In recent years, space division multiplexing (SDM) technology has been extensively utilized to enhance the capacity of optical fiber communications, with SDM fibers serving as the key enabling foundation. Among them, weakly coupled few-mode optical fibers reduce inter-mode coupling, effectively suppress mode crosstalk, and significantly lower the complexity and cost of the transmission system. In this paper, different structural types of weakly coupled few-mode fibers, and the work undertaken by our research team on center-assisted ring-core fibers are presented. Center-assisted ring-core fibers achieve the design objective of significantly separating spatial mode degeneracy while keeping polarization degeneracy unseparated under the conventional core-cladding refractive index difference level of communication fibers, and exhibit good tolerance to fabrication errors. Furthermore, design rules for controlling the separation of spatial mode degeneracy in center-assisted ring-core fibers are identified, unveiling the principle of regulating specific spatial mode separations by altering the symmetry of the fiber’s refractive index profile. The results provide valuable guidance for the design of similar fiber structures.

Based on a 100 m long home-fabricated bismuth-doped silica fiber with high germanium content, the U-band broadband amplification of 1 650~1 750 nm was achieved with a maximum gain of 28.34 dB at 1 720 nm. At the same time, the influence of different pump schemes on the amplification performance of bismuth-doped fiber amplifier was systematically studied through both experiments and numerical simulations. Results show that bidirectional pumping is more conducive for achieving high gain and low noise figure at low pump power. Furthermore, the gain is greatly affected by the input signal power under forward pumping, followed by bidirectional pumping, and is most stable under backward pumping. This study can provide a reference for the design optimization and application of U-band bismuth-doped fiber amplifiers.

The point cloud semantic segmentation model that integrates multi-source data has significantly improved the classification accuracy of point clouds in areas with mixed ground objects. How to effectively fuse features of different modalities is a key and difficult issue in multi-modal point cloud semantic segmentation. Aiming at urban ground objects, this paper proposed a monocular aerial image-guided LiDAR point cloud semantic segmentation network (IG-Net). This network extracted multi-scale and multi-level features and contextual information from aerial images and LiDAR data and utilized the aerial image features to perform attention-guided weighted fusion on the LiDAR point cloud features, thereby enhancing the expression ability of point features and optimizing the semantic segmentation results of LiDAR point clouds. The proposed model achieved favorable results on the experimental dataset. Compared with the benchmark model RandLANet, its overall accuracy increased by 2.32%, mean intersection over union by 2.58%, and mean F1 score by 2.13%.

Visible watermarks are widely adopted as an important tool for copyright protection. However, as visible watermarks follow fixed embedding rules, they are hardly resistant to destruction by neural networks, which poses significant threats and challenges to existing copyright protection methods. To solve this problem, this paper proposed an active defense method based on recoverable adversarial watermarks, which improved the anti-removal ability of visible watermarks by introducing adversarial noise, thereby forming a new and more effective copyright protection method. In addition, to address the problem that watermarks may cover important areas of the host image after embedding, a recoverable adversarial visible watermark scheme was proposed. This scheme assisted authorized users in recovering the host image by embedding the important regions of the host image as secret information into non-watermark regions, thereby improving the recoverability of adversarial visible watermarks. Experimental results demonstrate that this method can effectively attack watermark removal networks while maintaining favorable recoverability.

Currently, the low-light image enhancement methods mainly adopt a single feature to reconstruct the target image. Among these methods, stacked upsampling-downsampling operations inevitably cause irreversible loss of high-frequency information when performing feature scaling, ultimately resulting in blurred detailed information in the enhanced image. To address this issue, this paper proposed a low-light image detail enhancement method based on edge feature guidance. The method consisted of three components: an edge feature extract module (EFEM), an enhancement module, and an edge-aware feature guidance module (EFGM). By leveraging Transformer and guided by edge features, it progressively generated high-quality enhanced images in a coarse-to-fine manner. First, the EFEM acquired edge features from low-light images via a parallel window transformer block (PWTB), which guided the image enhancement process. Second, the enhancement module employed a coarse-to-fine transformer block (CFTB), which included a channel transformer block (CTB) and a PWTB. These two components extracted global coarse-grained features and local fine-grained features respectively, and modifications were made to the feed-forward network (FFN) in the Transformer. Finally, the EFGM embedded edge features into the image feature space, mitigating the severe loss of details in dark regions. The experimental results show that the proposed method achieves peak signal-to-noise ratio (PSNR) of 24.97 dB, 23.20 dB, and 25.92 dB, and structural similarity index measure (SSIM) of 0.873, 0.865, and 0.941 on the LOL-v1, LOL-v2-real, and LOLv2-synthetic datasets, respectively. All these metrics outperform those of the current mainstream methods. In terms of subjective quality, the enhanced images well preserve the image detail information.

The Greater Khingan Mountains serve as a crucial ecological barrier in northern China. Accurately assessing its long-term forest disturbance dynamics is vital for regional ecological supervision and evaluating the effectiveness of the “natural forest protection project”. However, traditional bi-temporal remote sensing change detection methods struggle to capture the complex intra-annual and inter-annual dynamics of large-area forests. This study used the Google Earth Engine (GEE) platform to construct a Landsat time-series stack for the period 2000—2021. To improve monitoring sensitivity to sub-pixel changes such as forest degradation and selective logging, this study first adopted spectral mixture analysis (SMA) to perform per-pixel unmixing of Landsat imagery, generating a normalized difference fraction index (NDFI) sequence. By taking NDFI as the input for the continuous change detection and classification (CCDC) algorithm, a harmonic model was established for each pixel to fit its change trend, and breakpoints in the time series were automatically identified based on statistical thresholds to capture and refine the occurrence time and location of forest disturbances. The results indicate that: 1) During 2000—2021, the total area of forest disturbances was 28 958 km², with disturbance hotspots concentrated in the northeastern part (Huma County) and northwestern part (Mohe County). 2) The annual disturbance area showed significant fluctuations, with peaks in 2002 (4 092 km²) and 2013 (4 120 km²). 3) Accuracy verification shows that the CCDC algorithm has an overall accuracy of over 91% and a Kappa coefficient of 0.85, which is highly consistent with the results of manual interpretation. This study realizes the monitoring of subtle disturbances of forest degradation in the Greater Khingan Mountains region and can provide important data support for the ecological environment monitoring of this region.

