Research On Water Segmentation Method Of Near Infrared Remote Sensing Image Based On Improved U-Net

Water bodies are the resources on which humans and all kinds of organisms on the earth depend for survival.Fast and effective identification and segmentation of water bodies is an important way to grasp the spatial distribution of water resources.The use of remote sensing technology for the acquisition of water information has important characteristics such as fast,real-time,rich,and accurate.The water body in the nearinfrared band has a strong absorption capacity for light,so the gray value presented in the image has a large gap with other features.Using this band information can clearly reflect the boundary between water and land.Compared with traditional segmentation methods,deep learning technology can extract deeper and abstract information.Therefore,this paper proposes a thermal infrared image water body segmentation method based on deep learning,and improves the U＿Net image segmentation method to improve the accuracy of water body segmentation.The specific contents of the paper and the problems to be solved are as follows :(1)As a classical segmentation model,U-Net has a relatively simple model structure.In the process of skip connection,the computing resources of information allocation with different importance are the same,resulting in no emphasis on feature extraction.In the process of propagation,the problem of network degradation that may occur is not considered,and there is room for optimization.This paper proposes a U＿Net model that integrates attention mechanism and residual module.In the process of jump connection,SAM spatial attention and attention gate are added to adaptively allocate the information weight of jump connection.Residual connections are added to the convolution process to optimize the problem of network degradation during propagation.Through quantitative analysis of ablation experiments and visual analysis of segmented images,it is proved that the introduction of the three modules has better performance in the near-infrared water segmentation task than the original U-Net model,and the Io U is improved by 3.76 %.(2)The U-Net model with attention mechanism has achieved good accuracy and detection effect in water detection of near-infrared images,but there are still some problems of false detection of shadow parts,missed detection of small targets and gradient disappearance.In view of the above situation,this paper proposes a multi-task assisted method to solve and optimize the corresponding problem : optimize the shadow detection problem through the RGB recognition task;the problem of missing detection of small targets is optimized by small target recognition task,and the problem of gradient disappearance in the model is optimized by deep supervision task.Through ablation experiments and comparative experimental analysis,it is proved that the introduction of three types of auxiliary tasks has a positive effect on the detection effect of the model,and Io U is increased by 3.64 %.Compared with the classical model of image segmentation,the model in this paper has the best comprehensive performance,which is 1.04 % higher than the best U-Net++in Io U.