A multi-neural-network weighted ensemble model for landslide displacement prediction is constructed by integrating three machine learning models: support vector regression (SVR), back propagation neural network (BPNN), and linear regression neural network (LRNN). The model development process primarily consists of two components: individual model prediction and multi-neural-network weighting. Instead of emphasizing prediction errors from all historical moments preceding the current prediction point, this ensemble model focuses on the proximity period around the current prediction point. Optimal weight coefficients for individual prediction models are calculated using the least squares method based on observed data, followed by weighted summation of predictions from individual models at the current point using these weights, ultimately achieving high-precision forecasting results. Comparative tests conducted on Beidou/GNSS monitoring data from the Dangchuan landslide mass at Heifangtai in Gansu province—a typical loess landslide area in China—demonstrate that this proposed ensemble model exhibits superior prediction accuracy compared to single-model approaches, simple averaging ensemble models, and multi-neural-network switching models. The proposed methodology shows significant potential for practical applications in landslide displacement prediction.

This study integrates the high spatial-resolution monitoring advantages of InSAR with physics-based reduced-order modeling to invert sliding surface parameters. Focusing on the progressive Landslide 5# in Dangchuan, Heifangtai (Gansu province)—a typical loess landslide-prone area—we adapted the fault dislocation model from seismology to develop a novel inversion framework. By incorporating time-series InSAR deformation as surface constraints into the dislocation model, the method progressively inverts sliding surface geometry and position from a geodetic perspective. The results indicate that the sliding surface of the 5# landslide mass is located approximately 4.83±0.3 m beneath the slope surface, with a length of about 136.43±0.5 m, a width of around 47.82±0.3 m, and an aspect of approximately 272°±1.0°. The model inversion achieved a fitting degree of 99%, which is generally consistent with the results of geological surveys and statistical analyses.

Based on multi-frequency GNSS observations within a 450 km radius of the 2025 M_W6.8 Dingri earthquake epicenter, including low-frequency (30 s), high-rate (1 Hz), and ultra-high-rate (50 Hz) data, we derived both coseismic static displacement fields and dynamic waveforms using precise point positioning (PPP) technology. Our findings reveal that: 1) The low-frequency GNSS coseismic static displacement field shows a maximum permanent westward displacement of 40.0 mm in the epicentral region, consistent with regional tectonic stress field patterns; 2) High-frequency GNSS waveforms captured significant dynamic displacement peaks at near-field stations, with the XZZF (Everest) station exhibiting 28.83 mm east-west displacement, though a 12.5 mm deviation from low-frequency results was observed due to dynamic effects; 3) Ultra-high-rate GNSS waveforms recorded Love and Rayleigh waves in the far field, identifying double-pulse characteristics of fault segmented rupture. Spectral analysis of high- and ultra-high-rate coseismic waveforms demonstrated that near-field stations (such as XZZF) concentrated energy primarily at 0.5 to 5 Hz, reflecting initial fault rupture processes, while far-field stations (such as XZRK) were dominated by low-frequency surface waves (0.1 to 1 Hz), validating the complementary nature of multi-frequency GNSS data. Notably, ultra-high-rate observations proved irreplaceable for capturing high-rate components (>5 Hz) and transient rupture processes of seismic waves, providing critical constraints for earthquake source parameter inversion.

Based on gravity data, topographic information, and the Crust1.0 global Crustal structure model, this study employs discrete wavelet decomposition, analytical continuation, the Parker-Oldenburg iterative inversion algorithm, and gravity edge detection methods to investigate the characteristics of gravity anomalies and microplate tectonics in northeast Asia. This study divides northeast Asia into 15 microplates and classifies them into three categories including island arc type, cratonic type, and accretionary type. Since the Mesozoic, the tectonic and magmatic activities of the microplates in northeast Asia have been dominated by the subduction of the (Paleo-) Pacific plate, exhibiting significant spatial variations in crustal thickness. The Moho discontinuity is shallower in subduction zones, while it is deeper beneath orogenic belts and cratonic regions. Key tectonic boundaries such as the Mudanjiang suture zone and the Xinlin-Xiguitu suture zone archive the multi-stage amalgamation of microplates in northeast Asia. Both deep-seated geophysical anomalies and surficial magmatic activities collectively unravel the geodynamic linkage between (paleo-) Pacific subduction and the present-day tectonic architecture of northeast Asia.

