To develop an efficient method for quantifying tobacco plant types in the field, the three-dimensional (3D) point clouds of individual plant of five tobacco cultivars were reconstructed based on multi-view image sequences using the structure from motion method. According to the plant type characteristic indexes commonly used, ten phenotypic parameters such as plant height, top width, bottom width, and maximum width of leaf layer were automatically extracted based on the 3D point cloud of tobacco plant, and the calculation accuracy was evaluated based on the plant height and maximum width of leaf layer measured manually in situ in the field. The results indicated the coefficients of determination (R²) of the plant height and maximum width of leaf layer extracted from the 3D point cloud were all greater than 0.97, and the root mean square errors were 3.0, 3.1 cm, respectively. Meanwhile, the extracted phenotypic parameters of tobacco plants were analyzed by different methods. The results of intergroup correlation analysis showed that 16 pairs of traits were extremely significant positive correlations, while one pair of traits was extremely significant negative correlation. The results of one-way multivariate analysis of variance showed that there were highly significant differences among the plant types. The first three principal components were extracted by principal component analysis, and their cumulative contribution rate to the overall variance was 81.6%. The accuracy of plant type discrimination was 93.7% using Stacking ensemble learning method, which was significantly higher than those using random forest, support vector machine and naive Bayesian. This study can provide a method basis for phenotypic characteristics and plant type recognition of field-grown tobacco plants.

In view of actual situations such as light change, soil, branch and leaf background and fruit overlap in the natural environment, which causing the problem of red tomato recognition during the robotic picking process was not accurate, a tomato image recognition method based on circle fitting algorithm was proposed. We collected the images of tomato by camera, used the red, green, blue (RGB) color space based Matlab as simulation experiment, and preprocessed the tomato images with red-green (R-G) color component. Then, edge detection algorithm, threshold segmentation and watershed segmentation methods were adopted to segment tomato target and the background, respectively. The Otsu segmentation method of threshold segmentation was adopted, which was best to segment target. We used the back propagation-artificial neural network (BP-ANN) and circle fitting algorithm to recognize the tomato fruit. Finally, the contour, centroid and radius of the red tomato were obtained. The results of red tomato images were statistically analyzed, and the recognition rate of circle fitting algorithm was as high as 90.07%. This algorithm not only has a high recognition rate for single fruit, but also solves the problem of multiple fruit overlapping in a complex environment, which lays a good foundation for the following robotic picking work.

Light stress can restrict the normal growth and development of rapeseed seedlings. In order to realize the early diagnosis of light stress in rapeseed seedlings, a 21 d experiment was conducted on rapeseed seedlings of two leaves and one heart stage using hyperspectral imaging technology. After preprocessing the collected canopy leaf spectra, the light-stress-sensitive bands were extracted through spectral reflectance and continuous wavelet transform. Then successive projection algorithm was used to extract characteristic wavelengths, and the continuous wavelet transform-stepwise discriminant analysis method was used to extract wavelet features. To further improve the accuracy of the stress detection model, a total of four features including the area under curvein the 939-978 nm band, the tangent value of the characteristic angle (tan θ), the reflectances at 984 and 1 408 nm were selected by analyzing the characteristics of the spectral band and the evolution of the spectral characteristics at the seedling stage of rapeseed to establish a multi-feature fusion Fisher discriminant model. The results showed that the average classification accuracy of the model was 86.88%, which achieved the best classification effect in the d₂₀ family, with an accuracy of 95.00%. The research provides a powerful reference for the rapid diagnosis of light stress in rapeseed based on hyperspectral imaging technology.

Rice species identification and the retrieval of key phenological stages were processed in the area around Jinhu County, Huai’an City, Jiangsu Province, based on multi-temporal full-polarization synthetic aperture radar (SAR) data. By extracting and analyzing the change characteristics of the time-series curve of the polarization characteristic parameters of rice, the polarization characteristic parameters that are sensitive to the changes of rice phenology were screened out, and a radar phenology index (RPI) that can reflect the key phenological changes of rice growth was constructed, and then was reconstructed by Savitzky-Golay (S-G) filter. The key phenological stages of rice were retrieved by the RPI. The results showed that the response of the polarization parameter Shannon entropy was quite different between japonica rice and indica rice, indicating that Shannon entropy could be used to identify indica rice and japonica rice with precisions of 92.38% and 95.10%, respectively; the curve derivative method was used to extract the characteristic points of the time-series RPI curve of rice, and three key phenological stages of rice were identified. The dates of the identified key phenological stages of rice were all within ±16 d from the date obtained in the field survey. The above results show that the use of RPI can more accurately retrieve the key phenological stages of rice.

