Search

CN-122023501-A - Farmland boundary construction and operation area calculation method, device, system and storage medium

CN122023501ACN 122023501 ACN122023501 ACN 122023501ACN-122023501-A

Abstract

The invention discloses a farmland boundary construction and operation area calculation method, device, system and storage medium, which comprise the steps of obtaining farmland images and GNSS/IMU metadata thereof, extracting semantic features for distinguishing crops, ridges and non-cultivated lands and capturing detailed features of fine trend of the ridges according to the farmland images and the GNSS/IMU metadata thereof, simulating walking and dotting of a surveyor along the ridges according to the semantic features and the detailed features to obtain ridge polygon vertexes, obtaining offset fields from each pixel to geographic coordinates according to the GNSS/IMU metadata, obtaining geographic polygons according to the ridge polygon vertexes and the offset fields, obtaining predicted areas according to the geographic polygons, and outputting vector boundaries and credible faces of geographic references based on end-to-end training optimization. By adopting the technical scheme of the invention, the precision of the operation area of the agricultural machinery is greatly improved, and the field map distribution is accurately obtained.

Inventors

  • LIU ZHAOPENG
  • OUYANG BAOHUA
  • Ning Bengpi
  • LIU XIAN
  • GUO ZHIDA
  • LIU MUHUA

Assignees

  • 江西农业大学

Dates

Publication Date
20260512
Application Date
20260211

Claims (8)

  1. 1. A farmland boundary construction and operation area calculation method is characterized by comprising the following steps: S1, acquiring farmland images and GNSS/IMU metadata thereof; s2, extracting semantic features for distinguishing crops, ridges and non-cultivated lands and detail features for capturing fine trends of the ridges according to farmland images and GNSS/IMU metadata; step S3, simulating a surveyor to walk and dott along the ridge according to the semantic features and the detail features to obtain polygonal vertexes of the ridge; S4, obtaining an offset field from each pixel to a geographic coordinate according to GNSS/IMU metadata; s5, obtaining a geographic polygon according to the ridge polygon vertexes and the offset field; s6, obtaining a predicted area according to the geographic polygon; And step S7, based on end-to-end training optimization, outputting a vector boundary and a trusted surface of the geographic reference.
  2. 2. The method of claim 1, wherein the loss functions of the end-to-end training optimization include a vertex sequence loss L_vertex, a geographic coordinate regression loss L_geo, and an area consistency loss L_area.
  3. 3. The method for constructing a boundary of a farmland and calculating an operation area according to claim 2, wherein in step S2, semantic features for distinguishing crops, ridges and non-cultivated lands and detailed features for capturing fine orientations of ridges are extracted through a modified BiSeNet according to farmland images and GNSS/IMU metadata thereof.
  4. 4. A farmland boundary construction and work area calculation apparatus, comprising: The first processing module is used for acquiring farmland images and GNSS/IMU metadata thereof; The second processing module is used for extracting semantic features for distinguishing crops, ridges and non-cultivated lands and capturing detail features of fine trends of the ridges according to farmland images and GNSS/IMU metadata; the third processing module is used for simulating a surveyor to walk and dott along the ridge according to the semantic features and the detail features to obtain polygonal vertexes of the ridge; the fourth processing module is used for obtaining an offset field from each pixel to the geographic coordinates according to the GNSS/IMU metadata; the fifth processing module is used for obtaining a geographic polygon according to the ridge polygon vertexes and the offset field; The sixth processing module is used for obtaining a predicted area according to the geographic polygon; And the seventh processing module is used for outputting a vector boundary and a trusted surface of the geographic reference based on the end-to-end training optimization.
  5. 5. The farm boundary construction and work area calculation apparatus of claim 4, wherein the loss function of the end-to-end training optimization comprises a vertex sequence loss L_vertex, a geographic coordinate regression loss L_geo, and an area consistency loss L_area.
  6. 6. The apparatus according to claim 5, wherein the second processing module extracts semantic features for distinguishing crops, ridges and non-cultivated land from detailed features for capturing fine orientations of ridges according to the farmland image and GNSS/IMU metadata thereof through the improved BiSeNet.
  7. 7. A farmland border construction and work area calculation system comprising a memory and a processor, said memory having stored thereon a computer program to be run by said processor, said computer program, when run by said processor, performing the farmland border construction and work area calculation method according to any of claims 1-3.
  8. 8. A storage medium having stored thereon a computer program which, when run, performs the method of farmland boundary construction and work area calculation according to any of claims 1-3.

