深度卷积神经网络特征提取用于地表覆盖分类初探

目的 地表覆盖监测是生态环境变化研究、土地资源管理和可持续发展的重要基础,在全球资源监测、全球变化检测中发挥着重要作用。提高中等分辨率遥感影像地表覆盖分类的精度具有非常重要的意义。方法 近年来,深度卷积神经网络在图像分类、目标检测和图像语义分割等领域取得了一系列突破性的进展,相比于传统的机器学习方法具有更强的特征学习和特征表达能力。基于其优越的特性,本文进行了深度卷积神经网络对中分辨率遥感影像进行特征提取和分类的探索性研究。以GF-1的16 m空间分辨率多光谱影像为实验数据,利用预训练好的AlexNet深度卷积神经网络模型进行特征提取,以SVM为分类器进行分类。分析了AlexNet不同层的特征以及用于提取特征的邻域窗口尺寸对分类结果的影响,并与传统的单纯基于光谱特征和基于光谱+纹理特征的分类结果进行对比分析。结果 结果表明在用AlexNet模型提取特征进行地表覆盖分类时,Fc6全连接层是最有效的特征提取层,最佳的特征提取窗口尺寸为9×9像素,同时利用深度特征得到的总体分类精度要高于其他两种方法。结论 深度卷积神经网络可以提取更精细更准确的地表覆盖特征,得到更高的地表覆盖分类精度,为地表覆盖分类提供了参考价值。     

长江三峡库区Landsat7 ETM+数据的处理方法探讨

Landsat7 ETM+是陆地卫星系列的最新数据,其突出优势在于提高了多光谱波段的空间分辨率,增加了一个分辨率为15 m的全色波段,丰富了数据的信息量。选取长江三峡库区的两景ETM+原始数据,基于不同的地学遥感应用要求,试验性地研究了Landsat7 ETM+新型数据的数字处理方法,并比较了这些处理方法的应用效果。结果表明,选择适当的数字方法对新型Landsat7ETM+遥感数据进行处理,能够快速、有效地获得同时具备较高的光谱分辨率和空间分辨率的遥感数字图像,显示出较好的应用优势。     

HSRS-SC:面向遥感场景分类的高光谱图像数据集

目的 场景分类是遥感领域一项重要的研究课题，但大都面向高分辨率遥感影像。高分辨率影像光谱信息少，故场景鉴别能力受限。而高光谱影像包含更丰富的光谱信息，具有强大的地物鉴别能力，但目前仍缺少针对场景级图像分类的高光谱数据集。为了给高光谱场景理解提供数据支撑，本文构建了面向场景分类的高光谱遥感图像数据集（hyperspectral remote sensing dataset for scene classification，HSRS-SC）。方法 HSRS-SC来自黑河生态水文遥感试验航空数据，是目前已知最大的高光谱场景分类数据集，经由定标系数校正、大气校正等处理形成。HSRS-SC分为5个类别，共1 385幅图像，且空间分辨率较高（1 m），波长范围广（380~1 050 nm），同时蕴含地物丰富的空间和光谱信息。结果 为提供基准结果，使用AlexNet、VGGNet-16、GoogLeNet在3种方案下组织实验。方案1仅利用可见光波段提取场景特征。方案2和方案3分别以加和、级联的形式融合可见光与近红外波段信息。结果表明有效利用高光谱影像不同波段信息有利于提高分类性能，最高分类精度达到93.20%。为进一步探索高光谱场景的优势，开展了图像全谱段场景分类实验。在两种训练样本下，高光谱场景相比RGB图像均取得较高的精度优势。结论 HSRS-SC可以反映详实的地物信息，能够为场景语义理解提供良好的数据支持。本文仅利用可见光和近红外部分波段信息，高光谱场景丰富的光谱信息尚未得到充分挖掘。后续可在HSRS-SC开展高光谱场景特征学习及分类研究。     

