# Astronomical image time series classification using CONVolutional attENTION (ConvEntion)

**Authors:** Anass Bairouk, Marc Chaumont, Dominique Fouchez, Jerome Paquet, Frédéric Comby, Julian Bautista  
**arXiv:** [2304.01236](https://arxiv.org/abs/2304.01236)  
**Date:** 2023-04-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.01236v1/

## Key points

- Treatment of astronomical image time series is gaining attention due to numerous surveys following up on transient objects
- Need for robust automatic tools to detect and classify celestial objects is growing steadily
- Anass Bairouk et al. propose a novel approach called ConvEntion based on deep learning for classifying different types of space objects directly using images
- ConvEntion combines spatial and temporal feature extraction in one step, improving the network's capacity to recognize objects
- The solution integrates spatio-temporal features and can be applied to various types of image datasets with any number of bands
- The authors propose a solution for the missing observations problem which demonstrates a significant improvement in the accuracy of the model
- Tested with actual data from the Sloan Digital Sky Survey (SDSS) Supernova Survey, showing an increase in accuracy compared to state-of-the-art approaches that use image time series and light curves.
- Provides an improved supervised transformer-based architecture for astronomical image time series classification
- ConvEntion provides robust automatic tool to detect and classify celestial objects with high accuracy

## Summary

In recent years, the treatment of astronomical image time series has gained increasing attention due to numerous surveys following up on transient objects that are in progress or under construction. With such a large amount of data available, the need for robust automatic tools to detect and classify celestial objects is growing steadily. In this paper, Anass Bairouk et al. propose a novel approach based on deep learning for classifying different types of space objects directly using images. The authors named their approach ConvEntion, which stands for CONVolutional attENTION. It is based on convolutions and transformers, which are new approaches for the treatment of astronomical image time series. Unlike other works that separate spatial and temporal feature extraction, ConvEntion combines these two steps by performing spatio-temporal feature extraction in one step. This improves the capacity of the network to recognize objects. The solution integrates spatio-temporal features and can be applied to various types of image datasets with any number of bands. The authors also propose a solution for the missing observations problem which demonstrates a significant improvement in the accuracy of the model. To illustrate the performances of their model they tested it with actual data from the Sloan Digital Sky Survey (SDSS) Supernova Survey (Sako et al., 2014), one of three components (along with Legacy and SEGUE surveys) of SDSS-II, a three-year extension of the original SDSS that operated from July 2005 to July 2008. The Supernova Survey is a time-domain survey involving repeat imaging of the same region of sky every other night weather permitting. In this work various problems that datasets tend to suffer from were solved and new results were presented for classifications using astronomical image time series with an increase in accuracy compared to state-of-the-art approaches that use image time series and light curves. Overall this study provides an improved supervised transformer-based architecture for astronomical image time series classification. The associated scientific topics are extensive ranging from the study of objects in our galaxy to observation of most distant supernovae for measuring expansion of universe. With such a large amount of data available ConvEntion provides robust automatic tool to detect and classify celestial objects with high accuracy.