Conference Agenda

This report will present the progress of the project: Electromagnetism Anomaly Detection and Deformation Monitoring by Generative and Predictive AI Approaches, studying the development of viable deep learning (DL) methods for generating synthetic electromagnetic data and detecting anomalies within synthetic data generated and observed by the SWARM and Zhangheng satellites. The study also investigates the correlation between detected anomalies and earthquakes. These two data sources offer a unique opportunity to explore potential causes and effects embedded in the measured magnetic fields, and how they can be used for generating synthetic data, particularly in relation to earthquake preparation and seismic precursors.

The project has developed a data-driven, non-physics-informed DL approach targeting earthquake-prone areas, treating electromagnetic data as time series prediction and reconstruction tasks. Two models were proposed: one using Long Short-Term Memory (LSTM) networks and another based on Generative Adversarial Networks (GANs). To ensure the quality over the five-year period of observed electromagnetic data for the targeted areas, extensive pre-processing was performed. For predictions, time series data were processed by extending the given look-back time window using an appropriate frequency interpolation. For reconstructions, the original data window was recovered by the interpolating the frequency representation of its down-sampled part. The pre-processing represents initial steps toward generating synthetic electromagnetic data observed.

Among the two methods, the LSTM-based approach demonstrated superior performance in long-term predictions. Additionally, using equal input and output window lengths significantly enhanced the models' effectiveness. A rigorous evaluation was conducted using five-year electromagnetic data from both SWARM and Zhangheng satellites. The results underscore the suitability of deep learning techniques for these tasks and provide a solid foundation for future efforts to detecting anomalies within observed electromagnetic data incorporated by synthetic electromagnetic data.

In the second year of the Dragon 6 program, progress has been achieved in the following aspects.

l SAR2LiDAR: A deep learning-based cross-modal matching framework between SAR images and LiDAR point clouds

As two fundamental active remote sensing technologies, SAR and LiDAR can be synergistically used to significantly enhance the effectiveness of remote sensing applications. However, existing studies primarily focus on product-level integration, and direct matching remains challenging due to their cross-dimensional modality differences. UPC team developed a deep learning-based matching framework (SAR2LiDAR) that enables direct matching between SAR images and LiDAR point clouds. The framework includes a dataset generation method that creates accurate 2D-3D correspondences between SAR images and LiDAR point clouds. It explicitly models the geometric and scattering visibility relationships of SAR imaging, and employs an autoencoder-based deep learning model with specific global feature representation and cross-modal feature alignment capabilities. The proposed model is trained and evaluated using airborne LiDAR datasets from three urban regions in Spain: Barcelona, Tarragona, and Girona. Extensive qualitative and quantitative experiments demonstrate that the proposed SAR2LiDAR framework achieves accurate and robust matching performance across diverse urban scenarios, and it shows strong generalization capability and potential for large-scale remote sensing applications.

l Efficient detection of active landslides across wide area by combining InSAR phase gradient stacking and YOLO network

In recent years, modern InSAR technology has demonstrated its unique capability of detecting potential landslide hazards by measuring subtle ground surface deformation. Nevertheless, its application in wide-area landslide detection is usually limited by the sensitivity and reliability of InSAR measurements, heavy computation burden due to massive data volume, and subsequent labor-intensive manual interpretation. In this study, we developed a multi-channel YOLO network model that integrates improved InSAR phase gradient stacking (IPGS) results with auxiliary slope data for automatic wide-area detection of active landslides. In total 5,032 ascending and 4,222 descending Sentinel-1 images covering the mountainous regions of Southwest China that were acquired from 2018 to 2022 are processed. Using more than 5000 labeled training samples, the model achieved a good performance in terms of recall, precision and average precision being 0.84, 0.93 and 0.95, respectively. The effectiveness and reliability of the proposed approach were validated against previous studies based on InSAR and optical imagery in Guizhou Province and the upper reach of the Jinsha River.

l Urban surface deformation monitoring with RCM and PAZ satellite SAR datasets

In terms of RCM data applications, our team for the first time applied time-series polarimetric InSAR technology to fully polarimetric RCM data, using Vancouver, Canada as the study area for exploratory research. Cross-validation against Sentinel-1 time-series monitoring results confirmed the reliability and accuracy of the time-series RCM measurements, demonstrating the feasibility of this approach for urban surface deformation monitoring. Furthermore, comparative experiments showed that fully polarimetric-optimized RCM data can significantly improve interferometric phase quality and increase the density of valid monitoring pixels compared to HH single-polarization data.

In terms of PAZ data applications, our team conducted research on differential tomographic SAR technology based on polarimetric phase optimization, achieving the estimation of building height parameters and deformation rates. Comparative evaluation against single-polarization results demonstrated that the three-dimensional point clouds generated by the proposed method exhibit a marked improvement in imaging quality, more completely preserving the structural details of buildings while maintaining good information integrity in low-backscatter regions, validating the superiority of the proposed method in enhancing reconstruction quality.

l Deformation monitoring and prediction of the Lupu Bridge via InSAR and deep learning

In this study, we explored the approach of combining InSAR analysis and LSTM network model to perform deformation monitoring and prediction of large-scale urban bridges. First, 18 ascending and 18 descending COSMO-SkyMed images covering the Shanghai Lupu Bridge are processed by PSI to obtain millimeter-level surface deformation field. Second, the response law of deformation upon temperature variation is effectively identified by cross-correlation analysis combined with dynamic lag characteristics. A temperature-deformation coupling model is constructed to quantify the thermal expansion coefficient of the whole bridge. A 3D finite element model is established to verify the structural displacement response under temperature load. The results show that the arch crown and mid-span region of the bridge deck exhibited maximum displacements, while relatively small displacement can be detected over the arch springing region. Meanwhile, four deep learning models including LSTM, KCC-LSTM, LSTM-Attention, and KCC-BiLSTM-Attention were employed to carry out deformation prediction for the Lupu Bridge. Comparative experimental results suggest the KCC-BiLSTM-Attention model achieved the best performance among the four models and it can accurately capture the structural deformation characteristics of the arch springing and side spans of arch bridges, and the prediction effect is particularly prominent for key stressed components of the bridge.

l Coupled modeling of groundwater flow and land subsidence across the Beijing Plain based on InSAR measurements and groundwater level data

Taking the Beijing Plain as the study area, by integrating unconfined and confined groundwater level monitoring data from 2005 to 2024 and InSAR observations of land subsidence from 2015 to 2024, a coupled groundwater flow and land subsidence model was employed to achieve numerical simulation and parameter optimization. The grid search method was used to calibrate the lag time constant and the total inelastic specific storage coefficient. The model outputs showed good agreement with InSAR observations, with an RMSE of less than 2 cm and an R² close to 1, validating the model's reliability. The optimal parameter calibration and subsidence simulation for Beijing have been completed. A preliminary risk assessment model for infrastructure damage has been established by integrating the spatial distribution of infrastructure. The next step is to extend the study to Tianjin, where optimal parameters will be calibrated based on local groundwater level and subsidence data, and future land subsidence will be simulated using predicted water levels. Damage risk assessments for infrastructure such as buildings and high-speed railways will also be conducted, providing scientific and technological support for land subsidence prevention and control in the Beijing-Tianjin-Hebei region.