Scholarlink

<p>Automated change point detection in Interferometric Synthetic Aperture Radar displacement time series is pivotal for characterizing crustal deformation, particularly in volcanic environments exhibiting complex spatiotemporal instabilities. Although InSAR provides extensive spatial coverage with millimetric precision, substantial phase noise, pronounced seasonal oscillations, and intricate signal architectures frequently impede robust change point identification. To mitigate these complexities, this study develops a hybrid Convolutional Neural Network–Long Short-Term Memory framework to autonomously detect change points by integrating hierarchical spatial feature extraction with long-term temporal dependency modeling. The architecture was trained on 5000 synthetic time series incorporating nonlinear trends, stochastic noise, and seasonal constituents, and subsequently deployed on Sentinel-1 datasets from the Campi Flegrei caldera (Italy) spanning the period 2015–2025. Performance assessment using synthetic data yielded a high F1-score of ~ 92.91%. Real-world application delineated change points concentrated proximal to the caldera centroid, coinciding with documented bradyseismic zones. Detection frequency markedly escalated post-2021, consistent with accelerated ground deformation and seismic swarms. Cross-validation against Global Navigation Satellite System geodetic data demonstrated high temporal correlation (<i>R</i> &gt; 0.85) and sub-centimeter spatial accuracy (Root Mean Square Error &lt; 5&#xa0;mm). The framework’s robustness and transferability were verified on an independent volcanic dataset (Mount Etna, F1-score &gt; 82%), while its systemic generalization was validated through large-scale testing on heterogeneous, non-volcanic time series from the European Ground Motion Service. The proposed methodology establishes a robust, automated, and transferable instrument for monitoring geodynamic instabilities, exhibiting significant potential for integration into operational early warning systems.</p> Graphical abstract <p></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

Hybrid deep learning for automated detection of spatiotemporal change points in InSAR time series: a case study on the Campi Flegrei volcano using a CNN–LSTM model

  • 期刊论文

  • 发表日期：

    2026年3月30日

查看全文
  • Seyed Arya Fakhri,
  • Mehran Satari

摘要

Automated change point detection in Interferometric Synthetic Aperture Radar displacement time series is pivotal for characterizing crustal deformation, particularly in volcanic environments exhibiting complex spatiotemporal instabilities. Although InSAR provides extensive spatial coverage with millimetric precision, substantial phase noise, pronounced seasonal oscillations, and intricate signal architectures frequently impede robust change point identification. To mitigate these complexities, this study develops a hybrid Convolutional Neural Network–Long Short-Term Memory framework to autonomously detect change points by integrating hierarchical spatial feature extraction with long-term temporal dependency modeling. The architecture was trained on 5000 synthetic time series incorporating nonlinear trends, stochastic noise, and seasonal constituents, and subsequently deployed on Sentinel-1 datasets from the Campi Flegrei caldera (Italy) spanning the period 2015–2025. Performance assessment using synthetic data yielded a high F1-score of ~ 92.91%. Real-world application delineated change points concentrated proximal to the caldera centroid, coinciding with documented bradyseismic zones. Detection frequency markedly escalated post-2021, consistent with accelerated ground deformation and seismic swarms. Cross-validation against Global Navigation Satellite System geodetic data demonstrated high temporal correlation (R > 0.85) and sub-centimeter spatial accuracy (Root Mean Square Error < 5 mm). The framework’s robustness and transferability were verified on an independent volcanic dataset (Mount Etna, F1-score > 82%), while its systemic generalization was validated through large-scale testing on heterogeneous, non-volcanic time series from the European Ground Motion Service. The proposed methodology establishes a robust, automated, and transferable instrument for monitoring geodynamic instabilities, exhibiting significant potential for integration into operational early warning systems.

Graphical abstract