%A Bo Zhu,Fangjie Zhong,Junsuo Zhao %T Fractal Dimension and Angular Second Moment-assisted Convolutional Neural Network for Cloud and Snow Recognition %0 Journal Article %D 2022 %J Remote Sensing Technology and Application %R 10.11873/j.issn.1004-0323.2022.6.1328 %P 1328-1338 %V 37 %N 6 %U {http://www.rsta.ac.cn/CN/abstract/article_3590.shtml} %8 2022-12-20 %X

For space-based earth observation imaging missions, the cloud cover contained in an remote sensing image often determines whether the data is available. However, the spectral characteristics of cloud and snow are similar, which makes it difficult to distinguish them. The purpose of the research on cloud and snow recognition is to improve the ability to judge the validity data. So an novel method is designed to solve it. Firstly, an improved lightweight convolutional neural network, which is built based on the proposed DCP (Double Convolution Parallel) structure, is used as the backbone to classify the quasi-cloud (cloud, snow and highly reflective ground objects) and other ground objects. Secondly, the textures and gray features of cloud and snow and ground objects are analyzed by a binary tree network formed by fractal dimension and angular second moment for fine recognition in further. The network weight layers are only six (four convolutional layers and two full connection layers). The proposed method is trained on the data sets containing cloud, snow and cloud-snow with different ground sample resolutions from Tianzhi-1 and SPOT4/5/6 and Pleiades. When compared on the accuracy with reference methods, such as random forest, SVM, traditional methods and binary tree methods, our method provide an increasing accuracy to 89.08%. The experimental results shows: (1) The comparative experiment between network structures shows that the DCP could effectively improve the model ability of feature information extraction and promote faster convergence; (2) Texture features analysis of remote sensing images makes the recognition process not completely dependent on convolutional neural network, so as to reduce the network depth and weight parameters; (3) The combination of traditional remote sensing analysis method and neural network is better than one of them alone, which can improve the accuracy of cloud and snow recognition. The proposed method is suitable for cloud and snow recognition on panchromatic, multispectral and hyperspectral remote sensing imagery.