Table of Content

25 August 2024, Volume 63 Issue 8

Previous Issue    Next Issue

Overview of Smart Agriculture

Research on the current situation and problems of smart agriculture development in Hubei Province

CHEN Ping-ting, SHEN Xiang-cheng, LUO Zhi-qing, MA Hai-rong, ZHENG Ming-xue, TU Jing, GUAN Bo

2024, 63(8):  1-4.  doi:10.14088/j.cnki.issn0439-8114.2024.08.001

Asbtract ( 191 )   PDF (1497KB) ( 134 )  

References | Related Articles | Metrics

The current situation and issues of smart agriculture development in Hubei Province were analyzed, and the future paths and countermeasure suggestions were explored. The results indicated that Hubei Province had achieved phased results in areas such as the construction of smart agriculture policy systems, infrastructure, and intelligent equipment. However, challenges existed including inadequate planning and deployment of construction, uneven development of infrastructure, insufficient technology research and application, low utilization rate of smart agricultural machinery and equipment, lack of compound talents, and insufficient funding support. In response to the bottlenecks in the development of smart agriculture in Hubei Province, recommendations were proposed to strengthen the top-level design of policy systems, expand the scope of policy support, unify planning and deployment of infrastructure construction, establish a diverse investment pattern, improve the efficiency of innovative applications of intelligent equipment, enhance talent training and introduction mechanisms, and promote technology dissemination and application.

The potential and pathways of carbon reduction in climate-smart agriculture in Hubei Province

LI Zhi-hui, LI Liang, HOU Jun-rui, LI Zhi-ru, BAI Zhen-zhong

2024, 63(8):  5-9.  doi:10.14088/j.cnki.issn0439-8114.2024.08.002

Asbtract ( 149 )   PDF (1507KB) ( 54 )  

References | Related Articles | Metrics

The development of climate-smart agriculture in Hubei Province not only responded to the global challenge of climate change, but also met the national strategic needs to promote the construction of a strong agricultural province. By adopting innovative concepts and adaptive technologies, climate-smart agriculture could effectively address the triple challenges of food security, climate change and greenhouse gas emissions. However, Hubei Province faced problems such as lack of supporting policies, lagging infrastructure, immature market development and difficulties in data measurement in promoting climate-smart agriculture. In order to achieve the carbon emission reduction target, measures were proposed, such as strengthening the top-level design and mechanism construction of agricultural carbon emission reduction trading, encouraging and promoting the carbon emission reduction demonstration projects of climate-smart agriculture, establishing a platform for sharing the carbon emission reduction results of climate-smart agriculture, and enhancing the agricultural products market value of climate-smart agriculture. These measures were aimed at solving existing problems, promoting the transition of Hubei Province’s agriculture to green, low-carbon and sustainable development, and improving the competitiveness of agriculture in domestic and international markets.

Image and Graphic Recognition

Depth recognition of branch obstacles of apple picking robot based on improved YOLOv4

HUANG Zhe, TANG Shi-xi, SHEN Guan-dong, GAO Xin-yue, WANG Shi-lian

2024, 63(8):  10-16.  doi:10.14088/j.cnki.issn0439-8114.2024.08.003

Asbtract ( 151 )   PDF (7397KB) ( 28 )  

References | Related Articles | Metrics

In computer vision, it was difficult to train and recognize objects with unclear features, and improve the detection in many fields. In order to identify the branches with less obvious characteristics, mainly the branches that covered the apple position when the manipulator picked the apple, a method to obtain the branch semantic skeleton and identify the branch position box by combining semantic segmentation and YOLOv4 algorithm was proposed. Before using the data set for training, the method of semantic segmentation to divide the rectangular envelope of branches, eliminate the small branches and branches that affected the effect of branch recognition, and then label the data set with labelimg and labelme tools was used; Three layers of maximum pooling layer were added to the trained network model, and the CIOU of YOLOv4 was improved in terms of regression loss. A confidence correlation function BIOU was proposed to improve the prediction accuracy according to the complex characteristics and suitable range. The final experiment showed that the F1 and AP of the tree branches with occluded apple positions trained by the improved YOLOv4 network model were 20.00 precentage and 23.36 precentage higher than those of all the trees trained by the original network.

Detection of weeds in paddy field at the seedling stage based on improved YOLOv8 convolutional neural network

LIN Zong-miao, MA Chao, HU Dong

2024, 63(8):  17-22.  doi:10.14088/j.cnki.issn0439-8114.2024.08.004

Asbtract ( 166 )   PDF (15162KB) ( 37 )  

References | Related Articles | Metrics

Aiming at the problems of complex background of paddy field, small size of weed image at the seedling stage, inaccurate identification and positioning under field natural environment photography, an improved YOLOv8 convolutional neural network method for weed detection at the seedling stage was proposed. A dedicated dataset based on the PASCAL VOC dataset format was constructed. First, DenseNet in the network convolution process was added to better extract weed features and optimize the vanishing gradient problem. Then, CBAM（Convolutional block attention module）attention mechanism was added to improve the model’s sensitivity to small sizes. Finally, the WIOU（Weighted intersection over union） loss function was used to optimize the loss function in the original network and improve the positioning ability of the model to the detection target. In the experiment, the performance of the improved algorithm was compared with algorithms such as Faster R-CNN, SSD （Single shot multiBox detector） and the original YOLOv8. The results showed that the improved algorithm was significantly superior to other algorithms, achieving an average precision of 97% and a detection speed of 100.3 frames/s on the test set, respectively. This high-precision and rapid detection capability met the demand for rapid and accurate detection in precision agriculture. This algorithm provided important theoretical and technical support for mechanical equipment to quickly identify weeds during the seedling stage and accurately spray pesticides.

