The Global Navigation Satellite System Interferometric Reflectometry (GNSS-IR) technique is an effective method of monitoring snow depth. The detrended signal-to-noise ratio (dSNR) series is analyzed by a Lomb–Scargle periodogram (LSP) to extract the characteristic frequency, which can be converted to the snow depth. However, the dSNR data are greatly affected by noise in the observation environment, which leads to an abnormal characteristic frequency and low accuracy of snow depth retrieval. In order to reduce the influence of noise and to ensure the correct extraction of the characteristic frequency, we present an improved adaptive retrieval method for the multi-constellation retrieval scenario. First, the dSNR sequences are decomposed adaptively into several singular spectrum components (SSCs) with different frequency scales by singular spectrum decomposition. Then, the corresponding SSCs are selected, according to the empirical scope of snow depth, to reconstruct the ‘pure’ dSNR series. Finally, the reconstructed signals are analyzed by LSP to derive the characteristic frequency, in order to obtain the snow depth. Multi-GNSS observations of site SG27 and site P351 from the plate boundary observation network in a representative period from winter 2019 to spring 2020 were used to validate the proposed method. The snow depths were estimated from individual signals, individual constellations and multi-GNSS combination using both the traditional and improved methods. The experimental results show that compared with the traditional method, the snow depth trend of the improved method is more consistent with the measured snow depth trend, especially in the early stage of snowfall. Furthermore, the proposed method shows a universal applicability to various signals of GPS, GLONASS, Galileo and BDS and the retrieval accuracy of all signals is improved to different degrees. When using multi-GNSS combination signals, the mean bias and root mean square error (RMSE) of multi-GNSS snow depth retrieval at site SG27 are improved from 4.6 and 6.2 cm to 4.2 and 5.4 cm, respectively. The mean bias and RMSE at site P351 are improved from 10.5 and 12.4 cm to 9.5 and 11.5 cm, respectively.

Sea level monitoring is of great significance for studying global climate change, disaster monitoring, and water resource management. GNSS Interferometric Reflectometry (GNSS-IR) technology is considered an effective complement to traditional sea level monitoring methods and has gained significant attention in recent years. However, in the two main GNSS-IR methods for sea level retrieval (spectral analysis and inverse modeling), real-time performance is limited because each low-elevation angle trajectory corresponds to a single retrieval value, and post-processing is required to eliminate outliers. In this study, we propose a near-real-time sea-level retrieval method based on shared-frequency signals. A sliding time window was used, and different strategies were employed to combine dSNRs from various satellites within the time window using both spectral analysis and inverse modeling processing methods to accurately extract the reflector height, which was further converted into sea level. Through experiments conducted at two PBO stations, we verified that near-real-time spectral analysis and inverse modeling methods can stably output sea-level retrievals within 20/30-minute time windows. Subsequently, a comparison with the traditional method revealed that the near-real-time approach enables sea-level monitoring with higher accuracy and much more uniform time resolution. This study shows that GNSS-IR technology can achieve near real-time, high-precision sea-level monitoring in multi-system scenarios and further promotes its application in sea-level monitoring.

Accurate road surface and boundary information is vital for road damage detection, high-definition map construction, and traffic scene perception. To address the challenges posed by complex road scenes, irregular point cloud data, and class imbalance, this paper develops a novel dataset of mobile laser scanning (MLS) point clouds and proposes a deep learning-based method for road surface and road boundary segmentation. To handle the problem of large scenario spans and large data volumes, a time-based data partitioning strategy is adopted to preserve road geometry. And to tackle the class imbalance issue, a dynamic equilibrium sampling method is proposed. Moreover, the graph attention module is incorporated to enhance the model’s ability to capture structural features. Comprehensive experimental results demonstrated that the proposed model achieves outstanding performance in large-scale road scenes, with the overall accuracy exceeding 99%, mean accuracy surpassing 98%, and the mean intersection over union above 89%, significantly outperforming existing mainstream models. These results provide an efficient and reliable solution for the automated processing of MLS point clouds and key technical support for road extraction. Future research will focus on the study of cross-domain generalization capabilities across multiple scenarios.