Research On Remote Sensing Imagery Change Detection Method Based On Deep Learning

With the further development of aerospace technology and electronic information technology,remote sensing technology has broken through the bottleneck of data acquisition and has entered a new stage of comprehensive application.Change detection based on remote sensing imagery is the main method to study global surface change.It has important applications in land and resource management,agricultural and forestry monitoring,natural disaster monitoring and assessment.Faced with the increasing remote sensing data,we have accumulated a large amount of data of surface changes over time.We can fully process and utilize these remote sensing data to find the information of interest,and then get the change of the type of object of interest or the change of internal state.However,because remote sensing imagery change detection is a complex comprehensive processing process,remote sensing imagery taken for change detection in different periods are affected by seasonal changes,satellite sensors,solar height and atmospheric conditions,which reduces the achievable accuracy of change detection output.The degree of automation and accuracy of the inspection technology needs to be further improved.In recent years,deep learning algorithm has been rising in the field of artificial intelligence.Remote sensing data has the characteristics of being massive,diverse and multi-dimensional in nature.Deep learning has natural advantages in processing remote sensing data.It can automatically and multi-level extract abstract features of complex objects,learn and discover high-level semantic features of objects in images from massive data,and greatly improve the accuracy of remote sensing imagery interpretation.At present,the deep learning method has has been widely used in image recognition field.The same theory should also be applied to the field of remote sensing imagery change detection.Therefore,we urgently need to carry out the research in remote sensing imagery change detection methods based on deep learning,which not only greatly improves the accuracy of remote sensing imagery change detection,but also opens up a new automated and intelligent processing method for the application requirements of current efficient,fast and automated change detection.In this paper,aiming at the problem of bi-temporal optical remote sensing imagery change detection,we propose that remote sensing imagery change detection based on deep learning,according to different detection objects,can be divided into three categories: pixel-based,object-based and region-based.The main work includes the following aspects:(1)A pixel-based recurrent neural network method for remote sensing imagery change detection is proposed.Pixel-based method is a classical method.Its mathematical model is simple and fast,and it has advantages in processing medium and low spatial resolution remote sensing imagery.Based on the characteristics of recurrent neural network model which is suitable for processing time series data,this paper proposes three different input modes and six different recurrent neural networks for pixel-based recurrent neural network remote sensing imagery change detection model,and compares their effects on change detection accuracy through experiments.Among them,the paper proposes a recurrent neural network change detection model based on neighborhood and convolution neural network feature extraction.Compared with other single deep learning models,our model can further improve the accuracy of change detection.(2)An object-based high-resolution remote sensing imagery change detection method based on deep learning is proposed.For high-resolution remote sensing imagery,object-based change detection method is a mature traditional method.This paper combines deep learning method with traditional object-based method,and proposes two object-based change detection methods for remote sensing imagery.One is change detection method based on object feature and stack auto-encoder.We propose three different stack auto-encoder models,and discuss the influence of their model structure on the change detection accuracy through experiments.The other is masked and nonmasked object processing method based on the original object and the convolution neural network.We design two detection methods based on single channel convolution neural network and siamese convolution neural network.Then,through experiment,we compare the effects of different common convolutional neural network models on the accuracy of change detection.The method of combining the siamese convolutional neural network with the object-based method in this paper,can obtain high detection accuracy of remote sensing imagery change under the condition that the convolutional neural network model structure is simple.(3)A high-resolution remote sensing imagery change detection method for regionbased convolutional neural networks is proposed.In natural image target detection,detection method based on convolution neural network has been widely used to extract and recognize targets in complex scenes.In this paper,the "change" in remote sensing imagery is regarded as the target to be detected,and the changed parts are detected directly from a certain size of "image region".A high-resolution remote sensing imagery change detection model based on Faster-RCNN and Mask-RCNN is proposed.Based on this,two detection methods are proposed: one is to merge the bi-temporal images before detection;the other is to detect the difference between the bi-temporal images before detection.Through experimental analysis and comparison,it is concluded that our region-based convolutional neural network model for highresolution remote sensing imagery change detection has higher detection accuracy than other deep learning methods,and Mask-RCNN based change detection method has the best effect.In addition,whether based on Faster-RCNN or Mask-RCNN model,the detection accuracy of bi-temporal images difference is higher.It shows that the traditional the image deference method in the deep learning model is also helpful to improve the accuracy of change detection.(4)A prototype of remote sensing imagery change detection system based on deep learning is designed and implemented.Through this system,the experiment of remote sensing imagery change detection method based on deep learning can be carried out conveniently,and at the same time,it can provide reference for the follow-up application of remote sensing imagery change detection in related fields.