Cloud Detection In Domestic High-Resolution Remote Sensing Image Based Deep Neural Networks

With the rapid development of remote sensing technology,remote sensing images are widely used in various fields,such as military target recognition,environmental monitoring,meteorological analysis,mineral development,geographical mapping and so on.However,statistics show that at any time,the surface of the 50% earth is covered by clouds.In remote sensing images,the presence of clouds poses challenges for many subsequent analytical tasks.At present,the common cloud detection is mainly focused on two types of applications.One is the macro application of small scale large frame and large frame and global cloud detection,represented by MODIS.One is the detection of military and civilian targets based on QucikBird and IKONOS,which are superior to 1m in super high resolution.In our country,the high-resolution satellite resolution is in 2~10 meters.In this case,the scale image often faces the problem of "not seeing" in the target detection,so the target detection interference and the image quality decline are particularly obvious.Therefore,cloud detection is very interesting for such "less clear" remote sensing images.After a lot of data processing practice,this paper proposes an adaptive segmentation algorithm ASLIC and DCNN network for feature extraction and classification for the flexible boundary features of the cloud,the inhomogeneity of interior luminance and the color difference of thin clouds.Finally,the cloud detection target of ASLIC+DCNN is realized.The main contributions of this paper are two aspects.1.first,the linear iterative clustering(SLIC)method is improved,the vector distance calculation function of the CIE color space is modified,and the independent index IDD is introduced to describe the position relationship between the cloud pixels and the image,thus the initialization condition and iteration of the SLIC algorithm are adjusted adaptively.Step length,and then output the cloud candidate.2.using DCNN network to extract and classify more than 20000 patches of ASLIC segmentation result dataset,and finally trained the neural network.DCNN network is a double layer neural network model composed of coiling layer,pool layer and all connected layer.The feature map is obtained throughthe convolution layer.Then the multi-scale classification is realized by the pool layer and the full connection layer,and the nonlinear problem can be solved well.The drawback lies in the fact that a large number of samples need to be trained and the parameter adjustment process is more complicated,and there is no mature theory to guide them.In practical applications,we can use the DCNN model for applications with more training samples and higher real-time requirements.