Research On Semantic Segmentation Method Of Remote Sensing Image Based On Deep Learning

Remote sensing has been widely used in various fields,and its interpretation is a fatal part.Traditional interpretation methods have become restrictive factors for further application of remote sensing due to their low segmentation accuracy,slow efficiency,and poor generalization ability.Recently,although deep learning methods that provide pixel-level classification provide a large number of models for image segmentation,remote sensing images contain rich spatial and spectral information,so deep neural network models built for life scenes cannot be used well for remote sensing images.In order to solve the problems of traditional methods and improve the above status,this paper designs Dense Block with holes by using dilated convolution and dense connection,proposes a model fusion method integrating Dense Net,U-Net,and Denconv Net.Based on this method,this paper proposes an end-to-end semantic segmentation model.On the Potsdam dataset,the study trained this model and evaluated this method.Compared with U-Net,this model improves the evaluation indicators PA,mPA,mIoU by about 11.1%,14.0%,and 13.5%,respectively.Its segmentation speed is approximately 1.18 times that of U-Net,while the model size is only 59.0% of U-Net.Besides,This study also verified the effectiveness and limitations of the fully connected conditional random field to further improve the segmentation results of this model.Compared with advanced semantic segmentation models,this model owns superior performance at the preceding level.Experimental results show that using Dense Block to extract primary features and reduce network layers is an effective and superiors U-Net improved method for remote sensing image segmentation.These indicate that model fusion is an excellent way to explore deep learning models to solve classification tasks in specific scenarios.Due to its low cost,flexible maneuverability,and easy operation,agricultural unmanned air vehicles(UAVs)have been widely used in agricultural remote sensing.But the deep learning models of remote sensing image segmentation for agriculture are slightly insufficient,which will hinder the gridding and information management of agriculture.Recently,transfer learning based on “ fine-tuning ” has gradually emerged.Combining knowledge transfer ideas,this study transfers the proposed model and classic semantic segmentation model to the extraction of corn,barley,and flue-cured tobacco in remote sensing images.The proposed model shows effectiveness and good adaptability and robustness.The experimental results show that deep transfer learning can effectively improve the segmentation accuracy and increase the convergence speed by 59.0%.The experiment also finds the negative impact of short-hop connections on deep transfer learning.It means that transfer learning has potential and challenges in agricultural remote sensing image segmentation.This research provides a new artificial neural network model for efficient and accurate interpretation of remote sensing images and enriches the thinking of developing deep artificial neural network models and provides a reference for the segmentation of agricultural UAV remote sensing images using an artificial neural network algorithm.