To better understand the spatial heterogeneity of the lithospheric mechanical structure and its tectonic evolution in western Hubei and adjacent areas, this study calculates and analyzes the elastic lithosphere thickness (T_e) and deviatoric stress based on WGM2012 gravity field data, Crust1.0 model data, and ETOPO1 topographic data. The results indicate that: 1) The elastic lithosphere thickness exhibits significant spatial variation (5 to 50 km). The western margin of the Yangtze craton, including the Sichuan basin and Wuling overthrust (T_e>30 km), represents high-strength stable regions, whereas the Jiangnan orogenic belt (T_e < 20 km) constitutes a weak lithospheric zone; 2) The effective lithospheric thickness is negatively correlated with surface heat flow. The Wuling overthrust, as a cold lithospheric block, is characterized by low heat flow values (< 50 mW/m²) and thick T_e; 3) Seismic activity is controlled by the coupling of T_e and deviatoric stress. The western margin of the Sichuan basin exhibits low T_e values and a high deviatoric stress background (10 to 15 MPa), corresponding to earthquake clustering, whereas the Jiangnan orogenic belt, with low T_e and deviatoric stress generally below 3 MPa, experiences fewer seismic events. Regional seismic hazard assessment requires a comprehensive consideration of the coupling between elastic lithosphere thickness and deviatoric stress field.

The frequency band of oceanic microseismic peaks lies outside the velocity flat response range of borehole short-period seismometers. This study employs a transfer function to restore the oceanic microseismic peaks recorded by low-noise short-period seismometers and uses these restored peaks as the signal source for calculating the installation azimuth angles of borehole short-period seismometers. To investigate the impact of preprocessing parameters on the qualification rate of the calculated installation azimuth angles of downhole short-period seismometers, six sets of filters with different bandwidths were first applied to preprocess 24 hours of continuous data. The continuous data were then divided into five different time lengths. Subsequently, the installation azimuth angles of the borehole short-period seismometers were calculated using a correlation method. Finally, the installation azimuth angles with an average correlation coefficient greater than 0.995 for both the east-west (EW) and north-south (NS) components, and with an angle between the EW and NS components not exceeding 1.2°, were selected. The results show that the most qualified azimuth angles were obtained when the data were preprocessed with a filter bandwidth of 0.2 to 0.6 Hz and a duration of 300 s. Moreover, the number of qualified azimuth angles was significantly higher during the period of 19:00 to 06:00 than during 07:00 to 16:00. A comparison of the Earth's background noise before and after the correction of the installation azimuth angles of surface and borehole short-period seismometers revealed that the EW and NS components did not coincide before correction but were perfectly aligned after correction, thereby validating the effectiveness of using a filter bandwidth of 0.2 to 0.6 Hz and a duration of 300 s for azimuth angle calculation. This set of preprocessing parameters can also be applied to the azimuth correction calculation of borehole very broadband seismometers and the alignment error testing of surface seismometers.

A thorough understanding of the static drift characteristics of CG6 relative gravimeter is crucial for obtaining high-quality relative gravity observation data. We utilize 65-hours observation data from CG6-90 relative gravimeter and iGrav-56 superconducting gravimeter installed on adjacent site at the Sanya seismic station. Three calculation schemes are designed to determine the static drift rate of CG6-90 relative gravimeter. The results show that CG6-90 relative gravimeter exhibits a static linear drift rate of -8.861±0.035 μGal/h, and there is a nonlinear drift phenomenon with gravity variation ranging from -1 to 1 μGal. When calculating the static drift rate of CG6 relative gravimeter, it is essential to apply corrections for gravity changes induced by solid Earth tides, ocean tides, and atmospheric pressure variations.

In this paper, a system identification method based on the combination of GNA and ARMA is proposed to model the transfer function of the pendulum tilt instrument system. The GNA algorithm dynamically adjusts the step size to optimize the estimation of system parameters and the ARMA model models the noise characteristics to suppress random fluctuations. The system identification results are comprehensively analyzed. Compared with the standalone GNA algorithm, the integration of the ARMA model improves the fitting accuracy of the system transfer function from 92.96% to 95.53%, and reduces FPE from 115.9 to 47.28. This enhancement significantly strengthens the system's ability to suppress external noise. Moreover, the zero-pole distribution and residual analysis further validate the improved stability and reliability of the identified transfer function. This method enhances the accuracy and robustness of transfer function identification for the pendulum tilt instrument system, providing a solid foundation for its calibration and performance optimization.

