基于深度学习的高清影像地物矢量化技术及应用

Deep learning-based vectorization techniques and applications for high definition imagery features

随着“两宽一高”技术的发展,勘探精度逐步提高,采集密度也随之提高,野外施工提质增效也尤为重要。在地震勘探采集过程中,地表障碍物是影响炮、检点布设的主要因素,需要详细的地表障碍物数据保障观测系统的详细设计。目前通过天地图提取矢量文件或者人工标定障碍物是保障观测系统精细设计的主要方法。但是,天地图数据通常情况下会滞后或需要配准后才能使用。人工标定障碍物耗时长、工作量巨大。本文提出一种利用天地图数据作为训练样本,以深度学习的方法训练模型,预测出建筑、道路、水域等障碍物矢量信息,通过多个勘探工区的应用,该方法地物识别精度达到80%以上,地物识别效率较人工标定方法大幅提高,为地震采集观测系统物理点的优化布设提供了快速而有效的支撑。

With the development of broadband and wide-azimuth technology,the exploration accuracy has been gradually improved,the acquisition density has been increased,and it is especially important to improve the qua-lity and efficiency of field construction.In the process of seismic exploration and acquisition,surface obstacles are the main factors affecting the layout of shot points,and detailed surface obstacle data are needed to guarantee the detailed design of the observation system.At present,vector files extracted from sky maps or manually calibrated obstacles are the main methods to guarantee the detailed design of the observation system.However,sky map data usually lags behind or needs to be aligned before it can be used.Manual calibration of obstacles is time-consuming and huge workload.In this paper,we propose a method of using sky map data as training samples,training models with deep learning methods,predicting obstacle vector information such as buildings,roads,waters,etc.Through the application of multiple exploration work zones,the accuracy of this method's feature recognition reaches more than 80%,and the efficiency of feature recognition is greatly improved compared with the manual calibration method,which provides a fast and effective support for the optimization of the placement of the physical points of the seismic acquisition and observation system.