Citrus fruit recognition in natural environment based on improved YOLOv8

YU Sheng-xin, WEI Ying-ying, FANG Hui, LI Min, CHAI Xiu-juan, ZENG Zhi-kang, QIN Ze-lin

2024, 63(8):  23-27.  doi:10.14088/j.cnki.issn0439-8114.2024.08.005

Asbtract ( 135 )   PDF (11492KB) ( 26 )  

References | Related Articles | Metrics

In order to achieve precise and fast identification of citrus fruits, an improved YOLOv8 was proposed. Firstly, certain traditional convolutions in the YOLOv8 were replaced with ODConv full-dimensional dynamic convolutions to enhance the model’s adaptability in complex natural environments. Subsequently, the CIoU loss function of YOLOv8 was substituted with the MPDIoU loss function to address the degradation issue of the CIoU loss function in specific scenarios. Furthermore, the effectiveness of ODConv full-dimensional dynamic convolutions and MPDIoU loss function was verified through a series of ablation experiments. The average recognition accuracy （mAP） of the improved models, YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x, was increased from 86.40%, 88.92%, 88.97%, 88.99%, 89.11% to 88.25%, 89.32%, 89.57%, 89.90%, 90.12%, respectively. Experimental results demonstrated that ODConv full-dimensional dynamic convolutions and MPDIoU loss function significantly enhanced the citrus fruit identification capability of the YOLOv8 in natural environments.

Fruit and vegetable classification and recognition method based on Depthwise Separable Convolution

YUE Zhen, LI Zhuo-ran, WANG Xu-qian, HOU Zong-sheng, MIAO Zhuang, ZHENG Yi, LIU Jie

2024, 63(8):  28-34.  doi:10.14088/j.cnki.issn0439-8114.2024.08.006

Asbtract ( 130 )   PDF (6727KB) ( 56 )  

References | Related Articles | Metrics

Aiming at the problem that the settlement process in agricultural trade markets and fruit and vegetable supermarkets was not intelligent enough and the difficulty of deploying heavy neural network models, the lightweight recognition method of fruit and vegetable classification model was studied. Firstly, in response to the large differences in the environment where the intelligent recognition equipment for fruits and vegetables was located, and the problem of fuzzy features in fruit and vegetable bagging, a multi-scene collection scheme was used to collect 170 kinds of fruits and vegetables and 136 000 pictures in the fruit and vegetable supermarket, and an image preprocessing scheme for weakened bagging was formulated to further enhance the data. Secondly, aiming at the difficulty of deploying the heavyweight neural network and the high cost, a fruit and vegetable classification recognition model based on Depthwise separable convolution was designed, trained and tested. Its Top-1 success rate had reached 96.8%, and the Top-5 success rate had reached 100%. Compared to Mobilenetv2-224, the amount of computation had been reduced by 70%, compared to Mobilenetv3-224, the amount of computing had also been reduced by 60%, and the recognition ability was higher than Mobilenetv2-224 and lower than Mobilenetv3-224. Finally, the problems faced by the designed fruit and vegetable classification model in the actual deployment were analyzed.

Fruit classification recognition methods based on improved deep confidence network

GUO Ying-di, ZHAO Chao-yu

2024, 63(8):  35-38.  doi:10.14088/j.cnki.issn0439-8114.2024.08.007

Asbtract ( 116 )   PDF (1976KB) ( 32 )  

References | Related Articles | Metrics

In order to solve the problems of low recognition accuracy in existing fruit classification recognition methods, based on the fruit classification recognition system, an improved deep confidence network for different fruit classification recognition was proposed. Different feature images were taken as input through 2-channel deep confidence network, and the output was classified using SoftMax. Compared with the conventional classification recognition methods, the proposed method could more accurately achieve the classification recognition of different fruits, and the multi-feature fusion recognition accuracy was the highest, with the recognition accuracy of 98.75%, which met the needs of fruit classification recognition. By optimizing the existing deep learning method, the performance of this method could be effectively improved.

Identification of branches of fruit trees and determination of coordinates of pruning points

SONG Zhen-shuai, ZHOU Yan, ZHONG Ling, YI Jie, SONG Long, HE Lei

2024, 63(8):  39-46.  doi:10.14088/j.cnki.issn0439-8114.2024.08.008

Asbtract ( 149 )   PDF (8833KB) ( 13 )  

References | Related Articles | Metrics

In order to realize the pruning recognition of dormant fruit trees, a network model based on semantic segmentation was studied to identify pruned branches and determine the coordinates of pruning points. A binocular camera was used to build a visual system to obtain the data set of fruit trees. VGG16 and RestNet-50, which were respectively integrated with pre-training weights and CBAM （attention mechanism）, were used as two deep learning models of U-Net backbone feature extraction network to segment pruned branches. At the same time, their effects were obtained and compared. Based on the obtained segmented image, two methods, skeleton extraction and pruning point clustering, were used to determine the coordinates of pruning points. The results showed that the U-Net model based on VGG16 feature extraction network had better recognition results. The mean intersection over union （MIOU）, mean pixel accuracy （MPA） and F scores during the training of the model were 84.80%, 91.83% and 92.679% respectively. By segmenting the model image of artificial simulated fruit trees and using the pruning point clustering method, the two-dimensional coordinates of pruning points could be determined quickly and in real time, which laid the foundation for pruning operations.

