Abstract:Seed micro-invasive sampling equipment is an intelligent tool designed for the seed slicing sampling process in biotechnological breeding. It utilizes technologies such as electromechanical control and machine vision to achieve a fully automated and comprehensive process for seed gene sampling. Its application and implementation effectively enhances the efficiency and quality of germplasm resource cultivation, thereby promoting the revitalization of the seed industry. To clarify the future development direction of seed minimally invasive sampling technology and equipment, a historical overview of seed slicing detection and the current research status of equipment was presented. Subsequently, the key technologies of seed micro-invasive sampling were systematically divided into seed separation, posture adjustment, clamping and transportation, sampling, and sample collection and cleaning technologies, which were systematically categorized and analyzed in terms of their research status and development trends. Building upon this foundation, combined with the development requirements and application scenarios of seed slice sampling equipment, insufficient cutting theory, limited multi-scale versatility, and system integration to be improved that seed slice sampling equipment faces were summarized and the future development direction was to strengthen the basic theoretical research of seed slicing equipment, develop a multi-scale universal seed slicing platform, and develop a smart sampling and detection system for the entire production process.

Abstract:The Beidou positioning system (BDS) can achieve centimeter level positioning, and has been widely used in agricultural machinery navigation systems. However, a simple satellite positioning system has some limitations, such as being vulnerable to external occlusion and low data frequency. Therefore, integrating inertial components and developing integrated navigation technology in navigation systems is an important trend. Considering the high cost of high-performance inertial navigation, which is not conducive to promotion and application, the use of low-cost inertial navigation and BDS was studied to constitute an integrated navigation system in agricultural scenarios. In order to improve positioning accuracy and solve the problem of error divergence caused by BDS interruption, zero speed correction was designed. At the same time, the causes and sources of unobservable heading angle errors in the BDS/INS integrated navigation system were analyzed, and a heading constraint method was designed. When BDS information was available, dual antenna heading angle information was used to suppress the accumulation of heading angle errors. When BDS was interrupted, zero angular velocity correction was used. Field experiments verified that when BDS was available, zero speed correction can improve the accuracy of position, velocity, and horizontal attitude by more than 20%, 40%, and 15%, respectively. Heading constraints can improve the accuracy of heading angle by more than 90%. When BDS was interrupted, zero speed correction can improve the accuracy of position and speed by more than 90%, horizontal attitude accuracy by more than 80%, and heading constraint can improve the accuracy of heading angle by more than 40%. 

Abstract:Seasonal changes of fruit tree crowns and changes of fruit tree characteristics caused by the growth and aging of fruit trees will affect the matching of the three-dimensional environmental map of the orchard. Therefore, an accurate construction algorithm of orchard two-dimensional environmental map was proposed based on improved Gmapping algorithm. In this algorithm, the front-end odometer and the back-end optimization part of Gmapping algorithm were improved respectively, so as to improve the construction accuracy of two-dimensional environment map of orchard. For the front-end odometer part, the improved R-GPF method was used to improve its initial positioning accuracy, and for the back-end optimization part, the BAT heuristic adaptive resampling method was used to improve its final positioning accuracy. Then, the comparative experiment of pear orchard environment was carried out. By comparing the improved R-GPF method with the original R-GPF method, the output frequency of the improved R-GPF LiDAR odometer can reach 15.58Hz, the maximum lateral deviation was less than 25cm, the average lateral deviation was 12.7cm, and the standard deviation was 13.4cm, its performance was superior to that of the original R-GPF LiDAR odometer. Comparing the proposed algorithm with the original Gmapping algorithm based on R-GPF, the distance deviation between pear columns obtained by the proposed algorithm was always within 20cm, and the average distance deviation between rows was 10.3cm, with a standard deviation of 6.3cm, which was 50%, 43.41% and 32.26% lower than that of the original Gmapping algorithm based on R-GPF, respectively. At the same time, the reduction of the distance deviation between pear rows relative to the lateral deviation of odometer reflected the effectiveness of the back-end BAT heuristic adaptive resampling method. The proposed algorithm can improve the accuracy of orchard two-dimensional map construction, and meet the accuracy requirements of subsequent relocation, navigation and other operations.

Abstract:Aiming at the problems that orchard roads have no obvious boundaries and there are shadows, soil and sand interference at the edges of the road, a recognition method of orchard unstructured roads based on feature fusion was proposed. The distortion parameters were obtained through camera calibration to correct the distortion of the acquired image, and a dynamic region of interest (ROI) extraction method based on the combination of filtering and gradient statistics was proposed to select the ROI of the S component of the HSV color space. The maximum value method was used to merge the color features with the S component mask for multidirectional texture features for binarization and noise reduction. The feature points were found according to the abrupt features of road edges, and a two-level pseudo feature points elimination method based on the dual constraints of distance and position was proposed. To better fit the irregular edges of unstructured road, the method of segmentation cubic spline interpolation was introduced to fit the road edges to realize road recognition. The experimental results showed that under the six working conditions of sunny day, cloudy day, straight light, backlight, sunny day in winter and rain and snow weather, the mean value of average longitudinal deviations of S component, texture image and fusion image were 2.43 pixels, 39.71 pixels and 1.36 pixels, respectively, and the mean value of average deviation rates were 0.99%, 18.02% and 0.54%, respectively. Compared with the S component and texture image, the average longitudinal deviation and average deviation rate of the fusion image constructed by this method were effectively reduced. The mean value of average deviations of least squares method, random sample consensus method (RANSAC) and segmentation cubic spline interpolation method for fitting edges were 2.64 pixels, 3.16 pixels and 0.66 pixels, respectively, the mean value of average deviation rates were 1.02%, 1.21% and 0.26%, respectively, and the average standard deviations of deviation rate were 0.23%, 0.31% and 0.09%, respectively. The mean value of average deviation, mean value of average deviation rate and average standard deviation of deviation rate of the algorithm were the minimum, which indicated that the fitting method had higher fitting accuracy and better stability. Under the six working conditions, the average processing time of a single image of this algorithm was 89.9 ms, which met the real-time requirements of agricultural robots in the process of operation. The method can provide a reference for agricultural robots to recognize unstructured roads in complex orchard environments.

Abstract:In view of low localization accuracy and poor map construction during the visual navigation for orchard spraying robot, a visual localization and dense mapping algorithm was proposed. The algorithm was based on the ORB-SLAM2 algorithm architecture, firstly, through the optimization of FAST corner points, descriptor thresholds, and adopting the image pyramid method and Gaussian filtering algorithm, poor quality ORB feature points were eliminated to improve the image key frame quality and feature matching accuracy. Secondly, the dense map building thread was introduced, the point cloud recovery algorithm and statistical filtering method were used to form the point cloud queue, the point cloud stitching technology and voxel filtering algorithm were adopted to output the dense point cloud maps, and the key frame output interface and position publishing topic were added in the ROS node of ORB-SLAM2 algorithm, and then the key frame generated by ORB-SLAM2 algorithm was selected through the NeedNewKeyFrame function to reduce the system computation. Finally, the RGB-D camera was used to realize the precise positioning and dense mapping of the orchard spraying robot. In order to verify the effectiveness and practicality of the algorithm, simulation analysis of TUM dataset and real scenario testing were conducted. The results showed that compared with that of ORB-SLAM2 algorithm, the absolute trajectory average error of this algorithm was reduced by 44.01%, the relative trajectory average error was reduced by 7.93%, the average number of ORB feature point matching was increased by 19.03%, and the positioning accuracy and running trajectory effect were improved significantly. In addition, the working scene information of orchard spraying robot can be obtained with high accuracy. The algorithm can provide a theoretical basis for the autonomous navigation of orchard spraying robot.

