Research On Change Detection Algorithm Of Remote Sensing Image Based On Deep Learning

Remote sensing image change detection technology monitors the changes that occur on the ground through remote sensing images taken by satellites,and can identify the type of change while locating the changed area.It has been widely used in assessing natural disaster,monitoring of illegal construction and other fields.In the early days,this task was mainly completed by manual interpretation,but the labor cost and time cost were too high.Later,a number of traditional remote sensing image change detection methods emerged,but they also have limitations.On the one hand,they have strict requirements on image preprocessing.On the other hand,the quality of image preprocessing directly affects the final result,leading to a lower detection accuracy.With the improvement of remote sensing image resolution and the development of deep learning,remote sensing image change detection methods based on deep learning have far exceeded traditional methods.The thesis first objectively analyzes the research status and difficulties of remote sensing image change detection,divides that into binary change detection problem and semantic change detection problem for research,and proposes improvements and new solutions on the basis of existing methods.The work involved is summarized as follows:1.The research focus of the thesis is the remote sensing image change detection algorithm based on deep learning.In order to facilitate the development of the follow-up content,the basic theoretical knowledge of remote sensing image change detection and the basic theoretical knowledge of deep learning are first introduced.2.In order to locate the change area more accurately,the thesis first proposes a siamese DeepLabv3+ network in combination with the siamese network.This network can solve the problem of binary change detection end-to-end.On this basis,the attention mechanism and flow alignment module are used to improve it.The attention mechanism can capture dense global context information and improve the association between pixels,and the specific implementation uses a recurrent criss-cross attention module.The flow alignment module can learn the semantic flow between features of different resolutions,and the generated offset field can transfer rough features to high-resolution refined features more effectively.In the end,the thesis get an improved siamese DeepLabv3+ network,which can obtain a better binary change detection reslut in remote sensing images.3.In order to more accurately locate the area where the remote sensing image change occurs and identify the type of change,the thesis studies the remote sensing image semantic change detection algorithm.The thesis chooses remote sensing image semantic change detection algorithm based on multi-task learning,and proposes a semantic change detection algorithm based on multi-task learning and the siamese deep Labv3+ network,and have made some improvements to that.First,the multi-task attention network is applied in the base algorithm,which allows each subtask to extract the most suitable features from the backbone network.Then the dynamic weight average of the loss function is introduced,which can dynamically adjust the ratio of the loss function of different tasks during the training process.Finally,the thesis introduces the projecting conflicting gradients method,which resolves conflicting gradients through projection during training.The accuracy of the network is improved by the above improvements.The experimental results show that the network finally proposed in the thesis has a better effect in the task of semantic change detection in remote sensing images.