This study aims to design and optimize the main structure of a stable and lightweight unmanned vehicle-based field crop phenotyping platform. In order to meet the requirement of high safety, high stability, and lightweight, Pro/Engineer Wildfire 5.0 software was used to design the main structure model of the platform, and HyperWorks 2020 software was employed to perform the finite element analysis and optimize the structure model. Meanwhile, the statics and dynamics analysis of the structure was implemented during the design process. Taking the main structural mass as the objective function, with the material yield limit and the first-order mode as the constraints, the design of experiment (DOE) method was applied to extract the structural parameters of parts with the high sensitivity to the first-order mode and stress under multi-working conditions as design variables, which greatly reduced the variable number. Then, the adaptive response surface method (ARSM) was applied for iterative calculation to obtain the optimal variables. Compared with the corresponding output response of the actual finite element model, the ARSM approximate model produced a low error of 3.79% and 4.32% in the main structure mass and the first-order modal frequency, respectively, which also obtained the maximum stress error of 4.24%, 4.14%, and 1.26% under the static and uniform speed conditions, starting conditions, and emergency shutdown conditions, respectively. These results show that the ARSM approximate model has a high accuracy and the error is less than 5%. Compared with the original structure, the final overall mass was reduced by 63.61% at maintaining the safety factor of each working condition above 5.0. As a result, the main structure of field crop phenotyping platform is obtained with high safety factor and meeting usage requirements.

To investigate the impact of soil physical properties and driving velocity on the driving performance of a tracked tractor and obtain better operating conditions on paddy soils in the plastic state, a paddy soil mechanical model and a tracked tractor physical model were established using the RecurDyn, a multi-body dynamics software. The method of the quadratic orthogonal rotating combination design of four factors and five levels was applied to determine the impacts of clay content, moisture content, density of soils and driving velocity on the driving resistance and subsidence depth. The modeling results indicated that the driving resistance was positively associated with clay content and moisture content of soils, but negatively associated with soil density and driving velocity. The contribution rates of the factors to the driving resistance from high to low followed the order as soil moisture content, driving velocity, soil density, and soil density combined with soil clay content. The subsidence depth increased with greater soil moisture content but decreased with higher soil clay content, soil density, and driving velocity. The contribution rates of the variables to subsidence depth were ordered as soil moisture content, soil density, soil clay content, clay content combined with moisture content, clay content combined with driving velocity, and moisture content combined with driving velocity. Overall, this study quantifies the relationships among soil clay content, soil moisture content, soil density, driving velocity and driving resistance, subsidence depth of a tracked tractor using the modeling approach; and according to the soil clay content and density, the model can be used to optimize the soil moisture content and driving velocity when the driving resistance is minimum.

As a perennial foliage plant, the changes of physiological and biochemical indexes of tea plant under lead aerosol stress and the lead accumulation effect need to be studied urgently. In the present study, the lead aerosol was used to simulate atmospheric pollution, and the lead accumulation in roots, stems, and leaves as well as the changes of photosynthetic pigments and antioxidants in leaves of ‘Wuniuzao’ and ‘Yingshuang’ tea plants were evaluated. Then the model for the rapid detection of each index was established based on Fourier transform infrared (FTIR) spectroscopy. The results showed that the lead content of tea plant leaves in the normal environment was very low, which met the national food safety standards. The lead content of roots was much higher than that of leaves, which proved that the soil-root pathway was the main way for tea plants to accumulate lead in the normal environment. With the increase of stress time, the lead content in the leaves of high concentration lead stress group was significantly higher than that in the stems and roots, which proved that there was an air-leaf absorption pathway, and high concentration lead stress group was up to 14 times that of no lead treatment group. In addition, the photosynthetic pigment and ascorbic acid contents increased initially and then decreased, whereas glutathione content basically increased during the entire 42 days. Support vector machine (SVM) and artificial neural network (ANN) were used to establish quantitative prediction models for monitoring the physiological and biochemical indexes based on the characteristic wave-band of the mid-infrared spectrum, proving that the mid-infrared spectrum could be a potential approach for the rapid detection of physiological and biochemical indexes in tea plants under the lead aerosol stress, and the ANN model showed better effects than the SVM model. The ANN quantitative model of chlorophyll a obtained the best prediction effect, of which the best correlation coefficient of prediction set (r_p) could reach 0.810, and the root-mean-square error of prediction set (RMSE_p) was 0.032 mg/g. The above results indicate that lead aerosol stress could cause the accumulation of lead and result in the significant changes of physiological and biochemical indexes in tea plants, and the FTIR spectroscopy is a reliable method for the rapid detection of physiological and biochemical indexes in tea plants under the lead aerosol stress.