联合快舟一号影像纹理信息的城市土地覆盖分类

仅依靠光谱信息无法满足高分辨率遥感分类的应用需求,辅之以纹理特征信息进行分类,可提高影像分类精度。利用KZ\|1卫星影像和Landsat\|8卫星影像数据,基于面向对象的影像分割法和灰度共生矩阵纹理分析法对新疆石河子市局部城区进行了地表覆盖分类实验,将不同空间分辨率的全色影像纹理信息、光谱信息构成多种影像特征组合进行分类比较研究,以选择最佳的分类特征集。结果表明:KZ-1影像能为城市区域的土地覆盖分类提供丰富的纹理信息,面向对象的影像分割可较好地利用高分辨率数据的几何结构信息实现优化的影像分割,从而提高多光谱影像的分类精度,总体分类精度为90.06%,Kappa系数为87.93%,比单纯利用光谱信息分类的总体精度提高了8.02%,Kappa系数提高了9.65%,表明KZ\|1数据可为光谱分类提供丰富的纹理信息,从而提高城市区域的土地覆盖分类精度。     

基于IDL的遥感影像大气与地形校正方法实现

介绍了利用交互式数据语言（Interactive Data Language,IDL）开发TM/ETM遥感影像大气与地形校正模型的详细过程,以2000年4月30日密云ETM影像为例,对大气与地形校正方法的有效性和实用性进行了验证。结果表明,该方法有效地消除了大气与地形影响,提高了地表反射率等地表参数的反演精度和数据质量,为进一步开展定量遥感研究提供了数据质量保障。     

基于多种变换的遥感图像新型融合方法

针对多光谱图像空间分辨率低这一特点,提出一种在PCA变换基础上,利用小波变换和高通滤波相结合的图像融合算法。实现了ETM+全色波段与ETM+多光谱波段图像的融合,并从空间纹理信息,光谱真实性两个方面进行定性和定量评价。研究表明,该融合算法产生的光谱失真较小,同时很大程度地保持了高分辨率全色波段的空间纹理细节信息,是一种较好的图像融合方法。     

复杂地形区土地利用/土地覆被分类机器学习方法比较研究

针对高海拔复杂地形区地貌类型复杂、多样,沟壑纵横、地形破碎等特点,研究快速、有效的土地利用/土地覆被分类方法对土地利用信息获取及更新是非常重要的。以位于黄土高原向青藏高原过渡带的湟水流域为研究区域,基于Landsat 8 OLI影像数据、DEM数据,并结合各种专题特征,在对研究区进行地理分区的基础上,采用人工神经网络、决策树、支持向量机和随机森林4种机器学习方法进行土地利用信息提取并进行精度评价,探索适合于复杂地形区最优的分类方法。研究结果表明:随机森林和决策树的分类精度明显高于支持向量机和人工神经网络。其中随机森林方法的分类精度最高,总体分类精度达85.65%,Kappa系数达0.84。在上述分类基础上,选择随机森林分类方法对Landsat 8全色与多光谱影像融合数据进行进一步的分类研究,总体分类精度达到86.49%,Kappa系数达0.85。这表明随机森林分类方法在保证分类精度的同时又能获得较高的分类效率,对于复杂地形区土地利用信息提取是非常有效的,数据融合在一定程度上提高了分类精度。     

面向GF-2遥感影像的U-Net城市绿地分类

目的 高分2号卫星（GF-2）是首颗民用高空间分辨率光学卫星，具有亚米级高空间分辨率与宽覆盖结合的显著特点，为城市绿地信息提取等多领域提供了重要的数据支撑。本文利用GF-2卫星多光谱遥感影像，将一种改进的U-Net卷积神经网络首次应用于城市绿地分类，提出一种面向高分遥感影像的城市绿地自动分类提取技术。方法 先针对小样本训练集容易产生的过拟合问题对U-Net网络进行改进，添加批标准化（batch normalization，BN）和dropout层获得U-Net+模型；再采用随机裁剪和随机数据增强的方式扩充数据集，使得在充分利用影像信息的同时保证样本随机性，增强模型稳定性。结果 将U-Net+模型与最大似然法（maximum likelihood estimation，MLE）、神经网络（neural networks，NNs）和支持向量机（support vector machine，SVM）3种传统分类方法以及U-Net、SegNet和DeepLabv3+这3种深度学习语义分割模型进行分类结果精度对比。改进后的U-Net+模型能有效防止过拟合，模型总体分类精度比改进前提高了1.06%。基于改进的U-Net+模型的城市绿地总体分类精度为92.73%，平均F₁分数为91.85%。各分类方法按照总体分类精度从大到小依次为U-Net+（92.73%）、U-Net （91.67%）、SegNet （88.98%）、DeepLabv3+（87.41%）、SVM （81.32%）、NNs （79.92%）和MLE （77.21%）。深度学习城市绿地分类方法能充分挖掘数据的光谱、纹理及潜在特征信息，有效降低分类过程中产生的"椒盐噪声"，具有较好的样本容错能力，比传统遥感分类方法更适用于城市绿地信息提取。结论 改进后的U-Net+卷积神经网络模型能够有效提升高分遥感影像城市绿地自动分类提取精度，为城市绿地分类提供了一种新的智能解译方法。     

