摘要
云检测是多光谱卫星云图分析的前提。传统云检测方法不能很好地对多光谱卫星云图进行特征表示,导致了云检测不是很准确。卷积神经网络虽然能有效地提取特征,但训练时会产生梯度扩散,训练效率低,优化困难等问题。针对这些问题,提出多维加权密集连接卷积神经网络模型实现对多光谱卫星云图的云检测。跨层连接能够实现网络中所有层之间的信息流,从而减少训练过程中的梯度消失导致收敛困难的问题。特征图之间连接的权值不同使得网络能够更高效地利用特征信息。通过实验结果对比,该模型可以很好地提取云图特征,提高多光谱云图检测的准确率,具有更好的泛化性能和优化效率。
Cloud detection is the prerequisite for multi-spectral satellite cloud image analysis. The traditional cloud detection method cannot well characterize multispectral satellite cloud images, which leads to a low accuracy in cloud detection. Although deep convolutional neural network can extract features effectively, it will have problems such as gradient diffusion, low training efficiency, hard optimization, and poor generalization. In order to solve these problems, a multidimensional weighted densely connected convolutional neural network is proposed to realize cloud detection of multi-spectral satellite imagery. Cross layer connection can realize the information flow between all layers in the network, thus reducing the difficulty of convergence caused by the disappearance of the gradient in the training process. The different weights of the connection between the feature graphs make the network more efficient use of the feature information. The simulation shows that the proposed model can extract the features of cloud images well and improve the accuracy of multi-spectral cloud image detection, it has better generalization performance and higher optimization efficiency.
引文
[1]宋小宁,赵英时.图象的云检测及分析[J].中国图象图形学报,2003,8(9):1079-1083.
[2]Paola J D,Schowengerdt R A.A review and analysis of backpropagation neural networks for classification of remotely-sensed multi-spectral imagery[J].International Journal of Remote Sensing,1993,16(16):3033-3058.
[3]孙劲光,赵欣.一种改进近邻传播聚类的图像分割算法[J].计算机工程与应用,2017,53(6):178-182.
[4]王根,张华,杨寅.高光谱大气红外探测器资料质量控制研究进展[J].地球科学进展,2017,32(2):139-150.
[5]李超炜,邓新蒲,赵昊宸.基于小波分析的遥感影像薄云去除算法研究[J].数字技术与应用,2017(6):137-139.
[6]Bai T,Li D,Sun K,et al.Cloud detection for high-resolution satellite imagry using machine learning and multi-feature fusion[J].Remote Sensing,2016,8(9):715.
[7]雒培磊,李国庆,曾怡.一种改进的基于深度学习的遥感影像拼接方法[J].计算机工程与应用,2017,53(20):180-186.
[8]Simonyan K,Zisserman A.Very deep convolutional networks for large-scale image recognition[J].Computer Science,2014.
[9]He K,Zhang X,Ren S,et al.Deep residual learning for image recognition[C]//CVPR,2016.
[10]周俊宇,赵艳明.卷积神经网络在图像分类和目标检测应用综述[J].计算机工程与应用,2017,53(13):34-41.
[11]Huang G,Liu Z,Maaten L V D,et al.Densely connected convolutional networks[C]//IEEE Conference on Computer Vision and Pattern Recognition,2017:2261-2269.
[12]李卫利,李伟杰,李丽苗.基于最速下降法和遗传算法求解函数的优化问题[J].科学与财富,2016,8(1).
[13]范秀娟.基于自适应误差反向传播信号调整的BP网络研究与应用[J].计算机应用与软件,2010,27(8):117-119.
[14]王同科,佘海艳,刘志方.分数阶光滑函数线性和二次插值公式余项估计[J].计算数学,2014,36(4):393-406.
[15]李海亮,汪秀华,罗红霞,等.基于HJ-1A/1BCCD数据的海南岛水体提取研究[J].中山大学学报:自然科学版,2016(1):149-154.
[16]Santos R B,Rupp M,Bonzi S J,et al.Comparison between multilayer feedforward neural networks and a radial basis function network to detect and locate leaks in pipelines transporting gas[C]//International Conference on Chemical and Process Engineering,2013:1375-1380.
[17]Huang G B,Bai Z,Chi M V.Local receptive fields based extreme learning machine[J].IEEE Computational Intelligence Magazine,2015,10(2):18-29.
[18]熊意超,徐哲,张娅.基于部分参数共享的深度卷积神经网络跨域服装检索方法:CN,106250423 A[P].2016.
[19]刘万军,梁雪剑,曲海成.不同池化模型的卷积神经网络学习性能研究[J].中国图象图形学报,2016,21(9):1178-1190.
[20]Sudicky E A,Cherry J A,Frind E O.Migration of contaminants in groundwater at a landfill:a case study:4.A natural-gradient dispersion test[J].Journal of Hydrology,1983,63(1/2):81-108.
[21]Szegedy C,Ioffe S,Vanhoucke V,et al.Inception-v4,inception-ResNet and the impact of residual connections on learning[J].arXiv:1602.07261,2016.
[22]刘欢,邵蔚元,郭跃飞.卷积神经网络在验证码识别上的应用与研究[J].计算机工程与应用,2016,52(18):1-7.
[23]华夏渝,郑骏,胡文心.基于云计算环境的蚁群优化计算资源分配算法[J].华东师范大学学报:自然科学版,2010(1):127-134.