In order to study the change rule of the internal flow field of the self-cleaning mesh filter with different inlet flow rates during the filtration and sewage processes, we used Fluent software to carry out the numerical simulation of the clear-water flow field for the filtration and sewage processes of the self-cleaning mesh filter, and obtained the pressure field and velocity field at the different inlet flow rates. Comparing the numerical simulation results with the physical test results, the relative error was within 10%, which confirmed the reliability of the numerical simulation. Meanwhile, the results of the numerical simulation showed that the higher the water flow rate was, the more violent the mixing of water was in the tank and the sewage system, which could reduce the sedimentation and help to improve the filtration and sewage effects. Moreover, with the increase of the inlet flow rate, the pressure difference between the inlet and outlet of the tank and the sewage system increased, and the head loss also increased, and the requirement for the stability of the filter structure was higher. The above research results can provide a reference for the optimization of the horizontal self-cleaning mesh filter structure.

The shape information of leaves from 39 tobacco varieties was extracted by using landmark method. The differences in leaf shapes were compared and analyzed among different varieties and different leaf positions at different growth stages. Principal component analysis was used to reduce the dimensionality of the data. The sources of differences were visualized among different leaf shapes. Decision tree, random forest and support vector machine were used to perform discriminant analysis on tobacco leaf shapes. The results of the principal component analysis showed that the first three principal components accounted for 42.7%, 21.3% and 10.7% of the total differences in tobacco leaves at the flowering stage, which were characterized by leaf width and the maximum width position, leaf torsion, and petiole size, respectively. The discriminant results of tobacco leaf shape based on machine learning showed that the discriminant accuracy based on landmark data was 52%-62%, while the value was 51%-54% for common leaf shape indicators. The discriminant accuracy on superior or medial leaves was about 10% higher than that of inferior leaves, representing more obvious characteristics of variety. Due to the growth of the leaves, the discriminant accuracy of the leaves at rosette stage was nearly 10% lower than flowering stage. The discriminant accuracy of landmark method increased to 77% after removing 12 atypical varieties. The effect of the landmark method on leaf shape information extraction is better than the common leaf shape indicators, which provides a new idea for the automated extraction of leaf shape information.

Taking the west lakeside oasis of Bosten Lake as the study area, using the measured soil organic carbon content and hyperspectral data, the successive projection algorithm (SPA) was used to filter the characteristic variables from the full-band spectral data, and then the full-band and characteristic bands were used to construct partial least square regression (PLSR) and support vector machine (SVM) models to estimate soil organic carbon content. The results showed that: 1) The soil organic carbon content varied from 0.75 to 48.13 g/kg, with an average value of 13.31 g/kg, showed moderate variability, with a coefficient of variation of 63.19%. 2) The soil organic carbon content and the original spectral reflectance showed a negative correlation, with －0.62＜correlation coefficient (r)＜－0.07. After the bands were preprocessed by Savitzky-Golay-standard normal variate-first derivative (SG-SNV-1st Der), the number of bands that passed the extremely significant test (P＜0.01) were 414, mainly concentrated in 487-575, 725-998 and 1 464-1 514 nm. The correlation between 788, 800 and 1 768 nm was the highest, with the correlation coefficients of more than 0.80. 3) After the spectra were preprocessed by SG-SNV-1st Der, the coefficient of determination (R²) of validation set of PLSR model constructed by SPA was 0.79; root mean square error (RMSE) was 3.58 g/kg; residual prediction deviation (RPD) was 1.99; and ratio of performance to interquartile distance (RPIQ) was 2.23. However, the validation set constructed by SPA combined with SVM was R²=0.81, RMSE=3.16 g/kg, RPD=2.25, RPIQ=2.53. It shows that the model constructed by SPA combined with SVM can better estimate soil organic carbon content in the study area.

