Research And Application Of Remote Sensing Image Change Detection Based On Neural Network

Change detection applied to remote sensing images is an important branch in the field of remote sensing,which has been widely studied in recent decades.A large number of change detection algorithms have been proposed by scholars as well as researchers at home and abroad,however,most of these algorithms are based on traditional methods,which require strict correction,alignment and other processing of the image,and then extract the feature information of the image,mainly including spectral,texture and other features Finally,by setting a suitable threshold or clustering method to divide the changed pixels and unchanged pixels,the change binary map is obtained after processing.It can be seen that the traditional method is complicated for data processing,extracts a single feature and lacks the consideration of spatial information and the relationship between pixels,and is easily affected by the good or bad threshold value.Therefore,the traditional remote sensing image change detection method is easily disturbed by noise,has poor generality,has high false detection rate and leakage detection rate,and other disadvantages.With the development of deep learning algorithms,deep learning algorithms are considered as a method choice for remote sensing image analysis.Due to the great success of deep learning in other fields,this thesis introduces neural networks into the change detection task to learn the semantic differences between images by extracting useful features such as spatial and spectral of remote sensing images through neural networks to achieve better results.The main research of this thesis includes the following three aspects.(1)For the change detection task and the semantic segmentation network in the different input methods,the image pairs are connected as the input of the U-network for training.In order to efficiently learn the deep information such as rich feature details in high-resolution remote sensing images,a residual structure is introduced to increase the depth of the network without adding additional network parameters,a convolution operation is used instead of pooling in downsampling to reduce the loss of key features,histogram matching is used for images to reduce the impact of contrast,and for the extremely unbalanced situation of changed pixels and unchanged pixels in the change detection task The corresponding loss function is designed for the change detection task.(2)To address the problem that the image pairs are not able to obtain the detail information in the original image by direct connection to adapt to the semantic segmentation network with jump connection,a two-branch network with shared parameters is proposed for remote sensing image change detection,which effectively learns the semantic difference between changed and unchanged pixels by sharing information,and introduces migration learning into the feature extraction network,and proposed a spatial attention mechanism based on difference information.which assists the network to focus on the learning of difference information between two images.(3)The input methods in dual-temporal remote sensing image change detection are summarized,and three feature fusion methods in remote sensing image change detection are analyzed.The impact of the three different feature fusion approaches in the dualbranch structure on the change detection task is experimentally compared,and a dualattention mechanism is introduced in the best-performing feature fusion approach to cascade channel attention with spatial attention to enable the network to quickly learn the key features in the images.Finally,the above research is applied in a practical way to design and implement a change detection system.