EN中文

地球信息科学学报 >

2017 , Vol. 19 >Issue 11: 1530 - 1537

DOI: https://doi.org/10.3724/SP.J.1047.2017.01530

遥感科学与应用技术

基于深度学习AlexNet的遥感影像地表覆盖分类评价研究

  • 党宇 , 1 ,
  • 张继贤 , 1, 2, * ,
  • 邓喀中 1 ,
  • 赵有松 2 ,
  • 余凡 3
展开
  • 1. 中国矿业大学,徐州221116
  • 2. 国家测绘产品质量检验测试中心,北京100830
  • 3. 中国测绘科学研究院,北京100830
*通讯作者：张继贤（1965-）,男,博士,研究员,主要从事遥感方向研究。E-mail:

作者简介：党 宇（1992-）,男,硕士生,研究光学遥感影像分类方向。E-mail:

收稿日期: 2017-07-07

  要求修回日期: 2017-09-13

  网络出版日期: 2017-11-10

基金资助

国家自然科学基金项目（41671440）

收起

Study on the Evaluation of Land Cover Classification using Remote Sensing Images Based on AlexNet

  • DANG Yu , 1 ,
  • ZHANG Jixian , 1, 2, * ,
  • DENG Kazhong 1 ,
  • ZHAO Yousong 2 ,
  • YU Fan 3
Expand
  • 1. China University of Mining and Technology, Xuzhou 221116, China
  • 2. National Quality Inspection and Testing Center for Surveying and Mapping Products, Beijing 100830, China
  • 3. Chinese Academy of Surveying and mapping, Beijing 100830, china
*Corresponding author: ZHANG Jixian, E-mail:

Received date: 2017-07-07

  Request revised date: 2017-09-13

  Online published: 2017-11-10

Copyright

《地球信息科学学报》编辑部 所有
Fold

摘要

地表覆盖分类信息是反映自然、人工地表覆盖要素的综合体,包含植被、土壤、冰川、河流、湖泊、沼泽湿地及各类人工构筑物等元素,侧重描述地球表面的自然属性,具有明确的时间及空间特性。地表覆盖分类信息数据量大、现势性强、人工评价费时,其自动化评价长期以来存在许多技术难点。本文基于面向对象的图斑分类体系,引入深度卷积神经网络对现有地理国情普查-地表覆盖分类数据进行分类评价,并通过试验利用AlexNet模型实现地表覆盖分类评价验证。试验结果表明,该方法可有效判读耕地、房屋2类图斑,正确分类隶属度优于99%,而由于数据较少、训练不充分,林地、水体图斑正确分类隶属度不高,分别为62.73%和43.59%。使用本文方法,经过大量数据充分微调的深度学习AlexNet可有效地计算图斑的地类隶属度,并实现自动地表覆盖分类图斑量化评价。

关键词： 深度学习; 地理国情普查; 地表覆盖分类; 质量评价; AlexNet

本文引用格式

党宇 , 张继贤 , 邓喀中 , 赵有松 , 余凡 . 基于深度学习AlexNet的遥感影像地表覆盖分类评价研究[J]. 地球信息科学学报, 2017 , 19(11) : 1530 -1537 . DOI: 10.3724/SP.J.1047.2017.01530

Abstract

As one of the important outcomes of the National Geographic Census of China, the land cover classification reveals the information of both natural and artificial coverage elements, including vegetation, soil, glaciers, rivers, lakes, marsh wetlands and various artificial structures. Obviously, it mainly focuses on profiling the natural characters of the land surface with temporal and distribution attributes, which has an obviously different classification system from other scene classification applications. In recent years, more and more high-resolution remote sensing platforms become available, it is possible to update and evaluate land cover classification quickly with the advantage of huge volume of data and more frequent data updates. Meanwhile, in practice we face with more and more challenges of the huge data. In this paper, we propose a novel approach for evaluating the land cover classification results by combining the object-oriented method with the Deep Convolutional Neural Network (D-CNN) model. With deeper structure and wilder receptive field, the deep neural network has the capability of abstract description from low-level features, and the deep learning has become one of the latest development trends in the artificial neural network field. The deep learning shows a completely different possibility in many fields, and it has been applied to the speech recognition, image recognition, information retrieval and so on. The newly-developed method of image recognition based on deep leaning has been preliminarily verified in the scene classification field. Traditionally, the land cover classification method is established on the pixel-based classifying. The latest improved method of the object-oriented classification frame has been proposed, but this new frame is hard to be achieved because of the lack of supports from efficient methods and algorithms. Nowadays, the deep neural network provides us an effective tool to achieve the object-oriented classification by clipping image spots from original images and inputting the clipped image spots to D-CNN. The D-CNN model can convolute and pool the image spots to realize the object-oriented classification of the land cover. By the combination of the object-oriented classification with the deep learning, the proposed method can extract more and better abstract features than the pixel-based approach, while the pixel-based method requires more manual interventions. When applying the deep learning method to land cover classification recognition, the prepared image spots as appropriate inputs will be automatically scored to its belonging classes. Thus, the score represent the degree of membership of the image spot matching to the corresponding class. By fine-tuning the D-CNN, we can obtain a new approach of judging the quality of the samples, in order to assure the reliability of the proposed approach. The fine-tuned D-CNN is required to be sufficiently robust, and we verify its robustness in the following experiment by employing the AlexNet. The experimental results show that the image spots of arable land and building can be recognized with the membership degree of 99.95% and 99.41%, but those of woodland and water area can be recognized only with the membership degree of 62.73% and 43.59%. Obviously, the proposed model can achieve the promising reliability that is related to the qualified and sufficient data set of the image spots which is used for fine-tuning of the net. The reason for poor robustness of the fine-tuned AlexNet in classifying the woodland and water area may be the insufficient size of data-set of these two classes. It shows that a fine-tuned deep convolutional neural network as a new model can be utilized in evaluating the land cover classification with high reliability.