基于时序Landsat数据的三江平原植被地表类型变化遥感探测研究

Landsat 卫星遥感数据具有分辨率较高,数据积累时间长的特点,在探测地表覆盖变化和地物分类中得到广泛应用。首先,对获取的Landsat TM/ETM+时间序列数据进行了定量化处理,获取了三江平原七台河市1989~2012年时间序列Landsat地表反射率图像。其次,设计了林地指数和湿地指数,提取了三江平原七台河区域地物光谱和时序特征,同时设计构建了地表覆盖分类和植被地表类型变化探测的决策树算法,实现了1989~2012年七台河区域的植被地表覆盖变化的动态监测,提取了森林覆盖变化的空间分布与变化时间。最后,对七台河区域地表覆盖与植被地表类型变化进行了精度检验,分类总体精度达到90.04%,Kappa系数达0.88。研究结果表明:基于定量化的Landsat时间序列数据的分类算法能克服单时相影像分类的缺陷,实现区域地物自动分类和地表覆盖变化的动态监测。
     

小波包信息熵特征矢量光谱角高光谱影像分类

目的 针对高光谱数据波段多、数据存在冗余的特点,将小波包信息熵特征引入到高光谱遥感分类中。方法 通过对光谱曲线进行小波包分解变换,定义了小波包信息熵特征矢量光谱角分类方法(WPE-SAM),基于USGS光谱库中4种矿物光谱数据的分析表明,WPE-SAM可增大类间地物的可区分性。在特征矢量空间对Salina高光谱影像进行分类计算,并讨论了小波包最佳分解层的确定,分析了WPE-SAM与光谱角制图(SAM)方法的分类精度。结果 Salina数据实例计算表明：小波包信息熵矢量能较好地描述原始光谱特征,WPE-SAM分类方法可行,总体分类精度(OA)由SAM的78.62%提高到WPE-SAM的78.66%,Kappa系数由0.769 0增加到0.769 5,平均分类精度(AA)由83.14%提高到84.18%。此外,通过Pavia数据验证了WPE-SAM分类方法具有较强的普适性。结论 小波包信息熵特征可较好地表示原始光谱波峰、波谷等特征信息,定义的小波包信息熵特征矢量光谱角分类方法(WPE-SAM)可增大类间地物可区分性,有利于分类。实验结果表明,WPE-SAM分类方法技术可行,总体精度及Kappa系数较SAM有一定的提高,且有较强的普适性。但WPE-SAM方法精度与效率有待进一步提高。     

A comparison of local variance,fractal dimension,and Moran's I as aids to multispectral image classification

The accuracy of traditional multispectral maximum‐likelihood image classification is limited by the multi‐modal statistical distributions of digital numbers from the complex, heterogenous mixture of land cover types in urban areas. This work examines the utility of local variance, fractal dimension and Moran's I index of spatial autocorrelation in segmenting multispectral satellite imagery with the goal of improving urban land cover classification accuracy. Tools available in the ERDAS Imagine^TM software package and the Image Characterization and Modeling System (ICAMS) were used to analyse Landsat ETM?+ imagery of Atlanta, Georgia. Images were created from the ETM?+ panchromatic band using the three texture indices. These texture images were added to the stack of multispectral bands and classified using a supervised, maximum likelihood technique. Although each texture band improved the classification accuracy over a multispectral only effort, the addition of fractal dimension measures is particularly effective at resolving land cover classes within urbanized areas, as compared to per‐pixel spectral classification techniques.     

