摘要
针对传统手工提取特征方法需要专业领域知识,提取高质量特征困难的问题,将深度迁移学习技术引入到高分影像树种分类中,提出一种结合面向对象和深度特征的高分影像树种分类方法。为了获取树种的精确边界,该方法首先利用多尺度分割技术分割整幅遥感影像,并选择训练样本作为深度卷积神经网络的输入。为了避免样本数量少导致过拟合问题,采用迁移学习方法,使用ImageNet上训练的VGG16模型参数初始化深度卷积神经网络,并利用全局平局池化压缩参数,在网络最后添加1024个节点的全连接层和7个节点的Softmax分类器,利用反向传播和Adam优化算法训练网络。最后分类整幅遥感影像,生成树种专题地图。以安徽省滁州市的皇甫山国家森林公园为研究区,QuickBird高分影像作为数据源,采用本文方法进行树种分类。试验结果表明,本文方法树种分类总体精度和Kappa系数分别为78.98%和0.685 0,在保证树种精度的同时实现了端到端的树种分类。
A tree species classification of high resolution image combining with object-oriented and deep feature is proposed to overcome the problem that traditional manual extraction features need professional knowledge and difficult to extract high quality features.In order to obtain the precise boundary of tree species,the method firstly uses multiscale segmentation technology to segment the whole remote sensing image,and selects the training samples as the input of the deep convolution neural network.In order to avoid over-fitting caused by a small number of samples,transfer learning method is used to initialize the deep convolution neural network with the parameters of VGG16 model trained on Image Net. Using global average pooling compression parameters,a 1024 nodes fully connected layer and 7 nodes Softmax classifier are added at the end of the network. The network is trained by back propagation and Adam optimization algorithm.Finally,the whole remote sensing image is classified and the tree thematic map is generated. The test site is located in the Huangfu Mountain National Forest Park in Anhui province. Quick Bird high resolution image is the data source. The results show that the overall accuracy and Kappa coefficient of this method are 78.98% and 0.6850 respectively,which can ensure the accuracy of tree species and achieve end-to-end tree species classification.
引文
[1]ZHANG Z,KAZAKOVA A,MOSKAL L,et al.Objectbased tree species classification in urban ecosystems using Li DAR and hyperspectral data[J].Forests,2016,7(6):122.
[2]WASER L T,GINZLER C,KUECHLER M,et al.Semiautomatic classification of tree species in different forest ecosystems by spectral and geometric variables derived from airborne digital sensor(ADS40)and RC30 data[J].Remote Sensing of Environment,2011,115(1):76-85.
[3]任冲,鞠洪波,张怀清,等.多源数据林地类型的精细分类方法[J].林业科学,2016,52(6):54-65.
[4]LOUARN M L,CLERGEAU P,BRICHE E,et al.“Kill two birds with one stone”:urban tree species classification using bi-temporal Pléiades images to study nesting preferences of an invasive bird[J].Remote Sensing,2017,9(9):916.
[5]王妮,彭世揆,李明诗.基于树种分类的高分辨率遥感数据纹理特征分析[J].浙江农林大学学报,2012,29(2):210-217.
[6]LECUN Y,BENGIO Y,HINTON G.Deep learning[J].Nature,2015,521:436-444.
[7]LECUN Y,BOTTOU L,BENGIO Y,et al.Gradient-based learning applied to document recognition[J].Proceedings of the IEEE,1998,86(11):2278-2324.
[8]龚健雅,季顺平.摄影测量与深度学习[J].测绘学报,2018,47(6):693-704.
[9]ZHANG F,DU B,ZHANG L.Scene classification via a gradient boosting random convolutional network framework[J].IEEE Transactions on Geoscience and Remote Sensing,2016,54(3):1793-1802.
[10]HUAN Y U,ZHANG S Q,KONG B,et al.Optimal segmentation scale selection for object-oriented remote sensing image classification[J].Journal of Image and Graphics,2010,15(2):352-360.
[11]孙坤,鲁铁定.顾及多尺度分割参数的FNEA面向对象分类[J].测绘通报,2018(3):43-48.
[12]DENG J,DONG W,SOCHER R,et al.Image Net:a large-scale hierarchical image database[C]∥IEEEConference on Computer Vision and Pattern Recognition.[S.l.]:IEEE,2009:248-255.
[13]LIN M,CHEN Q,YAN S.Network in network[C]∥International Conference on Learning Representations(ICLR).[S.l.]:Cornell University,2014.
[14]KINGMA D,BA J.ADAM:a method for stochastic optimization[C]∥International Conference on Learning Representations(ICLR).[S.l.]:Cornell University,2015.
[15]SIMONYAN K,ZISSERMAN A.Very deep convolutional networks for large-scale image recognition[C]∥ar Xiv preprint ar Xiv.[S.l.]:University of Oxford,2014.