Apple fruit detection method based on generative adversarial networks under occlusion conditions

LIU Shuai, XIAO Yi-tong, ZHANG Wu-ping, LI Fu-zhong, WANG Huan-chen

2024, 63(8):  47-53.  doi:10.14088/j.cnki.issn0439-8114.2024.08.009

Asbtract ( 143 )   PDF (8527KB) ( 25 )  

References | Related Articles | Metrics

Aiming at the problem that apple fruit was easily blocked by branches, leaves, and other obstacles in the natural environment, which led to the reduction of recognition accuracy, a method of apple fruit detection based on the fusion of generative adversarial networks （GAN） was introduced. The Faster RCNN model was used to detect the apple fruit and occlusion, mask the occluded apple fruit image, and then repair the occluded apple fruit image with the generative adversarial networks. Finally, the repaired image was transmitted to the target detection model for identification and positioning. The results showed that the combined model of GAN-Faster RCNN, which fused generative adversarial networks, had an mAP of 73.62% on the test set for apple fruits with a large area of occlusion, which was 8.76 percentage points higher than the original model; for the apple fruit with a small area of occlusion, the average precision on the test set was 90.67%, which was 9.54 percentage points higher than the original model. It solved the problem of low accuracy of apple fruit recognition under occlusion conditions with traditional target detection methods.

Research on adaptability of Panax notoginseng disease identification method based on deep learning

HE Heng, ZHOU Ping

2024, 63(8):  54-60.  doi:10.14088/j.cnki.issn0439-8114.2024.08.010

Asbtract ( 121 )   PDF (7150KB) ( 12 )  

References | Related Articles | Metrics

Based on deep learning technology, SSD, YOLOv5 and Fast RCNN models with the same basic network （vgg16-Fast R-CNN, darknet53 -Fast R-CNN） were built to detect round spot, gray mold, powdery mildew and viral diseases of Panax notoginseng under different conditions. The results showed that the m-weight model of YOLOv5 performed the best among all weight models of YOLO, with accuracy rate of 88.62%, recall rate of 89.59%, F1 precision of 89.10%, and average precision of 83.55%. The detection time of a single image was only 0.031 s. Compared with vgg16-Fast R-CNN, which performed better in the two-stage model, the accuracy rate, recall rate, F1 precision, and average precision were only reduced by 1.69 percentage points, 3.92 percentage points, 2.78 percentage points, and 3.47 percentage points respectively, but the detection speed of a single image was increased by 451.4%. Compared with the SSD model, the accuracy rate, recall rate, F1 precision, and average precision of YOLOv5m were improved by 1.06 percentage points, 1.32 percentage points, 1.19 percentage points, and 0.61 percentage points respectively, and the detection speed of a single image was improved by 83.52%. In addition, through the analysis of the confidence and robustness test, it could be seen that YOLOv5m had better disease detection ability in small areas and stronger anti-interference ability in complex environment, and was more conducive to deployment in embedded devices, which met the requirements of real-time detection of Panax notoginseng disease.

The text detection algorithm for agricultural materials image based on Ghost module and its application

YIN Chang-shan, YANG Lin-nan, LUO Shuang

2024, 63(8):  61-65.  doi:10.14088/j.cnki.issn0439-8114.2024.08.011

Asbtract ( 102 )   PDF (3690KB) ( 22 )  

References | Related Articles | Metrics

In response to problems such as slow detection speed of text in agricultural materials image and lack of mobile applications, based on the agricultural materials image dataset, a Ghost module-based text detection algorithm for agricultural materials image was proposed, which improved the DB network, used the MobileNetv2 network to extract the base features, introduced a multi-scale feature fusion module to obtain feature fusion between multiple layers, and used a differentiable binary post-processing algorithm to predict the text, making it possible to quickly detect the text in agricultural materials image. The accuracy of the algorithm on the agricultural materials image dataset was basically up to the standard of mainstream algorithms, with a detection speed of 18.6 img/s and a census count of 2.99 M, with lightweight features, and the algorithm was deployed to mobile devices and ran successfully.

Production and Growth Model

Estimation model of above-ground biomass of grassland in Tarbagatay Prefecture based on Landsat 8 and machine learning

YANG Yan-xiao, CAO Shan-shan, LI Quan-sheng, ZHANG Xian-hua, SUN Wei

2024, 63(8):  66-71.  doi:10.14088/j.cnki.issn0439-8114.2024.08.012

Asbtract ( 132 )   PDF (3997KB) ( 30 )  

References | Related Articles | Metrics

Taking Tarbagatay Prefecture of Xinjiang as the study area, using vegetation index, meteorological data and terrain data as independent variables, combined with the measured biomass data of sample plots in the study area, five machine learning models including k-nearest neighbors regression （KNN）, multiple linear regression （MLR）, gradient boosting decision tree （GBDT）, random forest regression （RF） and Gradient Boosting Decision Tree （GBDT） were analyzed and compared, as well as two ensemble learning models constructed using voting regressor and stacking methods. The results showed that the stacking ensemble learning model had the best performance, with R² of 0.764, RMSE and MAE of 23.29 g/m² and 16.8 g/m², respectively. The optimal model was then used to invert and map above-ground biomass （AGB） of grassland.

Evaluation of four neural network models optimized based on sparrow search algorithm for predicting the stem thickness of Panax notoginseng

SHANG Xiao-jian, ZHANG Rui

2024, 63(8):  72-77.  doi:10.14088/j.cnki.issn0439-8114.2024.08.013

Asbtract ( 117 )   PDF (5448KB) ( 15 )  

References | Related Articles | Metrics

Taking 1-year-old Panax notoginseng as the research object, the effects of light, water, and nutrients on the stem diameter of Panax notoginseng were investigated through orthogonal experiments, sparrow search algorithm （SSA） was used to optimize four models, namely back propagation neural network （BPNN）, Long short term memory （LSTM）, random forest （RF）, and general regression neural network （GRNN）, and these four models were applied to predict the stem thickness of Panax notoginseng. The results showed that non-biological factors such as light, water, and fertilizer had a significant impact on the stem diameter of Panax notoginseng. The degree of influence of each factor on the stem diameter of Panax notoginseng was shading layer>soil moisture content>potassium fulvic acid content from mineral sources>light duration. The SSA-GRNN model had the highest coefficient of determination, which was 0.865 6, followed by the SSA-RF model, SSA-BPNN model, and SA-LSTM model;the MAE and MSE of the SSA-GRNN model were 0.064 1 and 0.008 7, respectively, which were lower than those of the SSA-BPNN model, SSA-LSTM model, and SSA-RF model;the fitness of SSA-RF model and SSA-LSTM model was relatively high, and they were trapped in local optima, making it impossible to achieve a global optimal solution. SSA-GRNN model had the lowest fitness and achieved the best fitness with the least number of iterations.