To address the issue of rural hollowing and even the expansion of residential land, this paper extracted residential land from GF-1 satellite imagery using the maximum likelihood method, support vector machine method, and neural network classification method. On this basis, the paper superimposed NPP/VIIRS nighttime light data and point of interest (POI) data to estimate the vacancy rate of rural residential land in Fengnan District, Tangshan City. The results indicate: 1) Given the significant gaps in social security between urban and rural areas, rural laborers who take part-time jobs or work in towns and cities still return home to engage in agricultural work during the busy farming season, resulting in differences in the vacancy rate of rural residential land between the busy farming period and the slack farming period. The vacancy rate of residential land in Fengnan District aligns with the seasonality of agricultural production. In October, the busy farming season, the vacancy rate was 25.35%, which was slightly lower than the 26.73% recorded during the slack farming season. This validates the hypothesis proposed in this study and confirms that the idea of selecting remote sensing images for vacancy rate estimation based on the seasonal characteristics of agricultural activities is basically reliable. 2) The variation mechanism of the residential land vacancy rate in Fengnan District is consistent with that of the POI kernel density analysis results. This indicates that the high-precision estimation method for urban housing vacancy rate is applicable to the estimation of rural residential land vacancy rate in plain areas. The vacancy rate estimation results provide a reference for the selection of research areas in subsequent studies and for a deeper understanding of the vacancy phenomenon of rural residential land.

Helicopter blade low-light image enhancement is an indispensable preprocessing step for helicopter blade motion parameter measurement. Aiming at the problems of low image contrast, noise interference, and the scarcity of paired low-light/normal-light training data, this paper proposes a zero-shot network-based enhancement method. Firstly, an illumination correction module based on an exposure correction S-curve is constructed, where the illumination component of the image is directly enhanced using the optimal Scurve parameters estimated by the network. Secondly, an adaptive convolutional network is designed to extract and suppress image noise. Finally, a zero-shot-based loss function is designed to enable the adaptive training of the entire network without reliance on paired training sets. Experimental results show that the objective evaluation indicators and visual quality of the proposed method on multiple datasets are superior to the state-of-the-art algorithms. Moreover, the network computing capacity required is only 1.24 G, which meets the needs of real-time enhancement of helicopter blade low-light images under airborne conditions.

To address the problem of reduced segmentation accuracy in current ground-based tree point cloud segmentation methods caused by insufficient integration of local and global feature information, this paper proposes a tree segmentation method for forest sample plots based on spatial proximity relationship. This method first establishes the spatial relationship and separates the ground points by constructing an octree index. On this basis, a multi-stage random forest model is developed to achieve progressive segmentation of the tree point cloud. Specifically, the two-dimensional morphological features and spatial properties of tree trunk cross-section are used to accurately segment tree trunk point cloud. Subsequently, based on the segmentation results, the optimized ensemble of shape functions (ESF) feature description operator is used to obtain the spatial connectivity features between the tree trunk and the corresponding tree crown point cloud, enabling crown segmentation by incorporating dimensional properties of the tree crown point cloud. Finally, the structural parameters are extracted for single trees and validated against field measurements. Experiments using two sets of mobile laser scanning point cloud data show that the proposed method achieves tree segmentation recall rates of 90.57% and 90.05%, with accuracies of 93.20% and 95.47%, respectively.

Neural network steganography holds potential for copyright protection and covert communication. However, most existing methods rely on intermediate layer parameters of the model, which often leads to limited embedding capacity and insufficient robustness against noise. To address these issues, this paper proposes a steganographic scheme based on multi-level cascaded dynamic embedding. Firstly, a multi-layer weight distribution function is designed to adaptively allocate and embed secret data across multiple layers of the neural network using dynamic weights, thereby expanding the embedding capacity and enhancing robustness. Furthermore, a channel attention module is introduced into the model to counteract the performance degradation caused by secret data embedding, leveraging enhanced attention features to balance information embedding with model functionality preservation. Extensive simulation experiments demonstrate that the proposed scheme effectively addresses data hiding challenges and outperforms existing methods in terms of embedding capacity, robustness, and security.

To address the issues of fuzzy event description boundaries and insufficient utilization of syntactic and semantic information, this paper proposes a syntactic-semantic-enhanced model for joint extraction of entity-event relationships. The proposed approach comprehensively incorporates contextual, temporal, and syntactic structural information of entities and events. Firstly, it utilizes bidirectional long short-term memory networks (BiLSTM) to capture the contextual and temporal information of the text, generating more semantically enriched embeddings. Secondly, a syntactic-semantic-enhanced model is designed for joint extraction of entity-event relationships, considering the syntactic label semantics in the text. Finally, it establishes a connection between the extraction of entity-event relationships and entity-event matching, achieving the joint extraction of entity-event relationships. Experimental results demonstrate that the proposed joint extraction method not only improves model training efficiency but also enables parameter sharing and reduces error label propagation, thereby enhancing the accuracy of entity-event relationship pair extraction.