A Deep Learning Cloud Shadow Detection Method Based On Unified Sample For Multi-sensors

The high-precision detection of cloud and cloud shadows is of great significance for the application of optical satellite data.Due to the significant difference between spectral information and texture features,cloud detection can achieve a high precision,and it is generally difficult to detect cloud shadow due to the complexity of surface structure and different cloud types.Compared with traditional methods,deep learning algorithms can accurately identify cloud shadows in images,but because of the demand for a large number of learning samples,the method is difficult to promote in different sensors.A deep learning cloud shadow detection algorithm is purposed in this paper supported by a hyperspectral databset.The method constructs a pixel sample library through aeronautical hyperspectral data,transforms it into a multi-spectral pixel sample library to be detected by spectral simulation,and uses a unified sample to realize the sample library construction of different sensors,avoiding the traditional method of acquiring the sample library based on the remote sensing data to be detected,and uses BP neural network for cloud shadow detection,the main contents include:(1)Deep learning cloud shadow detection sample dataset construction based on hyperspectral aeronautical data.Based on AVIRIS data,the cloud shadow detection sample datasets of various types of surface coverage types are constructed by manual visual interpretation.The land cover types include cultivated land,vegetation,artificial land surface,water body and bare soil.Cloud shadow types include thick cloud shadows,thin cloud shadows,and broken cloud shadows.Based on the spectral response function of the sensor to be detected,the hyperspectral AVIRIS data is used to simulate the sample dataset which is suitable for the sensor to be detected.(2)Deep learning network architecture design,sample training and cloud shadow detection implementation.In this paper,BP neural network is selected.The simulated pixel sample library is classified into the neural network according to the type of surface coverage and trained to obtain various types of cloud shadow detection neural network models.The pixel sample library is divided into verification data and training data.The cross-validation method is used to optimize the number of hidden layer neurons,batch value,dropout value and momentum value during neural network training to improve the network generalization ability and obtain better recognition accuracy.Cloud shadow detection is performed on Landsat 8 OLI,NPP VIIRS and MODIS satellite images by matching the geographical coordinates of the GlobeLand30 feature type data.(3)Optimization of cloud shadow detection results.It is generally considered that the cloud shadow detection result conforms to the Markov random field(MRF)rule,that is,the value of each pixel on the image is related to its neighboring pixel,and is independent of other pixel values.The cloud shadow detection results are optimized by introducing spatial constraints to improve the detection of thin and broken cloud edge regions.(4)Verification and analysis of cloud shadow detection results.Accuracy verification of cloud shadow detection results of three satellite images,Landsat 8 OLI,NPP VIIRS and MODIS,combined with manual visual interpretation results,and compare with the result of Fmask cloud shadow detection algorithm.The results show that the proposed algorithm has good detection results in thick cloud shadows,thin cloud shadows and broken cloud shadows,and achieves high overall accuracy.