A prediction model for soil moisture content in blueberry root zone by integrating transformer and LSTM

WANG Yi, CAO Shan-shan, SUN Wei, HU Bo, Gulimila Kizilbek, KONG Fan-tao

2024, 63(8):  78-84.  doi:10.14088/j.cnki.issn0439-8114.2024.08.014

Asbtract ( 116 )   PDF (7162KB) ( 18 )  

References | Related Articles | Metrics

A deep learning prediction model for soil moisture content （transformer LSTM） was constructed, which integrated transformer and LSTM, to address the difficulties in solving nonlinear and complex features, as well as the tendency to fall into local minima in the soil moisture prediction model. Soil and meteorological data from the blueberry（Vaccinium spp.） root zone of two stations, cold shed and outdoor, in the blueberry production area of Dingjiazhai Village, Huangdao District, Qingdao City, Shandong Province, were collected as modeling data,based on Pearson correlation and partial autocorrelation analysis, the data input characteristics and input length of the selected model were compared and analyzed with a single transformer model and LSTM model to evaluate the predictive performance of the model on soil moisture content. The results showed that the transformer LSTM model outperformed both the single transformer model and the LSTM model in prediction accuracy. The mean absolute error （MAE）, root mean square error （RMSE）, mean absolute percentage error （MAPE）, and coefficient of determination （R²） of the transformer LSTM model were 0.245 9, 0.572 0, 0.012 1, and 0.960 6, respectively. The transformer LSTM model could more comprehensively extract feature information from the input sequence of blueberry planting environmental factors, effectively improving the accuracy and level of soil moisture factor prediction.

Simulation analysis and yield prediction of wheat growth based on ensemble learning algorithm and WOFOST model

LI Bo, ZHANG Jing-jing, LEI Jia-cheng, DU Yun

2024, 63(8):  85-91.  doi:10.14088/j.cnki.issn0439-8114.2024.08.015

Asbtract ( 113 )   PDF (5424KB) ( 22 )  

References | Related Articles | Metrics

In response to the limitations of traditional single crop growth models and machine learning models in prediction, the WOFOST model was combined with irrigation models, and an ensemble learning algorithm was used to establish a multi model coupling system （WOFOST coupling model）,simulated experiments were conducted using data from NASA from 1990 to 2020, and experimental results were presented in 2006 and 2018. The results showed that the leaf area index and total biomass of wheat in the WOFOST coupled model were higher than those in the WOFOST model, and the WOFOST coupled model was closer to actual production activities.The MAE and MSE of the coupled algorithm were lower than those of the Bagging, Boosting, and Stacking algorithms, with values of 2.836 and 7.581, respectively. The R² was higher than that of the Bagging, Boosting, and Stacking algorithms, with a value as high as 0.942. The WOFOST coupled model provided a more comprehensive and accurate simulation of crop growth status, improving the accuracy and credibility of yield prediction.

Allocation performance of kiwifruit estimated yield in planting areas based on improved local search algorithm

HUANG Zhen, JING Yue-lou

2024, 63(8):  92-95.  doi:10.14088/j.cnki.issn0439-8114.2024.08.016

Asbtract ( 101 )   PDF (2525KB) ( 9 )  

References | Related Articles | Metrics

In order to further improve the allocation performance of kiwifruit（Actinidia chinensis Planch.） estimated yield in planting areas, local search algorithms were improved through sparrow search algorithm and variable helix factor,and the learning strategy of progressive lenses was used to accelerate the convergence speed of the improved local search algorithm. The results showed that when the estimated yield of kiwifruit remained unchanged, the allocation time gradually decreased as the value of the variable helix factor increased;when the value of the variable helix factor remained unchanged, the estimated yield of kiwifruit increased and the allocation time also increased. The estimated yields of kiwifruit were 10, 20, 30, 40, 50, and 60 tons respectively, with four planting areas of 500, 650, 700, and 850 m². It was recommended to set the variable helix factor value to 4 when the estimated yield of kiwifruit was 10~60 tons;when the estimated yield was 10~40 tons, planting area 3 could meet the optimal planting interval. When the estimated yield was 50~60 tons, planting area 4 could meet the optimal planting interval. Kiwifruit planting interval was reasonably allocated based on different estimated yields and planting areas to ensure that kiwifruit received sufficient nutrients. The improved local search algorithm had a faster convergence speed, and by 500 iterations, the algorithm had tended to converge. The convergence speed of deep learning, particle swarm optimization algorithm, and grey wolf algorithm was lower than that of the improved local search algorithm.

