The Detection Of Built-up Areas In High-Resolution SAR Images

Compared with ordinary optical image systems,Synthetic Aperture Radar(SAR)is characterized by its ability to perform observations regardless of time and weather.Because of its strong persistence and strong anti-interference characteristics,it is widely used in the field of earth observation,especially in urban planning applications,in which the extraction of construction area is even more important.In the SAR image,the overlaid,dihedral and other effects of the built-up area itself make the overall area have strong structural features,which also provides a strong research entry point for the follow-up research.Because of the problem of interference of speckle noise and instability of the underlying features in the SAR image,this paper proposes a detection algorithm based on the convolutional neural network with weighted input and a detection algorithm based on the improved fully convolutional neural network to detect built-up areas in SAR images.The main content of the study is as follows:1.The detection method of built-up areas based on the convolutional neural network with weighted input.In this paper,the convolutional neural network is used to detect built-up areas,the process of the Gaussian weighting reduces the effect to detect results by the edge structure of the image block.In addition,the pixel-to-pixel detection process is implemented by the voting decision method,and then the detection accuracy is improved while avoiding the difficulty of representing the features of built-up areas in the SAR image,and an end-to-end unified detection framework is formed.Finally,the effectiveness of the algorithm is verified on the high resolution TerraSAR-X image.2.The detection method of built-up areas based on the improved fully convolutional neural network.In this paper,a fully convolutional neural network is proposed to implement the detection of built-up areas in SAR images,and the following optimizations are made:(1)A context network is proposed.The context network is added to the original network to improve the receptive field and improve the detection accuracy.This is because in the training process,when using transfer learning to reduce the data requirements and training difficulties,shallow convolutions are trained using low resolution images,the semantic segmentation effect on high resolution SAR images is limited.Due to the limitation of the receptive field,the detection results are inconsistent in the local blurred region,which affects the final detection result.(2)Add the dense conditional random field as the recurrent neural network.Thus,it can strengthen the connection between the pixels of the detection result,improve the detection accuracy and form an end-to-end,pixel-to-pixel detection framework.Finally,the effectiveness of the algorithm is verified on the high resolution TerraSAR-X image.