Key words： deep learning; national geographic census; land cover classification; quality evaluation; AlexNet

1 引言

高分辨率遥感影像中地物几何结构与信息更明确,空间信息更丰富,影像中地物尺寸、形状及图斑间的拓扑关系更加清晰,显著提高了地表覆盖分类判读的准确度和效率,已广泛应用于地理国情普查地表覆盖分类中。
地表覆盖分类标准体系较多,普遍存在特征复杂,地类间区别较小的特点。以地理国情普查分类标准为例,地表覆盖分类信息包含10个一级类,50个二级类和96个三级类,拥有庞大、复杂的分类体系与标准[1],使得其在分类上难度较大。此外,高分辨率遥感影像地表覆盖分类分布区域广,地物纹理和结构特征复杂,且影像源多样、时相差异大,以光谱特征为例,同类地物内的光谱差异大,类间的光谱差异较少,同物异谱及同谱异物现象较为普遍。影像中存在大量细节和复杂的地物光谱特征,导致基于光谱统计特征进行分类的方法准确性不能满足高效的自动分类识别需求[2],极大似然法、最小距离法、K-均值聚类法等现有方法均存在以上问题。
深度学习是人工神经网络的最新发展趋势,通过对输入数据从低层到高层逐层提取更抽象的特征,形成最适应所需特征的网络权值结构,从而提升分类的准确性。Hinton等[3]采用深度学习模型实现数据的分类,得出深度神经网络结构较现有方法可学习到更抽象的特征,且具备更强的分类能力,同时具有很好的泛化能力。Hinton和Alex Krizhevsky等[4]提出AlexNet以远高于其它算法的TOP-5（即前5次判断正确）错误率15.3%实现了深度学习方法在图像识别方向的突破。目前,深度学习已经成功应用于语音识别、图像识别、信息检索等领域,这些应用证明了深度学习是一种具有很强泛化能力的分类识别方法。
近年来,新的神经网络方法在遥感领域的应用日益受到关注。在全球地表覆盖分类中,如美国EOS MODIS通过应用浅层神经网络方法,显著提高了生产效率及准确度[5]。国内学者在土地利用、目标识别、场景识别等领域也对深度学习方法进行了相关研究及应用[6-12],Cheng等[13]通过在已有AlexNet前端叠至数据旋转层增强AlexNet在目标识别领域对多角度目标的识别性能,提高了泛用计算机视觉方法在目标识别方向的适用性。Hu等[14]通过对现有网络结构的分析,及对各隐层中间输出量的对比分析,提出了深度学习方法应用在光学遥感领域中的新的特征提取的方法。在地表覆盖分类评价方面,基于浅层神经网络的算法已成功应用于第一次全国地理国情普查中。本文通过引入深度学习方法,利用AlexNet模型对地表覆盖分类图斑评价进行研究,探索深度学习方法应用在地表覆盖分类评价中的有效性。

2 AlexNet深度卷积神经网络

AlexNet深度神经网络已广泛应用于图像识别方向,是近年来计算机视觉领域取得的一项重要突破。AlexNet结构简洁,引入了诸多新方法以实现高效的训练和稳定的收敛速度,其网络结构如图1所示。AlexNet相较与传统神经网络方法的特点和优势如下：
Fig. 1 Schematic diagram of AlexNet（Dual GPUs）

