A package for the automated classification of images containing supernova light echoes

Abstract

Context. The so-called light echoes of supernovae -the apparent motion of outburst-illuminated interstellar dust -can be detected in astronomical difference images; however, light echoes are extremely rare which makes manual detection an arduous task. Surveys for centuries-old supernova light echoes can involve hundreds of pointings of wide-field imagers wherein the subimages from each CCD amplifier require examination. Aims. We introduce ALED, a Python package that implements (i) a capsule network trained to automatically identify images with a high probability of containing at least one supernova light echo and (ii) routing path visualization to localize light echoes and/or light echo-like features in the identified images. Methods. We compared the performance of the capsule network implemented in ALED (ALED-m) to several capsule and convolutional neural networks of different architectures. We also applied ALED to a large catalogue of astronomical difference images and manually inspected candidate light echo images for human verification. Results. ALED-m was found to achieve 90% classification accuracy on the test set and to precisely localize the identified light echoes via routing path visualization. From a set of 13,000+ astronomical images, ALED identified a set of light echoes that had been overlooked in manual classification. ALED is available via github.com/LightEchoDetection/ALED.

Toward automated detection of light echoes in synoptic surveys: considerations on the application of the Deep Convolutional Neural Networks

“…We note that the final results reported for our model are on the median across the 6 cross-validation sets, while it appears that ALED did not perform cross validation. On a single set, our precision and recall can be as high as 0.9 simultaneously (see Figure 9), but perhaps the most fair comparison is that with the model reported in table 4 of Bhullar et al (2021) as "ALED-m": a precision P = 1, and a recall R = 0.25. When taking the median across all cross-validation sets, at P median = 1 we measure R median = 0.3 for score and IoP thresholds of 0.7 and 0.2 respectively.…”

“…Both our model and ALED (Bhullar et al 2021), the only other automated LE detection method we are aware of, can achieves a high (P ∼ 90%) precision and high recall (R ∼ 90%) although, since the models were applied to different datasets, it is not straightforward to compare their performance at this time. We note that the final results reported for our model are on the median across the 6 cross-validation sets, while it appears that ALED did not perform cross validation.…”

“…When taking the median across all cross-validation sets, at P median = 1 we measure R median = 0.3 for score and IoP thresholds of 0.7 and 0.2 respectively. While both datasets used for training the models are small, our dataset contains only 28 576×576 images with LEs (before augmentation) compared to 175 200×200 images in Bhullar et al (2021). There are significant differences in the methodological approach.…”

“…Deep Neural Networks (DNNs) have recently been used in the detection and study of astronomical sources, including the detection of galaxy clusters (Chan & Stott 2019), gravitational lenses (Davies et al 2019), supernova remnants (Liu et al 2019), and more. DNNs have also been applied to the detection of LEs in Bhullar et al (2021). This work produced a slide window detection model, ALED, which is based on a capsule CNN architecture (Sabour et al 2017).…”

Toward the Automated Detection of Light Echoes in Synoptic Surveys: Considerations on the Application of Deep Convolutional Neural Networks

“…Deep neural networks (DNNs) have recently been used in the detection and study of astronomical sources, including the detection of galaxy clusters (Chan & Stott 2019), gravitational lenses (Davies et al 2019), supernova remnants (Liu et al 2019), and more. DNNs have also been applied to the detection of LEs in Bhullar et al (2021). This work produced a sliding-window detection model, ALED, which is based on a capsule CNN architecture (Sabour et al 2017).…”