Research On Farmland Segmentation Method Of UAV Images Based On Semantic Segmentation Networ

With the rapid development of computer vision,semantic segmentation as an important part of the technology has made some achievements in different applications.However,in the farmland segmentation scenario of remote sensing images,the capability of common semantic segmentation methods in restoring the farmland edge and identifying narrow farmland ridges needs to be improved.Meanwhile,the problem of losing target position information by semantic segmentation networks becomes more and more prominent.In this thesis,two semantic segmentation networks are proposed for the above problems,namely,Multiple Attention Encoder-decoder network and Global Feature Fusion network.The main work is as follows:(1)For the problem of remote sensing image farmland segmentation,three datasets of farmland segmentation from UAV images are produced in this thesis.Labelme software is used to label four UAV stitched images.Then,these UAV images are cropped to 9684 labeled images.In this thesis,the cropped images are divided into three datasets according to the different resolutions and different categories.All the datasets are randomly divided into training and testing sets in the ratio of 4:1.(2)To address the problem that common semantic segmentation networks have difficulty in recognizing narrow field ridges in farmland remote sensing images,this thesis uses a Multiple Attention Encoder-decoder network.The network consists of a DPECA-Res Net-50 network,Dual Feature Attention module,and Global-guidance Information Upsampling module.The network achieves fine segmentation of farmland boundary and ridge.The multiple attention codec network is trained using three self-made UAV farmland image datasets and compared with other semantic segmentation networks.The experimental results show that the proposed network in this thesis all obtains the best accuracy results.(3)This thesis uses a Global Feature Fusion semantic segmentation network to address the problem of losing target spatial position information in semantic segmentation networks.This network consists of GP module,global information module,and adaptive feature fusion module.In this thesis,the global feature fusion network is tested using the Barley remote sensing dataset,Self-made multi-classification farmland segmentation dataset,Deep Globe land cover classification dataset,and Aero Scapes aerial semantic segmentation dataset,respectively.The experimental results show that the method proposed in this thesis achieves great progress compared with other semantic segmentation models in terms of the ability to recover object boundaries.