Abstract:Accelerating the application of artificial intelligence (AI) and other modern information technologies in agriculture is an urgent need for the development of modern agriculture, which will help promote the development of national rural revitalization strategy, digital village construction and smart agriculture. To deeply analyze the potential and direction of smart agriculture driven by AI technology, we reviewed the key technologies of agricultural AI and the research status of agricultural AI for planting, poultry, animal husbandry and agricultural product traceability and classification, analyzed the gap of agricultural AI technology at home and abroad as well as the international situation and challenge of agricultural AI technology in China, and proposed the countermeasures and suggestions for the development of agricultural AI in the future.

Abstract:In modern agriculture, agricultural producers need to know the farmland environment and the growth state of crop in a real-time, accurate and comprehensive manner, and make corresponding analysis, induction and decision of obtained information. Intelligent sensing and analysis technology of agricultural information plays an indispensable role in modern agriculture. In this review, we discussed two aspects of agricultural intelligent sensing and information analysis technology, focused on the research progress of agricultural information intelligent perception technology and agricultural information analysis method based on agricultural internet of things and big data at home and abroad, introduced the application of intelligent decision-making technology based on agricultural information in agricultural machinery and equipment intellectualization. The problems existing in application of agricultural sensors were summarized. Some suggestions were put forward for the development of agricultural information perception, information analysis technology, agricultural database technology and intelligent decision-making technology to provide a reference for the development of intelligent agriculture in future.

Abstract:Weed is one of the main causes for decline of rice yield and quality. The application of chemical herbicide has brought many negative problems, such as crop toxicity, weed resistance and environmental pollution, etc. Mechanical weeding, as an environmentally-friendly weeding method, can effectively replace chemical weeding and alleviate the harm caused by herbicide. Aiming at the technical difficulties of mechanical weeding among rice plants, the research status of mechanical weeding devices among plants were systematically introduced from the perspective of root difference characteristics of weed seedlings. The types and characteristics of new mechanical weeding technology were introduced, and the unique features and advantages of several new mechanical weeding equipments for rice were summarized. It is pointed out that intelligent weeding technology with high precision and high weed localization function will be the inevitable development trend of mechanical weeding technology for rice in future.

Abstract:Objective In order to solve the problem of low automation and high difficulty of mechanical intra-row weeding in paddy field, a pneumatic paddy intra-row weeding device was developed based on the recognition and positioning technology of machine vision.Method The mechanism of pneumatic intra-row weeding device was designed by applying the principle of mechanical design, discrete element dynamics (DEM) simulation method and field test. Firstly, the structure of pneumatic intra-row weeding device was designed, and the geometric parameters of the mechanism were calculated using the kinematic equations. The feasibility of the mechanism was determined from kinematic simulation by the motion analysis module of Pro/E. Then, the interaction between weeding blade and paddy soil was simulated and the simulation results were confirmed through test. Finally, the field test was carried out to evaluate the working performance of the whole machine, and the working parameters affecting weeding rate and seeding injury rate were analyzed using a three-factor and five-level quadratic rotation orthogonal test.Result The connecting rod length of the pneumatic intra-row weeding device was 35.00 mm and the oscillating rod length was 72.24 mm. The horizontal distance from the weeding part to the rotary center was 84 mm and the vertical distance was 191 mm. The DEM simulation results showed the better structure was the blending blade claw in 10°, which contact resistance was relatively low with an average value of 3.12 N when contacting with soil and the soil was disturbed to a great extent at this angle with affected area of 149.69 cm². The field test results showed the device achieved the optimum working performance with the average weeding rate of 83.91% and seedling injury rate of 3.63% when the machine forward speed was 0.25 m/s, the cylinder protruding speed was 0.45 m/s and the weeding depth was 2.5 cm.Conclusion The pneumatic paddy intra-row weeding device satisfies the requirements of above 80% weeding rate and below 4% seedling injury rate. It could meet the requirements of intra-row weeding and seedling avoiding in paddy field.

