一种面向土地覆盖分类的卷积神经网络模型
|
摘要
针对卷积神经网络在土地覆盖分类中卷积层尺寸过大问题,研究了一种适用于土地覆盖分类像素级分类的土地覆盖分类模型。以陆地卫星中分辨率影像和快鸟高分辨率影像为实验数据,对比了不同样本尺寸大小和不同分辨率影像对模型分类结果的影响,并与传统的基于光谱特征以及光谱加纹理特征的方法进行对比分析。结果表明,陆地卫星中分辨率影像最佳训练样本尺寸大小为5像素×5像素,过大的样本尺寸在分类结果上会产生较强的滤波效应,减少了分类结果的细节信息,而过小的样本尺寸由于包含信息太少,导致误分严重;陆地卫星中分辨率影像分类结果细碎图斑少,一致性好,可有效减少分类后处理环节;快鸟高分辨率影像最佳训练样本尺寸大小为7像素×7像素,相比陆地卫星中分辨率影像滤波效应得到缓解,细节信息保存更好,精度提升更大,对训练样本尺寸选择更为鲁棒,在总体分类精度上优于基于光谱特征和光谱加纹理特征的分类方法,可以很好地应用于土地覆盖分类。
Aiming at the problem that the convolution layer size of convolution neural network is too large in land cover classification,apixel-level classification model Land Cover Convolutional Neural Network(LCNet)for land cover classification is studied in this paper.TM medium resolution image and QuickBird high resolution image are utilized as the experimental data.The proposed LCNet is compared with the traditional methods based on spectral features and spectral textures.Experimental results show that the training sample size of 5×5 pixels is the best for LCNet in TM medium resolution image.Bigger sample size will yield obvious filtering effect to classification results and reduce the detail information;however,smaller size might cause serious misclassification because of containing too fewer pixel information.Besides,the classification results of the proposed LCNet show good consistency with less broken patches,effectively reducing the post-classification process.For QuickBird high resolution image,the training sample size of 7×7 pixels is the best for LCNet;different from the TM image,the filtering effect is relieved;the detail information is better preserved and the precision is significantly increased,since LCNet becomes more robust to the size of training samples.The classification accuracy of LCNet is higher than those of the traditional methods based on spectral features and spectral+texture features,which indicates that the proposed method can be well applied to deal with land cover classification tasks.
Aiming at the problem that the convolution layer size of convolution neural network is too large in land cover classification,apixel-level classification model Land Cover Convolutional Neural Network(LCNet)for land cover classification is studied in this paper.TM medium resolution image and QuickBird high resolution image are utilized as the experimental data.The proposed LCNet is compared with the traditional methods based on spectral features and spectral textures.Experimental results show that the training sample size of 5×5 pixels is the best for LCNet in TM medium resolution image.Bigger sample size will yield obvious filtering effect to classification results and reduce the detail information;however,smaller size might cause serious misclassification because of containing too fewer pixel information.Besides,the classification results of the proposed LCNet show good consistency with less broken patches,effectively reducing the post-classification process.For QuickBird high resolution image,the training sample size of 7×7 pixels is the best for LCNet;different from the TM image,the filtering effect is relieved;the detail information is better preserved and the precision is significantly increased,since LCNet becomes more robust to the size of training samples.The classification accuracy of LCNet is higher than those of the traditional methods based on spectral features and spectral+texture features,which indicates that the proposed method can be well applied to deal with land cover classification tasks.
引文
[1]张锦水,何春阳,潘耀忠,等.基于SVM的多源信息复合的高空间分辨率遥感数据分类研究[J].遥感学报,2006,10(1):49-57.
[2]温小乐,姜兴强,徐涵秋,等.基于TM影像的长乐市植被覆盖度变化研究[J].遥感信息,2013,28(6):56-62.
[3]梁继,王建,王建华,等.基于光谱角分类器遥感影像的自动分类和精度分析研究[J].遥感技术与应用,2002,17(6):299-303.
[4]吴小君,吴鹏海,刘紫涵,等.利用时序Landsat数据的杞麓湖面积年变化时空分析[J].遥感信息,2016,31(4):89-94.
[5]GONG P,WANG J,YU L,et al.Finer resolution observation and monitoring of global land cover:first mapping results with Landsat TM and ETM+data[J].International Journal of Remote Sensing,2013,34(7):2607-2654.
[6]黄昕,张良培,李平湘,等.融合形状和光谱的高空间分辨率遥感影像分类[J].遥感学报2007,11(2):193-200.
[7]郑淑丹,郑江华,石明辉,等.基于分形和灰度共生矩阵纹理特征的种植型药用植物遥感分类[J].遥感学报,2014,18(4):868-886.
[8]YANG W,YIN X G.Learning high-level features for satellite image classification with limited labeled samples[J].IEEETrans.Geosci.Remote Sens.2015,53:4472-4482.
[9]ZHANG X,WU B,LI Q.Research on classification of high-resolution remote sensing image[J].Proc.SPIE 6786,2007:1023-1031.
