深度卷积网络在航空高光谱岩性识别中的应用——以塔木素铀矿床北部地区为例  

Application of deep convolutional networks in airborne hyperspectral lithology identification:A case study of the northern Tamusu uranium deposit

作　　者：张川[1] 易敏 童勤龙[1] 叶发旺[1] 徐清俊[1] 李泊凇 ZHANG Chuan;YI Min;TONG Qinlong;YE Fawang;XU Qingjun;LI Bosong(National Key Laboratory of Remote Sensing Information and Image Analysis Technology Beijing Research Institute of Uranium Geology,Beijing 100029,China;Beijing Zhixin Remote Sensing Geographic Information Technology Co.,Ltd,Beijing 100032,China)

机构地区：[1]核工业北京地质研究院遥感信息与图像分析技术国家级重点实验室,北京100029 [2]北京智信遥感地理信息技术有限公司,北京100032

出　　处：《世界核地质科学》2024年第1期33-46,共14页World Nuclear Geoscience

基　　金：核能开发项目(编号:遥2001)资助。

摘　　要：岩矿信息识别是高光谱遥感在地质勘探领域的主要应用方向。传统高光谱遥感方法尽管在矿物识别中取得了良好效果,但对于岩性识别存在瓶颈。深度学习是当前人工智能领域的研究热点,卷积神经网络是适用于图像识别的重要网络架构。以巴音戈壁盆地西部塔木素铀矿床北部区域为试验区,以SASI航空高光谱影像为数据源,将深度卷积神经网络引入航空高光谱遥感岩性识别,测试和评估其应用效果。基于预处理后的SASI航空高光谱影像,以试验区地质图及野外调查为参考,制作了8类样本,包括:印支期花岗岩、华力西晚期花岗岩、华力西晚期花岗闪长岩、华力西中期石英闪长岩、石炭系碎屑岩、中下侏罗统火山凝灰岩、第四系沉积物和绢云母化蚀变岩。构建了基于光谱特征的一维卷积神经网络、基于图-谱联合特征的一维+二维卷积神经网络和三维卷积神经网络3种模型结构,分别进行模型训练、测试和试验区岩性分类应用。模型测试结果表明:一维卷积神经网络、一维+二维卷积神经网络和三维卷积神经网络的总体精度分别为82.13%、86.46%和90.90%。通过评价分析三种卷积神经网络模型的岩性分类识别结果,三维卷积神经网络的识别结果与真实参考最为接近,对试验区各类岩性的区分识别效果最优,一维+二维卷积神经网络的识别效果次之,表明利用卷积神经网络引入高光谱图像空间信息,进行图-谱特征的联合挖掘,有利于提高影像的识别精度和实际应用效果。同时,一维卷积神经网络和一维+二维卷积神经网络的识别结果因航空高光谱影像拼接后的条带效应,影响了它们的实际应用效果,而三维卷积神经网络较好地克服了这种影响,表明其对于大面积航空影像处理具有相对较好的应用前景。Rock and mineral identification is a key application of hyperspectral remote sensing in geological exploration.Although traditional hyperspectral remote sensing methods perform well in mineral identification,lithology identification is still a challenge.Deep learning,particularly convolutional neural networks(CNNs),is a hot topic in artificial intelligence research which provides an important framework for image recognition.This study focuses on the lithological identification in the northern region of Tamusu uranium deposit in the western Bayingebi basin with SASI airborne hyperspectral images as a data source.The paper introduced deep CNNs method in the lithology identification with airborne hyperspectral remote sensing and evaluated their application effectiveness.Based on the preprocessed SASI airborne hyperspectral images and referencing the geological map and field surveys of the test area,eight types of samples were collected:Indosinian granite,Late Variscan granite,Late Variscan granodiorite,Middle Variscan quartz diorite,Carboniferous clastic rocks,Middle-Lower Jurassic volcanic tuff,Quaternary sediments,and sericite-altered rocks.Three model structures were developed:one-dimensional CNN based on spectral features,one plus two dimensional CNN based on combined graph-spectral features,and three-dimensional CNN.These models were trained,tested,and applied to classify lithologies in the test area.The testing results of the models indicated that the overall accuracies of the one-dimensional CNN,the one plus two dimensional CNN and the three-dimensional CNN were 82.13%,86.46%and 90.90%,respectively.The evaluation and analysis of the lithology classification recognition results from the three CNN models showed that the three-dimensional CNNs results were the closest to the actual reference,providing the best differentiation and recognition performance for various lithologies in the test area.The one plus two dimensional CNN ranked second in performance,suggesting that incorporating spatial information from hyp

关 键 词：航空高光谱遥感 深度学习 卷积神经网络 岩性识别 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]