Research On Automatic Extraction Alogorithm Of Airport Runway Area Based On High Resolution SAR Image

Target extraction is a hot issue in the field of high resolution Synthetic Aperture Radar（SAR）.Automatic extraction of airport runway from high resolution SAR image has important research significance in military and civil fields.However,distinguishing the airport from the surrounding objects remains a challenging problem in SAR images.In this paper,the semantic segmentation method in deep learning is used to extract the runway area of the airport.The main research work is as follows:（1）Aiming at the problem that the extraction of airport runway from high-resolution SAR images is easily disturbed by background information,this paper proposes a Multi-resolution Dense Dual Attention Network（MDDA）architecture to realize the extraction of airport runway from high-resolution SAR images.In view of the large similarity between the target gray level and texture in the background information and the runway area,this paper introduces the dual-attention mechanism to realize feature screening,and adopts the dense connection to strengthen the mutual transmission and fusion of information among features.Firstly,the high-resolution SAR image was sampled by 5 times,and then the dataset was made to make the sample contain as many characteristics of the runway area as possible.Then,the processed images are input into the Multi-resolution Dense Dual Attention Network（MDDA）architecture to realize the learning and extraction of runway features.Finally,five-fold up-sampling is used to obtain the airport extraction results of high-resolution SAR images.The experiments on three sets of different types of high-resolution SAR images show that the mean pixel accuracy（MPA）and mean intersection over union（MIOU）of the airport are up to 0.98 and 0.97 respectively,and the proposed method can realize the high-precision automatic extraction of the airport runway in high-resolution SAR images.（2）Aiming at the problem that the current methods of airport detection are not fast and efficient,this paper proposes a Geospatial Contextual Attention Mechanism（GCAM）of airport from high-resolution SAR images based on the attention mechanism of geospatial context.In view of the lack of public SAR airport datasets at present,small sample datasets can greatly save manual annotation time.This method effectively combines the geoscience domain knowledge,context information and attention mechanism of SAR image analysis to adapt small sample data sets.Firstly,the SAR sample was subsampled by 5 times and then made into a small sample dataset.Then input feature extraction network,which contains the encoder and decoder blocks,contained in the code block to improve backbone ResNet₁01,multiscale pyramid and edge thinning module,coding block input decoding block after extracting multi-scale feature fusion,decoder block to edge thinning decoding to extract the runway area edge features and global features of all modules are concise and less;Then coordinate mapping is used to get the result of runway area extraction.Finally,a binary contour detection method is used to realize airport detection.,which further reduces the training burden of the network.The experimental results show that the MPA and MIOU of the proposed method are up to 0.9850 and 0.9536,the training time of the small sample data set is only 2.25h,and the average test time of the three airports is only 18.15s.Both the accuracy and efficiency of the proposed method are superior to the current dominant semantic segmentation methods,and the fast automatic detection of the airport in the high-resolution SAR image is realized.