摘要
离子吸附型稀土矿是重要的战略资源,长期的粗放式开采导致矿区地表覆盖遭到严重破坏,生态环境面临严重挑战。准确精细的土地利用信息是矿区生态恢复和过程监管的重要基础,利用高光谱影像获取土地利用信息被认为是准确监测大范围矿区的有效手段。然而,稀土矿区的地物复杂性和高光谱图像的信息冗余给其精细分类带来了挑战。本研究构建了一种基于面向对象思想和多层注意力卷积神经网络的稀土矿区精细分类方法。首先利用尺度参数估计模型定量分析了稀土矿区影像的多层次最优分割尺度,并获取了分割影像中的光谱、指数、纹理、几何4类影像特征,然后基于距离可分性分析得到了最优特征组合,在此基础上应用多层注意力卷积神经网络(OCTC)模型完成分类,该模型由一维卷积神经网络(1D-CNN)改进而来,通过引进Transformer和CBAM提升模型的特征提取能力和整体分类精度。为验证方法的有效性,以“珠海一号”高光谱遥感影像作为数据源,以江西赣南岭北稀土矿区作为研究区域进行实际验证,并与KNN、RF和1D-CNN分类方法进行精度对比分析。结果表明,该分类方法有效避免了椒盐现象的出现,分类整体性好,并且改进后的多层注意力卷积神经网络模型获得了最佳的分类精度,其总体精度可达88.11%,较其他分类方法提高1.22%~8.84%,Kappa系数提高了0.0159~0.1090。该方法能为稀土矿区的土地利用精细化分类与生产监测、环境保护管理提供方法借鉴与科学参考。
Ion-adsorption-type rare earth minerals are important strategic resources.Long-term extensive mining has led to severe surface damage in mining areas,posing significant challenges to the ecological environment.Accurate and detailed land use information is a critical foundation for ecological restoration and process monitoring in mining areas.Hyperspectral imagery is considered an effective means for large-scale monitoring of mining areas to obtain land use information.However,the complexity of the land cover and the information redundancy in hyperspectral images pose challenges for fine classification.This study proposes a fine classification method for rare earth mining areas based on object-oriented thinking and a multi-layer attention convolutional neural network(OCTC).First,a scale parameter estimation model was used to quantitatively analyze the optimal segmentation scale at multiple levels of the rare earth mining area images.Four types of image features—spectral,index,texture,and geometric—were extracted from the segmented images.Then,an optimal feature combination was obtained through distance separabil classif ity analysis.Based on this,a multi-layer attention convolutional neural network model(OCTC)was used for ication.This model is an improved version of the 1D-CNN,integrating the Transformer and CBAM to enhance the model's feature extraction capabil ities and overall classification accuracy.To verify the method,s effectiveness,Zhuhai-1hyperspectral remote sensing imagery was used as the data source,and the Jiangxi Gan,nan Lingbei rare earth mining area served as the study region.The proposed method was compared with KNN,RF,and 1D-CNN classification methods for accuracy analysis.The results demonstrate that the proposed method effectively mitigates salt-and-pepper noise,maintains good overall classification integrity,and achieves the highest classification accuracy.The overall accuracy reached 88.11%,representing an improvement of 1.22%to 8.84%compared to other classification methods,with the Kappa coefficient increasing by 0.0159 to 0.1090.This method can provide valuable reference and scientific insights for the fine classification of land use and production monitoring,as well as environmental protection management in rare earth mining areas.
作者
范晓勇
李恒凯
刘锟铭
王秀丽
于阳
李潇雨
FAN Xiao-yong;LI Heng-kai;LIU Kun-ming;WANG Xiu-li;YU Yang;LI Xiao-yu(Jiangxi Provincial Key Laboratory of Water Ecological Conservation in Headwater Regions,Jiangxi University of Science and Technology,Ganzhou 341000,China;Geospatial Information Engineering Team,Jiangxi Provincial Geological Bureau,Nanchang 330000,China;School of Economics and Management,Jiangxi University of Science and Technology,Ganzhou 341000,China)
出处
《光谱学与光谱分析》
北大核心
2025年第9期2666-2675,共10页
Spectroscopy and Spectral Analysis
基金
国家自然科学基金项目(42161057)
江西省自然科学基金重点项目(20232ACB203025)资助。
关键词
面向对象-卷积神经网络
珠海一号
高光谱遥感
离子型稀土
土地利用
Object-oriented convolutional neural network
Zhuhai-1
Hyperspectral remote sensing
Ion-adsorption rare earth
Land use