Abstract:Objective To explore the influences of different adjuvants and nozzles on the distribution and drift of droplets when applied with UAV in the field environment.Method The droplets were dyed by pigment staining. DJI MG-1S was used for spraying. The droplet reception cards were collected, scanned and analyzed. The solutions of different adjuvants (φ=1%) and different nozzles were used to compare the influences of adjuvant and nozzle on distribution of droplets.Result In the lab spraying experiments, the solutions of Yuyan oily and Heda adjuvants (φ=1%) showed better effect on increasing droplet size. The IDK 120-01 nozzle increased droplet size most significantly. In the field test, compared with clear water control, all adjuvants reduced drift, most droplets were within two meters of the spraying route. The amount of droplets at 80 cm above the ground was 40%-60% less than that at 50 cm above the ground. Yuyan oily adjuvant had more deposition in the target area. The IDK 120-01 nozzle showed the most deposition in the target area with less droplet number per unit area.Conclusion Using adjuvants and large droplet size nozzles can significantly reduce drift and increase deposition in the target area. The anti-drift effects of different adjuvants are significantly different.

Abstract:Objective Chlorophyll content is an important indicator of the growth status of japonica rice. This study was aimed at obtaining chlorophyll content of japonica rice in a regional scale in time with UAV hyper-spectral remote sensing technology.Method This study was based on the UAV remote sensing test data of japonica rice in Liaozhong Experiment Station of Shenyang Agricultural University from 2016 to 2017. The successive projection algorithm (SPA) was used to extract the effective bands including 410, 481, 533, 702 and 798 nm. The extracted characteristic bands were used as the input, and the inversion models of chlorophyll contents in japonica rice canopy were established respectively using the extreme learning machine (ELM) and particle swarm optimization-extreme learning machine (PSO-ELM). In the PSO-ELM model, five parameters of PSO algorithm including proportion of population (p), inertial weight (w), learning factors (C₁, C₂), and velocity position correlation coefficient (m) were optimized.Result The optimal parameters were determined: p was 80, w was from 0.9 to 0.3 with a linear decline, C₁ and C₂ were 2.80 and 1.10 respectively, and m was 0.60. For the established models of chlorophyll content in japonica rice using the optimized ELM and PSO-ELM, the determination coefficients were 0.734 and 0.887 respectively, and the mean square error were 1.824 and 0.783 respectively.Conclusion The inversion model for chlorophyll content in japonica rice based on the optimized PSO-ELM has higher precision compared with the model based on ELM, and has better inversion ability of chlorophyll content in japonica rice. This study provides data support and application basis for the diagnosis of chlorophyll content in japonica rice by UAV hyper-spectral remote sensing technology in Northeast China.

Abstract:Objective To obtain and analyze the low altitude remote sensing image of rice field, acquire the weed distribution map, and provide a reference for the precious pesticide application of weeds in the field.Method Three machine learning algorithms including support vector machine (SVM), K-nearest neighbor (KNN) and AdaBoost were used to classify and compare the weed visible light images in rice field captured by UAV after color feature extraction and principal component analysis (PCA) dimensionality reduction. A convolutional neural network (CNN) which can automatically obtain the image features without feature extraction and dimensionality reduction was introduced to classify the weed images and improve the classification accuracy.Result The run time of test set based on SVM, KNN and AdaBoost were 0.500 4, 2.209 2 and 0.411 1 s, and the classification accuracies were 89.75%, 85.58% and 90.25% respectively; The classification accuracy of image based on CNN was 92.41%, which was higher than those of three machine learning algorithms. All machine learning algorithms and CNN could effectively recognize rice and weed, acquire weed distribution information, and generate distribution map of weed in rice field.Conclusion The classification accuracy of weed in rice field based on CNN is the highest, and the weed distribution map generated by CNN is the best.