A three-dimensional reconstruction method for tomato plant in greenhouse environment based on ORB-SLAM3

YIN Shu-lin, DONG Luan, YOU Yong-peng, LI Jia-hang

2024, 63(8):  96-103.  doi:10.14088/j.cnki.issn0439-8114.2024.08.017

Asbtract ( 123 )   PDF (11657KB) ( 16 )  

References | Related Articles | Metrics

A tomato plant three-dimensional reconstruction method based on ORB-SLAM3 was proposed to address the difficulty of precise three-dimensional reconstruction of plants in the current production environment,by using a depth camera to capture RGB-D image information, pose estimation was performed based on the feature point information of the foreground and background frames. A point cloud dense reconstruction module was designed to achieve three-dimensional reconstruction of the tomato plant in a greenhouse environment. The results showed that this research method performed well in trajectory estimation as a whole, with no significant drift in the estimated trajectory. Compared with Elasticfusion and BadSlam methods, the trajectory estimated was more closely related to the real trajectory. Pose tracking had a certain degree of robustness, and the number of keyframes used was relatively small, reducing the interference of redundant information on the algorithm;the average absolute error between the reconstructed point cloud fruit diameter and the actual fruit diameter using this research method was 1.48 mm, which was very close to the actual situation,the point cloud had a high degree of restoration and good reconstruction quality,the filtering algorithm did not cause damage to the fruit phenotype information, and the information was preserved intact;this research method could obtain accurate pose information in a greenhouse environment and generate a three-dimensional model of the tomato plant. The three-dimensional reconstruction accuracy was high, which could meet the needs of three-dimensional reconstruction of the tomato plant in a greenhouse environment and target positioning of tomato harvesting robots.

Construction and application of a flue-cured tobacco plant factory planting model based on digital twin technology

YANG Rui, SHANG Xiao-jian, WANG Jing

2024, 63(8):  104-108.  doi:10.14088/j.cnki.issn0439-8114.2024.08.018

Asbtract ( 118 )   PDF (4765KB) ( 29 )  

References | Related Articles | Metrics

In order to study the effects of different growth conditions on the growth status of flue-cured tobacco, the growth characteristics of flue-cured tobacco in the plant factory environment were thoroughly studied, and real-time monitoring of its growth process was achieved through plant factory environmental control technology. Based on the physiological basis of the flue-cured tobacco growth environment, a flue-cured tobacco digital twin model was constructed to simulate and predict the planting process in a virtual environment. The results indicated that the flue-cured tobacco digital twin model, relying on digital twin technology, could simulate the growth status of flue-cured tobacco in real time and accurately, and make corresponding analysis and decisions based on feedback information, in order to accurately manage and optimize the growth process of flue-cured tobacco, and improve the yield and quality of flue-cured tobacco.

County level rice yield prediction model based on CNN-BiLSTM and residual attention

LIANG Ze, CAO Shan-shan, KONG Fan-tao, SUN Wei

2024, 63(8):  109-115.  doi:10.14088/j.cnki.issn0439-8114.2024.08.019

Asbtract ( 120 )   PDF (4099KB) ( 24 )  

References | Related Articles | Metrics

A county-level rice yield prediction model （CNN-BiLSTM-RA） was proposed, which integrated convolutional neural network （CNN）, bidirectional long short term memory network （BiLSTM）, and residual attention （RA） mechanism, key spatial features were effectively extracted from county-level rice meteorological data through CNN layers, the dynamic changes of time series data were deeply analyzed using BiLSTM layers, and RA mechanism was introduced to enhance the recognition and capture of key features in meteorological data. Using historical rice yield and meteorological data from 81 counties in Guangxi from 2015 to 2017 as samples, the prediction accuracy and effectiveness of the CNN-BiLSTM-RA model were compared with CNN, TRANSFORMER, BiLSTM, CNN-BiLSTM, and BiLSTM-RA models. The results showed that the R², MAE, RMSE, and MAPE of the CNN-BiLSTM-RA model were 0.986 1, 0.121 9, 0.224 8, and 0.864 8, respectively, indicating a high degree of fit between the predicted and actual values of the model. The CNN-BiLSTM-RA model fully utilized the spatial feature extraction ability of CNN, the time series data analysis advantages of BiLSTM, and the RA mechanism’s ability to enhance key feature capture. It was a new method suitable for high-precision prediction of rice yield in counties.

A small sample based method for predicting nitrogen application rates in wheat

DU Yun, ZHANG Jing-jing, HAN Bo, LU Zi-ao

2024, 63(8):  116-120.  doi:10.14088/j.cnki.issn0439-8114.2024.08.020

Asbtract ( 119 )   PDF (2560KB) ( 20 )  

References | Related Articles | Metrics

A XGBoost algorithm prediction model based on SBS （SMOTE+Bootstrap） data augmentation method was proposed to address the problem of limited data on fertilization experiments during the growth cycle of wheat （Triticum aestivum L.） and difficulty in effectively predicting fertilization using traditional prediction methods. Based on the original 135 nitrogen application data, the training set （80%） and the test set （20%） were divided. The SMOTE method was used to balance the training and test sets to obtain more feature information. Then, the Bootstrap method was used to expand the balanced data. Finally, the XGBoost prediction model was used for training and compared with other machine learning models. The results showed that using the SMOTE method to balance data significantly improved the prediction accuracy of the SBS-XGBoost model. MSE decreased from the original data of 66.802 to 13.027, MAE decreased from the original data of 6.711 to 2.393, and R² increased from the original data of 0.390 to 0.912. SBS-XGBoost not only performed well in predicting nitrogen application rates in this study, but also provided reference and guidance for scientific prediction of other small sample data.

