Research On Port Identification Method Based On Deep Learning And Spatial Analysis

As the transportation hub node between land and water,port is not only the distribution center of industrial and agricultural products and materials in import and export trade,but also the supply place in war.It plays a very significant role in economy and military.Due to the diversity of port shapes and complex features,it is difficult to identify ports automatically based on remote sensing images.In view of the difficulty of automatic port identification,this paper proposed a method of port recognition based on deep learning and geospatial analysis in the high-resolution optical remote sensing image.Firstly,this paper analyzed the elements of the port,then the ship-harbor-port progressive recognition model was proposed.The design first used the pattern recognition of the most significant harbor with the most significant features on remote sensing images,and then used spatial clustering and hot spot analysis to synthesize the port area according to the prior knowledge of the basic shape of the port.Secondly,the harbor detection model was trained by the common remote sensing harbor dataset,and three mainstream deep learning target detection algorithms,Faster R-CNN,YOLO v3 and SSD,were used in the training.The test results were generated according to the test data for contrast.The average precision was used for the evaluation of models.The results showed that YOLO v3 is more suitable for the detection of harbors,therefore the detection algorithm was determined.Thirdly,in view of the problem of the small amount of existing harbor samples of the remote sensing,the dataset of the harbor was constructed,and the sample was based on the collections and annotations.The strategy of image augmentation was developed with rotation,flipping,and retracting.The generalization ability of YOLO v3 model was improved.The average accuracy of the model on the test set reached 93.07%.The harbor detection model based on YOLO v3 and sliding window detection strategy was studied to solve the difficulty of detecting large remote sensing images.The non maximum suppression was used to enhance the multi-scale detection ability of the model.The accuracy of the model was improved by 13.69%,and the recall was improved by 34.74%.In the practical application of harbor detection,in the face of losses and errors in complex situations,the harbor of Japan was selected as the experimental area,and Google tiles were used to splice the remote sensing images along the coastline.The bottom features of the image were obtained,and the harbor category and pixel coordinates were calculated.Finally,the harbor points were transformed into geographical coordinates.Getis-Ord Gi*Statistical index were used for hot spot analysis of the harbor points.The classical density clustering method was used to realize the recognition and extraction of port location and range.Then the ports were detected.The proportion of the improved model for port basins recognition reached 82.79%.