Research On Road And Farmland Extraction From High Resolution Remote Sensing Images With Shape Prior

Roads and farmlands are common ground features in remote sensing images,which play a major role in monitoring emergency response,road network planning,agricultural insurance,land and resources planning and other fields.With the rapid development of remote sensing technology,the spatial resolution of optical remote sensing satellite images continues to increase,providing detailed spectral and spatial structure information for the precise extraction of roads and farmlands.In recent years,deep neural networks have repeatedly achieved good results in the field of image processing,and have gradually been applied to the task of extracting roads and farmlands from high-resolution remote sensing images.In the task of road extraction,the road is usually in the shape of a slender strip,with a large distribution span and a small width.Local detail information and global spatial distribution information are likely to cause conflicts.At the same time,the road has the characteristics of continuous network distribution,while remote sensing imaging trees,shadows,etc.will block the road at time,causing the road network in the image to be discontinuous.In the task of farmland extraction,the farmland is in the shape of regular polygons.In areas where farmlands are relatively densely distributed,natural boundaries such as fine ridges are easily covered,causing farmlands to stick together,and the internal features of the farmlands are relatively consistent.Different farmland areas will have different characteristics due to different crops,which will cause difficulties in the extraction of farmland types.Aiming at the above extraction difficulties,this thesis proposes two extraction algorithms with shape prior and implements software design.The main research contents of this thesis are as follows:1)To solve the problem of large road span but small proportion and occlusion,this thesis proposes multi-path feature aggregation network road extraction method with strip shape.This method uses multiple convolution kernels of different sizes to extract the local detail information of strip-shaped roads in different orientations,and then aggregates the multi-path features to obtain the global spatial distribution information of strip-shaped roads.At the same time,multiple samples with increasing sampling rates are used.Dilated convolution extracts road context information to predict occluded roads through dense connections.Through comparative experiments,it can be concluded that the method has good effect on the overall and partial extraction of roads,and can identify the occluded roads.2)Considering the problem of easy adhesion and inconsistent characteristics of farmlands,this thesis proposes boundary guided network farmland extraction method with polygon.This method uses the spatial relationship constraints of the farmlands and boundaries to extract them at the same time,and combines the linear boundary information of the polygonal farmlands to complement the extracted boundary to further separate the cohesive farmlands.In addition,the spatial relationship between boundary and normalized difference vegetation index is used to eliminate the interference of non-farmland featurese,and the farmlands with different features are identified by the consistency of features within the farmlands.The comparative experiment results show that the farmlands extracted by this method have less adhesion,and the ability to recognize farmlands with different characteristics is stronger.3)To meet the actual road and farmland extraction needs,this thesis uses the Py Qt framework to develop software functions,uses automatic extraction algorithms and manual correction to accurately extract roads and farmlands,and counts the width,length,area and number of farmlands.Such information can provide important supporting basis in applications such as emergency route planning and agricultural insurance loss determination.