面向土地管理的深度学习与线索知识图谱耦合智能现场照片解译

摘要

To strengthen land resource governance, China has established a satellite-ground integrated land use law enforcement (LULE) system, in which remote sensing imagery is used to identify potential anomalies and on-site photographs provide essential evidence for subsequent administrative inspection. However, the large-scale interpretation of on-site photographs remains highly dependent on manual inspection, resulting in substantial administrative workloads, potential inconsistency in enforcement judgments, and limited transparency in evidence interpretation. To address these challenges, the study proposes an intelligent evidence interpretation framework based on deep neural networks and a clue knowledge graph to support on-site photograph inspection in LULE. The framework establishes a hierarchical evidence-processing workflow, in which a cross-scale attention-convolution fusion (CAF) model performs preliminary scene interpretation, a positional prior-constrained object detection (PPOD) method extracts enforcement-related visual clues, and a clue knowledge graph connects these observations with land-use compliance rules for traceable reasoning. Experimental results on 102,519 real-world on-site photographs covering eight LULE scenarios in Guangdong Province demonstrate that the proposed framework substantially improves the reliability of compliance-related evidence interpretation compared with conventional scene interpretation approaches, while providing traceable evidence chains for administrative review. The proposed framework provides an intelligent evidence-processing approach that supports more efficient, consistent, transparent, and evidence-based land-use law enforcement practices.