Abstract:The path trajectory planning of orchard spray robot affects the smooth line of robot driving route and the reliability and smoothness of the driving process which needs more comprehensive consideration and more comprehensive planning. Aiming at the problems that the reference trajectory at the turn is not smooth enough and the curvature is large in the path planning of orchard spray robot, a trajectory optimization method of cubic non-uniform B-spline curve for orchard spray robot based on kinematics multiple constraints of orchard spray robot was proposed. The prior map was used to obtain the position information of the tree rows, and the path points between the rows were fitted to ensure that the orchard spray robot driving on the center line of the tree row met the requirements of the spray operation. The objective function of minimizing the path curvature was constructed by considering the minimum turning radius, the constraint of the first and end points, the delay constraint of the steering mechanism, and the continuity of curvature. The curve parameters to be optimized were solved by the optimization algorithm, and the global path that met the driving requirements of the orchard spray robot was generated. Finally, the pure tracking algorithm was used to verify the driving accuracy of the robot. The simulation and test results showed that the maximum curvature of the planned trajectory was 0.31m-1, and the average curvature was 0.15m-1, which met the driving requirements of the orchard spray robot. The average error of the trajectory tracking driving was 0.225m, and the mean square error was 0.031m, which met the requirements of the orchard spray robot for driving accuracy when spraying in the orchard.

Abstract:Automatic navigation control of agricultural machinery is the basis of precision agriculture. Realizing automatic navigation operation of agricultural machinery can reduce labor intensity of agricultural machinery operators and improve work efficiency, which has been widely used in various links of agricultural production. Aiming at the problems of the traditional sprayer in the process of turning and wrapping, such as limited turning space, large turning radius and easy rolling of crops, a control method was proposed to realize the wrapping operation by using parallel vehicle movement. A navigation control system for four-wheel steering sprayer was designed based on translation and line feed mode. The control system adopted the positioning module of real time kinematic (RTK) and attitude sensor for integrated navigation. The position information and attitude information of sprayer were taken as input. The automatic navigation and tracking control of non-turn turn line feed of the sprayer was realized by combining the kinematic solution. The automatic operation strategy based on finite state machine was designed according to the requirements of spraying operation. A field comparison test between traditional proportion integration differentiation (PID) controller and single-neuron PID was carried out. In the conventional square hard flat block, the maximum tracking deviation and average absolute deviation of the springer equipped with conventional PID controller in the translation and line wrapping process were 7.63cm and 4.27cm. The maximum tracking deviation and average absolute deviation of the sprayer equipped with single-neuron PID controller in the translation and line feeding process were 6.48cm and 3.24cm. In the conventional square field test plots, the maximum tracking deviation and average absolute deviation of the sprayer equipped with conventional PID controller in the translation and line wrapping process were 11.01cm and 6.66cm. The maximum tracking deviation and average absolute deviation of the sprayer equipped with single-neuron PID controller in the translation and wrapping process were 8.60cm and 4.47cm. The experimental results showed that compared with the traditional controller, the single-neuron PID controller had better control accuracy and adaptability. It solved the problems of inflexible and low land utilization rate due to the large turning radius and large turning space of the traditional line feed mode, and provided a solution for the ground turning and line feed of the wide-width sprayer and provided a reference for the automatic navigation technology of the sprayer.

Abstract:The introduction of unmanned autonomous systems to realize intelligent inspection of canal networks is of great significance to the construction, monitoring and maintenance of water conservancy projects. When unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGV) are used to cooperate in the inspection of canals, UAVs carry out patrol inspection work over the canals, and unmanned vehicles can be used as UAV carrier platforms and energy supply stations, which is helpful to realize rapid autonomous inspection of large-scale canal networks. However, the dual constraints of canal network and road network bring great difficulty to the path planning of unmanned systems. In view of the above problems, aiming to minimize the time to complete the entire inspection task. Firstly, based on the degree constraint, a canal network segmentation method was proposed to allocate the inspection task to the UAVs, so that the UAV did not need to take off or land to recharge when inspecting each canal segment. Then the optimal movement path for UAVs and UGV was calculated based on genetic algorithm. Finally, through the real-world example verification, when the UAVs were operating at a constant speed of 60km/h and the UGV was operating at a speed of 40km/h, the inspecting speed of the unmanned system was 8.4~9.8 times that of the human inspection based on the regular speed of 2km/h.

Abstract:Aiming at the problems of complex motion planning, multiple degrees of freedom and difficult control of industrial manipulator apple picking, human hand picking was simulated and a lightweight 3-DOF apple picking manipulator was developed. Firstly, the structural design and kinematic analysis of the manipulator were completed for the requirements of apple picking. The mechanical arm adopted a parallelogram structure, which reduced the rotational inertia of the whole machine through a rear power source, and had a long arm span, a large working space, and small branch interference during movement, which was more suitable for apple picking. Secondly, the Newton-Euler equation was used to establish the dynamic model, and the apple picking simulation of the manipulator was completed. Through the theoretical data of the dynamic model, the stress and strain of the arm and its key components were analyzed to reduce the mass of the manipulator itself. The stress and strain under different lightweight schemes were calculated to select the optimal lightweight scheme. By comparing the simulation data of the manipulator before and after lightweight, the peak driving torque of the bone rod lightweight scheme was reduced by 21N·m and 15N·m, respectively, both of which were reduced by about 20%. The weight of the whole machine was reduced by 1.8kg, which was reduced by 32.1%, and the lightweight manipulator maintained good working ability. According to the optimization results, a physical prototype of a 3-DOF apple picking manipulator was built. The maximum driving torques of the large and small arms were 92N·m and 63N·m through experiments, which basically conformed to the simulation results and verified the correctness of the dynamic model.

Abstract:Aiming at the problems of poor passability and adaptability, low tillage rate, grass entanglement of the knife roller, and poor straw burying performance in the traditional rotary tiller when tillage operations in the planting pattern of rapeseed under rice-rapeseed or rice-rice-rapeseed rotation planting, a driven disc plow and double-edged rotary tillage combined tiller was developed. A working method of active plow followed by double-edged rotary tillage and ditching on both sides was proposed. The main structural parameters of driven disc plow and the layout of driven disc plow-ditch device were determined. A double-edged rotary tillage device for driven disc plow and double-edged rotary tillage combined tiller was designed. The key structural parameters of double-edged rotary blade with long blade and short blade were determined according to sliding cutting principle. The arrangement of the double-edged rotary tillage blades were determined according to the structure layout of the driven disc plow group. The DEM simulation method was used to analyze the straw burial performance and soil exchange performance of the driven disc plow and double-edged rotary tillage combined tiller. The experiment showed that the average straw burial rate of the whole machine operation was 94.69% and the soil layer was evenly mixed after the whole machine worked. The field experiment showed that under the two conditions of high and low straw stubble, the average straw burial rate was 96.45% and the average soil crushing rate was 95.30% after the operation of the driven disc plow and double-edged rotary tillage combined tiller with double-edged rotary tillage device and knife roller did not twine grass. The field sowing test showed that the rapeseed emergence was uniform, and the indexes met the requirements of rapeseed direct seeding bed preparation in rice stubble field.

