High-resolution Remote Sensing Image Cloud Detection Method Based On Multi-scale Depth Features

Most remote sensing images have been polluted by clouds of varying degrees,and cloud pollution has caused image information attenuation or even loss.The detection of thick and thin clouds in remote sensing images is of great significance to cloud removal and meteorological detection.Most existing algorithms are rule-based,which means that their ability to detect clouds depends on specific intensity thresholds and carefully tailored rules for a given satellite platform and sensor.Due to the spectral heterogeneity of the cloud and the spectrum and temperature changes of the underlying surface,the size,texture,brightness,and shape of the cloud are quite different.The thin cloud is semi-transparent,and its spectral information appears as a mixture of the underlying surface and the cloud spectrum.Light and thin clouds are easily missed in the detection,and the spectra of some of the highlighted ground are similar to those of thick clouds.This brings challenges to the detection of thin and thick clouds.This thesis focuses on the problems in the accurate cloud detection of high-resolution remote sensing satellite images,and proposes a high-resolution remote sensing image cloud detection method based on multi-scale depth features.This method divides the remote sensing image into thick clouds,thin clouds and no clouds.When performing context integration in deep learning,shallow feature maps tend to have relatively small perception fields,which can easily cause some light and thin clouds to be missed.Therefore,when the shallow feature map and the deep feature map are combined with contextual information,the method of this thesis designs a jump connection structure that takes into account the receptive field,which takes into account the receptive field of the feature map and the importance of different feature maps at the same time.At the same time,in order to realize the detection of different sizes of clouds,a multi-scale module is designed to extract the multi-scale global features of thick clouds and thin clouds.In addition,thin clouds have always been a problem in cloud detection due to the complexity of their spectrum.According to the spatial distribution of clouds,thin clouds are generally between thick clouds and no clouds,and the probability distribution of pixels of thin clouds always shows that certain two types of probability values are relatively close.Therefore,to solve the problem that thin clouds are more difficult to detect,this thesis proposes a thin cloud detection optimization module based on the spatial distribution characteristics of clouds.In the comparative experiment,the method of this thesis is compared with SVM,FCN-8s,U-net and MFCNN.Without any prior information,the method of this thesis not only increases the F1?score of thick clouds by 1.78%,but also increases thinness.The F1?score of the cloud has also increased by 3.58%,and it has the best detection effect in the cloud area and the entire image.In the ablation experiment of the model in this thesis,it is proved that each module can effectively improve the accuracy of cloud detection.