CN-122024103-A - Tobacco transplanting machine operation quality detection method based on remote sensing and deep learning

CN122024103ACN 122024103 ACN122024103 ACN 122024103ACN-122024103-A

Abstract

The invention discloses a tobacco transplanter operation quality detection method based on remote sensing and deep learning, which relates to the technical field of agricultural machinery operation quality detection, and comprises the following steps: identifying the characteristics of seedling holes and missed planting holes through a target detection model, wherein the target detection model comprises a second-order attention mechanism SOCA module and a SimSPPF module, the second-order attention mechanism SOCA module captures the correlation of multi-scale characteristics in video images, the SimSPPF module fuses the multi-scale characteristics, background noise information is removed through twice matching of the characteristics, the seedling holes which are effectively sown during operation of the tobacco transplanting machine and the missed planting holes which are not effectively sown are positioned and extracted from video data synchronously acquired during operation of the tobacco transplanting machine, and the missed planting rate of the operation of the tobacco transplanting machine are tested through calculation of the effective sowing operation efficiency and the missed planting rate of the operation of the tobacco transplanting machine, so that the operation quality of the tobacco transplanting machine is quantitatively detected while the detection precision of the seedling holes and the missed planting holes is improved.

Inventors

GAO LEI
SU RUI
CHEN DU
YU BEI
WANG LING
GUO SHUJIN
ZHU ZHENTAO

Assignees

中国农业大学
中国烟草总公司湖南省公司

Dates

Publication Date: 20260512
Application Date: 20260203

Claims (8)

1. The tobacco transplanting machine operation quality detection method based on remote sensing and deep learning is characterized by comprising the following steps of: acquiring remote sensing video data of a field area of the tobacco transplanter in a field operation period; Inputting the remote sensing video data into a target detection model based on a YOLOv5s model framework, wherein the target detection model comprises a second-order attention mechanism SOCA module and a SimSPPF module, the second-order attention mechanism SOCA module identifies acupoint plots in the remote sensing video data frame by frame and extracts multi-scale features including morphology, color and texture of each acupoint plot, the SimSPPF module fuses the multi-scale features of each extracted acupoint plot based on time scale, and identifies seedling-bearing and load-missing cavities in each frame of remote sensing video data based on the fused features and preset classification logic and outputs a detection frame of acupoint contours corresponding to each seedling-bearing and load-missing cavity; a multi-target tracking algorithm based on DeepSORT algorithm allocates a tracking ID to each frame of the remote sensing video data, namely the detected frames with the seedling holes and the detected frames with the missing holes, and carries out cross-frame correlation based on time sequence; Setting a virtual counting line in the remote sensing video data, counting the types and the quantity of tracking IDs crossing the virtual counting line based on the tracking IDs of the seedling-bearing hole detection frame and the missing-carrying hole detection frame in each frame, and calculating the total number of accumulated marked seedling-bearing holes and missing-carrying holes in the remote sensing video data through a line crossing counting logic; And comprehensively calculating the missing planting rate of the tobacco transplanting machine according to the total number of seedling holes and the missing planting holes in the remote sensing video data and combining the duration of the tobacco transplanting machine in the field operation period and the preset reference planting rate corresponding to the duration, so as to quantitatively evaluate the operation quality of the tobacco transplanting machine.
2. The tobacco transplanter operation quality detection method based on remote sensing and deep learning according to claim 1, wherein the remote sensing video data are collected in a field operation area through an aerial unmanned aerial vehicle, a route of the aerial unmanned aerial vehicle covering all the field operation areas is set, and remote sensing video data of continuous operation of the tobacco transplanter in the field operation area are synchronously obtained through the aerial unmanned aerial vehicle.
3. The tobacco transplanting machine operation quality detection method based on remote sensing and deep learning according to claim 1 is characterized in that remote sensing video data are input into a target detection model, the target detection model comprises a second-order attention mechanism SOCA module, the second-order attention mechanism SOCA module captures correlation of multi-scale features in the remote sensing video data through a covariance matrix, multi-scale features including morphological features, color features and texture features of each acupoint block in a field area are extracted, feature values of the morphological features, the color features and the texture features of each acupoint block are calculated respectively based on second-order statistics, the feature values of each feature are output through a covariance channel of the second-order attention mechanism SOCA module, and feature values of each feature are normalized, wherein the second-order attention mechanism SOCA module adjusts channel directions of the covariance channels through the second-order feature statistics and maps the relation of the multi-scale features.
4. The tobacco transplanter operation quality detection method based on remote sensing and deep learning according to claim 3, wherein the target detection model further comprises a SimSPPF module, the SimSPPF module fuses the multi-scale features of each acupoint plot, the pooling features corresponding to the multi-scale features are obtained by compressing the channel number of the multi-scale features, the pooling features with the same scale are spliced based on time sequence to obtain a fused feature map of the same scale features, and the SimSPPF module identifies seedling holes and missing holes in each frame of remote sensing video data based on the fused feature map and preset classification logic.
5. The method for detecting the operation quality of the tobacco transplanting machine based on remote sensing and deep learning according to claim 1, wherein the target detection model performs frame selection on tracks of seedling-bearing holes and load-missing holes in each frame of remote sensing video data based on the time sequence of the remote sensing video data according to the extracted seedling-bearing holes and load-missing holes, and sets the boundary of a detection frame to cover and frame the seedling-bearing holes and the load-missing holes.
6. The method for detecting the operation quality of the tobacco transplanting machine based on remote sensing and deep learning according to claim 5, wherein the detection frames with holes and the detection frames with holes missing in each frame of remote sensing video data are input into a multi-target tracking algorithm, a tracking ID is respectively allocated to each independent detection frame with holes and each independent detection frame with holes missing, and the detection frames with holes missing in each frame are subjected to cross-frame association; The multi-target tracking algorithm acquires the central point positions of each seedling-bearing hole detection frame and each missing-carrying hole detection frame between the previous frame and the current frame based on the time sequence of the remote sensing video data, calculates the average value of the same seedling-bearing hole and the same missing-carrying hole corresponding to the current frame and the previous frame as a reference displacement vector, corrects the track of the seedling-bearing hole detection frame and the missing-carrying hole detection frame in the previous frame respectively according to the reference displacement vector, acquires the predicted position of the corresponding detection frame of the next frame, and correlates the predicted position with the characteristics of the seedling-bearing hole detection frame and the missing-carrying hole detection frame in the current frame.
7. The method for detecting the operation quality of the tobacco transplanting machine based on remote sensing and deep learning according to claim 1, wherein the multi-target tracking algorithm introduces an adaptive measurement noise covariance adjustment mechanism based on detection confidence in a DeepSORT algorithm, the confidence quantifies the reliability of the multi-target tracking algorithm, and the adaptive measurement noise covariance adjustment formula is as follows: ; Wherein, the A measured noise covariance matrix in a Kalman filtering algorithm of DeepSORT algorithm, wherein conf is a target confidence score of a detection object, and k is a preset adjustment coefficient; And the measurement noise covariance matrix is subjected to self-adaption adjustment.
8. The method according to claim 1, wherein a virtual counting line is set in the remote sensing video data, the seedling-bearing detection frames and the missing-planting-hole detection frames exceeding the virtual counting line are respectively accumulated by counting logic crossing the virtual counting line, the total number of the seedling-bearing detection frames and the missing-planting-hole detection frames are calculated, and the total number of seedling-bearing holes and missing-planting holes in the remote sensing video data is counted and counted based on the tracking IDs of the seedling-bearing-hole detection frames and the missing-planting-hole detection frames.