Abstract:No tillage sowing with straw mulch on the surface has social, ecological and economic benefits such as water storage and moisture conservation, improving soil fertility, improving soil structure, controlling soil erosion, reducing production costs and increasing crop yield. In order to solve the problems of poor operation quality and low efficiency of the straw cleaning device under the condition of heavy straw coverage and high speed operation of the no-tillage seeding unit in service, an improved straw cleaning device with the function of straw axial acceleration was designed. The mechanism of the straw cleaning device was clarified, the key components were designed, and the main parameters affecting its working performance and the value range were determined. Using the quadratic regression orthogonal rotation center combination test method, taking the operating speed, operating deflection angle, spiral rise angle and spiral blade number as the test factors, and the straw cleaning rate and working resistance as the performance evaluation indicators, the parameter combination optimization test was carried out on the constructed EDEM-ADAMS joint simulation test platform. The results showed that each factor had a significant impact on the straw cleaning rate, and the significant factors were working deflection angle, operating speed number of spiral blades and spiral rise angle. Each factor had a significant impact on the working resistance, and the significance from large to small was the working speed, working deflection angle, number of spiral blades, and spiral rise angle. The Design-Expert software was used to optimize the parameter combination of the test results. When the helix angle was 40°, the number of spiral blades was 4, the operating speed was 7.5~10.7km/h, and the operating deflection angle was 20.0°~32.5°, the straw removal rate was more than 85%, and the working resistance was less than 110N. Under the operating speed of 8km/h, 9km/h and 10km/h, the field performance test was conducted on the straw cleaning device with a spiral angle of 40°, a number of spiral blades and a working deflection angle of 30°. The straw cleaning rate was more than 82%, and the working resistance was less than 112N, which proved that the simulation test results were credible. At the operating speed of 10km/h, the straw cleaning rate was increased by 33.5% compared with that of the non optimized straw cleaning device, and there was no significant difference in the working resistance.

Abstract:The vibratory excitation of the sieve by the sieve-cleaning device is a fundamental reason for increasing the probability of material penetration, reducing seed clogging and effective sieve cleaning. Existing sieve-cleaning devices mostly use randomly bouncing rubber balls to clean the sieve, but the sieve-cleaning effect is easily restricted by the structure and operating parameters. In order to solve the problem that rubber balls’ impact force is difficult to accurately control, an electromagnetic variable frequency excitation sieve-cleaning device was developed, the overall structure and operating principle was introduced, the excitation sieve-cleaning monomer and frequency conversion excitation control system was designed, and the vibration excitation mechanism was analysed;using acceleration as an indicator, the effects of spring pre-compression and excitation frequency on the vibration excitation law were studied, the results showed that both were positively related to the vibration excitation;sixteen sets of maize seed cleaning trials were carried out by using purity, screening efficiency, screening time and number of clogging seeds as indicators. The results showed that the sieve-cleaning device had a good operating effect when the spring pre-compression was 2mm, the operating frequency was 3.5Hz and the sieve cleaning frequency was 50Hz, with seed purity of 99.1%, screening efficiency of 88.6%, screening time of 70s, and clogging seed of 0. By setting the operating frequency and the sieve cleaning frequency in stages, the sieve-cleaning device could be precisely adjusted to meet the requirements of normal sieving and strong vibration sieve cleaning under different operating conditions. The research could provide reference for the development of intelligent sieve-cleaning equipment.

Abstract:When corn is sown at high speed (12~16km/h), the initial speed of the seeds leaving the dish is high, and the seeds collide with the seed cavity wall of the belt-type seed guide device, resulting in collision and dislocation, which leads to the low precision of the seeds entering the seed cavity. The belt-type high-speed corn seed guide device with seed receiving mechanism was taken as the research object, and the dynamic model of clamping, transportation, and discharge of the seeds was established. The main factors that affect the seed receiving stability and the precision of the seeds entering the seed cavity were identified, and the improvement method of adding herringbone lines on the surface of the finger was put forward. Single-factor comparison tests and multi-factor optimization tests were carried out by using high-speed camera and image target tracking technology. The single factor test showed that the seed acceptance index and variation coefficient of seed cavity spacing of the finger wheel with improved herringbone lines were obviously better than those of the finger wheel without herringbone lines when the sowing speed was fast. In order to obtain the best performance parameters of the improved seeding mechanism, taking the wheel center distance, the rotation speed of the finger wheel, and the finger length as test factors and the qualified index of seed acceptance and the variation of seed cavity spacing as evaluation indexes, a quadratic orthogonal rotation combination test with three factors and five levels was carried out. By using the multi-objective optimization method, it was determined that when the wheel center distance was 36.8mm, the rotating speed of the finger wheel was 584.97r/min, and the finger length was 10.8mm, the qualified index of seed acceptance was 98.23%, and the coefficient of variation of seed cavity spacing was 0.24%. The optimization results were verified, and the verification results were basically consistent with the optimization results. Under the same conditions, the bench comparison test showed that the sowing performance with a high-speed seed guide device was better than that without high-speed seed guide device.

Abstract:An air-suction metering device for broad beans with a flat belt auxiliary seed-filling device was designed to address the difficulties of sowing large-sized and highly different three-axis dimensional broad bean seeds. The motion mechanism of the flat belt auxiliary seed-filling device and seeds was elucidated through a dynamic analysis of the seed-filling process. A one-factor test was conducted by using the computational fluid dynamics and discrete element method bidirectional coupling simulation method (CFD-DEM) to determine the main component parameters affecting the seed-filling performance and clarify the mechanism of the flat belt auxiliary seed-filling device. An experimental platform was constructed, and a quadratic regression orthogonal combination test was conducted with operating speed, flat belt input shaft speed, and negative pressure as the test factors and qualified index, reseeding index, and miss-seeding index as the test indicators. The test results showed that the primary and secondary factors affecting the qualified index of the air-suction metering device were operating speed, negative pressure, and flat belt input shaft speed. A multi-objective optimization of the test results yielded the optimal parameter combination of the operating speed at 5.69km/h, flat belt input shaft speed at 395r/min, and negative pressure at 3845Pa. The air-suction metering device performance was verified through sowing tests with a qualified index of 91.6%, reseeding index of 3.8%, and miss-seeding index of 4.6%, which met the requirements for broad bean planting.

Abstract:In order to solve the problem that when the corn planter works at high speed and precision, the seed throwing point is high and the seed collides violently, which leads to the poor uniformity of grain spacing, a seed guiding device assisted by positive pressure air flow was designed based on Venturi principle. the main structure and key parameters of the seed guiding device were determined. The mechanism of air-assisted delivering seeds to realize “zero-speed seeding” was analyzed. The DEM-CFD coupling simulation method was used to simulate the working process of the seed guiding device. By comparing and analyzing the airflow field and the seed exit velocity, it was determined that the constriction angle of the intake chamber was 70°, and the length of the constriction section of the intake chamber was 8.2mm. The speed matching test, bouncing test, operation performance test and comparison test were carried out on the performance test platform of the seed metering device. The results showed that when the operating speed was 8~16km/h and the grain spacing was 20~25cm, the qualified rate was not less than 85.7%;and the coefficient of variation of particle spacing was not more than 15.8%. Compared with the gravity type seed guide tube, the higher the operating speed was, the more outstanding the excellent operating performance of the positive pressure airflow assisted seed guide device was. When the operating speed was 16km/h, the qualified rate of particle spacing was increased by 13.6 percentage points and the coefficient of variation of particle spacing was decreased by 7.4 percentage points, which met the requirements of precision seeding under high-speed conditions and was conducive to improving the overall performance of high-speed precision seeders.

Abstract:In order to solve the problem of uneven fertilizer discharge flow of screw fertilizer feeder affecting precision control fertilization, the reason for uneven discharge of fertilizer is determined based on the simulation analysis of fertilizer movement during the discharge process，the structure design of oblique opening fertilizer discharge port is adopted to improve the uniformity of fertilizer discharge. Using EDEM to establish simulation model of oblique opening screw fertilizer distribution apparatus. Taking the length of oblique opening x1, angle of oblique opening x2, width of opening x3 as the test factors, and the coefficient of variation of fertilizer discharge as the test index, a secondary general rotary combination design experiment was conducted. The test results showed that the order of influence of test factors on test indexes was x3、x2、x1, and when x1 was 105mm, x2 was within range of 30°~44°, x3 was within range of 40.05~55.00mm, the variation coefficient of fertilizer discharge discharge was less than 15%, and the uniformity of fertilizer discharge was better. The bench test was used to compare the traditional and oblique opening screw fertilizer distribution apparatus. The results showed that the coefficient of variation of fertilizer discharge of the oblique opening screw fertilizer distribution apparatus was 13.59% at speed of 60r/min, which was consistent with the theoretical optimization value, and the oblique opening screw fertilizer distribution apparatus were better than the traditional screw fertilizer distribution apparatus. At the same time, based on the measured fertilizer discharge speed flow curve of the fertilizer discharger, a fertilizer discharge controller was developed and bench test was carried out. The results showed that it can achieve precise fertilization. The research result can provide some reference for improving the design of screw fertilizer distribution apparatus.

