Abstract: The problem of agricultural labor shortage and high cost has become increasingly prominent, and the research and development of unmanned agricultural machinery technology are imperative. In order to meet the needs of unmanned agricultural machinery for high-precision farmland maps, a method of high-precision farmland map construction was proposed. Since the obstacles, facilities, and boundaries were complex infield, it was different to construct the high-precision map in the field than on the road. The main work of this paper includes five tasks: 1) defining different map layers by agricultural features; 2) collecting the field photograph via UAV in Heinanzhai Town, Miyun District, Beijing; 3)constructing the dense point cloud and 3D model through aerial slice stitching and processing; 4) annotating the dense point cloud data with different layer types to construct final maps; 5) The feature recognition, semantic definition and map annotation are carried out based on Autoware. For the different levels of unmanned agricultural machines and various scenarios needs, the high precision maps should be designed and implemented in layers, a five-layer high precision map construction was proposed, which included a field boundary layer, static obstacle layer, operational information layer, dynamic perception layer, and brain-like layer. The field boundary layer and static obstacle layer which can satisfy the demand for the tracing operation in a closed environment was built. For the field boundary layer, the boundary lines, the entrance and exit data structure, and the corresponding topological relationships were considered. For the obstacle layer, shapes such as curves, polygons, circles, and rectangles were used to describe various obstacles with different geometric properties. The tests were conducted in clear weather, and 511 images were obtained from the aerial survey. All the aerials were positioned in a fixed position. 12 target inspection points were evenly arranged in the field, entrance, exit, and roadside before the flight, and the precise location was obtained by GNSS equipment. The results showed that the absolute accuracy of the in-field high-precision map constructed through the method proposed in this paper was better than 7 cm, and the variance of the coordinate difference of 12 checkpoints was less than 2 cm. The average error and standard deviation of line elements were better than 2 cm; the average error rate of surface elements was better than 0.2%. As the prior knowledge of unmanned agricultural path planning and control systems, high precision maps reduced the network bandwidth without redundant sensors, reduce the requirements of computing capacity and data processing difficulties, not only meet the requirements of unmanned agricultural machines for automatic driving and field operations but also provide prior information for farmland management, path planning, and perceptual assistance.