碳酸盐岩高光谱数字化特征及深度学习智能解译

Hyperspectral Digitization Characteristics and Deep Learning Intelligent Interpretation of Carbonate Rocks

碳酸盐岩是工程地质勘察关注的重点,特别是在地形复杂的山区,传统的地面调查方法存在效率低、周期长、工作难度大等问题。高光谱数据真实且丰富的记录着岩石的数字信息,可以实现大范围、多时相、高效率的数据获取。采用深度学习方法识别高光谱图像中的碳酸盐岩,具有效率高、周期短、工作易开展的优势,对辅助工程地质勘察选线,提高工程地质勘察的精度具有重要的意义。以某峡谷周边为研究区,对划定的调查区进行野外工程地质调查和采样,分析现场采集的碳酸盐岩高光谱曲线的数字特征,对碳酸盐岩进行了波段分析和波段选择。选取资源一号02D高光谱影像数据,以调查区的碳酸盐岩高光谱影像为深度学习样本,训练卷积神经网络深度学习模型,采用训练的模型对研究区碳酸盐岩进行识别。深度学习识别碳酸盐岩的总体精度为96.73%,Kappa系数为0.729,生产者精度为59.92%,使用者精度为98.55%,识别结果比传统的光谱角识别方法更加准确、可靠,具有更高的解译效率。深度学习解译结果与区域地质资料、目视解译、论文文献、传统光谱解译等资料对比分析,修正区域地质资料的结果,可以支撑线性工程的地质调查工作和选线工作,为工程勘察提供了新的数字化勘察手段。

Carbonate rocks are a key focus of engineering geological investigation,especially in the difficult mountainous area.Traditional ground-based investigation method has some problems,such as low efficiency,long period and work difficulty.Hyperspectral data is a real and abundant record of digital information of rocks,which can achieve large-scale,multi-temporal and efficient data acquisition.The deep learning method has the advantages of high efficiency,short period and easy work to identify carbonate rocks in hyperspectral images.This is of great significance for assisting in engineering geological exploration and improving the accuracy of engineering investigation.This article focuses on the surrounding area of Canyon as the research area,where field engineering geological surveys and sampling were conducted on designated investigation areas.The digital features of carbonate rocks hyperspectral curves collected on-site were analyzed,and band analysis and selection of carbonate rock were performed.Select the 02D hyperspectral image data of Resource No.1,take hyperspectral images of carbonate rocks in the experimental area as samples,train the convolutional neural network deep learning model,and use the trained model to identify carbonate rocks in the study area.The overall accuracy of deep learning in identifying carbonate rock was 96.73%,with a Kappa coefficient of 0.729.The producers accuracy was 59.92%,and the users accuracy was 98.55%.The identification results of deep learning are more accurate,reliable,and have a higher interpretation efficiency than traditional spectral angles identification methods.Comparative analysis between the interpretation results of deep learning and regional geological data,visual interpretation,literature,and traditional spectral interpretation can be used to correct regional geological data.The interpretation results obtained from deep learning can support geological survey and engineering geological route selection.It provides a new digital survey method for engineering exploration and improves the efficiency of engineering surveys.