Abstract:In order to solve the problems of small size of seedling picking jaws of automatic vegetable pot transplanting machine, the structure and installation method of picking force detection sensor interferes with the normal picking action of picking jaws and affects its accuracy and service life, selecting polydimethylsiloxane (PDMS) film as the dielectric layer of sensor and a built-in pot picking force sensor was developed. The PDMS film was selected as the dielectric layer of the sensor, and a built-in potting force sensor was developed. Firstly, a simulation model of cavity, pot substrate and seedling jaw was established, and LS-PrePost software was applied to simulate the coupling of seedling extraction process, obtain the maximum force area in the contact area between seedling jaw and pot substrate, and determine the structure and size of seedling jaw and sensor;the clamping force signal detection system was designed, and the hardware circuit and acquisition software were combined to complete the capacitance-voltage conversion, signal amplification, noise filtering, and realize the acquisition of clamping force signal. In order to realize the functions of acquisition, processing, display and storage of the gripping force signal, the system was designed. In order to verify the performance of the sensor, calibration test and indoor validation test were conducted;the calibration test showed that the average sensitivity of the clamping force sensor was 0.3728N/V, the average linear coefficient of determination was 0.9892, the accuracy was 7.548%, and the range was 7N, which satisfied the accuracy requirement of clamping force detection in the transplanting process;the indoor validation test showed that the clamping force detection sensor had good stability and adaptability, and can be used for real-time and accurate detection of the clamping of transplanting machine pick-up mechanism.

Abstract:To explore the spatio-temporal variation of drought characteristics in the Loess Plateau from 2001 to 2020 and its influencing factors, MODIS enhanced vegetation index (EVI) and land surface temperature (LST) data was used to establish the temperature vegetation dryness index (TVDI) model. The driving factors of TVDI in the Loess Plateau from 2001 to 2020 were analyzed by using Geodetector Model. The results showed that from 2001 to 2020, the spatial distribution of TVDI in the Loess Plateau had a strong spatial heterogeneity, and the drought increased gradually from west to east. The average TVDI of the Loess Plateau for the past 20 years was 0.522, indicating a light drought on the whole. According to the variation trend of TVDI, more than 64% of the regions showed a drying trend, and there was an obvious regional differentiation. The drought situation in Inner Mongolia, northern Ningxia and parts of Shanxi was mostly intensifying, while the areas of alleviating drought were concentrated, mainly distributed in central Shaanxi, southern Ningxia and northern Gansu. The annual change of TVDI of all land use types showed a rising trend in varying degrees, and the annual average TVDI of each land use type was significantly different, in order from large to small as follows: unused land (0.571), grassland (0.554), cultivated land (0.503), forest land (0.473) and construction land (0.462). The spatial differentiation of TVDI in the Loess Plateau was mainly affected by three factors: elevation, soil type and vegetation type, whose q values were all exceeding 0.3, which were the main driving factors of drought in the Loess Plateau. Under the interaction of multiple factors, the combination of elevation and SIF had the strongest influence on the occurrence of drought in the Loess Plateau, with q value reaching 0.709.

Abstract:Plant potassium content (PKC) of potato plants is an important indicator for monitoring potato nutrition status. Obtaining PKC quickly and accurately has guiding significance for field fertilization and production management. RGB images of potato plants during the tuber formation period, tuber growth period, and starch accumulation period were obtained by using an unmanned aerial vehicle (UAV) remote sensing platform equipped with an RGB sensor, and PKC was measured. Firstly, the average spectral and texture features of each plot were extracted from the RGB images of each growth period. Then vegetation indices and texture indices (NDTI, RTI, and DTI) were constructed based on the spectral and texture features of the canopy, and their correlations with the measured PKC were analyzed. Finally, multiple linear regression (MLR), partial least squares regression (PLSR), and artificial neural networks (ANN) were used to construct models for estimating potato PKC. The results showed that the correlations between NDTI, RTI, DTI and PKC were higher than those of single texture features during each growth period. Combining vegetation and texture indices can improve the reliability and stability of the model. MLR and PLSR models were superior to ANN. The research result can provide scientific references for monitoring PKC in potato plants.

Abstract:The traditional method of artificial seedling phenotype measurement has some problems, such as low efficiency, strong subjectivity, large error and damaged seedlings. A method for nondestructive detection of cucumber seedling phenotype by using the RGB-D camera was proposed. An automated multi-view image acquisition platform was developed, and two Azure Kinect cameras were deployed to simultaneously capture color, depth, NIR, and RGB-D images from the top view and side view. The Mask R-CNN network was used to segment the leaves and stems in the NIR image, and then mask them with the RGB-D image to eliminate the background noise and ghost in the RGB-D images and obtain the RGB-D image of the leaves and stems. The category and number of segmentation results of the Mask R-CNN network were the numbers of cotyledons and true leaves. The CycleGAN network was used to process the RGB-D image of a single leaf, repair the missing and convert it into 3D point clouds, and then filter the point clouds to achieve edge-preserving denoising. Finally, the point clouds were triangulated to measure the leaf area. In the stem RGB-D image obtained by Mask R-CNN segmentation, the approximate rectangular feature of the stem was used to calculate the length and width of the stem respectively, and then the depth information was combined to convert the hypocotyl length and stem diameter. YOLOv5s was used to detect the growing point of cucumber seedlings in the RGB-D image, and the height difference between the growing point and the substrate was used to calculate the plant height. The experimental results showed that the system had good flux and accuracy. The mean absolute errors of key phenotypes of cucumber seedlings at cotyledon, 1 true-leaf and 2 true-leaf stages were all no more than 8.59% and R2 was no less than 0.83, which can well replace the manual measurement method, and provide key basic data for seed selection and breeding, cultivation management, growth modeling, and other research.

Abstract:Crop disease recognition is a prerequisite for rational pesticide application and a powerful guarantee for promoting healthy and stable agricultural development. Existing deep learning-based crop disease recognition methods mainly use classical networks such as VGG and ResNet or networks that use attention mechanisms for disease recognition. Although these deep learning-based crop disease recognition methods have achieved better disease recognition results than traditional methods, they do not sufficiently mine the discriminative information contained in the shallow, middle and deep features of networks, and most of the extracted saliency features of crop disease images are insufficient. To extract discriminative features in crop disease images more effectively and improve crop disease recognition accuracy, a crop disease recognition network based on multi-layer information fusion and saliency feature enhancement (MISF-Net) was proposed. Specifically, MISF-Net mainly consisted of a ConvNext backbone network, a multi-layer information fusion module (MIFM), and a saliency feature enhancement module (SFEM). The ConvNext backbone network was mainly used to extract features of crop disease images. The multi-layer information fusion module was mainly used to extract and fuse the discriminative information from the shallow, medium and deep layers of the backbone network. The saliency feature enhancement module was mainly used to enhance the saliency discriminative features in crop disease images. The experimental results on the crop disease dataset AI challenger 2018 and the homemade dataset RCP-Crops showed that the crop disease recognition accuracies of MISF-Net reached 87.84% and 95.41%, and the F1 values reached 87.72% and 95.31%, respectively.

