基于多特征U-Net网络的工程活动图斑快速识别

为有效监测具有填挖、采剥等行为的工程活动情况,本文提出了一种用于像素级露天工程活动图斑提取的遥感多特征语义分割模型。该模型以高分二号(GF-2)光学遥感影像为数据源,采用U-Net深度神经网络架构,通过人工标注构建了反映露天工程活动的影像样本集,并提取样本的多维特征投入模型进行训练,从而实现了工程活动图斑的快速识别。试验结果显示,本文方法对露天工程活动图斑的总体识别精度可达87.36%,平均精度达86.78%,优于KNN、SVM两种传统分割方法,为工程活动自动化监管提供了技术参考。

Fast semantic segmentation of surface engineering activity area based on multi-feature U-Net

In order to effectively monitor the engineering activities with filling, digging, mining and stripping behaviors, this paper proposes a remote sensing multi-feature semantic segmentation model for pixel-level surface engineering activity extraction. In this model, GF-2 optical remote sensing images are used as the data source, and U-Net is used as deep neural network. By manual labeling, the surface engineering activity image samples are constructed for model training with multi-dimensional features, so as to achieve the effect of surface engineering activity extraction. The experimental result shows that the overall extraction accuracy of this method is 87.36%, and the average accuracy is 86.78%, which is better than KNN and SVM. The proposed method provides a technical reference for the automatic supervision of engineering activities.