Research On Cloud Detection Algorithm Of Remote Sensing Image Based On Convolution Neural Network

Cloud image detection in remote sensing image is of great significance for the subsequent application of remote sensing image.At present,remote sensing satellite image is more and more widely used in agricultural production,weather forecasting,natural disaster prediction,military technology,geographic mapping,change detection,water conservancy and transportation and other fields.Cloud image detection in remote sensing image is a prerequisite for efficient use in the later period.In recent years,with the rapid development of artificial intelligence,the cloud image detection of remote sensing image based on convolution neural network has also attracted much attention.Cloud detection methods based on such algorithms have stronger robustness and accuracy than traditional algorithms,and can greatly improve detection.effectiveness.This paper studies the algorithm in depth and does the following work:1.Aiming at the problem of large amount of remote sensing image data and high resolution,resulting in low cloud detection efficiency,a cloud detection algorithm based on superpixel segmentation and convolution neural network is proposed.First,manually select the cloud and clear sky image blocks.Then,convert the RGB color remote sensing image to color space,and then use the super pixel method to divide the remote sensing image into super pixels as the basic unit,build a convolution neural network,and select the optimal model,Input the image into the model to get the two-category image of cloud and clear sky.2.Aiming at the lack of obvious color distribution and texture pattern of clouds in RGB color images,cloud detection is prone to misdetection and details are seriously lost.An improved M-net model(RM-Net)is proposed to achieve end-to-end Pixel-level semantic segmentation.First,the original data set is enhanced and the corresponding pixel-level labels are labeled;second,the hollow space pyramid pooling is used to extract the multi-scale features of the image without losing information,and the residual unit is used to make the network less prone to degradation;finally Using the encoder module and the left path to extract the global context information of the image,the decoder module and the right path to recover the image spatial resolution,determine the class probability of each pixel based on thefused features,and input the classifier for pixel-level cloud and non-cloud segmentation.By training and testing Landsat8 and high-score WFV RGB color images,the experimental results show that the method in this paper can detect cloud edge details well under different conditions,and achieve higher accuracy cloud shadow detection,which proves that the method in this paper has Good generalization and robustness.3.In optical high-resolution images,it is difficult to distinguish clouds from some bright features(such as snow and white buildings)using only spectral features.Due to the limited spectral range of clouds and cloud shadows and the complexity of the underlying surface,it is difficult to accurately detect clouds and cloud shadows for optical high-resolution images.Therefore,a detection method based on Multi-scale Feature Fusion Network(MFFN)is proposed.The model mainly includes Res.block module,Multi-scale Convolution Module(MCM)and Multi-scale Feature Module(MFM),capable of extracting rich spatial and semantic information,and performing pixel-level detection of clouds and cloud shadows in remote sensing images,and can obtain more precise cloud and cloud shadow edge detection results.