Abstract:Plant diseases are one of the main causes of crop yield reduction, however, traditional manual diagnosis methods are costly and inefficient, which are difficult to adapt to the demands of modern agricultural production. Recognizing crop diseases automatically and accurately is hence of great importance. Currently, most studies have focused on images taken by professionals for academic purposes, rather than by farmers in actual agricultural production. However, images taken in real applications by farmers are with far more complex backgrounds and hence alleviating the performance of many state-of-art methods. A grape leaf disease dataset were construted under natural complex environments where images were taken by farmers in actual agricultural production. And a network architecture named MANet was proposed for efficient recognition of grape leaf diseases under natural complex environment. The inverted residual module was embedded to build the model, which significantly lowered the number of model parameters. Moreover, the attention mechanism SENet module was used to improve the ability of the model to extract key disease features from complex background images and suppress other irrelevant information. In addition, a multi-scale convolution (MConv) module was designed to extract and fuse multi-scale features of disease images. The experimental results indicated that the proposed model presented a superior performance relative to other most advanced methods. On the public crop disease dataset, MANet achieved the highest average recognition accuracy of 99.65%. And even on the complex background crop disease dataset of the construction, the average recognition accuracy of grape diseases reached 87.93%, which was still better than other state-of-the-art models. Therefore, the proposed model can effectively recognize grape leaf diseases and has certain potential for practical applications.

Abstract:Hyperspectral image has continuous spectral information of ground objects, which is an essential means of remote sensing monitoring. On this basis, the endmembers of the features can be extracted by decomposing the mixed pixel spectrum and exploring the degree of each endmember participates in the mixing. However, specific spectral changes cause trouble for spectral unmixing due to the sensor and the image’s resolution. To solve this problem, an endmember bundle extraction method based on multi-modal and multi-objective particle swarm optimization by special crowding distance (MOPSOSCD) was proposed. Firstly, for a three-dimensional hyperspectral image, the label coding was carried out pixel by pixel, and the index-based ring topology was used for individual interaction in different neighborhoods. Secondly, for particle velocity and position update, the position update method of PSO was adopted and the particle swarm velocity update method and the integer particle position update were improved through neighborhood optimization. The objective function selection was measured by two RMSEs, that was, the unconstrained least squares method was used to solve the RMSE of the abundance map anti-mixing and the original map, and the fully constrained least squares method was used to solve the RMSE of the abundance map anti-mixing and the original map. At the same time, according to the spatial characteristics of hyperspectral images, decision space diversity was improved by improving the crowded distance of decision space. Finally, the crowding distances of the decision space and the target space were combined and sorted, and the particles were updated according to the sorting results. When the particle directional movement probability was 0.2, the number of particles was 30, and the number of iterations was 400, the results of RMSE and mSAD on the MUUFL dataset were 0.0088 and 0.1112, respectively. Through the comparative test, the method had higher extraction accuracy and efficiency than VCA and DPSO, providing a more accurate end beam extraction method for hyperspectral unmixing.

Abstract:In actual production, the number of wheat seedlings plays a key role in estimation of emergence rate, yield prediction, and grain quality. Timely and accurate estimation of number of wheat seedlings is very important for wheat production. Due to the complex growing environment in the field, imaging of wheat seedlings is easily affected by factors such as illumination, occlusion and overlapping, which results in poor performance when existing target object counting methods were directly used for wheat seedling counting. In order to reduce negative impacts of these factors and further improve counting accuracy, an improved wheat seedling counting model was proposed by enhancing local contextual supervision information based on existing target object counting network, P2PNet (Point to point network). Firstly, wheat seedling images were preprocessed, and a private wheat seedling data set was built by using point labeling method. Secondly, a wheat seedling local segmentation branch was introduced to improve the architecture of P2PNet, so as to extract the local contextual supervision information of wheat seedling. Then an element-by-element point multiplication mechanism was designed to fuse global and local contextual supervision information of wheat seedling. Finally, per-pixel weighted focal loss was introduced to construct the overall loss function, and the model was optimized. Experimental results on the self-built dataset showed that the mean absolute error (MAE) and root mean square error (RMSE) of P2P_Seg were 5.86 and 7.68, respectively, which were 0.74 and 1.78 lower than those of P2PNet. Compared with other state-of-the-art counting models, P2P_Seg exhibited better counting performance. In the actual field environment, the application test analysis, error counting and missing counting analysis were conducted. P2P_Seg was more suitable for complex field environments, and it provided a method for automatic wheat seedling counting.

Abstract:There is a low efficiency of automatically measuring and analyzing plant anatomic phenotypes currently, which makes it difficult to well deal with the issue of extracting and recognizing the complex anatomical phenotypes. In order to solve this problem, a mask region convolutional neural network (Mask R-CNN) based instance segmentation model for microscopic images of the citrus main leaf veins was proposed. In this model, the deep residual network (ResNet50) and the feature pyramid network (FPN) were used as the backbone feature extraction network. In addition, a new region of interest Align (RoI-Align) layer was added to the Mask branch to improve the segmentation accuracy. The results showed that the network can accurately identify and segment pith, xylem, phloem and cortical cells, respectively, in the citrus main leaf veins. The average precision (IoU was 0.50) of the model for segmentation of pith, xylem, phloem and cortical cells was 98.9%, 89.8%, 95.7% and 97.2%, respectively, and the overall average precision (IoU was 0.50) for segmentation of the four tissue regions was 95.4%. The mean average precision of Mask R-CNN with adding RoI-Align to the Mask branch was improved by 1.6 percentage points compared with that without. The results showed that Mask R-CNN model presented good performance of recognition and segmentation of various tissue regions of citrus main leaf veins, which can provide technical support for citrus microscopic phenotyping.

Abstract:Discrimination of ripe asparagus and accurate location of the picking hand is a challenge in the selective harvesting process of asparagus harvesting robots. To address this challenge, an improved you only look at coefficients (YOLACT++) based algorithm was proposed, which was used to detect and discriminate ripe asparagus and locate harvesting cuts. Improving the traditional YOLACT++ backbone feature extraction network, specifically including the introduction of a convolutional block attention module (CBAM) attention mechanism and a spatial pyramid pooling (SPP) module, to improve the effectiveness of the network for feature extraction and enhance its detection segmentation results. Asparagus have different sizes and postures, by designing different anchor frame sizes to ensure that they were covered, the adaptability of the anchor frame to the aspect ratio of the asparagus was improved, thus improving the detection accuracy and speed of the network. The skeleton was then fitted to asparagus with varying growth forms. Determination of asparagus maturity after calculating asparagus length and basal diameter in segments. Finally, the location of the cutting point in the bottom area of the mature asparagus was calculated, and its spatial location was determined by quantifying the roll angle and pitch angle to locate the final harvesting cutting surface. The results of the harvesting robot field trials showed that the detection accuracy of the trained improved YOLACT++ model was 95.22%, the average accuracy of the mask was 95.60%, the detection time of 640 pixels×480 pixels size image was 53.65ms, the accuracy of mature asparagus discrimination was 95.24%, the error of cutting point positioning in X, Y and Z directions was less than 2.89m, and the maximum error in rotation and pitch angles was 7.17°. Compared with that of the Mask R-CNN, SOLO and YOLACT++ models, the average accuracy of the mask was improved by 2.28, 9.33 and 21.41 percentage points, respectively;the maximum positioning errors were reduced by 1.07mm, 1.41mm and 1.92 mm, respectively, and the maximum angle errors were reduced by 1.81°, 2.46° and 3.81°, respectively. The harvesting success rate of the trial asparagus harvesting robot was 96.15%, and that the total time taken to harvest a single asparagus was only 12.15s. The detection-discrimination-location method proposed had high detection and location accuracy, which ensured detection speed on the premise. It can provide technical support for optimizing and improving the asparagus harvesting robot based on machine vision.

