Research On Unsupervised Change Detection Of Remote Sensing Image Sequences With Unequal Time Intervals

Change detection is defined as the process of identifying differences in an object or phenomenon in coregistered remote sensing images collected at different times.It plays an important role in several applications,such as land surface monitoring,disaster evaluation,and urban expansion.With the increasing accumulation of remote sensing image data and the development of remote sensing satellite technology,the spatial resolution and temporal resolution of remote sensing images have been greatly improved.The long-term accumulation of remote sensing image time series data provides data support for the research of time series change detection,which also makes the use of time series remote sensing images to realize the detection of land cover changes as a feasible and accurate technical means.Compared with the traditional two-phase change detection,remote sensing image time series has natural advantages in change detection.However,the time series of remote sensing images containing more information also brings new challenges to change detection.First,more images mean that manual visual interpretation to obtain labels is more cumbersome,limiting the practical application of time-series change detection algorithms.Secondly,remote sensing image time series contains rich spatial information,how to fully consider the spatial correlation between pixels in the time series has become a key issue in the field of remote sensing image time series change detection.In addition,the time series of remote sensing images have very obvious characteristics,such as inconsistent time interval,seasonal cycle change and so on.The purpose of this paper is to improve the accuracy of remote sensing image change detection,and to optimize the label acquisition method to achieve unsupervised change detection,which has higher application value.The research contents of this paper are as follows:(1)Due to the fact that time series training samples are difficult to obtain and traditional time series change detection methods are difficult to make full use of spatial features,this paper proposes a time series change detection method for remote sensing images based on spatiotemporal-spectral features.The method can automatically generate changed and unchanged samples for training,then combine spatial information(texture and statistical features)with spectral and temporal information in change detection.Finally,we use SVM to train and analyze samples containing spatiotemporal spectral features.The accuracy of the method proposed in this paper has been compared with the widely used time series change detection methods.Compared with mainstream change detection methods,its accuracy is improved by1.25%-6.56% in different scenarios.The experimental results show that this method has excellent performance in various evaluation indicators,which verifies the effectiveness of this method in solving the problem of time series change detection in remote sensing images.(2)In order to solve the problem that the existing time series change detection methods do not consider the influence of the time interval between different phases on the change detection,this paper proposes a time series change detection method based on irregular time distance in remote sensing images.This method uses a weight-sharing multi-layer convolutional neural network to construct a deep feature extraction network to extract the deep spatial spectral features of remote sensing images.Then use the irregular time distance long short-term memory network to extract the spatiotemporal spectral features in the time series,and finally input the extracted features into the fully connected layer to detect land changes.The method in this paper is compared with the current mainstream time series change detection methods.Compared with mainstream change detection methods,its accuracy is improved by 7.18%-10.82% in different scenarios.The experimental results show that this method has excellent performance in both qualitative and quantitative comparisons,which verifies the effectiveness of this method in the problem of time series change detection in remote sensing images with irregular time distances.(3)In order to reduce the influence of pseudochanges factors such as intra-annual and inter-annual changes in time series on the change detection results,this paper proposes an unsupervised change detection method based on temporal distance model-guided convolutional recurrent networks,which can be achieved by adopting a novel temporal distance-guided long short-term memory cells to suppress the effects of pseudochanges.A novel temporal distanceguided long short-term memory unit to suppress the effects of spurious changes.It adds a temporal modulation model guided by temporal distance to the input gate and forgetting gate in traditional long-short-term memory network to adapt to irregular temporal distance.Furthermore,we propose a weighted pre-change detection model to automatically extract the most reliable training samples,and use a new focal weighted cross-entropy loss function during training to address the changed/unchanged sample imbalance and hard/easy Sample imbalance problem.We compare our method with a variety of advanced change detection methods on the Landsat8 time series dataset of nine typical scenes collected from 2013 to 2021.Compared with the traditional LSTM change detection method,its accuracy is improved by 1.23%-4.90%.The experimental results demonstrate the excellent performance of our method in time series change detection in remote sensing images.,which verifies the effectiveness of this method.