Conference Agenda

In the past year, the project team has conducted a range of studies on AI approaches to analyseelectromagnetic satellite data (Swarm and CSES), several machine learning algorithms have been developed, including predictive models (Long Short-Term Memory, LSTM) and generative models (Generative Adversarial Networks, GANs) for generating synthetic time series data based on Swarm satellite data, as well as STL (Seasonal-Trend decomposition using LOESS) algorithm, EMD (Empirical Mode Decomposition) algorithm, etc. for extracting anomalous disturbance signals and obtaining the spatiotemporal distribution characteristics of anomalies under seismic and non-seismic conditions. These provide new insights into the analysis of electromagnetic anomalies from Swarm and CSES satellites, and Sentinel imaginary.

These studies expose the limitations of current machine learning methods and help us propose a novel research direction, such as, integrating LSTM-based and GAN-based models to enrich seismic datasets through synthetic data generation. This hybrid approach not only helps in accurately reproducing authentic temporal trends in seismic activities but also offers a promising avenue for further research in seismic data analysis.

Based on ground-based extremely low-frequency (ELF) observations, we also developed a differential ant colony inversion algorithm to invert anomalous data from multiple frequency points and multiple response functions, which can be used to obtain relevant parameters of equivalent radiation sources in multiple underground media. The Preliminary results show that the anomalous azimuth has a good indicative effect on future epicenters.

In addition, using remote sensing data, the tectonic deformation characteristics of earthquakes were inverted, the results reveal the rupture characteristics and disaster effects of thrust faults under deep-seated thick-skinned geological structures.

This report will present the progressing results from both European and China partners sides, highlighting the potential and challenges of applying AI-approaches to detect seismic anomalies and generate synthetic electromagnetic satellite data, and underscoring the contribution of such methods to seismic anomaly detection from electromagnetic satellite data.

In the first year of the Dragon 6 program, our researches are focused on the development of efficient time-series InSAR algorithms and their applications in infrastructure deformation monitoring. Progress has been achieved in the following aspects.

• PSSformer: persistent scatters selection method for SAR interferometry based on temporal-spatial vision transformers

The selection of persistent scatters (PS) candidates is a crucial step for the classic PSI method, and the result quality has a strong impact on the final deformation measurements. Existing approaches failed to fully exploit the contextual relationships between phase, amplitude, temporal, and spatial dimensions of SAR datasets, leading to a suboptimal PS selection. Therefore, we propose a novel deep learning method for persistent scatters (PS) selection that leverages the temporal-spatial context features of amplitude images and interferometric phase. Specifically, a temporal-spatial vision transformer (TS-ViT) architecture is employed to process input amplitude and phase time-series stacks simultaneously. Then, a phase consistency attention mechanism (PCAM) is adopted to improve the model's perception. Furthermore, an amplitude and phase interaction (API) module is embedded to achieve multi-scale fusion. After training on the TerraSAR-X datasets over Barcelona, the proposed method achieved good results in urban areas and tall buildings.

• Interferometric phase optimization by joining polarimetric and temporal dimensions

Most existing MT-PolInSAR methods optimize interferometric phase separately in the temporal and polarimetric dimensions, failing to leverage the inter-dimensional relationships fully. Aiming at this problem, a novel multi-polarization optimization method is proposed to achieve one-step phase optimization by combining the temporal and polarimetric dimensions using a joint probability density function and maximum likelihood estimation, named JPTPO. The effectiveness of the proposed approach is validated using both simulated and real Radarsat-2 quad-pol datasets, and the comparison with a few state-of-the-art phase optimization methods show its advantage of improved phase quality in terms of a significantly higher coherence.

• Review of Phase Linking methods in time-series InSAR data processing

The performance of InSAR technology is severely limited by decorrelations arising from temporal changes in scatterer characteristics. The Phase Linking (PL) methods, which can recover the systematic phase series through multi-temporal interferometric phase analysis, have become a core technique for addressing the challenge of monitoring low-coherence areas in time-series InSAR. The motivation, fundamental model, methodological advancements, performance comparison, and future trends of the PL methods have been reviewed systematically. In total, five state-of-the-art PL methods, including PTA, PCA, EMI, StBAS, and AdpPL, are compared with each other using both simulated and real Sentinel-1 SAR datasets. The experimental results show the advantages of the AdpPL method in terms of cumulative phase standard deviation and deformation rate accuracy.

• Progressive sequential polarimetric phase optimization concerning land surface variations

In the era of SAR big data, time-series DInSAR processing approaches, which are capable of near-real-time dynamic monitoring of geological hazards, are essentially needed. However, mountainous areas covered by dense vegetation in dynamic changes are prone to suffering from severe decorrelation, imposing higher requirements on the accuracy, efficiency, and dynamic processing capabilities of phase optimization. For this purpose, a novel progressive sequential polarimetric phase optimization (PSPPO) method based on the previously proposed Sequential Polarimetric Phase Optimization Estimator (SETP-EMI) has been developed. By considering the temporal variations of PS/DS targets and fully utilizing a weight scheme based on polarimetric information for progressive sequential phase optimization, better phase optimization results have been achieved. The progressive polarimetric sequential phase optimization method adopts a recursive processing strategy that incrementally optimizes phases for DS and PS targets in SAR images according to land surface variations. Through adaptively weighting among multiple polarization channels subject to the equivalent scattering mechanism constraint, this method enables refined phase optimization processing and effectively enhances phase optimization accuracy.

• Time-series polarimetric InSAR analysis based on adaptive coherence matrix decomposition

We propose a novel efficient polarimetric time-series InSAR method based on adaptive polarization coherence matrix decomposition (ADCMD-PolMTI), which enables efficient adaptive polarimetric optimization for phases of both PS and DS pixels. Experimental results in Southern California demonstrate that compared with the single VV-polarization method, ADCMD-PolMTI significantly enhances the interferometric phase quality, providing a 494% increase in density of high-quality monitoring points. Validation against GPS data has shown that the proposed method exhibits a lower average root mean square error (RMSE) than the single VV-polarization method, demonstrating a higher in deformation monitoring accuracy. Compared with the ESPO algorithm, it improves the processing speed for PS and DS targets by 235 times and 13 times, respectively. Therefore, the proposed method demonstrates superior adaptive optimization capability and computational efficiency, making it suitable for large-scale surface deformation monitoring.

• Deformation monitoring and prediction of the Hong Kong-Zhuhai-Macao Bridge

A deformation prediction method based on ARIMA-PF (autoregressive integral moving average particle filter) is proposed for modeling the complex deformation characteristics of the Hong Kong-Zhuhai-Macao Bridge (bridge-island-tunnel) structure measured by time-series InSAR analysis. Firstly, the ARIMA model is applied to preliminarily model the InSAR deformation monitoring data. Then, the PF algorithm is introduced for dynamic parameter optimization to enhance the predictive ability of the model under nonlinear conditions and improve the capture effect of complex deformation trends. Selecting the artificial islands of Zhuhai and Macau, Hong Kong Port (affected by soft soil consolidation and construction disturbance), and Qingzhou Channel Bridge (a cross-sea bridge structure) as typical research targets, the predictive performances of ARIMA and ARIMA-PF models are compared. The results showed that ARIMA-PF model exhibited big advantages and significantly improved the prediction accuracy.