Research On Cloud Detection Algorithm Of Remote Sensing Images Based On Deep Learning Attention Mechanism

In recent years,with the rapid development of optical remote sensing satellite technology,high-resolution satellite images have been widely used in various research fields.Satellite images are often covered by clouds,which are significantly important to research fields such as atmosphere,climate and natural disasters.However,it is unfavourable for the research in the fields of surface observation,marine development and urbanization.Because the clouds on the satellite image will lead to the reduction or even loss of image information and affect the spectral and texture information of optical sensor imaging,the image containing clouds dose not meet the conditions of information automatic processing.When the sensor observes the surface along the non vertical direction,the cloud shadow will interfere with the follow-up research.Therefore,cloud and its shadow detection is an indispensable preprocessing process in satellite image analysis,which improves the availability of satellite images for subsequent research and application in various fieldsTraditional cloud detection algorithms need a lot of prior conditions such as manually setting thresholds and manually designing features in order to achieve good results in specific data sets.With the rapid development of deep learning,a large number of cloud detection algorithms based on deep learning have been proposed.These algorithms usually regard remote sensing image cloud detection as a semantic segmentation task.At present,in the field of semantic segmentation,the algorithm based on deep learning has achieved much better performance than the traditional algorithm.In recent years,researchers have paid close attention to adding attention mechanism to the deep learning model.This mechanism can make the model pay more attention to the region of interest,so as to further improve the performance of the algorithm.It has achieved great success in the research fields of computer vision,natural language processing and medical imaging.However,in the task of cloud detection in remote sensing images,the research on the mechanism of attention is still not deep enough.Therefore,this thesis focuses on the research of attention mechanism in cloud detection task,and proposes two cloud detection algorithms based on deep learning attention mechanism.1.Aiming at the two kinds of cloud detection tasks of cloud and non cloud,an asymmetric encoder decoder model convolution activate attention-UNet(CAA-UNet)based on attention mechanism is proposed.CAA-UNet introduces a new and effective activation function,activate or not(ACON),which learns to activate the neurons or not.Compared with traditional activation functions,ACON remarkably improves the performance.The asymmetric encoder-decoder structure helps our model discovery more discriminative features,reduce the amount of parameters and improve the computational efficiency.CAA-UNet modifies an attention gate which is embedded into each skip connection.The modified attention gate suppresses irrelevant regions and highlight salient features,which makes the model pay more attention to the cloud feature.At last,we employ two binary cloud datasets to verify the effectiveness of our proposed method.2.Aiming at the multi-class cloud detection tasks of cloud,cloud shadow and background,proposed is a cloud detection model convolution net with squeeze and excitation(ConvSENet)containing large receptive field,multi-scale features and lightweight attention.ConvSENet introduces Conv Ne Xt,a pure convolutional neural network backbone network with the advantages of Transformer architecture,as the encoder,and unified perceptual parsing net(UPerNet),which integrates multi-scale features,as the decoder and lightweight attention mechanism spatial and channel squeeze & excitation(scSE)module in the model.Secondly,in the process of training,a new training method using multi feature maps to calculate the loss function is proposed.Finally,the effectiveness of ConvSENet and training methods are verified on a multi classification cloud dataset.