Abstract:Aiming at the problem of low accuracy of corolla detection and position during field operation of safflower picking robots, a deep learning-based object detection and position algorithm, mobile safflower detection and position network，MSDP-Net, was proposed. For object detection, an improved YOLO v5m model was proposed. By inserting the convolutional block attention module, the model precision, recall and mean average precision were improved by 4.98, 4.3 and 5.5 percentage points, respectively, compared with those before the improvement. For spatial position, a camera-moving spatial position method was proposed, which kept the position accuracy in the best range and avoided the missed detection caused by the obstructed corolla at the same time. The experimental verification showed that the success rate of mobile camera-based positioning was 93.79%, which was 9.32 percentage points higher than that of fixed camera-based positioning, and the average deviation of mobile camera-based positioning method in X, Y and Z directions was less than 3mm. The MSDP-Net algorithm had better performance compared with five mainstream object detection algorithms and was more suitable for the detection of safflower corolla. The MSDP-Net algorithm and the camera mobile position method were applied to the self-developed safflower picking robot for picking experiments. The indoor test results showed that among 500 replicate tests, totally 451 were successfully picked and 49 were missed, with a picking success rate of 90.20%. The field test results showed that the success rate of safflower corolla picking was greater than 90% within the selected monopoly length of 15m.

Abstract:The stem moisture status of living tree is an effective manifestation of plant life state. Stem water content (StWC) and stem sap flux density (SFD) are important parameters to study the variation of water in plants. Stem water content is a fundamental parameter to correctly detect the thermal equilibrium point or zero-flux conditions and measure the sap flux density. The water content at different heights and the sap flux density in different orientations of the stems of the living tree may differ significantly. The plant growth status can be evaluated comprehensively and the relationship between the water content and sap flux density can be analyzed effectively with accurate detection of the two parameters at the same spatial position of living tree stem. The stem water content detection method based on standing wave ratio (SWR) principle and the stem sap flow detection method based on heat ratio method (HRM) principle were combined to design a composite detection system for stem water content and sap flux density of living trees. The water content detection unit and the sap flow detection unit of the composite detection system reused one set of three-needle probes, which could accurately detect water content and sap flow in the same spatial position of the living tree stems in real time. The output voltage of the water content detection unit had a good linear relationship (R2=0.9701) with the dielectric constant (in the range of 6～53.3, corresponding to the stem water content range of 0～85%), and the static stability was good (with maximum fluctuation of 0.6% of the full scale for a long time test). The measuring results of the water content detection unit and BD-IV plant stem moisture sensor were consistent（R2=0.9800）in a comparative test taking poplar as the research object. The comparative test between the sap flow detection unit and the ST1221 thermal dissipation plant sap flow meter showed a highly significant linear relationship between the value of sap flux density detected by both (R2=0.8991), and the mean value of sap flux density detected by the ST1221 sap flow meter was 1.1cm/h lower than that of the sap flow detection unit, mainly because the thermal dissipation sap flow meter could not accurately determine the zero flow conditions leading to its underestimation of sap flux density, while the heat ratio method used by the sap flow detection unit can accurately detect low-speed sap flow. The long-term monitoring results of poplar stem water content and sap flow by the composite detection system were consistent with previous studies and in line with plant physiological laws. There was a significant negative correlation between stem water content and sap flux density (Pearson correlation coefficient was -0.7951). A high-performance and low-cost device for plant life state monitoring was provided.

Abstract:Existing stemflow sensors based on the thermal equilibrium method are not accurate in measurement, and the stemflow response is not sensitive to transient changes when transpiration is not significant or when the external temperature is low. Therefore, an adaptive stemflow detection system of heat source power was proposed. Taking camphor stalks as the object, a nested experiment based on the thermal equilibrium method of stemflow calibration was designed by comprehensively considering the trend of the proportional change of stemflow in heat source energy, and the sample set of stemflow rates with multi-gradient under different environmental factors such as external temperature, stemflow rate and cross-sectional area were collected. A combined prediction model of heat source power based on support vector regression (SVR) and genetic algorithm (GA) was established. The results showed that the GA-SVR had good accuracy and robustness, its root mean square error (RMSE), mean absolute error (MAE) and determination coefficient (R2) were 0.015W, 0.012W and 0.989, respectively. The accuracy verification test suggested that the average relative error of the system was 2.64 percentage points (6℃), 2.53 percentage points (11℃) and 3.68 percentage points (16℃) smaller than that of the FLOW-32KS sensor in the low-temperature section. The adaptive model had a small effect on the correction of the results in the high-temperature section which was similar to FLOW-32KS. It was demonstrated that the stemflow detection system improved the accuracy of the heat balance stemflow measurement after embedding the GA-SVR heat source power adaptive model.

Abstract:In the cage mode, the elimination and death of laying hens will lead to changes in the number of hens and eggs production in the cage, so it is necessary to update the number of laying hens in the cage in a timely manner. Traditional machine vision methods recognized poultry by morphology or color, but their detection accuracy was low for complex scenarios such as uneven lighting in the cages, hens obscured by cages and the eggs adhesion. Therefore, based on deep learning and image processing, a lightweight network YOLO v7-tiny-DO was proposed for hens and eggs detection based on YOLO v7-tiny, and an automated counting method was designed. Firstly, the JRWT1412 distortion-free camera and the inspection equipment were used to build an automated data acquisition platform, and a total of 2146 images of caged hens and eggs were acquired as data sources. Then the exponential linear unit (ELU) was applied to the YOLO v7-tiny network to reduce the training time of the model;the regular convolution was replaced in efficient layer aggregation network (ELAN) with depthwise convolution to reduce the number of model parameters, and on this basis, a depthwise over-parameterized depthwise convolutional layer (DO-DConv) was constructed by adding a depthwise over-parametric component (depthwise convolution) to extract the deep features of hens and eggs. At the same time, coordinate attention mechanism (CoordAtt) was embedded into the feature fusion module to improve the model’s perception of the spatial location information of hens and eggs. The results showed that the average precision (AP) of YOLO v7-tiny-DO was 96.9% and 99.3% for hens and eggs respectively, and compared with that of YOLO v7-tiny, the AP of hens and eggs was increased by 3.2 percentage points and 1.4 percentage points, respectively. The model size of YOLO v7-tiny-DO was 5.6MB, which was 6.1MB less than the original model, and it was suitable to be deployed in the inspection robot which lacked computing power. YOLO v7-tiny-DO could achieve high-precision detection and localization under partial occlusion, motion blur and eggs adhesion, and outperformed other models in dim environment, with strong robustness. YOLO v7-tiny-DO recognized that the F1 score of hens and eggs were 97.0% and 98.4% respectively. Compared with the mainstream object detection networks such as Faster R-CNN, SSD, YOLO v4-tiny and YOLO v5n, the F1 score of hens were increased by 21.0 percentage points, 4.0 percentage points, 8.0 percentage points and 1.5 percentage points, respectively, and the F1 scores of eggs were increased by 31.4 percentage points, 25.4 percentage points, 6.4percentage points and 4.4 percentage points, respectively. And frame rates were increased by 95.2f/s, 34.8f/s, 18.4f/s and 8.4f/s, respectively. Finally, the algorithm was deployed to the NVIDIA Jetson AGX Xavier edge computing device and 30 cages were selected for counting tests in a real-world scenario for 3d. The results showed that the average precision of counting hens and eggs for the three test batches were 96.7% and 96.3%, respectively, and the mean absolute error were 0.13 hens and 0.09 eggs per cage, respectively, which can provide a reference for digital management of large-cale farms.

