HIUFE: Hybrid intelligence-based unauthorized farmland excavation scene cognition

Résumé fourni par la source

Unauthorized farmland excavation refers to activities such as digging, mining, and related resource development within farmland boundaries, conducted without legal authorization or in violation of relevant regulations. These activities directly contribute to the destruction and functional degradation of farmland, posing significant threats to national food security and social stability. Existing farmland monitoring methods utilizing video recognition exhibit limitations, including high false positive rates, and low levels of automation. To address these challenges, this paper proposes a hybrid intelligence-based cognitive approach to video scene analysis for unauthorized farmland excavation activities. At the data level, a video dataset capturing the behavioral interactions of construction machinery in unauthorized farmland excavation scenes is constructed, incorporating temporal and spatial dimensions to comprehensively depict interaction features among the machinery. At the algorithmic level, considering the frequent motion of objects and the high timeliness requirements in video scenes, expert knowledge is integrated to enhance YOLOv8, specifically proposing a hybrid intelligence-based object behavior recognition model that accurately captures subtle feature differences in the same object under different behaviors. During the inference phase, a knowledge graph and reasoning mechanism are constructed to deeply integrate dynamic video information with domain knowledge, overcoming the challenge of incomplete recognition of object interaction and achieving precise identification of unauthorized farmland excavation activities. Comparative experiments thoroughly validate the model’s superiority in identifying subtle feature differences. Compared to the latest single-stage object detection model, YOLO11, the proposed object behavior recognition model improves the F1 score by 3.26% (from 85.17% to 88.43%). Ablation experiments further confirm the effectiveness of incorporating expert knowledge. For example, the CSPELAN module, enhanced with multi-scale feature knowledge, increases the F1 score by 3.75% (from 84.29% to 88.04%). The research outcomes not only provide efficient and reliable technical support for farmland protection, but also contribute valuable practical experience and methodological references to theoretical innovation and technological development in the field of geospatial intelligence analysis.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.