图1 AlexNet（双GPU）结构示意图

（1）ReLu激活函数
AlexNet采用ReLu激活函数取代之前普遍采用的Sigmoid非线性激活函数。由于Sigmoid函数的饱和效应,对较大和较小输入数据会造成梯度损失,且输出梯度不以0为中心,故会在梯度下降阶段造成收敛波动。通过ReLu激活函数可有效改善梯度消失及收敛波动2种缺陷。同时由于ReLu只需一个阈值即可得到激活值,且其为非饱和线性函数,故在随机梯度下降阶段的收敛速度比Sigmoid函数快很多。
（2）数据集扩展抑制过拟合
AlexNet为解决过拟合问题引入了数据集扩展方法,通过对已有影像数据进行变换增加数据集的样本。变换方式主要包括：平移变换、反射变换、光照和色彩变换3种。AlexNet在模型训练阶段,对影像进行随机平移及水平翻转,并通过对RGB影像数据集进行主成分分析,并对主成分进行（0,0.1）高斯扰动,以进一步降低过拟合现象[18]。
（3）DropOut方法抑制过拟合
AlexNet通过引入多种权值组合的DropOut方法控制过拟合。在训练过程中,隐层神经元的激活状态通过特定范数阈值进行控制,超过阈值的神经元在前向传播与反向传播中被抑制。
通过DropOut方法,网络每输入一组新数据,都会激活一组不同的隐层神经元,从而使网络结构发生显著变化,而网络的所有激活状态始终共享权值,从而显著降低了神经元间复杂的互适应关系,从而实现对过拟合的抑制。
（4）局部响应归一化（LRN）层增强泛化
根据先验知识改进函数模型权重是现阶段提升深度神经网络泛化能力的主流研究方向,由于张量在深度模型中的运行阻力会很大程度上影响模型在复杂数据环境下的正则化,故AlexNet模型引入了一种学术界广泛认同的隐式先验,即“平滑先验/局部不变性先验”,即所学习的函数不应在小区域内发生剧烈变化,从而维持网络内部权值的局部相对平滑。
AlexNet引入LRN层以提高模型对数据特征学习的泛化能力。网络中的LRN层,通过平滑约束前后层对应位置的权值,实现对当前层输出的平滑（图2）。
Fig. 2 Smooth processing of the output by LRN layer

图2 LRN层对输出的平滑处理

在网络中实现本层功能的数学模型如式（1） 所示：
b x , y i = a x , y i / ( k + a ∑ j = max ( 0 , i - n / 2 ) min ( N - 1 , i + n / 2 ) ( a x , y j ) 2 ) β （1）
式中：i表示第i个卷积核;（x, x）表示第i个核内的所在位置;n为同一位置所在的特征图数量;N为卷积核总数;参数k,a,b为超参数,默认值为k=2, a=1×（e-4）,b=0.75,本次试验采用默认值进行微调。

3 地表覆盖分类图斑评价模型构建及验证

3.1 试验环境及数据集采集

试验采用ubuntu操作系统下的caffe开源框架及CUDA-GPU加速方案,显卡采用的NVIDIA GeForce 1080Ti（11 G显存）进行GPU加速,其它主要硬件为英特尔Xeon E5-2620八核处理器、32G内存、PCI-E X8接口256 G固态硬盘等。
试验数据采集自地理国情普查地表覆盖分类成果数据,所采图斑来自1：10 000比例尺1 m分辨率光学遥感影像RGB全色波段数据,包含高质量的地表覆盖分类矢量层及分类字段信息,其中地表覆盖分类信息通过矢量层及套合的光学影像对网络进行先验知识输入。由于数据集构建所需图斑的识别和采集量大,本文选择程序自动采样方法,对分类矢量层进行掩膜处理,并进行全幅影像高重叠度分块（步进为20像元）及识别,对目标地类所占比例高于特定阈值（优于80%）的分块进行提取并标记分类信息。由于本次试验只采集一省的影像数据,数据集相对单一,为了保证训练的有效性,数据集经过筛选,筛选的标准为：特征明确、影像清晰、地类表征正确。
最终选取试验数据1875幅图斑进行训练和验证,所采图斑为地理国情普查中4个一级类耕地、林地、房屋和水体。数据集分为训练集（train）和验证集（test）,数据集采用统一的4个标签进行标记,0为耕地、1为林地、2为房屋、3为水体。训练集耕地图斑700个,林地图斑700个,房屋图斑100个,水体图斑120个。验证集耕地图斑100个、林地图斑100个、房屋图斑15个、水体图斑40个。本文数据集图斑样例如图3-6所示。
Fig. 3 Samples of arable land

图3 耕地图斑示例

Fig. 4 Samples of woodland

图4 林地图斑示例

Fig. 5 Samples of building area

图5 房屋图斑示例

Fig. 6 Samples of water area

图6 水体图斑示例

为便于输入和处理,数据集的所有影像图斑均统一为256像元×256像元进行提取,经过网络的规范化预处理转换为227×227×3的格式以进一步处理。

3.2 训练微调AlexNet模型

本试验采用AlexNet模型共8层,由超过600万个权值构成,前5层为卷积层,后3层为全链接层。
本试验通过林地、耕地、水体及房屋4类图斑对网络进行训练和评价。故最后端的全链接层具有4个输出。网络最后的优化目标是最大化平均的多元逻辑回归。训练过程输出结果如图7、8所示。
结合图7、8可以看出,AlexNet对本次试验采集的数据集学习效果较好,经过20个批次准确率已经到达较高水平,损失函数快速收敛。学习率在10个批次之后逐步下降,在20个批次后逐渐平稳,可判定网络对数据集进行了有效的学习,收敛速度快,准确率高。
Fig. 7 The accuracy and loss function curve of the training period

