Sea-land Segmentation Of High Resolution Remote Sensing Image Based On Residual Structure And Spatial Pyramid Pooling

Sea-land segmentation of remote sensing images is a basic work of coastline information extraction.High-resolution remote sensing images have high imaging quality,are characterized by rich texture information,wide coverage and short revisit time period,and are widely used in agriculture,forestry and environmental protection.The efficient and accurate pixel-level sea-land segmentation of high-resolution remote sensing images is of great significance for the research on the monitoring of coastline dynamic changes and the analysis of the macro-trends of shorelines and beaches.When traditional sea-land segmentation method performs sea-land segmentation on high-resolution remote sensing images,it ignores the relationship between adjacent pixels and contextual semantic information,which may easily lead to misclassification of salt water,breeding ponds and other coastal water bodies and seawater with high sediment concentration.When the contrast between the land and the sea is not large enough,the phenomenon that the boundary between the sea and the land is blurred is likely to appear.In response to the above problems,this paper proposes a deep neural network method RSNet based on residual structure and spatial pyramid pooling.RSNet method uses encoding-decoding structure and end-to-end method to achieve high-resolution remote sensing image sea and land segmentation.RSNet mainly has two innovations:1)For the misclassification of coastal water bodies such as salt pans and breeding ponds and seawater with high sediment concentration,RSNet uses the optimized Residual Convolution Block proposed in this paper ORC_Block)Extract image details and semantic information.The optimized residual convolutional block introduces batch normalization operations and double jump connections to alleviate the gradient disappearance problem,improve network convergence speed,improve network learning and generalization capabilities,and improve misclassification of coastal water bodies.2)In order to obtain more accurate land and sea segmentation boundaries,RSNet adds an intermediate layer to achieve multi-scale feature fusion based on the traditional encoding-decoding structure.Its core structure is the multi-scale feature fusion block based on spatial pyramid pooling proposed in this paper.(Multi-scale feature fusion block,MF_Block).MF_Block can fuse the semantic feature information of different levels of the encoder to improve the difficulty of recovering the detailed information of the feature map in the upsampling operation.At the same time,the global context information can be obtained by expanding the spatial pyramid pooling,and the sea and land features of high-resolution remote sensing images can be fully learned.Clear and accurate boundary dividing land and sea.In order to verify the effectiveness of the sea-land segmentation method proposed in this paper,the paper used 12 remote sensing images of the coastal zone of Lianyungang City,Jiangsu Province,China,taken from 2017 to 2019,as the training and test images.Experimental results show that the RSNet method proposed in this paper can effectively improve the misclassification of coastal land water bodies,ports and other features,and make the boundary between sea and land more accurate.Compared with traditional image segmentation methods such as NDWI and current advanced deep learning methods such as PSPNet and SegNet,RSNet has better segmentation performance in land and sea segmentation tasks.