Cloud And Snow Classification In Plateau Area Based On Deep Learning Algorithms

Cloud/snow recognition is an important remote sensing application technology.It plays an important role in resource survey,natural disaster prediction and environmental protection.However,the cloud and snow regions have very close pixel values and very high similarity in satellite remote sensing images.Aiming at addressing the problems that feature extraction method for the traditional cloud snow classification algorithm is too complicated,the extracted feature are single and low utilization of spectral information which cannot effectively distinguish the cloud area and the snow area.A deep learning technology is proposed to automatically extract more abundant cloud/snow features and integrate effectively spectral information,and the deepened convolutional neural network structure can extract deep features,which can improve the accuracy of cloud and snow recognition to a certain extent,but the vanishing gradient and model degradation caused by simply stacked deep network are inevitable.To solve the above problems,residual learning is borrowed to further deepen the structure of the network,thereby extracting higher level feature information,and constructing a multi-dimensional deep residual network can effectively extract the spatial information and the spectral information of the satellite image.Experiments on a large number of environmental and disaster monitoring and forecasting satellites(HJ-1A/1B)show that multi-dimensional deep residual networks can achieve more accurate cloud-snow recognition tasks than traditional algorithms.Compared to shallow convolutional neural networks and other deep algorithms such as multi-grained cascade forest,multi-dimensional deep residual networks have better generalization performance and robustness,can more effectively integrate spectral information for more accurate identification of cloud/snow free,cloud only area,snow only area,cloud/snow mixed area,which is more suitable for cloud/snow classification tasks.In addition,a large number of experimental results show that multi-spectral satellite remote sensing data is more suitable for cloud-snow recognition tasks than single-spectral satellite remote sensing data.Therefore,the algorithm has important guiding significance for achieving accurate cloud-snow monitoring.