Abstract:Objective To accurately, efficiently and non-destructively identify the weeds in rice field using deep convolutional neural network, obtain the optimal network model, and provide a theoretical basis for rice field planting management and variable drone spraying.Method The weeds in rice field were taken as the main research object, and weed image samples were collected by CCD photosensitive camera to construct weed data set (PFMW) in rice field. The deep convolutional neural network with multiple structures was used to automatically extract the features of the PFMW data set, and then to model and test.Result VGG16 model achieved the highest precision among all the deep learning models, the F values in Bidens, Goose Starwort, Gomphrena, Sprangle, Eclipta, Wedelia were 0.957, 0.931, 0.955, 0.955, 0.923 and 0.992 respectively, and the average F value was 0.954. The VGG16-SGD model achieved the highest precision in setted deep model optimizer experiments, the F values in each weed mentioned above were 0.987, 0.974, 0.965, 0.967, 0.989 and 0.982 respectively, and the average F value was 0.977. In the equilibrium experiments of sample category quantity in the dataset, the accuracy of the VGG16 model trained by the balanced weed dataset was 0.900, while those of the models trained by the 16.7%, 33.3% and 66.6% category imbalance dataset were 0.888, 0.866 and 0.845 respectively.Conclusion The machine vision and other advanced technologies can accurately identify weeds in rice field. It is of great significance for promoting fine cultivation of rice field and variable drone spraying, etc., and the technology can effectively assist weed control in the process of agricultural planting.

Abstract:Objective To realize intelligent diagnosis of maize leaf diseases with similar spots and complicated background in real field conditions by introducing and improving a regional convolutional neural network algorithm, Faster R-CNN.Method We obtained 1 150 maize leaf images with complicated background for nine kinds of common diseases from maize field and public dataset websites. After manual annotation of the original images, offline data augmentation was used to enlarge the image data. The Faster R-CNN algorithm was introduced and improved for adaptive application by adding batch normalization processing layer and introducing center cost function to improve the identification accuracy of similar disease spots. We used the stochastic gradient descent algorithm to train and optimize this model. Four pre-trained convolution structures for feature extraction were selected and compared in Faster R-CNN training and testing to get the most optimal model. During the test, the trained model was used to select test sets under different weather conditions for comparison, and improved Faster R-CNN was also compared with unimproved Faster R-CNN and SSD algorithm.Result In the framework of improved Faster R-CNN, VGG16 convolutional feature extraction network had better performance than others. The testing image data set was used to verify the model performance, and the average precision of final recognition result was 0.971 8, the average recall rate was 0.971 9, F1 was 0.971 8, and the overall average accuracy reached 97.23%. The recognition effect under sunny conditions was better than that of cloudy conditions. The average precision of improved Faster R-CNN increased by 0.088 6 and the detection time per image decreased by 0.139 s compared with unimproved Faster R-CNN algorithm. The average precision of proposed method was 0.0425 higher than that of SSD algorithm, and the detection time per image decreased by 0.018 s. The results indicated that the improved Faster R-CNN algorithm was superior to unimproved Faster R-CNN and SSD algorithm in the field of intelligent detection of maize diseases under complex field conditions.Conclusion It is feasible to introduce improved Faster R-CNN algorithm into the intelligent diagnosis of maize diseases under complex field conditions, and it has higher accuracy and faster detection speed, which can avoid the subjectivity of traditional artificial identification. The proposed method lays a foundation for precise prevention and control of maize disease in field environment.

Abstract:Objective To realize remote diagnosis of banana diseases.Method Deep learning method was used to diagnose seven common diseases of banana plant. A total of 5 944 images of diseased and healthy banana plants were collected and divided into training set, validation set and testing set according to the ratio of 7∶1∶2. Transfer learning was used to train GoogLeNet which is a deep convolutional neural network for obtaining the diagnosis model. A software system including a mobile application (APP) and a remote server was further developed.Result By comparing different iteration times and optimizers, the model of MomentumOptimizer with 10000 iteration times was finally selected, and the average test accuracy was 98%. The designed mobile APP could acquire banana images in situ, and communicate with the remote server which was integrated with a diagnosis model via the network to obtain diagnosis results in real time.Conclusion The disease diagnosis model can identify the main diseases with high accuracy. The online diagnosis system is simple and easy to operate, it can diagnose common banana diseases online quickly and effectively, and therefore it has a wide application prospect.