Description

Tobacco transplanting machine operation quality detection method based on remote sensing and deep learning Technical Field The invention relates to the technical field of agricultural machine operation quality detection, in particular to a tobacco transplanting machine operation quality detection method based on remote sensing and deep learning. Background Tobacco is an important special economic crop in China, tobacco planting is particularly common in southwest hilly mountain areas in China, but the terrain of the areas is complex, cultivated lands are scattered with zero stars, and the tobacco is mainly planted mechanically, but the mechanization level of transplanting (planting) links is still weak. The method mainly derives from the strict requirements of special agronomy such as tobacco pit-type transplanting, wherein the transplanting is to make a 'top-like' pit with specific size (such as diameter of 8-10cm and depth of 18-20 cm) on the ridge, and create a proper 'micro greenhouse' environment for tobacco seedlings. The harsh agricultural standard provides extremely high requirements for a cellar opening device, a seedling throwing mechanism and cooperative control of the cellar opening device and the seedling throwing mechanism of the transplanter, so that the problem of excessively high miss-planting rate of the existing transplanter in field operation is caused, the large-scale popularization and application of the existing transplanter are severely restricted, and the detection of the operation quality of the tobacco transplanter is very important. In the research and development testing stage of the transplanting machine, the key performance index of the miss planting rate is evaluated rapidly and accurately, and the method is a core link for finding out the defects of the machine, optimizing the design scheme and improving the operation quality. In recent years, with the development of deep learning, the deep learning technology is gradually applied to the aspect of tobacco transplanter operation quality detection, when the existing deep learning model is combined with unmanned aerial vehicle video images to detect the tobacco transplanter operation quality, when detecting and identifying planting points, the accuracy of identifying planting point plots and the calculation accuracy of missing planting rate are insufficient due to interference of field bare soil, white coating, green weeds and other complex background environments in background noise, so that the accuracy of detecting tobacco transplanter operation results is reduced. Disclosure of Invention 1. Technical problem to be solved Therefore, the invention provides a tobacco transplanter operation quality detection method based on remote sensing and deep learning, which can solve the problem of accurately identifying the planting points and types of the tobacco transplanter from field background noise. 2. Technical proposal In order to achieve the purpose, the invention provides the technical scheme that the tobacco transplanting machine operation quality detection method based on remote sensing and deep learning comprises the following steps: acquiring remote sensing video data of a field area of the tobacco transplanter in a field operation period; Inputting the remote sensing video data into a target detection model based on a YOLOv5s model framework, wherein the target detection model comprises a second-order attention mechanism SOCA module and a SimSPPF module, the second-order attention mechanism SOCA module identifies acupoint plots in the remote sensing video data frame by frame and extracts multi-scale features including morphology, color and texture of each acupoint plot, the SimSPPF module fuses the multi-scale features of each extracted acupoint plot based on time scale, and identifies seedling-bearing and load-missing cavities in each frame of remote sensing video data based on the fused features and preset classification logic and outputs a detection frame of acupoint contours corresponding to each seedling-bearing and load-missing cavity; a multi-target tracking algorithm based on DeepSORT algorithm allocates a tracking ID to each frame of the remote sensing video data, namely the detected frames with the seedling holes and the detected frames with the missing holes, and carries out cross-frame correlation based on time sequence; Setting a virtual counting line in the remote sensing video data, counting the types and the quantity of tracking IDs crossing the virtual counting line based on the tracking IDs of the seedling-bearing hole detection frame and the missing-carrying hole detection frame in each frame, and calculating the total number of accumulated marked seedling-bearing holes and missing-carrying holes in the remote sensing video data through a line crossing counting logic; And comprehensively calculating the missing planting rate of the tobacco transplanting machine according to the total number of seedling h