图7 AlexNet训练阶段正确率与损失函数曲线

Fig. 8 The learning rate curve of the training period

图8 AlexNet训练阶段学习率变化曲线

3.3 AlexNet模型图斑评价策略

经过微调阶段,深度卷积神经网络AlexNet已经具备了判读4类地表覆盖分类遥感影像图斑的能力,并可通过输出的图斑隶属度实现对图斑的量化评价。以一幅随机抽取的无标签房屋图斑为例,对网络正向计算过程进行分析（图9）。
Fig. 9 Initialization of the input data of AlexNet

图9 AlexNet网络输入数据初始化

第1个卷积层采用96个11×11×3的卷积核。在步进为4的情况下对224×224×3的图像进行滤波。即以11×11卷积模板在3个通道上以间隔4个像元的采样频率对影像进行卷积操作。在经过卷积层处理后,对数据进行ReLu激活及规范化变换后进行池化处理,并作为输出传递到下一层。
Fig. 10 The first regularization results

图10 第1次正则化结果

通过对比第1次卷积（图10、11）和第4次（图12）卷积的输出结果可以发现,AlexNet通过多次卷积池化获得了对输入遥感影像图斑的局部激活特征结构及均值进行了有效的离散化及分类。
Fig. 11 The first pooling results

图11 第1次池化结果

Fig. 12 The fourth convolution results

图12 第4次卷积结果

第一个全链接层是叠至在最后一个卷积和池化层后的,在最终池化层后,特征图的数量减少一半。最后一个全链接层输出融合了标签的softmax结果。本层节点4个,对应耕地、林地、水体、房屋4个分类。在经过全连接层处理后,网络比较准确的对4个地类的特征进行了区别,并准确地对输入的房屋图斑进行了分类（图13）。
经过训练的AlexNet网络准确的对输入的房屋图斑进行打分,由于本次试验数据集中房屋类图斑的特征较为简单明确,故最终网络对本幅样本的分类置信度高于99%。
Fig. 13 Final output

图13 最终输出结果

3.4 地表覆盖分类图斑评价试验结果及分析

试验结果表明,应用深度卷积神经网络模型,结合地表覆盖分类数据集,将先验知识输入到网络中,并利用已训练网络对图斑进行打分,可实现对图斑的地表覆盖分类进行自动评价。对于简单特征图斑,网络可以进行高置信度自动评价,而当图斑中含有未被充分训练的特征时,评价结果不稳定。试验后选取了4类12个验证图斑进行验证。具体结果见表1。
Tab. 1 Evaluation results(%)

表1 4类评价样本混淆矩阵（%）

分类数据 0-（耕地） 1-（林地） 2-（房屋） 3-（水体）
耕地 99.95 0.05 0.00 0.00
林地 5.67 43.59 50.16 0.58
房屋 0.00 0.59 99.41 0.00
水体 37.25 0.02 0.00 62.73
通过试验的验证测试可以看出,当图斑含有未被训练的新特征时,网络模型评价结果会出现明显的异常,如将稀疏林地判读为房屋,将水体边缘识别为耕地等（图斑见图14、15）,存在欠拟合现象。
Fig. 14 Dense woodland and sparse woodland

图14 稀疏林地与密集林地

Fig. 15 The center of water and the edge of water

图15 水体中心与水体边缘

深度卷积神经网络应用于地表覆盖分类评价中,光学遥感影像的先验知识是依靠完备的训练数据集输入网络的。本次试验中,训练阶段由于图斑样本较少,且图斑经过筛选,特征较为单一,故出现了稀疏的林地被识别为房屋,水体的边缘被识别为耕地等情况,其主要原因是网络对不同地类的特征学习不充分。通过分析试验中出现的欠拟合现象,我们认为构建一个具有更强可靠性的网络,需要数据集包含不同卫星平台、传感器在多时相下获得的不同数据,从而对不同硬件、不同时相、不同光照、不同大气条件下的影像进行充分学习。同时,由于地域性差异,对可能出现的模型参数过拟合问题,也应结合大量相关数据进行试验研究。
试验同时发现,对于整幅影像的分块逻辑及评价策略也有待进一步完善,如通过构建面对对象的图斑评价策略,对图斑内及图斑边缘进行取块自动评价,或进行多尺度分层评价,对不同尺度的图斑特征进行区别处理。未来将进一步探索符合现有评价体系的技术和方法。在此基础上通过研究自动评价打分与图斑分类精度间的关系,进一步完善深度学习方法在遥感影像地表覆盖方向的应用。