Abstract:Objective Combined with the advantages and disadvantages of traditional and modern agricultural pest monitoring, the method of monitoring pest and disease were discussed, which detected the diseased citrus plants by UAV hyperspectral remote sensing technology and judged the disease species and disease degree by artificial field investigation.Method The original hyperspectral images were obtained by UAV. After spectral preprocessing and feature engineering, continuous projection algorithm was used to extract the feature wavelength combination which contributed the most to the classification of citrus diseased plants. Finally, the BP neural network and XgBoost algorithm were used based on the full band, and the logistic regression and support vector machine algorithm were used to establish the classification model based on the characteristic band.Result The AUC scores of BP neural network and XgBoost were 0.8830 and 0.9120 respectively, and the accuracy rates of both methods were over 95%. The feature wavelength combination of 698 and 762 nm was extracted. Based on this characteristic band，the recall rates of logistic regression and support vector machine algorithm were 93.00% and 96.00% respectively.Conclusion The model based on characteristic band shows high accuracy in the classification of disease samples, which proves the effectiveness of characteristic wavelength combination. This result can provide some data and theoretical support for monitoring diseases and pests in citrus plantations.

Abstract:Objective Corn silk is the pollination organ of maize, its growth state will affect the yield of corn. In order to identify the corn silk in real-time and accurately in corn growth state monitoring and yield prediction, a corn silk detection model based on multi-feature fusion SSD (MF-SSD) convolutional neural network was proposed.Method Corn silk detection was based on feature images. The MF-SSD network model was modified from VGG16-SSD through replacing feature extractor by MobileNet and integrating multi-layer feature fusion structure. By optimizing and adjusting the network, three kinds of MF-SSD with different network structures (MF-SSD-cut-3, MF-SSD and MF-SSD-add-3) were tested, and the structure with the best detection performance was selected for corn silk detection. Based on the image data set of corn silk, two kinds of data augmentation techniques (randomly rotating original image from 0 to 180°, horizontally rolling over or translating original image) were applied to improve the training effect of the model. Whether using secondary training strategy and Focal loss to solve sample number imbalance was investigated, and the decrease process of Loss was compared and analyzed.Result The improved SSD model added with multi-layer feature fusion structure could improve the detection ability and recognition speed of network. Compared with VGG16-SSD, the average accuracy of intersection over union increased by 7.2%, the average recall of small target detection of corn silk increased by 19.6%, and the detection speed increased by 18.7%. In the embedded environment having high demand for storage space and run time, MF-SSD-cut-3 obtained shorter run time with smaller storage space in the premise of satisfying detection effect. MF-SSD obtained better detection effect in the condition of taking no account of storage space and run time. The secondary training strategy improved the network convergence speed and model stability. Focal loss effectively solved number imbalance problem of positive and negative samples, and made the training of network model more convergent.Conclusion The detection effect of the proposed MF-SSD model for small targets can meet the needs of real-time detection of corn silk in agricultural production. It can be used for automatic monitoring of corn growth state and yield prediction.

Abstract:Objective In view of the problem that the existed rapid and non-destructive maize seed germination rate testing methods are easily affected by the color of seed skin, the photoacoustic spectroscopy deep scanning technology was proposed to improve the detection accuracy of maize seed germination rate.Method Six maize cultivar seeds with three different colors were selected and treated using artificial aging method to obtain eight kinds of maize seeds with different aging time. The photoacoustic spectrum information with seven different depths was obtained by modulating the spectral frequency. The best scanning frequency and characteristic spectrum were determined by principal component analysis method. Different modeling approaches including partial least squares regression, back propagation neural network, generalized regression neural network and support vector regression were applied for comparing the prediction accuracy to optimize maize seed germination rate model.Result The best scanning frequency of photoacoustic spectrum was 500 Hz. The prediction model accuracy of support vector regression was the highest, and the correlation coefficients were all over 0.980 0. The prediction correlation coefficients of germination rates of six maize cultivar seeds were 0.983 8, 0.984 7, 0.983 6, 0.987 8, 0.983 3 and 0.994 7 respectively, while that of the mixed six cultivar maize seeds reached 0.942 1.Conclusion Through expanding the spectrum, sound and depth information, the photoacoustic spectrum depth scanning technology has a good generalization ability, and is suitable for high-precision germination rate detection of maize with different colors.

