基于数据融合的无人机影像碎屑岩岩性识别

不同类型岩性影像纹理相似性高,基于单一的二维影像进行岩性识别精度较低。本文针对这一问题,开展了顾及影像深度信息的岩性智能识别方法研究。利用无人机影像具有多模态的特性,采用通道叠加、IHS变换、小波变换以及多模态融合四种影像融合方式,将深度信息融入到影像数据中,运用深度卷积神经网络DeepLabv3+进行碎屑岩岩性识别。经人工解译结果对比分析,结果表明,实验区内基于多模态融合影像的岩性识别精度最高,Kappa系数可达76.17%,总体识别精度可提升到91.05%;分析认为,顾及影像深度信息的岩性智能识别方法针对岩层表面不平整,高差落差大的砾岩识别效果有明显提升,但表面平整、高差表现不明显的泥岩和砂岩地层识别效果有待提升,总体为野外碎屑岩露头岩性快速识别提供了新思路。

Lithology identification of clastic rock in UAV images based on data fusion

Because of the similarity of texture of different lithology images, the accuracy of lithology identification based on a single two-dimensional image is low. In order to solve this problem, an intelligent lithology identification method based on image depth information is studied in this paper. Firstly, the depth information is fused with the image data by using four image fusion methods: channel superposition, IHS transform, wavelet transform and multi-modal fusion. Then, based on the converted image data fused with depth information, the intelligent identification method of clastic lithology is studied by using deep convolutional neural network DeepLabv3+ technology. The results of intelligent lithology identification by different fusion methods and manual interpretation were compared and analyzed. The results showed that different fusion methods had different lithology identification effects. In the experimental area, the lithology identification accuracy based on multi-modal fusion images was the highest, with kappa coefficient up to 76.17%, and the overall identification accuracy up to 91.05%. The analysis shows that the intelligent lithology identification method taking into account image depth information can significantly improve the identification effect of the gravel with uneven rock surface and large height difference, but the identification effect of the mudstone and sandstone with flat surface and not obvious height difference needs to be improved. This method provides a new technical idea for rapid lithology identification of clastic rock outcrop in a large area in the field.