This study employs the probability density distribution method to extract and analyze high-frequency anomalies from digital water level and temperature records preceding the 2016 Cangwu M5.4 earthquake in Guangxi. The findings reveal: 1) Approximately 18 months pre-earthquake, five monitoring wells exhibited distinct high-frequency water-level anomalies demonstrating significant temporal synchronicity; 2) During 1-2 months prior to the event, two wells displayed pronounced short-term high-frequency anomalies, likely indicating a dynamic transition from regional stress accumulation to impending instability; 3) Spatially, anomalous points clustered predominantly along the SW direction of the epicenter, aligning with regional seismotectonic features. Most anomalies showed high spatial concentration consistent with the NW-SE oriented regional compressive stress field. This research establishes a viable methodology for identifying precursory anomalies in fluid monitoring data before moderate-to-strong earthquakes in south China.

In order to better promote the application of time-varying measurement results of ambient noise wave velocity in earthquake tracking and earthquake prediction, this paper comprehensively processes the time series of direct Rayleigh wave travel time migration in 43 stations in western Yunnan in the past five years, and extracts the time-varying anomaly identification index (comprehensive index) of wave velocity suitable for M_S≥5.0 earthquakes in western Yunnan. Combined with the seven M_S≥5.0 earthquake events in western Yunnan during the study period, the R-value scoring method was used to evaluate the prediction efficiency of the comprehensive index. The results show that: 1) The comprehensive index has a good indication for the occurrence time of M_S≥5.0 earthquakes in western Yunnan since 2020. Under the premise of passing the significance test of 97.5% confidence level (R>R₀), the best prediction window length of the comprehensive index is 90 d. For the actual occurrence of 7 M_S≥5.0 earthquakes, the comprehensive index accurately predicts 7 earthquakes, omits 0 earthquakes, and falsely reports 5 earthquakes. 2) The corresponding relationship between the epicenter of M_S≥5.0 earthquakes in the study area and the distribution of abnormal stations before the earthquake (when the index appears) shows that the precursory behavior monitored by the time-varying measurement of ambient noise wave velocity should be more 'field precursor', rather than 'source precursor'. In the actual location prediction, it is necessary to combine the abnormal areas of seismic activity (such as empty areas and bands) and the abnormal areas obtained by field observations such as gravity and mobile geomagnetic to find possible earthquake locations, so as to improve the accuracy of location prediction.

This paper presents the theoretical relationships between two sets of precession models relative to MERS and GCRS, as well as a complete precession model that accounts for frame bias. The theoretical relationships between different precession angles are derived using vector methods, and two new expressions for the general precession in longitude are provided, one of which is suitable for long-term precession studies. Through numerical calculations, the spherical angle expressions between the north ecliptic pole at the reference and observation epochs and the GCRS polar axis are revised. It is pointed out that the P03 or IAU2006 precession models contain errors on the order of microarcseconds, indicating inconsistencies among different precession angles at the level of several tens of microarcseconds, necessitating a re-examination of the numerical calculations of various precession angles.

During the integrated navigation process, the traditional Sage-Husa adaptive filter often suffers from decreased filtering accuracy and even divergence due to external disturbances. To address these challenges, this paper proposes an improved Kalman filter-based integrated navigation algorithm using multiple fading factors. The algorithm dynamically adjusts the predicted state covariance matrix through multiple fading factors to suppress filtering anomalies in high-dynamic environments. Meanwhile, control factors are constructed based on the error characteristics of different state components and combined with the abnormal measurement noise detection conditions to correct the impact of abnormal noise on the integrated navigation system. Simulation results show that when applied to the GNSS/SINS integrated navigation system, the algorithm can accurately track the changes in measurement noise of each state component, effectively prevent filter divergence, and significantly enhance the system's stability and filtering accuracy.

Using 60 days of observational data from 233 monitoring stations of the crustal movement observation network of China (CMONOC) in 2023, the ionospheric delay modeling accuracy and standard point positioning (SPP) accuracy of BDGIM, Klobuchar, NeQuickG and BDSBAS models are analyzed. The results show that the BDSBAS ionospheric correction has the highest accuracy, with the ionospheric delay residual error of 1.0 m and an average correction rate of 87% within the region, and 68% of the 3D positioning errors reaching 1.95 m. The NeQuickG and BDGIM models have similar performance, with the ionospheric delay residual error of 2.0 m and an average correction rate of 73%, and positioning accuracy of 2.59 m and 2.82 m, respectively. The Klobuchar model has a correction rate of 67% and a positioning accuracy of 4.35 m. For single frequency users within China, the BDSBAS ionospheric model is recommended as the first choice, followed by the BDGIM model, which has a relatively simpler algorithm and better accuracy.