Abstract:Objective To obtain remote sensing image of sand sugar tangerine orchard by UAV, rapidly extract the distribution position of fruit trees, and provide references for growth monitoring and yield prediction of fruit trees.Method The visible light remote sensing images taken by drones were used as the research object. Six visible light vegetation indexes of excess red index, excess green index, excess blue index, visible band differential vegetation index, red-green ratio index and blue-green ratio index were calculated. We used the double peak threshold method to select the threshold for fruit tree extraction. Based on the spectral index identification, digital surface model was added as input variable of the identification model, and the comparative test was conducted.Result Compared with using a single spectral index, the addition of digital surface model improved the extraction accuracies of fruit tree and non fruit tree pixels. The total accuracies of six band fusions were all greater than 97%. The total accuracy of excess red index combined with digital surface model was the highest (98.77%) with Kappa coefficient of 0.956 7, and the vegetation extraction accuracies were superior to those of other five combinations of visible light vegetation indexes with digital surface model.Conclusion The combination of digital surface model with visible light vegetation index can excavate more deeply the information contained in the remote sensing data, and provide a reference for the extraction of similar tonal features in the image.

Abstract:Objective To build an internet of things (IoT) system for transmitting the environmental information of citrus orchards in real time, establish prediction models of soil moisture content and electrical conductivity in citrus orchard based on IoT system and long short-term memory (LSTM), and provide references for irrigation and fertilization management as well as effect prediction and evaluation. Method Soil temperature, moisture and electrical conductivity sensors were applied in five IoT nodes and a weather station was set in citrus orchard. The meteorological data and soil moisture data collected in the orchard were transmitted to a remote server via ZigBee, a short range wireless communication technique, and GPRS, a long distance wireless transmission technique. The prediction models of soil moisture content and electrical conductivity were established using weather data based on the LSTM model. The root mean square error (RMSE) and coefficient of determination (R²) were calculated to evaluate the performance of the model.Result The IoT system was capable to transmit environmental data of the citrus orchard to a remote server. LSTM and general regression neural network (GRNN) model were built to predict soil moisture content and electrical conductivity. The performance of models in five nodes were as following: The RMSE of soil moisture content and electrical conductivity ranged from 6.74 to 8.65 and 6.68 to 8.50 respectively based on LSTM model, and ranged from 7.01 to 14.70 and 7.60 to 13.70 respectively based on GRNN model. With the generated LSTM model and meteorological data for predicting, regression analysis was conducted between predicted and measured values of soil moisture content and electrical conductivity. The R² of soil moisture content and electrical conductivity ranged from 0.760 to 0.906 and 0.648 to 0.850 respectively based on LSTM model, and ranged from 0.126 to 0.369 and 0.132 to 0.268 respectively based on GRNN model. The results indicated that the LSTM model performed better than the GRNN model.Conclusion The IoT system for citrus orchard environmental information transmission is established. The LSTM model has high accuracy in predicting soil moisture content and electrical conductivity, and the model can be helpful for guiding irrigation and fertilization management.

Abstract:Objective To simplify network deployment in orchards, extend the signal coverage, provide precise and real-time irrigation monitoring, and improve its compatibility with traditional equipment.Method Remote data transmission and extended coverage of base station signals were realized by combining narrow band internet of things (NB-IoT) and LoRa network. The circuit was examined using terminal electrical parameter and the power was calibrated, which was combined with the anomaly detection algorithm to accurately monitor the operation status of the equipment. The abnormal status was uploaded immediately, and the data upload frequency was reduced. Meanwhile, the main frequency of the processor was reduced to extend the standby time under the premise of ensuring the processing capacity.Result Abnormal status was uploaded within 150 ms and the frequency was limited to 20 000 times per year for the orchard real-time monitoring system. After calibrating the detection power, the determination coeffecient was 0.999 8 for the linear regression prediction of power. The process time of JSON data generated by macro was 10% of that of cJSON method, which further reduced the calculation requirement of MCU. On the premise of meeting the requirements of calculation and control, the main frequency of 2 MHz microprocessor and 200 mA·H lithium battery could meet the minimum requirements of calculation and continuous operation of orchard irrigation monitoring system. The use of low-power microprocessor could further extend the working time.Conclusion This monitoring system extends the coverage of NB-IoT network and realizes accurate, low-cost and real-time remote monitoring.

