Semantic segmentation of meter-scale remote sensing imagery is essential for detailed land-cover mapping, agricultural assessment, and ecological analysis. GaoFen-2 (GF-2) multispectral imagery, with a spatial resolution of 4 m, provides valuable spatial information for land-cover interpretation. However, accurate segmentation remains challenging because of severe class imbalance, fragmented minority-class regions, complex land-cover transitions, and heterogeneous spatial structures. To address these challenges, this study proposes a Class-Imbalance-Aware Edge-Graph Enhanced UNet (CAEG-UNet) for GF-2 semantic segmentation. Built upon the classical UNet, CAEG-UNet integrates class-frequency-aware weighting, weighted sampling, and composite loss optimization to alleviate majority-class bias. A Sobel-based edge-enhancement module provides supplementary gradient cues for land-cover transitions, while a graph-reasoning bottleneck captures long-range contextual dependencies. In addition, a minority-aware training phase and an auxiliary head are introduced to improve the representation of underrepresented categories. Experiments were conducted on three representative regions from the Gaofen Image Dataset, including the Beijing urban area, Anhui agricultural area, and Yunnan hilly area. Compared with UNet, CAEG-UNet improves mIoU by 31.81%, 5.83%, and 7.34% under the corresponding aggregated class settings, respectively. It also achieves the best overall performance among all evaluated models and delivers consistent improvements across the three regions. Additional validation on the LoveDA dataset yields an overall mIoU of 57.39%, providing further evidence of the applicability of the proposed architecture to different scene distributions. Visual comparisons and error maps show that CAEG-UNet produces more spatially coherent predictions with fewer errors, particularly for fragmented and minority land-cover categories. These results demonstrate that jointly modeling class imbalance, transition information, and global context is effective for GF-2 land-cover segmentation and supports reliable detailed land-cover mapping.
A comprehensive monitoring of urban drainage network (UDN) is essential for maintenance, management, and sustainable urban development. However, limited sensor deployment hinders the acquisition of sufficient information. Conventional deep learning methodologies can predict and correct monitored data but struggle with unobserved data. Hydraulic models can simulate behaviors but face data collection challenges and low real-time performance. To address these issues, a novel spatiotemporal graph convolutional network (STGCN) model, based on graph neural networks, is proposed to reconstruct a real-time information system for UDNs. By extracting fundamental elements from limited monitoring data and UDN topology, the STGCN model effectively reconstructed unmonitored node data. The experimental results showed that the training efficiency and reconstruction accuracy of the model could be optimized by reducing the spatial data dimensionality to 0.6, adopting a passive-masked training strategy with a ratio of 4:3 for model-training sensors to loss-calculation sensors, and using a historical data input length of 3 h. This approach allowed for the reconstruction of water levels for 527 unmonitored nodes using only seven monitoring nodes, with a median mean absolute error of 0.038 m and an accuracy of 71.3 %. These results demonstrate that the STGCN model can accurately reconstruct unmonitored node data using low monitoring-node density and basic network topology, offering a practical solution to datadriven challenges in intelligent UDNs. The source code is available at https://github.com/holylove9412/ UDNs_STGCN_model.
Group excavations are composed of several individual excavations adjacent to each other with simultaneous or successive construction sequences (CS), which are distinctive from individual excavation in terms of the performance of excavation. In this study, a hyper-scale 3D finite element model was established to investigate the deformation behavior of a diaphragm wall system retaining a deep and oversized group excavation (DOGE) in Shanghai soft clay deposits. The numerical model simulated the practical construction stages and sequences, and it was verified by a series of comparisons with field measurements. Based on the numerical model, the spatial effect of the performance of DOGE in the process of excavation stages was investigated in this study, which cannot be addressed by limited field measurements. Furthermore, the effects of partition walls and CS on the deformation control were discussed to provide practical suggestions for oversized and deep excavations. The results indicate that the employment of bi-partition walls to divide the oversized excavation into several small pits and mono-partition walls and cross walls to further divide the pits near the metro lines into smaller ones, was proved to have significant effectiveness in controlling the wall deflection and protecting the adjacent metro line. For the partition wall, the magnitude and direction of the wall deflection primarily depended on the initial excavation, while the influence of subsequent excavation activities proved insignificant. Thus, it should be noted that the effect of the initial excavation should be especially concentrated. The findings can help optimize similar DOGE engineering.
With the rapid development of underground rail transit, subway spaces have become an essential part of modern urban life. While carefully designed subway spaces with appropriate scales can significantly enhance passengers' spatial perception and associated visual comfort, there has been a lack of quantitative approaches to systematically investigating the optimal scale for subway station hall (SSH) designs. In this study, we identified five common spatial morphology types of SSH from cross-sections: column-free rectangular, column-free curved, column-free vertical-wall and curved-ceiling, single-column rectangular, and double-column rectangular SSH. We then constructed representative models with different spatial scales under fixed viewpoints as experimental stimulus images from each spatial type of SSH, generating 84 scenes in total. The semantic differential (SD) method, eye-tracking technology, and associated statistical analyses were employed to investigate the relationships between participants' psychological perceptions and the spatial scale of SSH. The results indicate that the width-to-height ratio (D/H), total fixation time and the number of fixations in areas of interest (AOI) are key factors affecting passengers' visual comfort in SSH. While the column-free rectangular SSH was preferred by participants, better scenes that are more visually pleasing as well as optimal spatial scale ranges were also identified in each spatial type of SSH. This study provides strong evidence-based support for the design and optimization of subway stations through both subjective and objective methods.
For bridge structures, the mechanical systems in real physical space are high-dimensional, complex, and nonlinear, and the loads that bridges experience are random and time-varying. This makes it very difficult to conduct statistical analysis of the mechanical effects of bridge structures in real physical space. In view of this, this article redefines the P-function and L-function used to describe traffic flow loads through theoretical derivation of statistical steady-state mechanical effects analysis of bridges: The P-function represents the average weight of all vehicles passing through any position on the bridge deck; The L function represents the probability of a vehicle passing through any position on the bridge deck. Furthermore, a digital twin model for traffic flow loads was proposed, which is suitable for mechanical effect analysis under steady-state conditions of bridge structure big data statistics, in order to quantify the big data characteristics of traffic flow loads in physical space bridges. Starting from the traditional traffic flow survey, a measured traditional traffic flow load model was established. Using numerical simulation methods, the conversion relationship and parameter influence relationship between it and the proposed traffic flow load digital twin model were studied, solving the indirect detection problem of the entire bridge deck traffic flow digital twin model. In response to practical engineering problems, the digital twin model of traffic flow load was applied to the statistical steady-state strain effect calculation of actual bridge structures, and the calculation results confirmed the accuracy and effectiveness of the model.
