Current task allocation methods for UAV swarms typically ignore environmental uncertainty and overlook the dynamic risks caused by cooperative waiting, making them unreliable in complex operational environments. To address these limitations, this paper proposes a risk-aware segment-learning evolutionary algorithm (RASLEA) for heterogeneous UAV swarm task allocation considering risk uncertainty. First, considering the interval uncertainty of environmental hazard threats, an interval multi-objective cooperative task allocation model is established, and a deterministic conversion mechanism based on Hurwicz decision criterion is introduced. Subsequently, to improve the performance of RASLEA, several new strategies driven by problem characteristics, including decomposition-based population evolutionary mechanism, dual-layer hybrid encoding scheme, and dynamic cooperative repair mechanism based on incremental cost, are designed. Most importantly, a prospective repair decoding strategy is proposed to convert high-risk aerial loitering into safe base delays via reverse temporal calculation, fundamentally eliminating UAV exposure risks on dangerous area. Experimental results on four varying-scale operational scenarios demonstrate that, compared with existing classical algorithms, the proposed method can achieve a set of task allocation schemes with superior distribution and convergence; especially, it shows exceptional solution performance in large-scale scenarios.
Accurately identifying and predicting long-term deformation of tunnel surrounding rock is crucial for the stability assessment and safety control of underground engineering. However, conventional prediction approaches are often limited by idealized assumptions and single modeling frameworks, which hinders their ability to achieve high-precision forecasts. Utilizing over 8 years of deformation monitoring data collected from a drill-and-blast-excavated experimental tunnel at the Beishan exploration tunnel test platform, this study proposes a novel hybrid forecasting model ALA–BiTG by integrating the Artificial Lemming Algorithm (ALA), Bidirectional Temporal Convolutional Network (BiTCN), and Bidirectional Gated Recurrent Unit (BiGRU). In the proposed model, the BiTCN captures long-range temporal dependencies and bidirectional patterns in time-series data, the BiGRU enhances bidirectional temporal dynamics, and the ALA optimizes the hyperparameter configuration of the BiTG subnetwork. The results demonstrate that the proposed ALA–BiTG model significantly outperforms benchmark models including BiTG, CNN-LSTM-ATT, BiGRU, and LSTM on the testing set, achieving a Coefficient of Determination (R2) of 0.971 and a Root Mean Square Error (RMSE) of 0.0016. Furthermore, comparative experiments with several mainstream metaheuristic optimization algorithms reveal that ALA achieves superior performance in convergence speed, computational efficiency, and optimization accuracy, demonstrating its effectiveness and applicability in complex engineering time-series modeling. Finally, based on existing time-series data, reliable prediction of the deformation pattern of surrounding rock in the next 1000 days can be achieved.
The three-dimensional geological model of fold structures can facilitate a better understanding of the complex underground fold structures and provide guidance for resource exploration. To address the inability of traditional arc-segment cutting methods to construct three-dimensional geological models of upright and inclined fold structures that satisfy fold element constraints, this study proposes an integrated explicit–implicit regional modeling method that combines nonuniform rational B-splines (NURBS) and Hermite radial basis functions (HRBFs) based on fold element constraints. The specific procedures include (1) restoring the fold layers based on the NURBS with the fold elements as constraints, (2) performing cutting modeling by implicitizing the NURBS surface to construct a three-dimensional fold model that conforms to the laws of structural geology, and (3) completing the modeling of other geological structures and the implicit integration of the entire geological map based on the implicit HRBF modeling method. This study selects the Yeyahu area in Beijing and the typical fold structure areas in the northwest of China as the research objects to validate the three-dimensional modeling method for fold structures. A comparative study with the traditional arc-segment cutting method shows that the method proposed in this paper integrates mathematical parameters and fold elements for modeling. The modeling process is reasonable, the degree of automation is high, and it can accurately construct the three-dimensional geological models of various upright and inclined fold structures in regional geological survey maps.
Camouflaged Object Detection (COD) aims to accurately segment objects hidden in their surroundings from images. Due to the limited information provided by RGB images, recent studies have begun incorporating depth information to assist detection. However, there are inherent differences between RGB and depth modalities, and how to effectively integrate information from both modes has become a key issue in RGB-D COD. Existing methods often directly fuse dual-modal features extracted by backbone networks, ignoring their differences and complementarity in feature representation. To address this issue, a novel Asymmetric Enhancement and Interaction Fusion Network (AEIFNet) for RGB-D COD is proposed. Specifically, an Asymmetric Feature Enhancement (AFE) module is designed to asymmetrically enhance RGB and depth features, effectively reducing the differences between modalities. Additionally, a Multi-modality Interaction Fusion (MIF) module is proposed to explore complementary information between modalities through interaction mechanisms and adopt an adaptive fusion strategy to mitigate the impact of low-quality depth maps. Finally, a Cross-layer Refinement Decoder (CRD) composed of multiple Cross-layer Refinement Modules (CRMs) is constructed to promote interaction and refinement of features at different levels, effectively balancing global semantic understanding and local detail representation for precise boundary and region prediction. Extensive experiments demonstrate that the proposed method achieves state-of-the-art performance compared to 13 advanced COD models on three public datasets.
Object detection from aerial platforms under adverse atmospheric conditions, particularly haze, is paramount to robust drone autonomy. Yet, this domain remains largely underexplored, primarily hindered by the absence of specialized benchmarks. To bridge this gap, we present HazyDet, the first large-scale benchmark specifically designed for drone-view object detection in hazy conditions. Comprising 383 000 real-world instances derived from both naturally hazy captures and synthetically hazed scenes augmented from clear images, HazyDet provides a challenging and realistic testbed for advancing detection algorithms. To address the severe visual degradation induced by haze, we propose the depth-conditioned detector (DeCoDet), a novel architecture that integrates a depth-conditioned kernel (DCK) to dynamically modulate feature representations based on depth cues. The practical efficacy and robustness of DeCoDet are further enhanced by its training with a progressive domain fine-tuning (PDFT) strategy to mitigate synthetic-to-real domain shifts and a scale-invariant refurbishment loss (SIRLoss) to ensure resilient learning from potentially noisy depth annotations. Comprehensive empirical validation on HazyDet substantiates the superiority of our unified DeCoDet framework, which achieves state-of-the-art performance, surpassing the closest competitor by a notable +1.0% mAP on challenging real-world hazy test scenarios. Our dataset and toolkit are available at https://github.com/GrokCV/HazyDet