Agricultural product price prediction based on EMD-PSO-ARIMA model

SHANG Jun-ping, LI Wen-hao, XI Lei, LIU He-bing

2024, 63(8):  121-125.  doi:10.14088/j.cnki.issn0439-8114.2024.08.021

Asbtract ( 111 )   PDF (6442KB) ( 29 )  

References | Related Articles | Metrics

In response to the nonlinear characteristics of agricultural product price data, a price prediction model for agricultural products based on the EMD-PSO-ARIMA model was proposed. Firstly, the EMD algorithm was used to eliminate the instability of price data,secondly, the PSO algorithm was applied to optimize the lag parameters of the ARIMA model and predict the sequence after decomposing the original data,finally, multiple predicted values were accumulated to obtain the final result. Empirical research was conducted on the price data of bulb crops （using garlic as an example）, rhizome crops （using potatoes as an example）, and leafy vegetables （using cabbage as an example） at a farmer’s market in Henan Province from January 2004 to December 2021. The RMSE of the EMD-PSO-ARIMA model for predicting prices of garlic, potatoes, and cabbage was 0.029 5, 0.016 8, and 0.066 9, respectively,MAE was 0.027 4, 0.018 9, 0.059 8, respectively, and MAPE were 0.32%, 0.64%, and 2.54%, respectively;compared with ARIAM, PSO-ARIMA, and EMD-ARIMA models, the three evaluation indicators of the EMD-PSO-ARIMA model had all decreased to varying degrees, and the model had the highest prediction accuracy. The EMD-PSO-ARIMA model could effectively make accurate predictions on the prices of three agricultural products, improving the predictive performance of the model to a certain extent. It could provide decision support for agricultural producers, operators, and governments, and maintain the stability of the agricultural market.

Yunnan sugarcane yield prediction based on intelligent algorithm

WANG Yong-zhi, TIAN Peng, LI Fu-sheng, SUN Ji-hong, SUN Chen, LIU Zhen-yang, LIU Nian, QIAN Ye

2024, 63(8):  126-131.  doi:10.14088/j.cnki.issn0439-8114.2024.08.022

Asbtract ( 96 )   PDF (2364KB) ( 32 )  

References | Related Articles | Metrics

A sugarcane yield prediction model based on intelligent algorithm was constructed to predict sugarcane yield in eight sugarcane production areas in Yunnan Province. Daily meteorological and soil data and yield data of Lincang, Dehong, Pu’er, Wenshan, Honghe, Baoshan, Xishuangbanna, and Yuxi of Yunnan Province for the period of 2000 to 2020 were selected, and the meteorological and soil factors that had a greater impact on the yield of sugarcane were preliminarily screened by the expert scoring method. Stepwise regression analysis algorithm was applied to screen the key influence factors of climate and soil during the growth cycle of sugarcane. Based on the division of the data set and the screening of the key influencing factors, a BP neural network yield prediction model was established with the annual meteorological and soil factors as the input variables and the annual sugarcane yield as the output variable. A Long Short-Term Memory （LSTM） neural network yield prediction model was developed using daily and annual meteorological and soil factors as input vectors and sugarcane yield as the output variable. The results of the test set showed that the coefficient of determination （R²） of the BP neural network model was 0.916, the mean absolute error （MAE） was 286 500 tons, and the root mean square error （RMSE） was 408 300 tons, and the R² of the LSTM neural network model was 0.978, the MAE was 160 400 tons, and the RMSE was 207 200 tons. The prediction accuracy of the LSTM neural network model was high, and the model performance was excellent and could better predict the sugarcane yield in Yunnan.

The crop light temperature yield potential simulation optimization method based on random forest

XU Hao, SONG Hua-lu, ZHANG Hai-bo, ZHANG Xiao-hu, WANG Shuai

2024, 63(8):  132-139.  doi:10.14088/j.cnki.issn0439-8114.2024.08.023

Asbtract ( 111 )   PDF (10716KB) ( 15 )  

References | Related Articles | Metrics

In order to effectively reduce the amount of data required for crop simulation and improve computing efficiency, a model for estimating the light-temperature yield potential of winter wheat was established based on machine learning. Taking 129 agro-meteorological stations in the winter wheat region of China from 1980 to 2009 as the research object, the characteristic variables of temperature, sunshine hours, latitude and longitude, etc., which had a great influence on the simulation of photoperiod yield potential were selected. Based on the input and output data of WheatGrow model, the random forest model （RF_GS） and the random forest model （RF_Mon） with the variables of growing season and month were established. Finally, the performance of the random forest model was evaluated by root mean square error （RMSE）. The results showed that the random forest model could reduce the data requirement under the premise of ensuring the simulation accuracy, and the accuracy of RF_GS was better than that of RF_Mon. The results of the variable importance test and partial dependence plots showed that latitude, sunshine duration in the growing season, sunshine duration in May and minimum temperature in March had a great influence on photoperiod yield potential simulation. If the range of model validation data exceeded the range of training data, the random forest model’s accuracy could not be guaranteed.

Analysis of hyperspectral characteristics from different positions of flue-cured tobacco and construction of discriminating models

YAN Ding, ZHANG Yi-zhi, CHENG Sen, CAI Xian-jie, DONG Xiang-zhou, YANG Yue-zhang, YUE Yao-wen, WANG Da-bin, LIN Run-ying

2024, 63(8):  140-146.  doi:10.14088/j.cnki.issn0439-8114.2024.08.024

Asbtract ( 104 )   PDF (3341KB) ( 16 )  