4 结论与展望

本文将AlexNet模型应用于高分辨率遥感影像地表覆盖分类评价中,使用地理国情普查1：10 000比例尺1 m分辨率RGB波段遥感影像及高精度地表覆盖分类矢量数据信息,提取了1870个一级类地表覆盖分类影像图斑,并通过微调AlexNet模型得到适用于地表覆盖分类图斑评价的深度卷积神经网络,随后通过测试数据验证了深度学习方法在遥感影像地表覆盖分类评价中的有效性。
结合本文试验结果分析,对今后开展地理国情检测地表覆盖分类评价中应用深度卷积神经网络方法提出以下思路：使用样本数量更大,完备性、代表性更好的数据集,进一步提高深度卷积神经网络用于地表覆盖分类图斑评价的准确性和可靠性。

The authors have declared that no competing interests exist.

参考文献

原文顺序 | 文献年度倒序 | 文中引用次数倒序
[1]
胡艳,袁超.多方法的地理国情地表覆盖分类研究[J].地理空间信息,2014(1):6-9

Hu Y, Yuan C.Study on national geographic census land cover classification with multi-methods[J]. Geospatial Information, 2014,1:6-9. ]

[2]
Luo B, Zhang L.Robust autodual morphological profiles for the classification of high-resolution satellite images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2014,52(2):1451-1462.Morphological profiles are widely used for the classification of high-resolution remote sensing images. Images are degraded by morphological operators with different sizes. Profiles, such as the variation of intensities of the pixels, are extracted during the degradation. In this paper, two issues with morphological profiles are addressed. First, we propose using the topographic map of the image, which is autodual and does not require any structural element, for the degradation of the image. Second, we propose extracting other profiles that are more robust than the intensity variations. Classification experiments are carried out on two IKONOS remote sensing images with a 1-m resolution and one Ple虂iades 1-A image with a 2-m resolution. The efficiencies of the new profiles are validated by the experimental results.

DOI

[3]
Hinton G E, Salakhutdinov R R.Reducing the dimensionality of data with neural networks[J]. Science, 2006,313(5786):504-507.

DOI

[4]
Krizhevsky A, Sutskever I, Hinton G E.ImageNet classification with deep convolutional neural networks[C]// International Conference on Neural Information Processing Systems. Curran Associates Inc, 2012:1097-1105.

[5]
陈军,陈晋,宫鹏,等.全球地表覆盖高分辨率遥感制图[J].地理信息世界,2011,9(2):12-14.全球地表覆盖分布及变化是气候变化研究、生态环境评估、地理国情监测、宏观调控分析等不可或缺的重要基础信息.国际上现有全球五套地表覆盖数据产品的空间分辨率为1 km或300 m,数据精度、分类体系、时空分辨率等均存在不足.为了满足全球变化研究与地球模式模拟的需求,应该研制具有较高时空分辨率、更符合全球变化需要、精度较好的全球地表覆盖数据产品.本文简要介绍了全球地表覆盖遥感制图的情况和数据产品的不足,讨论了对新一代地表覆盖数据产品的需求,介绍了我国研制全球30 m地表覆盖数据产品的863重点项目.

DOI

[ Chen J, Chen J, Gong P, et al.Remote sensing mapping of global land cover with high resolution[J]. Geomatics World, 2011,9(2):12-14. ]

[6]
Li X F, Ju W M, Chen S.Influence of land cover data on regional forest leaf area index inversion[J]. Journal of Remote Sensing, 2010,14(5):974-989.In this study, six different land cover datasets were employed in conjunction with MODIS 1km reflectance data to inverse LAI of forests using an algorithm based on the 4-scale geometrical optical model in Jian City, Jiangxi Province, China. Land cover datasets used in this study include five global land cover datasets (Three were produced by the United States Geo-logical Survey (USGS), University of Maryland (UMD), and Boston University (BU), respectively. Two were constructed in Europe.) and a regional land cover map produced using Landsat TM images. For assessing the impact of land cover on the in-version of LAI, LAI images inversely produced with different land cover datasets were compared with LAI data sampled from a 30 m LAI map at 1 km and 4 km scales, respectively. The 30 m LAI map was produced with TM reflectance images and ground measurements of LAI. The results show that the land cover datasets of TM and GLOBCOVER which was created by European Space Agency are the best for the inversion of LAI in this study area. At 1 km scale, the R2 values of LAI inversed using TM and GLOBCOVER land cover datasets with TM LAI estimated using an statistical model are 0.44 and 0.40, respectively. At 4 km scale, these R2 values increase to 0.57 and 0.54. The MODIS land cover data of BU is the third best data for the inversion of LAI, the R2 values between LAI inversed using this land cover dataset and TM LAI are 0.38 and 0.51 at 1 km and 4 km scales, respectively. The land cover datasets of UMD and European GLC2000 resulted in large discrepancies between inversed LAI and TM LAI. The averages of LAI inversed using these two land cover datasets are about 20% lower than TM LAI at 1 km and 4 km scales. Sensitivity analysis shows that inversed LAI is sensitive to clumping index. This study proved that reliable land cover data is required for improving the accuracy of inversed LAI at regional/global scales.

