Intelligent Detection Methods For Building Change In Remote Sensing Image With Sample Imbalance

Building change detection based on remote sensing images is a core problem in government management such as natural resource monitoring,illegal building discovery,and farmland protection.At present,building change detection is in the process of transforming from artificial interpretation to intelligent recognition,and is in the critical stage of manmachine collaboration.Due to the sparse positive samples and unbalanced scene distribution in remote sensing datasets,existing methods have problems such as misrecognition,missed recognition,and geometric shape anomalies.There is an urgent need to develop automated and highprecision building change detection methods.The emergence of artificial intelligence technology provides new opportunities for high-precision building change detection.To this end,this thesis uses theoretical methods such as deep learning,GAN network,imbalance measurement,and domain adaptation,combined with scientific research and practical needs,to systematically study the data expansion method,training method and postprocessing method.It provides a complete theoretical method support for the detection of building changes in remote sensing images under the condition of unbalanced data set samples.The main research content of this thesis includes:(1)Facing the problem of sample balance and scene balance in remote sensing image datasets,this thesis proposes a building change detection dataset extension method that takes both foreground and background into consideration.Specifically,according to the data distribution of foreground and background,we first count the changes of buildings in the data set and the remote sensing scenes where they are located,and increase the number of positive samples and scarce backgrounds.Then,relying on the convolutional network feature visualization technology and reusing the feature map in the model to select the positive sample generation area,so that a reasonable location label can be obtained without manual intervention.In the end,the high-quality virtual change samples are generated through the confrontation generation network,the ratio of positive samples to scarce background samples in the data set is increased,the imbalance problem of the data set is improved,and the goal of providing a better data set for model training.(2)Aiming at the problem that the current building change detection method ignores the imbalance of precision and recall in the training process,an adaptive optimization method for building change detection based on the balance of precision and recall is proposed.Specifically,a method for dynamically measuring the degree of indicator imbalance during training is firstly designed,which is based on the modeling of precision and recall during training,and realizes the measurement of indicator imbalance during training.Then,the training process is fed back according to the measurement results,and the sampling strategy and loss function weight are dynamically adjusted to realize dynamic,accurate,and adaptive parameter adjustment while maintaining the stability of model training,and finally improve the recognition ability of the model.(3)Aiming at the problem that the current building change detection method relies too much on the feature extraction ability of the model and ignores the prior knowledge of building geometry and topology information,we propose a priori knowledge-guided building change anomaly map suppression method.Specifically,the prior knowledge of the rich geometric semantics and topology of the building is integrated with the deep learning method,and the geometric semantics of the building in the feature space and the output space are aligned in the model training stage to improve the feature representation ability of the model and reduce the Anomaly blobs in prediction results.In the post-processing stage of the prediction results,geographical knowledge such as area and topology is used to extract abnormal patterns of building changes,and a twin network with object-scene context synthesis capability is used to realize the postprocessing of abnormal patterns and further reduce abnormal patterns.By optimizing the training function and post-processing the results,the appearance of abnormal spots is suppressed,and the integrity and rationality of the spots in the prediction results are improved.(4)We have developed an intelligent system for detecting building changes in remote sensing images.Based on big data management and remote sensing image building change detection methods,massive data management and building change detection and analysis are realized.The system serves multiple remote sensing monitoring applications such as cultivated land protection and urban management law enforcement,and provides data support for government management.Finally,we summarize the paper,condense the innovation points of the paper,and look forward to some issues to be further studied in the future.55 pictures,19 tables,and 222 references.