Abstract:Objective Aiming at the characteristics of large-scale aquaculture environment covering a wide area and interaction of a variety of water environment monitoring factors, to design a device that can simultaneously monitor five water quality parameters including dissolved oxygen, salinity, pH, ammonia nitrogen and temperature. The device can realize long-distance wireless transmission of water quality data through long-distance wireless communication technology, and dynamically display the monitoring environmental factors on the host computer side visualization platform.Method The control core of the data acquisition terminal adopted 16-bit MSP430F149 microcontroller of TI company. The water quality information was collected by various sensors. Ammonia nitrogen collection terminal adopted NHN-202A ammonia nitrogen sensor with test range of 0-10 mg/L. Dissolved oxygen and temperature acquisition terminal adopted RDO-206 sensor with dissolved oxygen range of 0-20 mg/L and temperature range of 0-40 ℃. pH collection terminal adopted PHG-200 sensor with test range of 0-14. Salinity collection terminal adopted DDM-202I/C sensor with test range of 0-0.5%. The server side was built using Linux system and built by the IntelliJ IDEA development tool under JetBrains. The programming language was Java. The online platform used the SpringMVC framework, and the database connection was operated through the HiBernate object-relational mapping framework. The monitoring platform was deployed on the Linux system through Tomcat. The data display interface was realized by calling the visualization library Echarts.Result The absolute error of dissolved oxygen content measured by the system was 0.12 mg/L, while those of salinity, pH and temperature were 0.001%, 0.017, and 0.05 ℃, respectively. In the power consumption test of single acquisition device, 5 200 mA battery could continuously supply power to the terminal device for 28.5 h, and the online system was stable.Conclusion The combination of LoRa wireless communication technology and the data visualization platform on the host computer side in the device enhances the reliability of the long-distance water quality monitoring, and solves the problems of long-distance transmission of monitoring data in dynamic real-time measurement and display of data synchronization on the host computer platform.

Abstract:Objective To propose an individual cow recognition method based on the improved Mask R-CNN algorithm, and solve the problem of low efficiency and strong subjectivity of artificially identifying individual cows in traditional dairy farming.Method This method optimizes the feature extraction network structure in Mask R-CNN, adopts ResNet-50 network embedded in SE block as backbone, and selects image channels by weighting strategy to improve feature utilization. For the problem of inaccurate target edge positioning during instance segmentation, a boundary weighted loss function is added to construct a new Mask loss function to improve the accuracy of boundary detection. A total of 3000 cow images are trained, validated and tested. Result The improved Mask R-CNN model had an average precision (AP) of 100% and IoU_Mask of 91.34%. Compared with the original Mask R-CNN model, AP increased by 3.28% and IoU_Mask increased by 5.92%.Conclusion The proposed method has strong segmentation accuracy and robustness, and can provide a reference for accurate recognition of cow images under complex farming environment.

Abstract:Objective To realize high-precision segmentation of individual pigs under different conditions such as pig adhesion and debris shielding in a group breeding environment.Method A total of 45 group-housed pigs of 20 to 105 days from eight sheds in real farming scenes were studied. Mobile camera images were used as data sources, and data enhancement operations such as changing brightness and adding Gaussian noise were performed to obtain 3 834 annotated pictures. We explored multiple models with the cross-combinations of two backbone networks ResNet50, ResNet101 and two mission networks Mask R-CNN, Cascade mask R-CNN. We also introduced the idea of recurrent residual attention (RRA) into the two major task network models to improve the feature extraction ability and segmentation accuracy of the model without significantly increasing the amount of calculation.Result Compared with Cascade mask R-CNN-ResNet50, Mask R-CNN-ResNet50 improved AP_0.5, AP_0.75, AP_0.5-0.95 and AP_{0.5-0.95-large} by 4.3%, 3.5%, 2.2% and 2.2% respectively. Different numbers of RRA modules were added to explore the impact on the prediction performance of each task model. The experiment showed that adding two RRA modules had the most obvious improvement effect on each task model.Conclusion The Mask R-CNN-ResNet50 model with two RRA modules can more accurately and effectively segment group-housed pigs under different scenes. The results can provide a model support for subsequent identification and behavior analysis of live pigs.