In view of the low efficiency of picking high-level apples from fruit trees by fruit farmers, a pipeline transportation device for assisting manual picking was designed. In order to optimize the transportation parameters of the device, a test bench for impact force was established. Taking 'Fuji' apples as the research object, the impact force and mechanical damage of 'Fuji' apples with a fruit diameter of 80-90 mm from a height of 3 m along the pipeline to the fruit box were analyzed. Taking the type of pipeline lining, the lining thickness, and the crash pad thickness as the test factors, and the impact force and damage volume of apples when they fell into the fruit box as the indexes, the response surface test was carried out on the basis of the single factor test. The results of single factor test showed that the pearl cotton material had a relatively good protective effect on apples. The impact force and damage volume gradually decreased with the increase of the lining thickness, and gradually decreased with the increase of the crash pad thickness. The results of the response surface test showed that the optimal combination of transportation parameters was as follows: the lining type was pearl cotton, and the lining thickness was 10 mm, and the crash pad thickness was 8 mm. At the optimal combination conditions, the impact force when the apple fell into the fruit box was 4.99-5.47 N, and the damage volume was 275.02-300.52 mm³. The results of verification test showed that the errors of the impact force and the damage volume of apple were both less than 5%, indicating that the optimization results of pipeline transportation parameters are reliable.

Traditional rice lodging measurements are time-consuming and destructive to rice plants. This study thus developed an easy-to-implement and nondestructive mechanical platform for phenotyping analysis of rice lodging, which can monitor the lodging-resistant characteristics of rice in different growth periods. The lodging measurements were conducted at the jointing stage, booting stage and heading stage from August 15th to September 21st, 2019. The force and displacement were measured from two different directions using the lodging measuring platform with a force sensor, which were used to calculate the dynamic bending stiffness coefficient (K_EI) of rice. Meanwhile, RGB images were collected from the mechanical platform, which were applied to calculate the projected area and the center of force (CoF). The results showed that the K_EI values of lodging-resistant cultivars (Beidao 1 and Shennong 9816) were different from those of lodging cultivars (Yueguang and Qiuguang), which can reflect rice’s lodging resistance in the growth period. In addition, we found that the average distances between CoF and the root of lodging cultivars within the RGB images were larger than those of lodging-resistant cultivars and easily led to rice instability. This study can provide valuable information for rice lodging monitoring and precision breeding.

In order to classify the appearance quality level of Fritillaria thunbergii, the F. thunbergii dataset was constructed with the DigiEye system followed by an image annotation tool. Several statistical learning and object detection algorithms were selected to train and test the F. thunbergii dataset. The results showed that the model trained by the YOLO-X of YOLO (you only look once) series had relatively better performance. In addition, to optimize YOLO-X, according to the unique features of F. thunbergii dataset, a dilated convolution structure was embedded into the end of the backbone feature extraction network of YOLO-X as it could improve the model sensitivity to the dimension feature. The mean average precision (mAP) of the improved model was raised to 99.01%; the average precision (AP) for superfine, level one, level two, moth-eaten, mildewed, and broken F. thunbergii were raised to 99.97%, 98.33%, 98.47%, 98.71%, 99.73%, and 98.85%, respectively; and the weighted harmonic mean of precision and recall (F₁) were raised to 0.99, 0.92, 0.94, 0.97, 0.99, and 0.97, respectively. The tune-up in this study enhanced the detection performance of the model without increasing the number of parameters, computational complexity, or major changes to the original model. This study provides a scientific basis for the subsequent construction of F. thunbergii detection platform.

In order to solve the precision and real-time problems of parental discrimination in the processes of hybrid rice breeding and pollination, an improved DeepLabV3+ hybrid rice breeding parental discrimination semantic segmentation model based on a fully convolution neural network was proposed. The lightweight MobileNetV2 structure of the backbone network was used to replace the Xception structure of the original DeepLabV3+ backbone network, which is more suitable for the application on mobile devices. An extraction method of low-level features with close connection was proposed. The lower-level information and higher-level information were preliminarily concated as the input of the original lower-level information, which enabled the network to obtain more intensive information, thus enhancing the ability of the network to extract details. The results showed that the improved DeepLabV3+ network model had higher segmentation precision for parents of hybrid rice seed production than the original DeepLabV3+ network model, and reduced the model training time and image predictive time. Compared with other mainstream network models and advanced network models, it is found that the accuracy of different parameters of improved DeepLabV3+ network model is improved. This study provides a reference for the development of deep learning in the field of agricultural visual robots.