Discovering Potential Illegal Constructions Within Building Roofs Using Semantic Segmentation And Object-based Change Detection

Potential illegal construction(PIC)inspection and prevention are regarded as routine tasks of urban management.It is of great significance to monitor the temporal and spatial variation of PICs in time and accurately by means of remote sensing.PICs in high spatial resolution remote sensing images can be located by information extraction and change detection techniques.In this paper,unmanned aerial vehicles(UAV)images are used to study PICs extraction.The main works are as follows:(1)Automated sample generation method is designed based on UAV images and DSM.Extract the building contours from the DSM images by using the canny operator and find the internal points of buildings through the relative elevation of the neighborhood.Then calculate the visible light vegetation index based on the image spectral feature to mask the vegetation areas in the image.Then training sample library is formed through data argumentation.This method can quickly and effectively provide initial building samples.(2)A modified UNet semantic segmentation method based on multi-scale depth information fusion(MSDFUNET)is proposed for building extraction.The MSDFUNET model takes multi-scale depth images as input data,adds atrous spatial pyramid pooling(ASPP)structure into original UNet network to acquire multi-scale feature maps,and outputs recognition results by overlying multi-layer prediction results.Compared with common semantic segmentation networks,the proposed method can effectively extract buildings with different types and scales,reduce the false extraction and the confusion with bare lands and vegetation,and improve the roof extraction accuracy.(3)A PICs detection method based on semantic segmentation and object-based change detection is proposed.Image segmentation with roof-constrained is used to obtain sub-roof primitives and a two-step change detection scheme is then implemented for PIC detection,which includes building extraction,roof-constrained segmentation and object-based change detection.A PIC detection system including sample generation,deep learning,image segmentation,change detection and other functional modules is developed.Experimental results show that the proposed method can effectively improve the PIC detection accuracy.