[10]VEDALDI A,MATCONVENT K.Convolutional neural networks for matlab[J].Proc.Annual ACM Conference on Multimedia Conference(ACMMM).ACM,2015:689-692.
[11]WANG X C,LI M.Classification of high spatial resolution remote sensing image using SVM and local spatial statistics Getis-Ord Gi[J].Proc.SPIE 8006,2011:1146-1152.
[12]HELLAH K F M.Texture classification using dominant neighborhood structure[J].IEEE Transactions on Image Processing,2011:20(11):3270-3279.
[13]NOGUEIRA K,PENATTI O A B.Towards better exploiting convolutional neural networks for remote sensing scene classification[J].Pattern Recognition,2016,61:539-556.
[14]KRIZHEVSKY A,SUTSKEVER I,HINTON G E.ImageNet classification with deep convolutional neural networks[J].Proc.Adv.Neural Inf.Process.Syst.,2012:1097-1105.
[15]SIMONYAN K,ZISSERMAN A.Very deep convolutional networks for large-scale image recognition[J].ArXiv:1409.1556.2014:1049-1556.
[16]HE K,ZHANG X,REN S Q,et al.Deep residual learning for image recognition[J].ArXiv:1512.03385.2015:1512-1524.
[17]HU F.Transferring deep convolutional neural networks for the scene classification of high-resolution remote sensing imagery[J].Remote Sensing,2015,7(11):14680-14707.
[18]LANGKVIST M,KISELEV A.Classification and segmentation of satellite orthoimagery using convolutional neural networks[J].Remote Sensing,2016,8(4):329.
[19]张伟,郑柯.深度卷积神经网络特征提取用于土地覆盖分类初探[J].中国图像图形学报,2017,22(8):1144-1153.
[2]温小乐,姜兴强,徐涵秋,等.基于TM影像的长乐市植被覆盖度变化研究[J].遥感信息,2013,28(6):56-62.
[3]梁继,王建,王建华,等.基于光谱角分类器遥感影像的自动分类和精度分析研究[J].遥感技术与应用,2002,17(6):299-303.
[4]吴小君,吴鹏海,刘紫涵,等.利用时序Landsat数据的杞麓湖面积年变化时空分析[J].遥感信息,2016,31(4):89-94.
[5]GONG P,WANG J,YU L,et al.Finer resolution observation and monitoring of global land cover:first mapping results with Landsat TM and ETM+data[J].International Journal of Remote Sensing,2013,34(7):2607-2654.
[6]黄昕,张良培,李平湘,等.融合形状和光谱的高空间分辨率遥感影像分类[J].遥感学报2007,11(2):193-200.
[7]郑淑丹,郑江华,石明辉,等.基于分形和灰度共生矩阵纹理特征的种植型药用植物遥感分类[J].遥感学报,2014,18(4):868-886.
[8]YANG W,YIN X G.Learning high-level features for satellite image classification with limited labeled samples[J].IEEETrans.Geosci.Remote Sens.2015,53:4472-4482.
[9]ZHANG X,WU B,LI Q.Research on classification of high-resolution remote sensing image[J].Proc.SPIE 6786,2007:1023-1031.
[10]VEDALDI A,MATCONVENT K.Convolutional neural networks for matlab[J].Proc.Annual ACM Conference on Multimedia Conference(ACMMM).ACM,2015:689-692.
[11]WANG X C,LI M.Classification of high spatial resolution remote sensing image using SVM and local spatial statistics Getis-Ord Gi[J].Proc.SPIE 8006,2011:1146-1152.
[12]HELLAH K F M.Texture classification using dominant neighborhood structure[J].IEEE Transactions on Image Processing,2011:20(11):3270-3279.
[13]NOGUEIRA K,PENATTI O A B.Towards better exploiting convolutional neural networks for remote sensing scene classification[J].Pattern Recognition,2016,61:539-556.
[14]KRIZHEVSKY A,SUTSKEVER I,HINTON G E.ImageNet classification with deep convolutional neural networks[J].Proc.Adv.Neural Inf.Process.Syst.,2012:1097-1105.
[15]SIMONYAN K,ZISSERMAN A.Very deep convolutional networks for large-scale image recognition[J].ArXiv:1409.1556.2014:1049-1556.
[16]HE K,ZHANG X,REN S Q,et al.Deep residual learning for image recognition[J].ArXiv:1512.03385.2015:1512-1524.
[17]HU F.Transferring deep convolutional neural networks for the scene classification of high-resolution remote sensing imagery[J].Remote Sensing,2015,7(11):14680-14707.
[18]LANGKVIST M,KISELEV A.Classification and segmentation of satellite orthoimagery using convolutional neural networks[J].Remote Sensing,2016,8(4):329.
[19]张伟,郑柯.深度卷积神经网络特征提取用于土地覆盖分类初探[J].中国图像图形学报,2017,22(8):1144-1153.