References | Related Articles | Metrics

Hyperspectral images of three parts （upper B, middle C and lower X） of flue-cured tobacco leaves were obtained by scanning with hyperspectral imaging technique （400~1 700 nm）, and their hyperspectral data were extracted. The hyperspectral characteristics of the three parts of tobacco leaves were studied by correlation analysis, principal component analysis and variance analysis, and five discriminant models （SVM, KNN, RF, LightGBM and XGBoost） for identifying tobacco leaf parts were constructed. The results showed that the spectral reflectance of the three parts of tobacco leaves was C>X>B （400~750 nm）, B>C>X （750~1 400 nm）, and C>B≈X （1 400~1 700 nm）. The hyperspectral data of the three parts of tobacco leaves had a strong correlation. In general, the correlation between the visible light and near-infrared bands was strong in their respective regions, while the correlation between the two was weak. A total of 7 principal components with eigenvalues greater than 1 were extracted, and the cumulative contribution rate of variance was close to 1.00. The spectral reflectance of the three parts of tobacco leaves was significantly different in 450~550 nm and 750~1 400 nm regions. The middle leaves had significant differences from the upper and lower leaves at 550~850 nm and 1 400~1 700 nm, respectively. The upper leaves had significant differences from the middle and lower leaves at 400~450 nm, respectively. The lower leaves had significant differences from the upper and middle leaves at around 680 nm. SVM performed best in distinguishing tobacco leaves in different parts, with accuracy, precision, recall and F1 scores all reaching above 95%, LightGBM performed in the middle, with various indicators between 90% and 95%, RF, KNN and XGBoost performed relatively poorly, with various indicators below 90%.

System and Platform

Design and implementation of rural domestic sewage management platform based on GIS

XU Lan-sheng, JIANG Hua

2024, 63(8):  147-152.  doi:10.14088/j.cnki.issn0439-8114.2024.08.025

Asbtract ( 110 )   PDF (7152KB) ( 7 )  

References | Related Articles | Metrics

In order to improve the management efficiency and informatization level of rural domestic sewage management, a rural domestic sewage management platform design was proposed based on GIS and developed with related technologies. The platform designed a set of closed-loop management process including rural domestic sewage problems reported by farmers, treated by operation and maintenance units, and audited by management departments. The method of integrating GIS with various sensors of rural domestic sewage, and the key technical route of some functional modules were expounded. Taking Linxiang District of Lincang City as an example, the application of the platform in the local rural domestic sewage information management was introduced.Through the use of the platform, the rationality and scientificity of the platform design were demonstrated.

A lightweight deep learning network RepYOLO for embedded devices

ZHOU Gan-wei, CHEN Jia-yue, WU Jia-wei, ZHAO Ya-qi, ZHAO Yi-kai, ZHANG Xiao-ying

2024, 63(8):  153-157.  doi:10.14088/j.cnki.issn0439-8114.2024.08.026

Asbtract ( 123 )   PDF (7982KB) ( 19 )  

References | Related Articles | Metrics

A lightweight deep learning network model RepYOLO algorithm was proposed and transplanted to embedded device MCU/MPU. The network model RepYOLO took YOLOv4 as the base network model. By modifying YOLOv4’s backbone network CSPDarkNet to the RepBlock structure, introducing the CBAM attention mechanism in the Neck layer, and replacing the anchor-based detection head with an anchor-free detection head in the head layer along with integrating the ATSS algorithm, the computational load was reduced, and both inference speed and detection accuracy were improved. The experimental results showed that compared with the original YOLOv4 model, the network model RepYOLO showed more significant advantages in wheat spike detection, and its precision rate, recall rate, F1 value and average precision value were increased by 4.7, 3.6, 1.5 and 1.7 percentage points, respectively. In addition, RepYOLO reduced inference time on embedded devices MCU/MPU by 37.03% and 41.44%, respectively.

Design and implementation of a farmland scene visualization system based on drone photography technology

LIU Ming, LI Yong-ke, JIN Sheng, LI Yue, YU En

2024, 63(8):  158-163.  doi:10.14088/j.cnki.issn0439-8114.2024.08.027

Asbtract ( 110 )   PDF (14549KB) ( 15 )  

References | Related Articles | Metrics

In order to quickly and accurately obtain digital images and 3D image information of large-scale farmland, a method based on drone tilt photography technology was proposed to construct high-precision farmland images and real-life 3D models. Taking Huaxing Farm in Changji City, Xinjiang Uygur Autonomous Region as the research area, based on the angle image date images collected by drones, the high-precision digital orthophoto images and 3D real-life models of farmland were made by using panoramic map construction technology and 3D modeling technology. On the basis of this data, it was integrated with basic information data of farmland to build a visualization system based on farmland scenes by B/S architecture. The results showed that this visualization system could visualize and display two-dimensional and three-dimensional views of farmland scenes.

Research on the classification method of cigarette blend modules with SOM neural network considering alternatives

WANG Lin, ZUO Ping-cong, GUAN Yu-han, ZHU Yong-qi, ZHOU Hong-shen, WU Qing-hua

2024, 63(8):  164-170.  doi:10.14088/j.cnki.issn0439-8114.2024.08.028

Asbtract ( 100 )   PDF (3733KB) ( 18 )  

References | Related Articles | Metrics

In order to improve the decision-making efficiency of module substitution and the flexibility and production efficiency of the entire cigarette manufacturing system, a substitution degree based SOM neural network model was proposed to classify cigarette blend modules, and the effect of this model was compared with the historical substitution statistical results. The results showed that the substitution degree could better measure the degree of substitution between modules. The larger the substitution degree, the stronger the consistency of the quality indicators in each category, the more similar the quality of the modules, and the more recommended for mutual substitution. When classifying cigarette formula modules with different substitution degree standard values, the larger the value was, the finer the classification was. It was most appropriate to select the substitution degree standard value of 3.06 as the standard of strong substitution of cigarette formula modules for classification where the quality of cigarette blend modules in each category had a high similarity. The classification results of SOM neural networks based on substitution degree showed that the proportion of intra-class substitution was superior to general SOM neural network algorithms, two-stage clustering algorithms, and K-means clustering algorithms. When the substitution degree standard value was 3.06, the intra-class mutual substitution rate could reach 95.39%, while the inter-class substitution rate was less than 5.00%. The replacement rate of modules in the same class was excellent.

