Application Of Hyperspectral Data Supported Deep Learning Algorithm For Satellite Data Cloud Detection

Due to the influence of sensor's own parameters and observation conditions,cloud cover is ubiquitous in optical remote sensing images,and it is difficult to obtain cloud free data.Cloud is the main factor affecting the quality of optical remote sensing image,which seriously restricts the sensor's ability to observe the earth,and will greatly influence the reliability and accuracy of the quantitative inversion of the surface and atmospheric parameters.Therefore,fast and accurate cloud recognition of remote sensing images is an essential step in remote sensing data preprocessing The traditional method of cloud detection is based on the spectral reflectance difference of the object which uses a set of threshold combinations to divide the clouds and clear space pixels.It is time-consuming and hard to study the algorithms for different sensors.To solve this problem,a deep learning cloud detection algorithm based on hyperspectral data is proposed in this paper,which can support cloud detection of multiple spectral data.The AVIRIS data has the characteristics of hyperspectral and high spatial resolution.There are 224 continuous narrow bands in the wavelength range of 400-2500nm,which can detect the small spectral differences of the target.Based on AVIRIS data,we select enough samples of cloud and clear hyperspectral pixels to creat prior database.According to the parameters of the spectral response function of the sensor to be detected,the data of the hyperspectral are simulated to the pixels of multispectral sensorBased on the Keras deep learning framework platform,the BP neural network for cloud detection is designed.The multispectral sample data which is simulated by hyperspectral data are inputed to the network and the cloud detection rules are trained based on spectral characteristics.The parameters in the network are improved and optimized by the method of cross validation.The experiment shows that the optimization of the number of neurons of the hidden layer,batch value.dropout and momentum value can effectively increase the recognition accuracy of the network model,improve the training efficiency,and achieve a better recognition result.Markov random field(MRF)is a image segmentation algorithm based on statistical theory which makes full use of the spatial constraint relations among pixels and is widely applied.In this paper,based on the Markov random field model,the iterative conditional model(ICM)algorithm is used to optimize the results of neural network cloud detection.Experiments show that using statistical information can reduce misclassification of cloud detection and improve detection effect.What's more,different remote sensor data is used to validate the proposed cloud detection algorithm,of which the cloud detection algorithm obtains good identification results for Landsat 8 OLI,VIIRS and MODIS.The results show that the whole accuracy is above 90%,and can meet the requirements of application.