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

Synthetic Aperture Radar(SAR)is an active imaging radar.Compared to optical imaging systems,SAR has full-time,all-weather ground observation capability.Therefore,SAR is widely used in ground monitoring missions.In recent years,with the rapid development of cities and city constructions,the detection of urban areas is particularly important in the task of ground monitoring.Therefore,the detection of built-up areas in SAR images is of great importance.In high resolution SAR images,due to the special imaging mechanism,the effects of layover,dihedral corner reflection,shadow and specular reflection will appear in the built-up areas during the imaging process.For such reason,built-up areas presenting rich structural information in the SAR image,which provides the basis for the extraction of image building areas.In this paper,we proposed a series of algorithms for the detection of built-up areas in SAR images by combining with the powerful feature extraction and representation capabilities of convolutional neural networks.The main research contents are as follows:1.A detection method for built-up areas which is based on structured prediction convolutional neural network is proposed.In the conventional method of the detection of the built-up areas which based on convolutional neural networks,images are processed on the block level.Therefore,the edge representation of built-up areas is not accurate,and the category correlation information between pixels is not fully utilized.In consideration of such problem,we introduce the structured prediction convolutional neural network for the detection of the built-up areas.In addition,we improved the network structure to make the network gain multi-scale feature expression ability,which leads to the improvement of the detection result of the built-up areas in SAR images.2.A detection method for built-up areas which is based on attention mechanism of convolutional neural networks is proposed.Aiming at the problem that the detection rate and the false alarm rate are hard to be optimized simultaneously due to the speckle noise and the surrounding environment in the SAR image,we employed the weighted cross entropy loss function to obtain the detection results of high detection rate and low false alarm rate respectively.Then,the attention mechanism is employed to merge the results with high detection rate and low false alarm rate to obtain the final result.3.A detection method for built-up areas which is based on adversarial networks is proposed.When training the segmentation networks,the cross-entropy loss function is often optimized for the category information of each pixel,which makes it difficult for the network to learn the high-level information of the image(image space consistency,etc.).Aiming at this problem,this paper introduces the adversarial networks.In the process of segmentation network and discriminant network's training optimization,the segmentation network learns the high-level information of the images from the feedback of adversarial network,which improve the detection accuracy.All of the above work verified the effectiveness of the algorithm on high-resolution TerraSAR-X images.