| Being an imperative food crop,rice is widely planted in China,which is of great significance to national food security.With the development of aviation plant protection machinery,the precision pesticides application of rice is getting higher and higher demanded.However,the key equipment of the precision pesticides application of rice in China presently is still lagging behind,and the“advance”cannot be adaptively applied in combination with field information as needed,triggering large pesticide waste,low utilization and environmental pollution;the“afterwards”liquid deposition parameter detection means are extensive and inefficient,which cannot provide timely and accurate data support for the effective evaluation of droplet deposition distribution law.Paddy vacancies formed due to human factors,such as uneven sowing,missed sowing,and excessive fertilization,or natural disasters,such as floods and typhoons resulting in crop vacancies are common phenomena in the rice planting management stage.The characteristics of uniform application of traditional unmanned aerial vehicle（UAV）cannot be competent for chemical control tasks in such paddy,which makes it difficult for pesticide reduction targets.Over the above problems,this paper starts with exploring the acquisition of characteristic information of agricultural UAV in paddy empty area,liquid adaptive variable control,rapid detection of droplet deposition parameters,uses machine vision,deep learning and automatic control technology,combines the existing UAV spray characteristics to construct UAV adaptive variable spray platform and rapid detection system of droplet parameters,and improves the mechanization,informatization and intelligence level of paddy management.Specific studies are as follows:（1）The paddy at tillering stage and long heading stage was utilized as the experimental site to obtain the image of paddy empty area and study the empty area segmentation and identification method.After Lucy-Richardson debouncing and definition preprocessing of blurred images in empty areas,on the one hand,based on the idea of“first segmentation and then identification”,the object-oriented Kmean segmentation support vector machine（K＿SVM）model identified empty areas and optimized the optimal regression algorithm,which improved the recognition accuracy of KSVM classifiers;on the other hand,based on the deep learning full coiler network（FCN）algorithm,the optimal improved structure was explored through different network improvement methods.The experimental results of each algorithm comparison displayed that the improved FCN16s model excellently performed in comprehensive performance indicators.Dynamic network cropping was employed to compress FCN16s,eliminate redundant parameters,reduce the computational load,accelerate model regression,which made it possible for real-time identification of paddy empty areas by airborne embedded platform.（2）Of the key technology research and development of precision pesticides application equipment,a set of visually assisted UAV adaptive variable spray system was developed for the current problems that agricultural UAVs cannot apply drugs on demand following the condition of paddy airspace.Firstly,the key components of the system were selected and designed,especially the airborne sliding rail device is developed for mounting the camera and centrifugal nozzle.According to the derived spacing formula,the distance between the camera and the nozzle can be adjusted with the flight speed,which provided a new solution to reduce the spray error caused by the system delay;then the area ratio model was established,and the spray decision model of the relationship between area ratio p and target flow was derived on this basis;finally,the BP neural network PID algorithm,fuzzy control algorithm and PID algorithm are designed,as well as the indoor simulation and field operation comparison tests carried out.The results of comprehensive test illustrated that the three control algorithms can realize adaptive control,but the target flow deviation range of BP neural network PID algorithm was the smallest,the system response time the fastest,and the spray control more accurate and stable.（3）Targeting at low intelligence and time-consuming and laborious of current droplet parameter detection methods,a set of droplet deposition parameter detection system with high automation was designed,including deposition droplet circulation collection device and upper computer interaction software platform.In the study,the correlation between theoretical frame rate and actual frame rate of image transmission with different resolutions,and the relationship between RSSI and packet loss rate and image transmission distance were tested respectively.The results disclosed that the image resolut ion should not over 307200Pix,and the communication distance should be less than 136 m in paddy environment;the effect of droplet deposition carrier thickness on droplet image acquisition and processing results was tested.The results revealed that the greater the ink thickness,the better the image processing effect;but when the ink thickness was greater than 1×10-2 cm,the achromatic ink presented interference cracks,affecting the droplet detection effect;the enhanced learning Sarsa algorithm was exploited so that the system can adaptively adjust the image contrast parameters with the change of light intensity c.The experimental results demonstrated that the segmentation effect of droplet image processed by light adaptive algorithm was obviously better than that without adaptive algorithm.（4）The received droplet images were preprocessed by grayscale,filtering and maximum internal variance method on the upper computer interaction platform.The droplet area was obtained through the connected area,and the area threshold,shape size threshold and particle size threshold were used to determine the adhesion droplet and pseudo droplet.An improved marker-controllable watershed algorithm was used to segment adhesion droplets,background isolated connection areas,and filter out pseudo droplets;Faster-RCNN,YOLOv3,and Mobile Net＿SSD models were established based on depth neural networks to process droplet images and counted the relevant droplet parameters.The comparative experimental results found that the Mobile Net＿SSD network with width multiplier a of 0.75 and resolution multiplier b of 0.875 had the best performance and better balances the accuracy rate and lightweight of the model;when the performance of Mobile Net＿SSD for different density droplet image detection was tested on the embedded device,it was found that the high,medium and low density droplet images had excellent performance,good adaptability,and the average detection efficiency was as high as 0.009s/frame,realizing the rapid detection of deposited droplet at the mobile end. |
PDF Full Text Request