Research On Building Detection Technology Of The Highresolution Remote Sensing Images

The effective application of the remote sensing images can provide significant information support for the geographic information monitoring,scientific management and policy-making.As the remote sensing images gradually possess the characteristics of high spatial resolution,high spectral resolution and high time resolution,it has become a key problem that how to make use of these images in the application fields.Whereas,the intelligent interpretation of remote sensing buildings is one of the significant tasks,especially for the building detection and change detection.Those technologies can be widely used in the national great demands,such as urban construction planning,disaster survey,land cover change,and resource management.Therefore,this dissertation mainly focuses on the technologies of building detection and building change detection of the high-resolution remote sensing images.Although these technologies have achieved remarkable achievements,there are still some problems and deficiencies.1)The object's semantic lack problem of the large-size remote sensing image partition.2)The weak structural information problem of the clustered buildings with complex backgrounds.3)The insufficient robustness problem of the change detection triggered by the large variance between the multi-temporal images.For the problems and deficiencies mentioned above,this dissertation performs corresponding research.The main contributions and innovations are as follows.1)To address the object's semantic lack problem of the large-size remote sensing image partition,this dissertation proposes a building detection algorithm based on the rotation correction mechanism.In order to support the tilt correction of the buildings,a tilt angle estimated method is established,by using linear edge detection and statistical histogram.At the same time,an evaluation criterion is defined to calculate the cost of building partition.The experimental results show that the proposed method is beneficial to relieve the object's semantic lack problem and reduce the difficulty of building detection.Hence,the proposed method can improve the robustness of the object detection model.The proposed method achieves 0.517 of AP and 0.862 ofAP₅₀ on the building object detection dataset,respectively the best accuracy result.2)To address the weak structural information problem of the clustered buildings with complex background,this dissertation proposes a clustered building detection method based on the density saliency.In order to mine the dense and cellular structure information of the clustered buildings,this dissertation particularly establishes a density saliency guided method.To explore wider applications of the remote sensing images,this dissertation proposes a new estimation method to predict the population carrying capacity of specific areas,by combining the detection results and the saliency heatmaps.Besides,based on the images of Gaofen-2,this dissertation annotates a clustered building detection dataset named CBDD with rectangle bounding boxes.The experimental results show the proposed method is effective for the clustered building detection.The proposed method achieves 0.410 of AP and 0.797 ofAP₅₀ on the self-built dataset,respectively the best accuracy result.3)To address the insufficient robustness problem of the change detection because of the large variance between the multi-temporal images,this dissertation proposes a new framework named multi-sensory pathway network.This framework is inspired by the parallel processing mechanism of the human visual information,and promotes the technology exploration of the bionic and explainable network.Specifically,the framework utilizes three diverse but related sensory pathways,and designs two fusion strategies,namely average fusion and maximum fusion.Two public change detection datasets are utilized in the experiments,and the results demonstrate the effectiveness and robustness of the proposed method.The proposed method respectively achieves the best F1 scores with 84.55%and 88.14%on the two testing dataset.Based on the data characteristics of the high-resolution remote sensing images,this dissertation deeply analyzes the physical attributes of the building objects.Through the organic combination of computer vision technology and deep neural network technology,this dissertation also proposes several promising schemes of the building detection and change detection with strong practicability and robustness.These works can provide scientific theoretical exploration and reliable technical support for establishing a China's comprehensive application platform of the remote sensing information.