Satellite Cloud Image Cloud Detection Based On Deep Learning

Satellite cloud image cloud detection is the key to the analysis and application of various remote sensing data,however,due to the high complexity of certain specific categories of features and remote sensing data in different bands has many correlations,which improves the difficulty of the model in feature learning.The utilization of multi-spectral image features in the existing models is low,which results in low accuracy in cloud detection.On the basis of investigation and analysis of domestic and foreign scholars,it is found that convolutional neural network has a very good learning ability for image features,and the paper focuses on the application of Deep Learning in cloud detection of satellite cloud images.In the process of deep learning,with the deepening of deep learning network,the deep learning network can extract features effectively,but there are some problems in training,such as vanishing gradient problem,low training efficiency and difficulty in optimization.In order to solve these problems,this paper uses Multidimensional densely connected convolutional neural network model to realize cloud detection of multi-spectral satellite cloud image.And from the network connection mode and the feature channel level,the densely connected convolutional neural network is improved respectively.And a weighted densely connection is proposed for the densely connected convolutional neural network of densely connected network redundancy,which improves the training efficiency of the model.Because it is not highly utilized in feature channels,this paper proposes a densely connected convolutional neural network model based on attention mechanism,which effectively combines the attention mechanism with the densely connected convolutional neural network,and it makes the use of features of the model reach the maximum.In order to improve the model,we adopt the way of shared storage space to reduce the appearance of the model and make it possible to use a deeper network structure in the case of limited computing resources;Combined with the improved optimizer,the learning rate is dynamically trimmed,which makes the model perform better.The experimental results show that the improved model based on attention mechanism also performs well in the application of cloud detection with complex spectral information.