δz	>0.15)

Conclusion: Including all redshift values — even the extreme ones — helped the model generalize better. It learned a wider range of galaxy types and redshifts.

Visualization

---

Acknowledgements

• Sloan Digital Sky Survey (SDSS)
• Original research and code structure inspired by peer collaboration.