Mapping health levels of Robinia pseudoacacia forests in the Yellow River delta,China, using IKONOS and Landsat 8 OLI imagery

The largest artificial Robinia pseudoacacia forests in the Yellow River delta of China have been infected by dieback diseases. Over the past several decades, this has caused a large amount of mortality of Robinia pseudoacacia forests in this area. Timely and accurate information on the health levels of the forests is crucial to improving local ecological and economic conditions. Remote sensing has been demonstrated to be a useful tool to map forest diseases over a large area. In this study, IKONOS and Landsat 8 Operational Land Imager (OLI) sensor data were collected for comparing their capability of accurately mapping health levels of the artificial forests. There were three health levels (i.e. healthy, medium dieback, and severe dieback) based on explicit tree crown symptoms. After the IKONOS and OLI images were preprocessed, both spatial and spectral features were extracted from the IKONOS and OLI imagery, and a maximum likelihood classification method was used to identify and map health levels of Robinia pseudoacacia forests. The experimental results indicate that the IKONOS sensor has greater potential for identifying and mapping forest health levels. Furthermore, texture features, especially texture variance, derived from the IKONOS panchromatic band, contributed greatly to the accuracy of classification results, achieving an overall accuracy (OA) of 96% for the IKONOS sensor and an OA of 88% for the OLI 2, which used both OLI spectral and IKONOS spatial features, compared with an OA of 74% for the OLI sensor alone. Our results indicate that the texture features extracted from high resolution imagery can improve the classification accuracy of health levels of planted forests with a regular spatial pattern. Our experimental results also demonstrate that classification of an image with a spatial resolution similar to, or finer than, tree crown diameter outperforms that of relatively coarse resolution imagery for differentiating living tree crowns and understorey dense green grass.     

Wetland classification using Radarsat-2 SAR quad-polarization and Landsat-8 OLI spectral response data: a case study in the Hudson Bay Lowlands Ecoregion

Circumboreal Canadian bogs and fens distinguished by differences in soils, hydrology, vegetation and morphological features were classified using combinations of Radarsat-2 synthetic aperture radar (SAR) quad-polarization data and Landsat-8 Operational Land Imager (OLI) spectral response patterns. Separate classifications were conducted using a traditional pixel-based maximum likelihood classifer and a machine learning algorithm following an object-based image analysis (OBIA). This study focused on two wetland classes with extensive coverage in the area (bog and fen). In the pixel-based maximum likelihood classification, accuracy increased from approximately 69% user’s accuracy and 79% producer’s accuracy using Radarsat-2 SAR data alone to approximately 80% user’s accuracy and 87% producer’s accuracy using Landsat-8 OLI data alone. Use of the Radarsat-2 SAR and Landsat-8 OLI data following principal components analysis (PCA) data fusion did not result in higher pixel-based maximum likelihood classification accuracy. In the object-based machine learning classification, higher bog and fen class accuracies were obtained when using Radarsat-2 and Landsat OLI data individually compared to the equivalent pixel-based classification. Subsequently, a PCA-data fusion product outperformed the individual bands of the Radarsat-2 and Landsat-8 imagery in object-based classification. Greater than 90% producer’s accuracy was obtained. The margin of error (MOE) was less than 5% in all classifications reported here. Further research will examine alternative data fusion techniques and the addition of Radarsat-2 SAR interferometric digital elevation model (DEM)-based geomorphometrics in object-based classification of different morphological types of bogs and fens.     

