摘要
为提高影像云识别精度,提出一种多尺度膨胀卷积深层神经网络云识别方法。结合卫星影像特征,设计云识别卷积神经网络结构,该结构包含深层特征编码模块、局部多尺度膨胀感知模块以及云区预测解码模块。首先,编码模块中通过基础卷积层获取深度特征;其次,联合多尺度膨胀卷积和池化层共同感知,每层操作连接非线性函数,以提升网络模型的表达能力;最后,云区预测解码模块中融合对应编码模块的特征,再利用L1正则化上采样算法实现端对端的像素级云识别结果。选用典型云遮挡区域影像进行云识别实验,并与Otsu算法和FCN-8S算法进行对比。结果表明,本文所提算法的检测精度较高,Kappa系数显著提升。
To improve the accuracy of cloud detection, we propose a multi-scale dilation convolutional neural network method. Combining with the characteristic of satellite images, we design the deep convolution network structure, which includes a deep-feature coding module, a local dilation perception module, and a cloud-dense decoding module. First, the deep-features of cloud are obtained by the basic convolutional layer in conjunction with the coding module. Second, multi-scale dilation convolution layers jointed with pooling layers are used to perceive corporately. A nonlinear function is employed in each block to improve the effectiveness of network model expression. Finally, the cloud-dense decoding module integrate the features corresponding to those included in the coding module and then utilize the L1 regularization upsampling algorithm to accomplish the end-to-end pixel-level cloud detection task. Cloud detection experiments are performed in the typical cloud mask areas;the results are compared with those of the Otsu algorithm and the FCN-8 S method. The results indicate that the accuracy of proposed method is higher and the Kappa coefficient is effectively improved.
作者
高琳
宋伟东
谭海
刘阳
Gao Lin;Song Weidong;Tan Hai;Liu Yang(School of Mapping and Geographical Science,Liaoning Technical University,Fuxin,Liaoning 123000,China;Satellite Surveying and Mapping Application Center,National Administration of Surveying,Mapping and Geoinformation,Beijing 100048,China)
出处
《光学学报》
EI
CAS
CSCD
北大核心
2019年第1期299-307,共9页
Acta Optica Sinica
基金
国家自然科学基金青年基金(61601213)
中国博士后科学基金(2017M611252)
辽宁省公益研究基金计划(20170003)
关键词
遥感
神经网络
膨胀卷积
云识别
资源三号卫星影像
全卷积网络
remote sensing
neural network
dilation convolution
cloud detection
ZY-3 satellite imagery
fully convolution network
