Change Detection Of High Resolution Remote Sensing Images Based On Deep Learning

Change detection on remote sensing images refers to extracting pixels whose feature category has been changed in the remote sensing images by comparative analysing remote sensing images with different times in the same range.The traditional change detection methods on remote sensing images mainly include post-classification comparison method,image algebra method,change vector analysis method and shape feature-based change detection method.Although these traditional methods of change detection have been widely applied in the fields of natural disaster early warning monitoring,cultivated land protection monitoring,land use dynamic analysis,land and resources conservation and monitoring,and social development analysis,there are still many shortcomings such as the strict requirements of image registration,low automation and weak on processing multi-source images.Deep learning algorithm(Deep Neural Network)has powerful learning ability,can fit complex nonlinear mapping and solve complex task scenarios.Therefore,this paper introduces deep learning algorithm into remote sensing field and utilizes powerful feature learning ability of deep neural network to solve complex problems of change detection on remote sensing images.The main works of this paper include the following three aspects:(1)Firstly,An iterated spatially selective noise filtration algorithm is proposed to solve the problem of that when the image is seriously polluted by noise,traditional spatially selective noise filtration algorithm cannot filter the noise completely.In addition,the least-square criterion is used as the termination condition of the iterative algorithm to achieve the optimal iteration number targeting at the over-filtering problem of useful information caused by the increasing of iterations.It is testified by the theoretical analysis and experimental results that the iterated spatially selective noise filtration algorithm outperforms the traditional spatially selective noise filtration algorithm in filtering the residual noise more thoroughly,and the least-square criterion can make the former algorithm terminate in the optimal iteration number so as to prevent excessive filtering.(2)In order to verify the feasibility of the deep learning algorithm applied to change detection and the superiority compared to the traditional methods,a change detection model based on FCN network structure is constructed and compared with the traditional method of change vector analysis by experimental analysing.The experimental results show that the change detection algorithm based on deep learning is superior to the traditional algorithm in image preprocessing precision requirements,intelligent recognition pseudo-change,discovery of weak change regions,and precision and recall rate of change prediction maps.(3)After verifying that the the change detection algorithm based on deep learning has great advantages over traditional methods,this paper proceeds to build a more powerful deep neural network model to improve the effect of change detection.And design comparison experiments to prove that due to the convergence of the context features by convolution learning,the combination of the shallower features and the further increase of the network depth make the UNet change detection model are all improved in precision rate,recall rate,the F value,and the edge segmentation accuracy of the change region than the FCN change detection model.(4)Based on the theoretical foundation that the deeper the neural network is,the stronger the ability of feature extraction is,the deep U-type neural network remote sensing image change detection model is constructed,which increases the depth of U-type neural network remote sensing image change detection model.At the same time,considering the degradation of deep neural network model,residual learning unit is introduced to improve the deep U-type neural network remote sensing image change detection model,and the deep residual U-type neural network remote sensing image change detection model is constructed.The residual block structure makes the network layers deeper while effectively avoiding model degradation and gradient disappearance.It can extract more advanced and complex features.At the same time,the network fuses context features and synthesizes multi-dimensional features,which further improves the prediction effect.