Main content area

Cloud/snow recognition for multispectral satellite imagery based on a multidimensional deep residual network

Xia, Min, Liu, Wan’an, Shi, Bicheng, Weng, Liguo, Liu, Jia
International journal of remote sensing 2019 v.40 no.1 pp. 156-170
disasters, forests, neural networks, pollution, remote sensing, snow, support vector machines, China
Cloud/snow recognition technology for multispectral satellite imagery plays an important role in resource investigation, natural disasters, and environmental pollution. Traditional feature based classification methods cannot make full use of the effective features and multispectral optical parameters of satellite imagery; the precision of cloud/snow recognition is not good enough. Although deep convolution neural network (CNN) can extract features effectively, it faces training gradient diffusion and model degradation, which lead to a low accuracy in classification. In order to solve this problem, an improved deep residual network with multidimensional input is proposed for the cloud/snow recognition. The multidimensional deep residual network (M-ResNet) can effectively extract the image features and spectral information of satellite imagery. The multispectral satellite imagery is divided into cloud/snow-free, cloud only, snow only and cloud/snow mixed using the proposed method. The experimental results of HuanJing-1A/1B (HJ-1A/1B) satellite imagery in China show that the M-ResNet performs a good distinction for the four kinds of images. The accuracy of the classification is higher than support vector machine (SVM), random forest, convolution neural networks, and multi-grained cascaded forest (GcForest).