Radiometric correction effects in Landsat multi‐date/multi‐sensor change detection studies

Radiometric corrections serve to remove the effects that alter the spectral characteristics of land features, except for actual changes in ground target, becoming mandatory in multi‐sensor, multi‐date studies. In this paper, we evaluate the effects of two types of radiometric correction methods (absolute and relative) for the determination of land cover changes, using Landsat TM and Landsat ETM+ images. In addition, we present an improvement made to the relative correction method addressed. Absolute correction includes a cross‐calibration between TM and ETM+ images, and the application of an atmospheric correction protocol. Relative correction normalizes the images using pseudo‐invariant features (PIFs) selected through band‐to‐band PCA analysis. We present a new algorithm for PIFs selection in order to improve normalization results. A post‐correction evaluation index (Quadratic Difference Index (QD)), and post‐classification and change detection results were used to evaluate the performance of the methods. Only the absolute correction method and the new relative correction method presented in this paper show good post‐correction and post‐classification results (QD index ≈ 0; overall accuracy >80%; kappa >0.65) for all the images used. Land cover change estimations based on uncorrected images present unrealistic change rates (two to three times those obtained with corrected images), which highlights the fact that radiometric corrections are necessary in multi‐date multi‐sensor land cover change analysis.     

Land‐cover classification in the Brazilian Amazon with the integration of Landsat ETM+ and Radarsat data

Land‐cover classification with remotely sensed data in moist tropical regions is a challenge due to the complex biophysical conditions. This paper explores techniques to improve land‐cover classification accuracy through a comparative analysis of different combinations of spectral signatures and textures from Landsat Enhanced Thematic Mapper Plus (ETM+) and Radarsat data. A wavelet‐merging technique was used to integrate Landsat ETM+ multispectral and panchromatic data or Radarsat data. Grey‐level co‐occurrence matrix (GLCM) textures based on Landsat ETM+ panchromatic or Radarsat data and different sizes of moving windows were examined. A maximum‐likelihood classifier was used to implement image classification for different combinations. This research indicates the important role of textures in improving land‐cover classification accuracies in Amazonian environments. The incorporation of data fusion and textures increases classification accuracy by approximately 5.8–6.9% compared to Landsat ETM+ data, but data fusion of Landsat ETM+ multispectral and panchromatic data or Radarsat data cannot effectively improve land‐cover classification accuracies.     

基于多时相Landsat8 OLI影像的作物种植结构提取

针对基于多时相遥感影像、多种特征量提取多种作物种植结构在我国研究较少的现状,利用多时相Landsat8OLI影像数据,根据温宿县不同作物的农事历,通过分析主要地物的光谱特征和归一化植被指数的时间变化信息,构建不同作物种植结构提取的决策树模型,实现了对温宿县多种作物种植结构信息的提取。结果表明:1水稻的最佳识别依据是5月20日影像的近红外波段和7月23日影像的NDVI值;棉花和春玉米的最佳识别依据是5月20日~9月9日影像的NDVI变化值;冬小麦—夏玉米和林果的最佳识别依据是5月20日~7月23日影像的NDVI变化值;2与单时相监督分类相比,多时相决策树法对多种作物种植结构的提取效果更理想,总体精度提高了7.90%,Kappa系数提高了0.10;3Landsat8OLI影像数据分辨率高、成本低、获取方便,是农作物遥感的良好数据源。     

基于超像素统计量的随机森林遥感图像分类

针对遥感图像地物覆盖分类方法对图像空间分布信息利用不足的问题,提出一种基于超像素统计量的随机森林遥感图像分类方法。以北京市海淀区为研究区,选用Landsat-8卫星为主要数据源,通过改进SLIC超像素分割方法,使之适用于多光谱遥感图像中超像素的分割,提取超像素常见的六个统计量(最小值、最大值、均值、标准差、上四分位数、下四分位数)用于随机森林在遥感图像中的分类。实验结果表明,本文对研究区遥感图像的总体分类精度为89.01%,明显改善了对地物的错分和漏分现象,能够推广到Landsat-8遥感图像的地物覆盖分类工作中。     

Comparison of Machine Learning Methods for Land Use/Land Cover Classification in the Complicated Terrain Regions

