Research On Change Detection Algorithm Based On High Resolution Remote Sensing Images

In recent years,change detection has become one of the popular research problems in the field of earth observation,and plays a very important role in various practical applications such as disaster assessment,ecological environment detection,urban development planning and map correction.Therefore,this paper studies the use of highresolution remote sensing images to identify change information quickly and accurately,which is of great significance for production and life.With the rapid development of space technology and information technology,remote sensing images are gradually developing in the direction of high resolution,hyperspectral and massive data,which brings some problems to the traditional remote sensing image change detection.Hyperspectral and high resolution data lead to more noise interference for the detection of specific targets.Especially,the pseudo-change caused by the natural growth of features,different shooting angles,etc.,is a great challenge for change detection by artificially constructing salient features.Therefore,in this paper,we study a deep learning-based change detection algorithm for remote sensing images.Through learning a large amount of data,the model can automatically extract deep-level features of images and achieve a better fitting effect on the data.This paper focuses on how to apply the traditional twin neural network architecture combined with Transformer framework to the change detection task of remote sensing images and improve and optimize it to be more suitable for the change detection task.Second,this paper replaces the traditional serial residual network as a feature extractor by introducing a parallel high-resolution network.The high-resolution network obtains the detailed information of the underlying features by repeatedly fusing the highresolution and low-resolution features several times,while maintaining the high resolution throughout;next,for the case of varying scales of change targets in remote sensing images,the atrous space pyramid module is introduced to obtain multi-scale target features and refine the detection edges by adding the max pooling module.;again,in order to obtain the dependencies between different pixels in the image further refine the detection targets.In this paper,we introduce the chunked position self-attention module to obtain more perfect features by enhancing the target pixels to suppress noise interference;finally,since most of the remote sensing data sets used for change detection have a very small portion of changed regions compared to unchanged regions,there is a problem of sample imbalance.In this paper,we solve this problem by introducing suppression factors to the traditional contrastive loss function.The final results on the three datasets of LEVIER-CD,WHU Building Dataset and SZTAKI are goods.It proves that the proposed algorithm in this paper has high detection accuracy and also has good generalization.