Recent progress in prompt-driven vision transformers has markedly enhanced remote sensing image (RSI) semantic segmentation. However, existing methods often fail to incorporate fine-grained geometric structure and to effectively address the spatial heterogeneity present in complex aerial scenes. To overcome these limitations, we propose GeoMoE, a geometry-driven adaptive mixture-of-experts framework tailored for precise and efficient semantic segmentation. Building upon a frozen vision transformer backbone, GeoMoE introduces three novel components, 1) a token-field hybrid adapter (TFH-Adapter) that enables geometry-aware feature adaptation without modifying backbone parameters; 2) a geometry-contextual prompt generator (Geo-Prompt) that integrates multi-scale shape prototypes and contextual cues into expressive prompt embeddings; and 3) a geometry mixture-of-complexity-experts (Geo-MoCE) decoder which dynamically routes spatial regions to specialized experts based on local geometric complexity. This unified architecture allows explicit modeling of geometric information and flexible allocation of decoding capacity, resulting in more accurate segmentation of structurally complex and heterogeneous regions. Extensive experiments on several benchmark remote sensing datasets demonstrate that GeoMoE achieves state-of-the-art performance in segmentation accuracy, model efficiency, and boundary delineation.
In complex and diverse power grid fault scenarios, constructing a systematic fault handling knowledge system is an effective approach to enhancing the intelligence of fault handling. Therefore, we propose a knowledge graph construction method for power grid fault handling based on prompt learning and Low-Rank Adaptation (LoRA) fine-tuning. The proposed method extracts fault handling knowledge from multi-source documents such as fault handling operational regulations and historical cases and represents it in a structured form. The method first segments long documents to satisfy the input length constraint of large language models (LLMs), then designs a stepwise prompting strategy with schema constraints to guide the LoRA-fine-tuned LLM in extracting entities and triples, and finally achieves cross-document knowledge fusion through entity alignment and consistency maintenance to complete knowledge graph construction. Experiments are conducted on four open-source LLMs, namely QwQ-32B, DeepSeek-R1-Distill-Qwen-32B, GLM-4-32B, and gpt-oss-20B, and the results show that QwQ-32B achieves the best performance. After LoRA fine-tuning, the QwQ-32B-tuned model achieves F1 scores of 91.9% and 84.8% for entity and triple extraction, respectively. The constructed knowledge graph can provide decision support for power grid fault handling and serve as a practical reference for LLM-driven knowledge graph construction.
Urban floods pose a serious threat to human life, property safety, and socio-economic development. Urban flood modeling plays a vital role in disaster prevention and resilient infrastructure planning. Traditionally, one-dimensional (1D) drainage network and two-dimensional (2D) surface flow coupling models based on physical processes have been widely used to simulate urban flooding. However, such models require substantial computational resources and long execution times, making them unsuitable for real-time flood forecasting over large areas. This study proposes a CNN–Transformer model as an efficient surrogate for physical models in rapid urban flood forecasting. The model is trained using high-resolution gridded precipitation data and inundation simulation outputs generated by a calibrated 1D/2D hydrodynamic model, covering 5,475,655 grid cells within the study area. By integrating the spatial features of rainfall and the temporal dynamics of flood evolution, the model enables fast and reliable predictions. The study shows that the CNN-Transformer model closely matches the physical model in predicting flood depth, inundation extent, and flood volume, achieving a mean relative error below 10%, NSE ≥ 0.956, and R2 > 0.958. Additionally, the model reduces computation time by approximately 287.4–308.8 times compared to the physical model, meeting the real-time requirements of flood emergency response. These results indicate that the proposed method significantly improves computational efficiency while maintaining high predictive accuracy. It contributes to the development of smarter and more sustainable urban decision-making systems and holds strong practical value for real-time flood risk management, infrastructure planning, and climate resilience enhancement.
The sim-to-real gap remains a fundamental challenge for deploying 6D object pose estimation in real scenes. Although existing synthetic data pipelines scale efficiently, they remain heavily dependent on predefined assets, simplified rendering assumptions, and weak grounding in real image context, thereby limiting illumination realism, appearance diversity, and viewpoint coverage. Recent generative foundation models provide a powerful basis for scalable data construction, but their outputs are often semantically rich while remaining under-constrained in geometry and unreliable in physical attributes such as material reflectance, transparency, and illumination response. To address these mismatches, we propose Real-Gen, a real-world-grounded generative framework for 6D pose data construction and augmentation that uses explicit pose supervision to anchor generative content inside real scenes. The framework combines RGB-only pose recovery, semantic 2D-3D-2D replacement, scene-consistent relighting, appearance augmentation from diverse priors, and video-driven pose-space expansion. Extensive experiments show that Real-Gen improves photometric harmony, provides controllable appearance diversity under fixed pose supervision, and enlarges viewpoint coverage under real-scene constraints. The downstream results further show consistent gains on Linemod-Occluded and improved robustness on challenge test sets with relighting, texture, and novel-view expansions, highlighting the practical value of Real-Gen for enriching 6D object pose benchmarks.
Transition metal sulfides (TMSs) hold immense potential for supercapacitors, yet insufficient energy density and limited cycling endurance impede their practical application. MOF-derived TMSs/carbon composites offer a promising solution by leveraging uniform nanoscale integration, chemically bonded interfaces, and hierarchical porosity. In this study, a one-step carbonization-sulfidation strategy for producing a Ni-Co sulfide/N-doped carbon composite from a bimetallic MOF (NiCo-MOF) precursor is presented. The optimized composite, denoted as NiCoS/NC-2, features a dual-phase NiCo2S4/NiS heterostructure intimately integrated within an N-doped carbon matrix. Together, these features provide abundant electrochemically active sites, enhance structural integrity, establish continuous conductive pathways, and promote swift electron transfer. Consequently, the NiCoS/NC-2 electrode exhibits pronounced battery-type charge-storage behavior, delivering a high specific capacity of 809.8 C g−1 at 1 A g−1 and retaining 62% of its capacity at 20 A g−1. Furthermore, the hybrid supercapacitor (HSC) assembled using NiCoS/NC-2 and activated carbon (AC) achieves a competitive energy density of 54.65 Wh kg−1 at a power density of 816.28 W kg−1 and retains 90.09% of its initial capacitance after 10,000 consecutive cycles at 20 A g−1. This research illustrates a systematic methodology for fabricating MOF-derived composites, establishing them as highly favorable options for next-generation energy storage applications.
