Research On Change Detection Of Medium Resolution Remote Sensing Image Based On Recurrent Neural Network

Continuous,repetitive,and wide-ranging satellite remote sensing observations have always been an important means of detecting surface cover change.With the rapid accumulation of remote time-lapse image data with high temporal resolution,high spatial resolution and high spectral resolution,based on massive remote sensing data,the change detection has become a hotspot of remote sensing,and various new methods of change detection have been proposed.The remote sensing image time series data contains a large amount of information on the time dimension and space dimension of the ground object.How to effectively mine the space-time information on the remote sensing image plays a vital role in the application of remote sensing image,but the current change detection method often cannot fully exploit it.The use of space-time information on remote sensing images has led to problems such as insufficient precision and low automation.Recurrent Neural Network(RNN)can effectively mine the information of time series data due to the existence of feedback mechanism.It has been successfully applied in the fields of machine translation,speech recognition and computer vision,and gradually extended to various comprehensive fields.Therefore,this paper introduces the Recurrent Neural Network into the remote sensing image processing,with the aim of solving some problems in the application of remote sensing change detection by means of advanced models in the industry.According to the characteristics of the Recurrent Neural Network and the research objectives,this paper uses a change detection method based on Long Short-Term Memory,aiming at the common low-resolution change detection problem and solving the change detection in the background of big data.This paper further explores the application of cyclic neural network in remote sensing change detection.This paper focuses on the Recurrent Neural Network and takes a series of researches on the problems involved in remote sensing change detection.A series of experiments are carried out on remote sensing images with different spatial scales and different spectral resolutions.The main research contents and innovations of the paper include the following two points:A model based on Recurrent Neural Network is constructed for change detection.In this paper,Recurrent Neural Network model for processing time series data is applied in the field of remote sensing change detection.Combined with massive remote sensing data,the Long Short-Term Memory(LSTM)with cell state is used as the basic unit of RNN.This paper construct multi-hidden layer neural network model change detection,and use backward propagation algorithm and random gradient descent to train the model to realize the processing of multi-temporal sequence image data.At the same time,we select the appropriate network structure and parameters by designing the verification set,tuning the model exploring the technical route and method framework for RNN for medium resolution change detection.In this paper,the feasibility of the method is verified by hyperspectral image and multispectral image.Two-class hyperspectral imagery and multi-class multi-spectral imagery are used to conduct RNN model experiments.Two basic problems of change detection are solved by level: calibration change area and identification change class.The experimental results show that the method of change detection using RNN model reduces the complexity of preprocessing,improves the automation degree and accuracy of change detection,and compares with other methods to prove the applicability of RNN model in a large space.It proves the feasibility of using RNN model for change detection.