Abstract:Animal face detection is of great significance to the intelligent management of animal farm. At present, goats have the characteristics of multi angle, random distribution and flexibility in the actual feeding environment, which greatly increases the difficulty of goat face detection. Therefore, a goat face detection model combined with coordinate information was proposed based on YOLO v5s target detection network. Firstly, indoor, outdoor, single and multiple goat images were obtained by using mobile devices to build sample data sets. Secondly, coordinate attention mechanism (CA) was integrated into the backbone network of YOLO v5s to make full use of target position information and improve the target recognition accuracy in the occluded area, small target and multi view sample images. The proposed YOLO v5s-CA based approach achieved a precision of 95.6%, a recall of 83.0%, an mAP0.5 of 90.2%, a frame rate of 69f/s and a model size of 13.2MB. Compared with that of the original YOLO v5s model, the detection precision of YOLO v5s-CA was increased by 1.3 percentage points, and the memory space was reduced by 1.2MB. And the overall performance of the YOLO v5s-CA was better than that of the Faster R-CNN, YOLO v4 and YOLO v5s. Experimental results showed that the proposed YOLO v5s-CA approach can improve the detection precision of occluding and small targets by introducing target coordinate information. In addition, datasets with different lighting and camera shake were simulated and constructed to further verify the feasibility of the proposed method. Overall, the proposed deep learning-based goat face detection approach can quickly and effectively detect and locate goat faces in complex scenes, providing detection ideas and technical support for target detection and recognition in intelligent animal farm.

Abstract:Respiration and rumination are the most basic physiological activities of dairy goats. The timely and accurate simultaneous obtain of the information on respiratory and rumination of dairy goats can provide data support for evaluating the health status of dairy goats. Aiming at the deficiency of the existing methods for simultaneous monitoring of respiration and rumination, a method for simultaneously monitoring respiration and rumination of a single lying dairy goats was proposed based on acoustic impulse response. Firstly, the multipath effect of acoustics in indoor space was used to realize omnidirectional acquisition of breast undulation during respiration and mouth chewing movement during rumination of dairy goats. Secondly, the impulse response of the received and transmitted signals was calculated to capture the characteristics of the periodic changes of multipath signals caused by respiration and rumination. Then the frequency difference of respiration and rumination was used to separate respiration and rumination signals. Finally, after amplitude normalization and phase synchronization, visualization of breath and rumination waveforms was realized. In order to verify the effectiveness of the method, the lying dairy goats were selected in different positions and orientations to conduct respiration and rumination monitoring experiments, and the influence of environmental noise on the test was analyzed. The results showed that for dairy goats at different orientations, the average relative error of this method was 2.60% for respiration, 3.51% for rumination, and 2.49% for frame leakage.The results of this study can provide technical support for the health monitoring of dairy goats with infectious diseases and other separately fed dairy goats.

Abstract:The body size parameters of live pigs are important criterion for evaluating the growth state of pigs. The manual measurement of body size is time-consuming and labor-intensive and easy to cause the stress response of pigs. The non-contact pig body size parameter measurement method was studied, referencing the manual measurement experience method, and the pig body size measurement method was proposed based on point cloud semantic segmentation. A non-contact pig point cloud collection platform was established to collect bilateral point cloud data of 3510 groups of pigs. The background point cloud was removed by the pass-through filter and random sampling consistent segmentation method. The outliers were removed by statistical filter. The point cloud was sparsed by voxel downsampling method to complete the pretreatment of pig point cloud. Based on PointNet network and combined with attention module the semantic segmentation model was constructed. The measurement method of pig body size was designed for different parts of segmentation. The experimental results showed that the accuracy of the improved semantic segmentation model was 86.3%, which was higher than that of PointNet, PointNet++ and 3D-RCNN. The maximum absolute error between the measured value and true value was 6.8cm, and the average absolute error was within 5cm, which had a high estimation accuracy. The method can be used for the measurement of pig body size. The research combined semantic segmentation with body size measurement, which can provide an idea for the non-contact measurement.

Abstract:There is a high application demand for accurate counting of disordered targets in agricultural environments, and such counting plays an important guiding role in their biomass and biological density management. In the process of larvae of black soldier fly target tracking, the tracking object has the characteristics of high speed and non-linearity, and the conventional algorithm has the problems of insufficient speed of tracking target and difficulty of re-identification after losing the target. To address these problems, an improved SORT algorithm was proposed, which improved the speed and accuracy of the target tracking algorithm by improving the Kalman filter model, and enhanced the counting accuracy. In addition, for the complex background problem caused by larval trait diversity and mixing in the process of black gadfly larval target recognition, the target recognition accuracy was improved by experimentally comparing the performance of multiple deep learning networks, which selected YOLO v5s algorithm to extract multidimensional features of images. The experimental results showed that in terms of delineation counting, the improved SORT algorithm improved the average accuracy by 4.19 percentage points compared with the original model, from 91.36% to 95.55%, and the effectiveness of the model was proved through simulation and practical application. In terms of target recognition, using the YOLO v5s model on the training set achieved a frame rate of 156f/s, mAP@0.5 value of 99.10%, accuracy of 90.11%, and recall rate of 99.22%. Its overall performance was better than other networks.

Abstract:Accurate, fast and non-destructive estimation of leaf area index (LAI) is of great significance for the production and management of winter wheat. Multi-spectral images were obtained by using the Prime ALTUM camera at the joining stage, booting stage, heading stage and filling stage of winter wheat, and the LAI was measured by using the LAI-2200C plant canopy analyzer. Totally twenty-five vegetation indices were selected based on the Pearson correlation analysis. And eight texture features were extracted: contrast (CON), entropy (ENT), variance (VAR), mean (MEA), homogeneity (HOM), dissimilarity (DIS), the second moment (SEM) and correlation (COR), and three color features: mean (M), variance (V) and skewness (S) were extracted as well. Then the multiple stepwise regression (MSR), support vector regression (SVR) and Gaussian process regression (GPR) models were used for winter wheat LAI inversion. The results showed that compared with single type variable-based models, models with combined texture and color features produced greater estimation accuracy;among the three types of models, GPR model outperformed the other two models in estimating winter wheat LAI;among all models, the GPR model with texture-color features and vegetation indices obtained the best estimation accuracy, with coefficient of determination (R2)of 0.94, root mean square error (RMSE) of 0.17m2/m2, mean absolute error (MAE) of 0.13m2/m2, and normal root mean square error (NRMSE) of 4.06%. The extraction of texture and color features can solve the oversaturation issue of vegetation indices under high-density canopy conditions, and more information can be derived for more accurate estimation of winter wheat LAI, which provided theoretical basis for winter wheat growth monitoring, production and management.

Abstract:Film-mulched drip irrigation with brackish water should avoid soil salt accumulation to maximize the benefits of water-saving. However, soil leaching by rainfall or flood irrigation is lack under greenhouses condition, the partial inhibition of surface evaporation by film mulching combined with the wetted volume intersection due to double-point source drip irrigation exacerbates the irregularity of water and salt distribution in the area between drip tapes (ABDT), which is not conducive to the effective implementation of salt control measures. Therefore, taking the field experimental plot of mulched drip irrigation with the layout of “two films and two rows” as the research object, and using the HYDRUS-2D model to simulate the dynamic distributions of soil water and salt in ABDT based on the two-dimensional simulation domains considering different drip discharge fluxes (0.5~3.0L/h) and bare soil spacing between films (0~50cm). The results indicated that the established model can accurately describe the water and salt distribution in ABDT, and reducing the horizontal distance from the emitter could obtain a high simulation accuracy. As the bare land between the films was decreased from 50cm to 0cm, the average moisture content of the soil in ABDT was increased from 25.12cm3/cm3 to 28.76cm3/cm3, and the average soil solute concentration in ABDT was decreased from 9.53g/L to 6.25g/L. The effect of drip discharge fluxes on soil water-salt distribution in ABDT was relatively low, the maximum differences in soil volume moisture content and salt mass concentration between treatments with discharge fluxes of 0.5h/L and 3.0h/L were only 0.14cm3/cm3 and 0.22g/L, respectively, and both occurred in the bare soil spacing between films of 50cm. Additionally, after long-term dry-wet alternation, salt accumulated on the outside of the drip belt might diffuse inward to ABDT, and with the decrease of bare soil spacing between films, the horizontal position of the lowest soil salinity would move from the location of dripper to ABDT. Findings of this research can provide a theoretical basis for selecting suitable low-salinity planting locations for crop cultivation in greenhouse condition.

