面向对象结合深度学习方法的矿区地物提取

Surface features extraction of mining area image based on object-oriented and deep-learning method

为了快速准确获取煤炭矿区的地物信息,以达到辅助安排和部署矿区安全生产工作的目的,采用无人机低空遥感拍摄的方式获取了矿区内的高清影像数据,并提出一种基于面向对象和深度学习的矿区无人机影像地物提取方法。首先利用面向对象的分类方法配合人工校正,制作用于深度学习语义分割的标签,再采用FCN-32s,FCN-8s和U-Net 3种深度学习语义分割模型提取图像特征,训练出3种不同的分类模型,并基于此提出多数投票和打分算法2种集成模型改进地物提取精度。实验结果表明,面向对象结合深度学习方法的地物提取准确率、Kappa系数较传统面向对象方法均有明显提升。其中打分集成模型识别效果最好,在测试集上的整体准确率为94.55%,高出面向对象方法5.96百分点;Kappa系数为0.8191。

Acquisition of surface features of the mining area is greatly helpful to safe mining operation and management.In this paper,the authors propose an object-oriented combined with deep-learning classification method to extract surface features of the mining area based on unmanned aerial vehicle(UAV)images.Firstly,images are segmented by object-oriented method with manual correction to make annotation data set for deep learning models.Secondly,prepared training image data set is used to train 3 deep learning models(FCN-32s,FCN-8s and U-Net)and obtain 3 trained deep learning models respectively.Thirdly,classification accuracy is improved,and 2 integrate algorithms,which are majority voting algorithm and scoring algorithm based on these deep learning models,are proposed.The experimental results show that,compared with the single object-oriented classification method,the proposed methods have higher surface feature extraction accuracy and higher Kappa coefficient,from which the scoring integrate model has the best recognition effect.The overall accuracy of feature extraction on the testing image data set is 94.55%,which is 5.96 percentage points higher than the single object-oriented classification method,with the Kappa coefficient being 0.8191.