Cloud Detection Of Remote Sensing Image Based On Deep Learning

In optical remote sensing,cloud as a ubiquitous object is an important factor affecting the accuracy of feature analysis and remote sensing image feature extraction,and has a significant impact on the efficiency of the data transmitted from the remote sensing image.In Landsat 5 and 7 images,the cloud coverage can reach 40%.High cloud coverage will significantly reduce the accuracy and application value of remote sensing data.Therefore,improving the accuracy of cloud detection has important practical significance for the application of remote sensing images.Since the time window for satellite data transmission is only 5 to 10 minutes,cloud detection preprocessing of the data before the satellite downloads the data can improve the quality and transmission efficiency of the remote sensing image.With the continuous development of remote sensing technology,the trend of miniaturization of satellite design has limited the computing power and power consumption of electronic systems.Therefore,a lightweight cloud detection model with low computing power requirements is necessary.This project aims at different application scenarios of cloud detection,based on semantic segmentation,combined with deep learning for cloud detection,achieved two methods of high precision and lightweight cloud detection.The main research contents and innovations of this article include the following two aspects:1.Aiming at the problems that traditional methods require manual feature selection and the inability to detect end-to-end,an improved U-Net based on residual network is proposed,which can achieve the pixel-wise segmentation of cloud in remote sensing images.The application of the residual module deepens the network and improves the expressive ability of the network.The pixel accuracy of this method reaches 93.33% on the Landsat 8 Cloud Cover Assessment dataset,which is 2.29% higher than that of the original U-Net,and 7.78% higher than that of the traditional Otsu method.The network can achieve high-precision end-to-end cloud detection.2.Aiming at the problem that the computing power of deep learning network is difficult to implement on mobile or embedded platforms,a lightweight cloud detection model based on depthwise separable convolution is proposed to reduce the model size and computation cost of pixel-wise cloud detection methods.The network uses depth separable convolutions to reduce the computation cost of the encoder and decoder,and uses depthwise separable convolutions with stride instead of maximum pooling to reduce the information loss of feature map caused by downsampling.The encoders continuously downsample the image to obtain feature maps of different scales.The decoders combine the multiscale features of the encoders to restore the image.Finally the network achieves lightweight end-to-end cloud detection.The pixel accuracy of this method exceeds 90%.The prarmeter and floating-point operations are 12.4% and12.8% of the U-Net.Inference speed on the CPU is about 5 times that of U-Net.The experiments show that the U-Net cloud detection model based on residual network has the advantages of high detection accuracy and fewer false detections,and can provide high-precision cloud detection results.The lightweight cloud detection model based on depthwise separable convolution reduces the amount of network model parameters and computation,and increases inference speed on the premise that the pixel accuracy only decreases slightly.