Preliminary Report on Mantis Shrimp: a Multi-Survey Computer Vision Photometric Redshift Model

Andrew Engel,Gautham Narayan,Nell Byler

2024-02-06

Abstract:The availability of large, public, multi-modal astronomical datasets presents an opportunity to execute novel research that straddles the line between science of AI and science of astronomy. Photometric redshift estimation is a well-established subfield of astronomy. Prior works show that computer vision models typically outperform catalog-based models, but these models face additional complexities when incorporating images from more than one instrument or sensor. In this report, we detail our progress creating Mantis Shrimp, a multi-survey computer vision model for photometric redshift estimation that fuses ultra-violet (GALEX), optical (PanSTARRS), and infrared (UnWISE) imagery. We use deep learning interpretability diagnostics to measure how the model leverages information from the different inputs. We reason about the behavior of the CNNs from the interpretability metrics, specifically framing the result in terms of physically-grounded knowledge of galaxy properties.

Instrumentation and Methods for Astrophysics,Artificial Intelligence

What problem does this paper attempt to address?

The main goal of this paper is to develop a multi-source computer vision photometric redshift model to improve the accuracy of redshift estimation for distant galaxies. Specifically, the research team created a model named "Mantis Shrimp," which integrates image data from different bands (including ultraviolet, optical, and infrared) to enhance the accuracy of photometric redshift (photo-z) estimation.

Key points mentioned in the paper include:
- **Background and Challenges**: Traditional spectroscopic redshift measurement methods, while accurate, are time-consuming and costly; photometric redshift methods, on the other hand, are fast and can be applied to more celestial objects but are less accurate. Current methods face additional complexity when combining data from multiple instruments.
- **Solution**: The research team utilized deep learning techniques, particularly convolutional neural networks (CNN), to develop a model capable of handling multi-modal data, aiming to improve photometric redshift estimation accuracy by analyzing images from different bands.
- **Model Architecture**: The Mantis Shrimp model is based on the ResNet50 architecture and has been modified to handle 9-band input images. Additionally, the model considers dust extinction factors along the line of sight.
- **Dataset**: The study used a large-scale dataset containing approximately 4.2 million galaxies, with data from multiple astronomical observation projects.
- **Results and Discussion**: Preliminary results show that the Mantis Shrimp model performs excellently in photometric redshift estimation, especially when integrating information from different bands. Interpretive analysis of the model reveals that its dependence on images from different bands varies with redshift, consistent with our understanding of galaxy spectral energy distribution.
- **Future Work**: The authors note that although the current results are promising, there is still much room for improvement. They plan to further optimize the model, including using the remaining 93% of the data for training and adjusting hyperparameters.

In summary, this paper introduces a new multi-source computer vision model aimed at improving photometric redshift estimation accuracy by integrating data from different bands, and demonstrates the effectiveness and potential of this approach through experiments.