DOI

[7]
曹云刚. 多时相ASAR数据的地表覆盖分类研究[J].测绘科学,2007,32(5):103-105.本文选择了位于念青唐古拉山脉 西段,覆盖范围大约100×100km2的区域,使用四个不同时期内的ASAR图像数据进行地表覆盖分类的研究。研究结果表明,虽然同种类型的地物在同一 景雷达图像上的后向散射系数存在一定的差异,但是其后向散射系数随时间的变化规律却是一致的。根据地物后向散射系数的这种时相特征,我们对研究区的地表覆 盖进行了分类,结果显示使用该方法能有效地区分草原、草甸、裸岩、水体、终年积雪等。

DOI

[ Cao Y G.Study on land classification using multi-temporal ASAR data[J]. Science of Surveying and Mapping, 2007,32(5):103-105. ]

[8]
谢宏全,王圣尧,刘善磊,等.面向地理国情普查的地表覆盖分类方法研究[J].测绘通报,2014,s2:245-247.立足于地理国情普查项目,探索适用于试验区的地表覆盖分类准则,建立地理国情普查分类方案,并以此方案为基础对现有测绘成果数据进行有效规整,得到具有参考价值的矢量参考数据,利用其辅助地表覆盖分类,得到完整的可进行统计分析的成果数据。

[ Xie H Q,Wang S Y, Liu S L, et al.Study on land cover classification for national geographic census[J]. Bulletin of Surveying and Mapping, 2014,s2:245-247. ]

[9]
汪权方,李家永,陈百明.基于地表覆盖物光谱特征的土地覆被分类系统——以鄱阳湖流域为例[J].地理学报,2006,61(4):359-368.

[ Wang Q F, Li J Y, Chen B M.Land cover classification system based on spectrum for Poyang lake basin[J]. Acta Geographica Sinica, 2006,61(4):359-368. ]

[10]
翁中银,何政伟,于欢.基于决策树分类的地表覆盖遥感信息提取[J].地理空间信息,2012,10(2):110-112.地处西南的渝北地区地表覆盖类型复杂、土地利用多元化,仅依赖于光谱特征的传统遥感信息提取方法难以获得较高的分类精度。利用决策树分类技术对渝北地区的TM遥感影像进行分类,除光谱信息外还结合地质、NDVI、PCI等多源数据进行实验。结果表明,总精度和Kappa系数分别为88.42%和0.854 7,较传统的监督分类和仅依赖于光谱特征的决策树分类方法有较大提高,这也表明基于多源数据的决策树分类技术对地表覆盖复杂地区的遥感影像分类比较适用,是遥感信息提取的一种有效手段。

DOI

[ Weng Z Y, He Z W, Yu H.Land cover information extraction based on decision tree[J]. Geospatial Information, 2012,10(2):110-112. ]

[11]
高志宏,周旭,程滔.地理国情普查中容易混分地表覆盖类型定量统计与分析[J].测绘通报,2015(6):32-34.地理国情普查是新形势下测绘地理信息领域一项新的重大工作,开展地理国情普查可以全面获取地理国情信息,为开展常态化地理国情监测奠定基础。地表覆盖分类数据是地理国情普查的重要组成部分和成果之一。本文针对目前地理国情普查工作实际数据采集过程中存在的混分问题,广泛收集了地表覆盖分类方面的典型样例和问题,从一级类内部混淆和跨一级类混淆两个方面统计和分析了容易混分的类型,探讨了容易混淆地表覆盖类型形成的原因和规律,并给出了提高分类精度的建议。本文研究结果可直接服务于地理国情普查工作的开展,为提高地表覆盖分类精度提供科学参考。

DOI

[ Gao Z H, Zhou X, Cheng T.Statistical analysis of the confusing land cover types in China Geography Census[J]. Bulletin of Surveying and Mapping, 2015,6:32-34. ]

[12]
陈利军,陈军,廖安平,等.30m全球地表覆盖遥感分类方法初探[J].测绘通报,2012(s1):350-353.全球地表覆盖分布及变化是气候变化研究、生态环境评估、地理国情监测、宏观调控分析等不可或缺的重要基础信息。由于全球范围内地表覆盖复杂多样,光谱差异大,单一的分类算法缺乏通用性,提出层次分类策略并进行试验分析。结果表明,层次分类策略要优于SVM、J48、RF和贝叶斯等分类方法。通过全球范围10个试验区的试验,除受大面积云影响的马来西亚试验区精度为65%之外,其他区域的总体分类精度均达到70%以上,具有较强的鲁棒性和通用性,可以用于30 m全球地表覆盖产品的研制。

