Temporal Stamp Classifier: Classifying Short Sequences of Astronomical Alerts
Daniel Neira O., Pablo A. Estévez, Francisco Förster·May 23, 2024
Summary
The paper introduces the Temporal Stamp Classifier, a deep learning model for classifying astronomical objects (AGN, SNe, VS) in ZTF alerts. The model combines CNNs and RNNs to process irregular time-series data, achieving an accuracy of around 98% with 2-5 detections. Enhancements include random rotations and varying image sizes, which improve test accuracy by 1.46%. The study compares different recurrent models and finds Conv1D to be a practical choice due to its speed and comparable accuracy. The model's performance is competitive with light curve classifiers and highlights the importance of metadata and image processing for multi-detection events. Future work involves refining categories, exploring time modulation, and deep attention models for enhanced outlier detection.
Introduction
Background
[ZTF ( Zwicky Transient Facility) and its importance]
[Challenges in classifying irregular time-series data]
Objective
[Development of Temporal Stamp Classifier]
[ Aim: High accuracy and multi-detection capability]
[Enhancements: Random rotations and varying image sizes]
Method
Data Collection
[ZTF alert data source]
[Inclusion of metadata and image data]
Data Preprocessing
[Image resizing and augmentation techniques]
[Temporal feature extraction from time-series data]
Model Architecture
Convolutional Neural Networks (CNNs)
[Role in extracting spatial features]
Recurrent Neural Networks (RNNs) - Conv1D
[Choice and comparison with other RNN models]
[Advantages: speed and accuracy]
Model Implementation
[Integration of CNNs and RNNs]
[Enhancements: random rotations and varying image sizes]
Performance Evaluation
[Accuracy of around 98% with 2-5 detections]
[Accuracy improvement with enhancements]
Comparison with Light Curve Classifiers
[Competitiveness and insights]
Future Research Directions
[Refining categories for better classification]
[Exploring time modulation in classification]
[Deep attention models for outlier detection]
Basic info
papers
instrumentation and methods for astrophysics
artificial intelligence
Advanced features