Scholarlink

<p>The exponential growth of photometric time-series data from space telescopes requires automated exoplanet vetting pipelines that are highly accurate, robust to noise, and physically interpretable. While traditional 1D Convolutional Neural Networks (CNNs) have advanced transit detection, they inherently struggle with long-range temporal dependencies and often operate as opaque black boxes, raising concerns about their reliance on instrumental artifacts rather than genuine astrophysical geometry. To address these limitations, we propose Astro-DBA, an Astronomical Dual-View DenseNet-BiLSTM-Attention architecture. By decoupling macroscopic orbital dynamics from high-resolution transit morphology, the dual-view inputs enable 1D Dense Convolutional blocks to preserve critical low-level geometric features while mitigating vanishing gradients. Subsequently, a Bidirectional LSTM coupled with Scaled Dot-Product Attention models the temporal characteristics of transit events.Evaluated on the Kepler DR24 catalogue, Astro-DBA achieves an accuracy of 96.38% and an <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(F_1\)</EquationSource> </InlineEquation>-score of 96.40%, outperforming the evaluated generic architectures and domain-specific baselines on the DR24 benchmark. Furthermore, Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated into the feature extraction layers to improve model interpretability. The resulting attribution maps demonstrate that Astro-DBA successfully localizes transit regions in noisy observations and focuses on physically meaningful ingress, trough, and egress structures, supporting transparent and reliable exoplanet candidate classification.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Astro-DBA: An interpretable dual-view DenseNet-BiLSTM architecture for exoplanet transit detection

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Ajay Waghumbare,
  • Palgun T,
  • Yogesh Bhore,
  • Mandira Mondal

摘要

The exponential growth of photometric time-series data from space telescopes requires automated exoplanet vetting pipelines that are highly accurate, robust to noise, and physically interpretable. While traditional 1D Convolutional Neural Networks (CNNs) have advanced transit detection, they inherently struggle with long-range temporal dependencies and often operate as opaque black boxes, raising concerns about their reliance on instrumental artifacts rather than genuine astrophysical geometry. To address these limitations, we propose Astro-DBA, an Astronomical Dual-View DenseNet-BiLSTM-Attention architecture. By decoupling macroscopic orbital dynamics from high-resolution transit morphology, the dual-view inputs enable 1D Dense Convolutional blocks to preserve critical low-level geometric features while mitigating vanishing gradients. Subsequently, a Bidirectional LSTM coupled with Scaled Dot-Product Attention models the temporal characteristics of transit events.Evaluated on the Kepler DR24 catalogue, Astro-DBA achieves an accuracy of 96.38% and an \(F_1\) -score of 96.40%, outperforming the evaluated generic architectures and domain-specific baselines on the DR24 benchmark. Furthermore, Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated into the feature extraction layers to improve model interpretability. The resulting attribution maps demonstrate that Astro-DBA successfully localizes transit regions in noisy observations and focuses on physically meaningful ingress, trough, and egress structures, supporting transparent and reliable exoplanet candidate classification.