Design And Experiment Of A Greenhouse Crop Variable Spraying System Based On Solid-State LiDAR

In the realm of greenhouse cultivation,a controlled environment is essential for optimal plant growth.However,due to factors such as improper temperature and humidity management,these facilities frequently experience pest and disease infestations.Although pesticide application remains the most effective approach to manage these challenges,traditional spraying methods are beset with issues such as low efficiency,suboptimal spraying effects,and difficulty in controlling pesticide quantities.These concerns have significant implications for both food and environmental safety.To address these challenges,the present study proposed a solidstate LiDAR-based variable spraying system for greenhouses.This system endeavors to automatically modulate pesticide application based on the volume of greenhouse plants,with the ultimate goal of enhancing spraying efficacy,minimizing non-target pesticide deposition,and curbing environmental pollution.Consequently,this project seeks to optimize pesticide utilization,improve application rates,and elevate the overall effectiveness of spraying practices.The key methodology employed in this research involves the utilization of solid-state LiDAR technology to detect point cloud information of the greenhouse plant canopy.This data is then supplemented with deep learning techniques to ascertain canopy volume.Subsequently,the spraying nozzle is manipulated to regulate the output of pesticide spray in relation to various canopy volumes and their corresponding rates.This approach aims to achieve the dual objectives of maximizing pesticide utilization and minimizing environmental pollution,offering valuable insights for the implementation of precise and intelligent variable spraying practices within greenhouse settings.The primary research components are presented as follows:(1)This study evaluates the accuracy of 3D dimensional measurements of solidstate LiDAR for detecting regular objects(five different sizes of cartons)and irregular objects(three simulated plants)under indoor conditions.The root mean square error(RMSE)and coefficient of variation(CV)were used to evaluate the accuracy of the lidar sensor in object 3D dimension detection.The velocity(0.1,0.3,0.6,and 0.9 m/s)factors are also introduced to verify the detection performance of the solid-state LiDAR sensor during motion,and the edge detection performance of the LiDAR sensor is verified by the edge similarity score(ESS).The experimental results show that the maximum RMSE(mm)and CV(%)in the X(length),Y(width),and Z(height)directions are 14.3 mm and 14.3%,14.3 mm and 14.3%,and 10.8 mm and 10.8%,respectively.The ESS is all greater than 0.90,proving that the solid-state lidar sensor and its supporting algorithm can effectively adapt to the complex target structure parameters and profile measurements,providing a useful reference basis for the design of accurate and intelligent variable spraying systems for greenhouses.(2)A deep learning-based point cloud complementation network,namely the multi-resolution complementation network(MRC-Net),is designed to complement the residual point clouds acquired by solid-state LiDAR sensors.First,a generator is used to generate a coarse complete point cloud of the residual point cloud,and the complete point cloud generated by the generator is sampled at different resolutions by an iterative farthest point sampling(IFPS)technique.Second,the generated point clouds are degraded at different scales by the designed multi-resolution degradation mechanism.Thus,the network’s ability to extract semantic information and geometric structure information is improved.Finally,drawing on the idea of multi-discriminators in 2D images,a multi-scale feature matching loss function is designed to operate on the basis of different resolutions and adjust the output distribution of the generator in an optimal way to generate complete point clouds that better match the geometric features of the actual objects.(3)In order to improve the problem that the point cloud data generated by MRCNet is small and the volume cannot be calculated directly from the point cloud data,the point cloud upsampling method and the point cloud triangular meshing method are used in this study,respectively.Among them,point cloud upsampling is based on the Voronoi diagram method.Then,the Poisson reconstruction algorithm and the greedy projection triangulation algorithm are used to triangulate the point cloud data of regular cartons and irregular simulated plants,respectively.The experimental results show that the Poisson reconstruction algorithm can generate smoother 3D mesh surfaces,while the greedy reconstruction algorithm is more suitable for triangulation of point clouds of irregular structures.Finally,the projection method is applied to calculate the point cloud volume of triangulated meshing,which provides a decision basis for the application volume calculation of a variable spraying system.(4)A variable application system for greenhouse crops has been designed,and its performance has been evaluated.The main components of the greenhouse variable spraying system include a microcontroller with an STM32F407 as the core,a photoelectric sensor,a spray control unit,a spray nozzle,and other components.The experimental results show that the solid-state LIDAR-based greenhouse variable spraying system can meet the requirements of application operations.Compared with the conventional constant volume spraying mode in greenhouses,the greenhouse variable spraying system can significantly reduce the amount of pesticides used while ensuring the application effect to meet the pest and disease protection target.In addition,the system exhibits excellent performance in spray droplet coverage and droplet deposition density,significantly improving application efficiency and pesticide utilization.