Aiming at the characteristics of varied and complex geomorphic types,crisscross network of ravines and broken terrain in high altitude complicated terrain regions,it is very important to study and find the rapid and effective land use/land cover classification method for obtaining and timely updating of land use information.Taking the Huangshui river basin located in the transitional zone between the Loess Plateau and the Qinghai-Tibet Plateau as acasestudy area,the objective of this study is to explore a kind of effective information extraction method from comparison of four kinds machine learning methods for complicated terrain regions.based on Landsat 8 OLI satellite data,DEM and combined with various thematic features,on the basis of geographical division of the study area,artificial neural network,decision tree,support vector machine and random forest four machine learning methods for land use information extraction were used to obtain land use data,and confusion matrix was constructed to evaluate classification accuracy.The results showed that the classification accuracies of random forest and decision tree are obviously higher than those of support vector machine and artificial neural network.The random forest method has the highest classification accuracy,the overall classification accuracy is 85.65%,the Kappa coefficient is 0.84.based on the above classification,Random forest classification method was chose to further classify Landsat 8 fusion datafrom panchromatic 15 meter and multispectral 30 meter image,the overall classification accuracy is 86.49% and the Kappa coefficient is 0.85.This indicated that the random forest classification method can obtain higher classification efficiency while ensuring the classification accuracy.It is very effective for the extraction of land use information in complicated terrain regions.Data fusion can improve the classification accuracy to a certain extent.     

Statistical model development and estimation of suspended particulate matter concentrations with Landsat 8 OLI images of Dongting Lake,China

Suspended particulate matter (SPM) is a dominant water constituent of case-II waters, and SPM concentration (C_SPM) is a key parameter describing water quality. This study, using Landsat 8 Operational Land Imager (OLI) images, aimed to develop the C_SPM retrieval models and further to estimate the C_SPM values of Dongting Lake. One Landsat 8 OLI image and 53 C_SPM measurements were employed to calibrate Landsat 8-based C_SPM retrieval models. The C_SPM values derived from coincident Landsat 8 OLI and Moderate Resolution Imaging Spectroradiometer (MODIS) images were compared to validate calibrated Landsat 8-based C_SPM models. After the best stable Landsat 8-based C_SPM retrieval model was further validated using an independent Landsat 8 OLI image and its coincident C_SPM measurements, it was applied to four Landsat 8 OLI images to retrieve the C_SPM values in the South and East Dongting Lake. Model calibration results showed that two exponential models of the red band explained 61% (estimated standard error (SE) = 7.96 mg l^–1) and 67% (SE = 6.79 mg l^–1) of the variation of C_SPM; two exponential models of the red:panchromatic band ratio obtained 81% (SE = 5.48 mg l^–1) and 77% (SE = 4.96 mg l^–1) fitting accuracy; and four exponential and quadratic models of the infrared band explained 72–83% of the variation of C_SPM (SE = 5.18–5.52 mg l^–1). By comparing the MODIS- and Landsat 8-based C_SPM values, an exponential model of the Landsat 8 OLI red band (C_SPM = 1.1034 × exp(23.61 × R)) obtained the best consistent C_SPM estimations with the MODIS-based model (r = 0.98, p < 0.01), and its further validation result using an independent Landsat 8 OLI image showed a significantly strong correlation between the measured and estimated C_SPM values at a significance level of 0.05 (r = 0.91, p < 0.05). The C_SPM spatiotemporal distribution derived from four Landsat 8 images revealed a clear spatial distribution pattern of C_SPM in the South and East Dongting Lake, which was caused by natural and anthropogenic factors together. This study confirmed the potential of Landsat 8 OLI images in retrieving C_SPM and provided a foundation for retrieving the spatial distribution of C_SPM accurately from this new data source in Dongting Lake.     

基于Landsat-8 OLI和GF-1 PMS数据融合的铁染蚀变信息提取

通过对Landsat-8OLI的多光谱波段与GF-1PMS的全色波段进行数据融合,并基于融合后的数据,利用主成分分析法对研究区的铁染蚀变信息进行了提取。为了验证融合后的数据在矿化信息提取中的优势,将数据融合前后分别提取的铁染蚀变信息结果进行了对比分析。结果表明:基于融合后的数据比不进行融合的数据能够提取更丰富的蚀变信息,尤其是对微小特征的提取效果更好。