[ Chen L J, Chen J, Liao A P.Preliminary discussion on global land cover classification of 30m-resolution using remote sensing images[J]. Bulletin of Surveying and Mapping, 2012,s1:350-353. ]

[13]
Cheng G, Zhou P, Han J.Learning rotation-invariant convolutional neural networks for object detection in VHR optical remote sensing images[J]. IEEE Transactions on Geoscience & Remote Sensing, 2016,54(12):7405-7415.Object detection in very high resolution optical remote sensing images is a fundamental problem faced for remote sensing image analysis. Due to the advances of powerful feature representations, machine-learning-based object detection is receiving increasing attention. Although numerous feature representations exist, most of them are handcrafted or shallow-learning-based features. As the object detection task becomes more challenging, their description capability becomes limited or even impoverished. More recently, deep learning algorithms, especially convolutional neural networks (CNNs), have shown their much stronger feature representation power in computer vision. Despite the progress made in nature scene images, it is problematic to directly use the CNN feature for object detection in optical remote sensing images because it is difficult to effectively deal with the problem of object rotation variations. To address this problem, this paper proposes a novel and effective approach to learn a rotation-invariant CNN (RICNN) model for advancing the performance of object detection, which is achieved by introducing and learning a new rotation-invariant layer on the basis of the existing CNN architectures. However, different from the training of traditional CNN models that only optimizes the multinomial logistic regression objective, our RICNN model is trained by optimizing a new objective function via imposing a regularization constraint, which explicitly enforces the feature representations of the training samples before and after rotating to be mapped close to each other, hence achieving rotation invariance. To facilitate training, we first train the rotation-invariant layer and then domain-specifically fine-tune the whole RICNN network to further boost the performance. Comprehensive evaluations on a publicly available ten-class object detection data set demonstrate the effectiveness of the proposed method.

DOI

[14]
Hu F, Xia G S, Hu J, et al.Transferring deep convolutional neural networks for the scene classification of high-resolution remote sensing imagery[J]. Remote Sensing, 2015,7(11):14680-14707.Learning efficient image representations is at the core of the scene classification task of remote sensing imagery. The existing methods for solving the scene classification task, based on either feature coding approaches with low-level hand-engineered features or unsupervised feature learning, can only generate mid-level image features with limited representative ability, which essentially prevents them from achieving better performance. Recently, the deep convolutional neural networks (CNNs), which are hierarchical architectures trained on large-scale datasets, have shown astounding performance in object recognition and detection. However, it is still not clear how to use these deep convolutional neural networks for high-resolution remote sensing (HRRS) scene classification. In this paper, we investigate how to transfer features from these successfully pre-trained CNNs for HRRS scene classification. We propose two scenarios for generating image features via extracting CNN features from different layers. In the first scenario, the activation vectors extracted from fully-connected layers are regarded as the final image features; in the second scenario, we extract dense features from the last convolutional layer at multiple scales and then encode the dense features into global image features through commonly used feature coding approaches. Extensive experiments on two public scene classification datasets demonstrate that the image features obtained by the two proposed scenarios, even with a simple linear classifier, can result in remarkable performance and improve the state-of-the-art by a significant margin. The results reveal that the features from pre-trained CNNs generalize well to HRRS datasets and are more expressive than the low- and mid-level features. Moreover, we tentatively combine features extracted from different CNN models for better performance.

DOI

[15]
Yu D, Deng L.Deep learning and its applications to signal and information processing[J]. IEEE Signal Processing Magazine, 2011,28(1):145-154.The purpose of this article is to introduce the readers to the emerging technologies enabled by deep learning and to review the research work conducted in this area that is of direct relevance to signal processing. We also point out, in our view, the future research directions that may attract interests of and require efforts from more signal processing researchers and practitioners in this emerging area for advancing signal and information processing technology and applications.

DOI

[16]
Jones N.The learning machines[J]. Nature, 2014,505(7842):146-148.Using massive amounts of data to recognize photos and speech, deep-learning computers are taking a big step towards true artificial intelligence.

