Research On Semantic Segmentation Of Remote Sensing Image Based On Encoder-Decoder Networks

With the continuous update of earth observation technology,the data volume of highresolution remote sensing images is increasing rapidly.How to efficiently extract valuable information from massive remote sensing images is an urgent problem for remote sensing practitioners.Semantic segmentation specifies a clear land cover category for each pixel in high-resolution remote sensing images.It has now become the mainstream method for interpretation of terrain information,and it plays a crucial role in the fields of land resource planning,environmental protection,and natural disaster monitoring.Important role.Due to the continuous expansion of application scenarios,traditional image segmentation algorithms that rely on hand-designed features cannot cope with increasingly complex remote sensing image segmentation tasks.Therefore,it is of great significance to study the remote sensing image semantic segmentation method based on deep learning.This thesis focuses on the difficulties of remote sensing image semantic segmentation tasks,and proposes a method to improve the existing depth model.The main research contents are as follows:（1）Aiming at the problem that the accuracy of the segmentation model is reduced due to the rich details of remote sensing image information and the complex background,a semantic segmentation model of remote sensing image based on Ushaped structure is proposed.The residual channel attention module is used to enhance the features that are effective for the segmentation task and suppress the features that are ineffective or small,so as to improve the classification accuracy of the model.Taking into account the differences between features at different scales,a multi-scale fusion module with residual structure is used to fuse the detailed information of lowlevel features with the semantic information of high-level features to improve the segmentation accuracy of the model.Finally,the refined segmentation of remote sensing images is realized.Experiments on public data sets show that the model’s segmentation accuracy has improved significantly.（2）Aiming at the problems of inaccurate target boundary segmentation in remote sensing image segmentation model and poor segmentation effect of small targets,a remote sensing image semantic segmentation model based on spatial pyramid pooling is proposed.First,the deep residual network and channel attention mechanism are used to construct a feature coding network to extract more expressive features.Secondly,the hollow space pyramid pooling module using the dense connection mechanism improves the receptive field of the model and captures more dense multi-scale context information to solve the problem of inaccurate segmentation of the segmentation target boundary.Finally,an efficient encoder is used to fuse low-level detail information and high-level semantic information to solve the problem of small target segmentation.The comparative experiment on the ISPRS Vaihingen dataset proves the effectiveness of the model.