Description

Farmland boundary construction and operation area calculation method, device, system and storage medium Technical Field The invention belongs to the technical field of operation of agricultural machinery, and particularly relates to a farmland boundary construction and operation area calculation method, a farmland boundary construction and operation area calculation device, a farmland boundary construction and operation area calculation system and a farmland boundary calculation storage medium. Background Along with the promotion of accurate agriculture and national agriculture modern construction, accurate mapping of farmland boundaries and accurate statistics of cultivated areas become important bases for agricultural subsidy distribution, land flow right determination and agricultural condition monitoring. Traditional farmland surveying and mapping mainly relies on manually carrying GNSS-RTK equipment to walk and dott along a ridge. The method has the advantages of high precision, extremely low efficiency and high labor intensity, is difficult to operate under complex terrains such as paddy fields, marshes or long-stalk crops in the growing period, and cannot meet the requirement of quickly acquiring the information of a large-scale farmland. In recent years, with the development of low-altitude unmanned aerial vehicle remote sensing technology and deep learning computer vision algorithm, farmland automatic extraction based on unmanned aerial vehicle images becomes a research hotspot. The prior art mainly focuses on classifying farmland orthographic images or single Zhang Hang pictures at pixel level by utilizing semantic segmentation networks (such as U-Net, deepLab series and the like), dividing images into binary masks (masks) of 'cultivated land' and 'non-cultivated land', and then calculating the area through image post-processing. However, despite the progress made in the prior art, the following significant technical problems and disadvantages still exist in practical applications: 1. The prior art generally adopts a step-by-step processing strategy for the severe accumulated error caused by the splitting of the processing flow, namely, a deep learning network is utilized to output a grid segmentation map of a pixel level, then the grid is converted into a vector boundary through the traditional image processing algorithms such as edge extraction, skeletonization, douglas-Peucker polygonal fitting and the like, and finally, the coordinate conversion calculation area is carried out. In the non-end-to-end process, edge noise (such as a sawtooth edge) in the segmentation stage can be brought into a subsequent step, and the polygon fitting process often depends on a manually set threshold value, so that the self-adaption is lacking, the finally generated vector boundary has larger deviation from a real ridge, and the errors are accumulated step by step. 2. It is difficult to cope with the "weak border" recognition in complex farmland environments in actual farmland situations, ridges (borders) tend to be very narrow (only 20-50 cm) and often covered by weeds or occluded by crops (e.g. lodged wheat, rice in growing period). The existing semantic segmentation network mainly relies on color and texture characteristics to carry out pixel classification, so that weed ridges similar to crop colors are easily misjudged as cultivated lands, or field shadows are misjudged as boundaries. Lack of a priori modeling of field geometry (e.g., linearity, closeness) results in extracted boundaries that are often subject to cracking, sticking, or shape distortion. 3. Neglecting imaging geometric distortion, area calculation accuracy is low in existing monocular vision area calculation methods, mostly estimating area simply from the number of pixels multiplied by ground resolution (GSD). This approach assumes that the camera is taking a full vertical downward picture and that the ground is absolutely flat. However, in the actual flight of the unmanned aerial vehicle or in the field collection process of the agricultural machine, pitching or rolling inclination is inevitably caused under the influence of airflow, so that perspective distortion of an image occurs (for example, a rectangular field appears as a trapezoid on the image), and meanwhile, projection errors are caused by topography fluctuation. The prior art lacks a mechanism for performing end-to-end geometric correction by using GNSS/IMU metadata, and the area calculated directly based on the distorted image often has larger deviation. 4. Deep learning networks that lack direct optimization of target mainstream for geometric and area accuracy typically use Cross entropy Loss (Cross-Entropy Loss) or IoU Loss (Dice Loss) when training. These loss functions focus on the accuracy of the pixel classification rather than the accuracy of the geometric position accuracy or final area value of the boundary. This results in a network that may classify well in the central