Rongping Zhu

Mine ecological restoration involves multiple tasks, including ecological problem identification, restoration measure selection, and restoration effect evaluation. Relevant knowledge is widely scattered across academic literature and policy and regulatory documents, making it difficult to directly support structured queries, knowledge association, and measure screening. To address the insufficient organization of knowledge in mine ecological restoration and the unstable quality of triples extracted by large language models, this study proposes a large language model-driven workflow for knowledge graph construction and restoration measure decision support in mine ecological restoration. First, a weak ontology schema containing eight entity types and eight relation types was constructed, and a manually annotated dataset was developed based on 30 papers on mine ecological restoration. Second, multiple large language models were compared on the triple extraction task, and Kimi, which achieved the best overall performance, was selected for large-scale knowledge extraction. On this basis, a weak-ontology-guided triple quality control method, named WOTQC, was proposed to check the structural consistency of candidate triples in terms of entity type, relation type, relation direction, and entity orientation. DeepSeek was then introduced to perform verification based on original textual evidence. Subsequently, entity normalization and relation aggregation were conducted on the verified triples to construct a knowledge graph for mine ecological restoration. Path mining, restoration measure decision support, and external case validation were further performed around the “ecological problem–restoration measure–restoration effect” path. The experimental results show that Kimi-based initial extraction achieved an F1 score of 0.7286. After WOTQC processing, the F1 score increased to 0.7561, and after DeepSeek-based evidence verification, it further improved to 0.7862. The final knowledge graph contains 18,783 entity nodes and 23,149 relations, and identifies 3048 “ecological problem–restoration measure–restoration effect” paths. The external case validation shows that the integrated ranking method achieved 1.00 under both strict Hit@3 and strict Hit@5, with a path explanation coverage@5 of 0.7500. The ranking results of the knowledge graph showed good correspondence with the main engineering measures in an actual mine ecological restoration design scheme. The results indicate that weak ontology constraints and verification based on original textual evidence can improve the reliability of triple extraction, while entity normalization and relation aggregation can enhance the structural consistency of the knowledge graph. The constructed knowledge graph can provide traceable and interpretable knowledge support for knowledge organization, candidate restoration measure screening, and restoration scheme formulation in mine ecological restoration, thereby supporting more informed and evidence-based restoration decision-making and contributing to the long-term ecological recovery and sustainable management of degraded mining areas.

Rapid and reliable building damage assessment (BDA) is crucial for post-disaster emergency response. However, existing methods face challenges such as complex background interference, the difficulty in jointly modeling local geometric details and global spatial dependencies, and adverse weather conditions. To address these issues, this paper proposes the Adaptive Difference State-Space Fusion Network (ADSFNet), capable of processing both optical and Synthetic Aperture Radar (SAR) data to alleviate weather-induced limitations. To achieve this, ADSFNet innovatively introduces the Adaptive Difference Attention Fusion (ADAF) module and the Hybrid Selective State-Space Convolution (HSSC) module. Specifically, ADAF integrates pre- and post-disaster features to guide the network to focus on building regions while suppressing background interference. Meanwhile, HSSC synergizes the local texture extraction of CNNs with the global modeling strength of Mamba, enabling the simultaneous capture of cross-building spatial relationships and fine-grained damage details. Experimental results on sub-meter high-resolution MultiModal (BRIGHT) and optical (xBD) datasets demonstrate that ADSFNet attains F1 scores of 71.36% and 73.98%, which are 1.29% and 0.6% higher than the state-of-the-art mainstream methods, respectively. Finally, we leverage the model outputs to construct a disaster-centric knowledge graph and integrate it with Large Language Models to develop an intelligent management system, providing a novel technical pathway for emergency decision-making.