DOI PMID

[17]
余凯,贾磊,陈雨强,等.深度学习的昨天、今天和明天[J].计算机研究与发展,2013,50(9):1799-1804.机器学习是人工智能领域的一个重要学科.自从20世纪80年代以来，机器学习在算法、理论和应用等方面都获得巨大成功.2006年以来，机器学习领域中一个叫“深度学习”的课题开始受到学术界广泛关注，到今天已经成为互联网大数据和人工智能的一个热潮.深度学习通过建立类似于人脑的分层模型结构，对输入数据逐级提取从底层到高层的特征，从而能很好地建立从底层信号到高层语义的映射关系.近年来，谷歌、微软、IBM、百度等拥有大数据的高科技公司相继投入大量资源进行深度学习技术研发，在语音、图像、自然语言、在线广告等领域取得显著进展.从对实际应用的贡献来说，深度学习可能是机器学习领域最近这十年来最成功的研究方向.将对深度学习发展的过去和现在做一个全景式的介绍，并讨论深度学习所面临的挑战，以及将来的可能方向.

[ Yu K, Jia L, Chen Y Q.Deep learning: Yesterday, today and tomorrow[J]. Journal of Computer Research and Development,2013,50(9):1799-1804. ]

[18]
Qin Q, Hobert J P. On the data augmentation algorithm for bayesian multivariate linear regression with non-gaussian errors[J].Statistics, 2015, arXiv:1512.01734v1.Let $\pi$ denote the intractable posterior density that results when the likelihood from a multivariate linear regression model with errors from a scale mixture of normals is combined with the standard non-informative prior. There is a simple data augmentation algorithm (based on latent data from the mixing density) that can be used to explore $\pi$. Hobert et al. (2015) [] recently performed a convergence rate analysis of the Markov chain underlying this MCMC algorithm in the special case where the regression model is univariate. These authors provide simple sufficient conditions (on the mixing density) for geometric ergodicity of the Markov chain. In this note, we extend Hobert et al.'s (2015) result to the multivariate case.

[19]
周星宇,张继贤,高绵新,等.高分辨率遥感影像下沿海地区地表覆盖信息的提取[J].测绘通报,2017(2):19-24.沿海地区地表覆盖信息是全国地理国情普查的重要内容,遥感影像分类技术为沿海地区地表覆盖信息提供了一种重要方法。本文基于GF-1高分辨率遥感影像,建立了沿海地区地表覆盖分类系统,采用中国测绘科学研究院自主研发的面向对象GLC决策树分类方法和软件进行了地表覆盖分类。通过对某试验区进行分类试验,并结合该区地表覆盖标准分类图进行精度评价,验证了基于高分辨率影像,面向对象GLC决策树分类方法在沿海地区地表覆盖信息提取上的有效性及优越性,其总体分类精度和Kappa系数分别为87.201 8%、0.840 6,均高于SVM分类法。最后提出基于高分辨率遥感影像的沿海地区地表覆盖信息提取流程。

DOI

[ Zhou X Y, Zhang J X, Gao M X.Land cover information extraction based on high-resolution remote sensing images in coastal areas[J]. Bulletin of Surveying and Mapping, 2017,2:19-24. ]

[20]
高常鑫,桑农.基于深度学习的高分辨率遥感影像目标检测[J].测绘通报,2014,s1:108-111.传统的目标检测识别方法难以适应海量高分辨率遥感影像数据,需要寻求一种能够自动从海量影像数据中学习最有效特征的方法,充分复挖掘数据之间的关联。本文针对海量高分辨率遥感影像数据下典型目标的检测识别,提出一种分层的深度学习模型,通过设定特定意义的分层方法建立目标语义表征及上下文约束表征,以实现高精度目标检测。通过对高分遥感影像目标检测的试验,证明了该方法的有效性。

[ Gao C X, Sang N.Deep learning for object detection in remote sensing images[J]. Bulletin of Surveying and Mapping, 2014,s1:108-111. ]

[21]
刘大伟,韩玲,韩晓勇.基于深度学习的高分辨率遥感影像分类研究[J].光学学报,2016(4):298-306.针对高空间分辨率遥感影像的分类问题,提出了基于深度学习的分类方法。该方法通过非下采样轮廓波变换计算影像的纹理特征,利用深度学习的常用模型—深度信念网络(DBN)对高分辨率遥感影像进行了基于光谱-纹理特征的分类,并与基于单源光谱信息的DBN分类方法、支持向量机(SVM)分类方法、传统神经网络(NN)分类方法进行了比较分析。研究结果表明:相对于单源光谱信息,利用影像的光谱-纹理特征能够有效提高高分辨率遥感影像的分类精度;相对于SVM、NN等分类方法,DBN能够更加准确地挖掘高分辨率遥感影像的空间分布规律,提高分类的准确度。

DOI

[ Liu D W, Han L, Han X Y.Study on remote sensing image classification of high spatial resolution based on deep learning[J]. Acta Optica Sinica, 2016,4:298-306. ]

[22]
张俊. 基于AlexNet融合特征的图像检索研究[D].重庆:重庆邮电大学,2016.

[ Zhang J.Research on image retrieval based on fusion feature of AlexNet[D]. Chongqing: Chongqing University of Posts and Telecommunications, 2016. ]

Options
文章导航

/

〈 〉