Remote Sensing Recognition Method Of Field Road Based On Deep Learning

Field roads are traffic hubs connecting residential and farmland plots,which play an important role in meeting farmers’ access,agricultural machinery cultivation and agricultural products transportation,and their rapid and accurate identification can provide basic data support for the assessment of farmland construction quality and effect evaluation.Previous road identification methods are mostly oriented to high-level roads with wide road surface and obvious spectral features,while field roads have typical features such as narrow road surface and easy to be shaded by vegetation,leading to problems such as difficult to accurately identify roads and existence of broken points when identifying field roads.To this end,this paper takes Dingzhou City of Hebei Province in the northern plain as the case area,uses the GF-2 remote sensing image as the basic data source,and carries out the research of field road recognition method based on deep learning by combining the characteristics of field roads,in order to provide a method reference for the recognition of fine and narrow roads.The main research contents and research results of this paper are as follows:(1)A field road scene object division method is constructed.In view of the phenomenon of mixing the field roads with other roads in the building area,and combining the difference of their distribution areas,the research adopts the object-oriented method to obtain the classification results of the farmland area and the building area.Firstly,the multi-scale segmentation method is used to segment the images layer by layer;then the feature parameters of different objects are calculated from spectral and geometric features;finally,the threshold method is used to achieve the classification of farmland and building elements by combining the feature differences of different objects: the total land area of Dongting Town,Dingzhou City is 49.17 km~2,and after the three-layer segmentation and classification process,the area of farmland area is about 42.01 km~2 and the area of building area is about 7.16 km~2.(2)Field road recognition models based on deep learning are studied and compared.Considering the typical characteristics of narrow field roads,the U-Net and D-Link Net deep learning networks with better performance in detail information retention are selected,and the applicability of both in field road recognition is tested.Firstly,according to the input data requirements of the deep learning network,the sample data of field roads were drawn by combining manual labeling and data enhancement techniques;then the model was trained,validated and tested;then the trained model was used to realize the recognition of field roads in Dongting town,and the recognition results were evaluated by using the accuracy evaluation model,and the results showed that the D-Link Net network performed better in the recognition of field roads;finally,the recognition results were optimized by combining the results of field road scene object division,and the final field road recognition result accuracy F1 score was improved by 2.22%.(3)The method of field road break detection and repair is constructed.For the problem of local breakpoints in the roads identified by deep learning method,we proposed the technical process of field road breakpoint repair based on "road refinement-breakpoint detection-breakpoint repair".Firstly,the tensor voting method is used to refine the road,and then the eight-domain algorithm is used to detect the road breakpoints,and the local connection method is used to repair the breakpoints according to the characteristics of the road breakpoints,which results in better continuity and more complete topological features of the repaired field roads.