The tropospheric delay information of the enhanced precise point positioning (PPP) comes from a small regional reference station network with station spacing of several tens of kilometers, which has the disadvantages of limited service scope and high construction cost. To investigate the impact of wide-area tropospheric delay models with larger station spacing on PPP augmentation performance, observational data from IGS tracking stations in Europe were used to establish tropospheric delay models with average inter-station spacings of approximately 200 km, 260 km, 320 km, and 380 km. These models were then compared with the GPT3 model to evaluate and analyze PPP enhancement performance. The results indicate that the wide-area tropospheric delay models with inter-station spacings of 200 km, 260 km, and 320 km all reduced PPP convergence time to varying degrees, with particularly notable improvements in the height component and its maximum enhancement can reach 39.8%. However, the enhancement effect diminished as average station spacing increased. When the station spacing reached 380 km, the PPP augmentation performance was inferior to that of the global GPT3 model.

We conduct a comparative analysis of RL06.3 time-variable gravity field models released by CSR, GFZ, and JPL, focusing on the degree one, C₂₀ and C₃₀ and their surface mass anomalies, mean degree variances, signal-to-noise ratios of global surface mass changes, and terrestrial water storage changes in typical regions. The results show that degree one, C₂₀ and their surface mass anomalies of three models exhibit significant differences in trend but little differences in annual terms. For the C₃₀, the trend shows considerable differences, while the annual terms are largely consistent. The C₂₀ and C₃₀ of RL06.3 models from the same institution with different truncation degrees have little impact on the trend and annual terms, but the differences between institutions are notable, particularly in the trend. The mean degree variances of three models are highly consistent in low-degree terms, while CSR RL06.3 shows the most gradual rise in noise fitting curves for high-degree terms. The terrestrial water storage inversion results from three models tend to be consistent, but the CSR and JPL models exhibit superior performance in terms of inversion accuracy and consistency. Conversely, the GFZ model generally presents a higher degree of uncertainty in its inversion results. When inverting terrestrial water storage changes, if the uncertainty of results is disregarded, it is recommended to use the GRACE time-variable gravity field models with higher truncation orders released by CSR or JPL. Otherwise, it is advisable to use the GRACE time-variable gravity field models with lower truncation orders released by CSR.

For the downward continuation method of gravity field, the frequency domain continuation method has different requirements for data noise and continuation distance, and there is no clear conclusion yet. This paper analyzes seven frequency domain downward continuation methods, including Milne method, Taylor series expansion method, integral iteration method, Tikhonov regularization method, CGNR method, iterative least squares method, and Landweber iterative method. The model was tested using simulation data and aviation measured data, and the results were compared and studied. The results show that the Milne method and the Taylor series expansion method can be used for continuation of point distances less than 14 times under noise-free conditions; the integral iteration method has better stability when continuation of point distances 1 to 20 times without noise, but is more sensitive to noise; the Tikhonov regularization method has significant advantages when it contains noise and is greater than 5 times of point distances, and can obtain higher accuracy; the CGNR method, iterative least squares method and Landweber iterative method have wide applicability, and can maintain high accuracy when continuation of point distances 1 to 20 times under different noise conditions. Among them, the noise resistance of the Landweber iterative method is the most outstanding, and the latter four methods show certain advantages in actual data experiments, and can obtain stable downtime results.

To address the low computational efficiency of serial programs in integral calculations for marine gravity anomaly inversion, this study takes the inverse Vening-Meinesz formula as a case study to optimize the spherical harmonic synthesis and integration processes. Two parallel computing methods, OpenMP and MPI, are compared, with an improved task allocation strategy proposed for MPI. MPI is ultimately selected to implement the rapid computation of marine gravity anomaly inversion using the remove-restore technique. The results demonstrate that parallel computing significantly enhances the efficiency of integral inversion calculations. OpenMP exhibits higher parallel efficiency, while MPI demonstrates stronger acceleration capabilities on cluster servers. By employing the MPI-based parallel method with optimized task allocation, the inversion process requires only 20.87 seconds under the following conditions: a cluster server with 80 processes (CPU model: Intel Xeon E5-2670 v2 @ 2.50 GHz), a computational region of 10°×10°, an inversion resolution of 1'×1', and an integration radius of 60'. The computational accuracy of this approach slightly surpasses that of the FFT method.