Cloud Detection Of Complex Surface Based On Multi-Scale Attention Residual Network

In recent years,with the increase of satellite sensors,the accuracy and efficiency of remote sensing image quality control are put forward higher requirements.As one of the main quality problems,cloud pollution needs a lot of manpower and energy to eliminate the problem of excessive cloud.Therefore,whether the cloud coverage area on the image can be extracted with high accuracy and automation,and then the cloud amount in the image can be calculated quickly to form a process remote sensing image quality inspection business system has become the most wanted problem for many quality inspection departments.Traditional cloud detection methods have difficulty in threshold selection,and when a large number of interference objects appear on the image,it is unable to segment the cloud area in large quantities and automatically.Therefore,this paper selects deep learning method for cloud detection experiment.At present,the cloud detection method based on simple background is relatively mature,and the extraction accuracy is high.However,when the background of ground objects is complex,most of them will have the problem of mixed ground objects false detection.Therefore,in this paper,the high-precision extraction of cloud area under complex background on resource3 remote sensing image is carried out.The main work of this paper is as follows:(1)In this paper,the spectral,texture and spatial characteristics of resource-3 image,cloud and complex objects in remote sensing image are analyzed in detail,so as to judge the cause of confusion between target objects and other objects.(2)In order to solve the problem of under fitting caused by too few data sets,based on the same neural network architecture and super parameter setting,the necessity experiment of data set expansion is carried out.According to the final accuracy evaluation,the test accuracy after sample enhancement is higher than that without processing,which is improved by nearly 3%.This proves that appropriate data set expansion can effectively improve the cloud detection accuracy and alleviate the problem of under fitting.(3)The traditional SGD optimizer is replaced by Nadam with momentum to improve the speed and accuracy of network fitting.By comparing the accuracy curve and loss curve,Nadam's fitting speed is fast,and the average loss value is less than 0.1,closer to 0,which is significantly better than SGD algorithm,and can better improve the network performance.(4)In order to reduce the impact of complex background on cloud detection,this paper selects resnet50 as the basic coding structure,and adds optimized attention mechanism(CBAM)in decoding,so as to design an end-to-end multi-scale attention residual network(MARN).In order to verify the effectiveness of this method,Resnet and U-Net methods are used to do cloud detection comparative experiments.The results show that the cloud detection method in this paper is better,and the cloud detection accuracy in complex background(snow,water and other bright background)is more than 90%,which is greatly improved compared with the two methods.It can avoid the large area interference of complex ground objects and extract more refined cloud boundary.