Abstract:There are many uncertain factors in the optimal allocation of water resources in irrigated areas, while the optimization models considering the uncertainties are often faced with the problems of complex structure, limited uncertain parameters, low calculation accuracy and efficiency. Therefore, a method for parameter sensitivity analysis of irrigation water optimization model as well as uncertainty optimization was developed through coupling the Latin hypercube-One factor at a time (LH-OAT) method with an irrigation water optimization model. Taking a typical irrigation district in the middle reaches of the Heihe River basin as the case study area, the sensitivity analysis method was conducted for 25 uncertainty parameters from six categories parameters of the model, and the uncertainty optimization of irrigation water use was then realized based on the highly sensitive parameters. The sensitivity ranking of 25 uncertainty parameters in the model was calculated, and 10 highly sensitive parameters were selected. Taking the highly sensitive parameters as uncertainty parameters input for the optimization model, the optimized results of irrigation water use under uncertainty were obtained. The case study indicated that the developed method can effectively find the highly sensitive key parameters in the optimization model, and can comprehensively consider the impact of uncertainty parameters on the optimization results. The method can greatly reduce the number of uncertainty parameters to be considered in an optimization model, which reduced the model complexity and effectively improved the efficiency and accuracy of the model. The study can provide important scientific reference and practical methods for the optimal allocation of water resources in irrigated areas.

Abstract:Concentrations of ammonia and carbon dioxide are important indicators for indoor environment control in pig house. Due to the time-varying and nonlinear coupling characteristics of gas concentration, the prediction accuracy of pig house environment prediction models is still relatively low. Aiming to achieve the precision control for gases concentration in pig house, a time-series data prediction model named ISSA-GRU-ARIMA for harmful gas concentrations was proposed based on gated recurrent unit (GRU), improved sparrow search algorithm (ISSA) fused with autoregressive integrated moving average model (ARIMA). Firstly, a GRU gas concentration time series prediction model was constructed, and Tent chaotic sequence, chaotic disturbance and Gaussian mutation were introduced to enhance the local optimization ability of ISSA algorithm and optimize the hyperparameters of GRU model;then the statistical learning ARIMA method was used to extract the linear features of the optimized ISSA-GRU model’s prediction residuals in order to improve the prediction accuracy of the model. A dataset with 1248 environment data that collected for 52d was used for model training and testing. It was shown that the RMSE, MAPE and R2 of ISSA-GRU-ARIMA model for ammonia concentration prediction were 0.263mg/m3, 8.171% and 0.928, respectively, and those for carbon dioxide concentration prediction were 55.361mg/m3, 4.633% and 0.985, respectively. The constructed ISSA-GRU-ARIMA had high predictive performance, it can provide scientific basis for accurate control of harmful gases in pig house.

Abstract:With the continuous development of blockchain technology in the field of traceability of agricultural products, the quality and safety of agricultural products have been effectively guaranteed. Due to the complex production process of Chinese seed potatoes, obvious physical form differentiation, long production cycle of each link, and many varieties, it is difficult to share the traceability data of all production links, which is prone to the problem of seed potato varieties, grades and other goods transmission. Seed potato production traceability cannot be effectively guaranteed, and the production base and relevant supervision departments cannot obtain all effective traceability data. When the problem of transshipment occurs and the final consumer traces the source of seed potato production, the positioning of the responsibility link is not clear, and it is difficult to find the exact responsible production link and the responsible person and other problems. Based on the above problems, a channe-proof traceability model of seed potato was proposed based on smart contract and digital signature. By using the characteristics of block chain technology, such as tamper-proof, data transparency and data sharing, intelligent contract was used to store the traceability data of the whole link of seed potato production and realize the highly sharing of the traceability data of the whole link of seed potato production. In addition, the smart contract and digital signatures were combined to solve the problem of cross-production easily occurring in the production process by using the public-private key pair verification and the highly autonomous blockchain network ecological environment of smart contract. Based on Hyperledger Fabric, an anti-channeling traceability model for seed potato production base was designed. The related test results showed that the model could realize the functions of seed potato production traceability, anti-channeling, channeling alarm information chain and query. The average link time of seed potato production traceability data was 2566ms, the average query time was 95ms, the average alarm trigger and alarm information link time was 2562ms, and the average query time of specific alarm information was 77ms. The model had high comprehensive performance, which can realize the safe storage of seed potato production traceability data, effectively solve the problem of seed potato production channeling, meet the link and query requirements of seed potato production traceability data, improve the seed potato production quality traceability guarantee, and provide reference for preventing seed potato production channeling to improve the overall efficiency and safety traceability.

Abstract:Traditional destructive detection methods have been unable to meet the requirements of rapid detection of quality content of beans. The existing non-destructive testing equipment has the problems of low stability and accuracy. In order to improve the performance of the device for detecting the quality content of beans, a non-destructive testing device for the quality content of beans was developed based on near infrared spectroscopy technology, which was small, portable and suitable for on-site detection. Based on the developed device, totally 30 samples of soybean, mungbean, red bean and black bean were taken respectively, and the same sample was measured 20 times by means of rotating static multi-spectral averaging and one spectral acquisition. It was concluded that with the increase of acquisition times, the average coefficient of variation of spectral reflectance was gradually decreased until it was flat. The selected bean acquisition times were 16, 8, 14 and 16, and the corresponding average coefficient of variation of spectrum were 2.9%, 2.435%, 2.763% and 3.019%, respectively. Taking soybean as an example, totally 80 samples were selected. Using different pretreatment methods, partial least squares prediction models for protein, crude fat and starch content of soybean were established respectively. The results showed that protein, crude fat and starch models were better than other pretreatments after SG-MSC, SNV and SNV pretreatment, respectively. The Rp were 0.9746, 0.9505 and 0.9607, and the RMSEP were 0.249%, 0.572% and 0.623%, respectively. Totally 40 soybean samples were taken to validate the device model. The Ri of protein, crude fat and starch were 0.9411, 0.9439 and 0.9334, respectively. The RMSEI were 0.465%, 0.604% and 0.673%, respectively. The AD of 20 repeated measurements were 0.409%, 0.623% and 0.637%, respectively. The results showed that the device had good prediction accuracy. Visual Studio 2015 was used as the software development platform to develop the real-time detection software for the quality of beans, which can realize the one-button operation detection of the quality of multiple beans. Elastic compute service and MySQL database were selected. Based on TCP/IP network communication protocol, the detection data were uploaded to the database automatically. Based on the development framework, a front-end network monitoring system was designed to facilitate the monitoring of bean quality and display the database information in real time.

Abstract:According to the topology design theory of parallel mechanism based on the position and orientation characteristics (POC), two three-degree-of-freedom (3-DOF) two-translation-one-rotation (2T1R) parallel mechanisms(PMs) with zero coupling degree and partial motion decoupling were designed, which had the same type and number of kinematic pairs, but the distribution order was different in the branches. Firstly, the main topological characteristics of these two PMs, such as orientation, DOF and coupling degree, were analyzed, and their topological analytical expressions were given. Secondly, according to the kinematic modeling principle based on topological characteristics, the symbolic forward and reverse position solutions of the two PMs were solved, and the workspace, singular conditions and configurations of the two PMs were analyzed respectively. At the same time, according to the single-open-chain method based on virtual work principle, the reverse dynamics of the two PMs were established, and the actuated forces of the two PMs were obtained respectively. Furthermore, the kinematics and dynamics performances of the two PMs were compared, and the optimal PM was suggested. Finally, the conceptual design of the application scenario of the optimal PM used for intelligent sorting and conveying in fruit deep processing was given. The research result can provide a technical basis for the structural design and practical application of this mechanism.