Design of visualization smart agricultural data management system based on multi-sensor

WAN Qing-song, LUO Xiao-jiao

2024, 63(8):  171-175.  doi:10.14088/j.cnki.issn0439-8114.2024.08.029

Asbtract ( 104 )   PDF (1741KB) ( 19 )  

References | Related Articles | Metrics

The protocol system based on WSNs and MQTT had accurate agricultural data collection results, but it was affected by rapid expansion data, resulting in poor data management effect. Aiming at this problem, a visual smart agricultural data management system based on Spring Boot was proposed. The Spring Boot micro service was implemented from three levels of presentation layer, business logic layer and data layer. The air temperature and humidity, light intensity and soil moisture data could be transmitted by using sensors of air temperature and humidity, light intensity and soil moisture. Processing and accessing data in Spring Boot micro service mode could avoid excessive data inflation. The observation distance between the two sensors was calculated, and the environmental parameters were fused to complete the design of a smart agricultural data management system. The experimental results showed that the system could accurately collect data of air temperature, air humidity, soil moisture and light intensity. The data visualization performance was good, indicating that the use of the system had good data management effects.

Research on intelligent agricultural monitoring and management system based on Internet of Things

HE Jian-qiang, ZHANG Ying, XU Xing

2024, 63(8):  176-181.  doi:10.14088/j.cnki.issn0439-8114.2024.08.030

Asbtract ( 117 )   PDF (12625KB) ( 30 )  

References | Related Articles | Metrics

In order to build a real-time monitoring system of crop environmental information and realize intelligent management of crop growth, an intelligent agricultural monitoring and management system was constructed. Firstly, based on the Internet of Things technology, the intelligent agricultural monitoring and management system was built with STM32 as the control core from three dimensions of perception layer, transmission layer and application layer. The environmental data such as environmental temperature and humidity, soil temperature and humidity, light intensity and air quality, etc, were collected through sensors, and transmitted to the server using a wireless module to achieve monitoring and remote control of environmental factors. Secondly, the image processing algorithm and SVM model based on OpenCV database were used to classify and recognize crop disease images. Finally, the upper computer UI interface was designed based on the QT platform. Through experimental tests, the system was accurate and reliable in information transmission and equipment control. The average accuracy rate of the system for disease image classification was 96.3%, and the accuracy rate of pomegranate brown spot recognition was the highest, reaching 97.4%.

Remote Sensing Technology

Risk zoning of hail disaster in tobacco area of Jianshi County based on GEE and Sentinel-2 images

ZHENG Xiang-tian, LUO Ju-ying, WANG Chuan-yi, TAN Yan-li, LIU Qiao, WAN Jun

2024, 63(8):  182-187.  doi:10.14088/j.cnki.issn0439-8114.2024.08.031

Asbtract ( 136 )   PDF (6108KB) ( 54 )  

References | Related Articles | Metrics

Sentinel-2 remote sensing images of Jianshi County in 2021 were selected, and based on GEE cloud platform, the parameters of NDVI, NDWI and NDBI were calculated. The random forest method was used to divide Jianshi County into five categories： water area, forest land, construction land, cultivated land and tobacco leaf. Combined with the results of the meteorological disaster risk survey in Jianshi County, the tobacco hail disaster risk regionalization model was established, and the risk grade of tobacco leaves subjected to hail disaster was evaluated. The results showed that the tobacco planting areas in Jianshi County were mainly distributed in the central and western parts of Guandian Town, most of Huaping Town, and the line from Gaoping Town to Maotian. Most of them were located in areas with low to medium risk of hail disasters in the tobacco growing areas. It was necessary to strengthen monitoring and early warning, and timely carry out artificial hail prevention operations. The risk level of hail disasters was low in the southern areas of Liangxiang, Sanli Township, and Gaoping Town, and the tobacco planting area could be appropriately increased. The risk level of hail disasters in the northeast and southern mountainous areas of Jianshi County was relatively high, and it was necessary to strengthen disaster prevention and appropriately reduce tobacco planting. The results of this study, as a typical case of the 2022 Enshi Prefecture disaster risk survey, had been applied in practical business, providing a good reference basis for tobacco planting production and meteorological disaster prevention and reduction.

The information extraction technology of farmland flood disaster based on SAR satellite remote sensing technology

FAN Bing, MA Liang, YUAN Xiu-zhen, LI Fu-lin, DUAN Zhou, WU Jia-mei

2024, 63(8):  188-193.  doi:10.14088/j.cnki.issn0439-8114.2024.08.032

Asbtract ( 103 )   PDF (15241KB) ( 25 )  

References | Related Articles | Metrics

In order to improve the extraction ability of farmland flood disaster information, the automatic extraction method of water body and farmland boundary information of SAR satellite remote sensing image was explored. Taking a heavy rainfall process in Fengcheng, Jiangxi Province as an example, the threshold segmentation method, radar and optical image fusion method was adopted, Sentinel 1 satellite image was used to extract water information before the disaster, and chaohu 1 satellite image was used to extract water information in the disaster. The two results were superimposed and the new water body range of the heavy precipitation was obtained. The satellite images of Sentinel 2 were used to superimpose the sky map images to extract the farmland boundary range of the study area, and this boundary was superimposed with the new water body range to obtain the scope of the farmland flood disaster area affected by the heavy rainfall. Through evaluation, this method could effectively improve the classification accuracy of ground scattering features, and the integrity rate of the extracted 11 flooded farmland verification plots was above 80%. SAR remote sensing image was not affected by cloud and rain weather, and could provide strong data support in the emergency monitoring of the flood disaster. This analysis method was conducive to the relevant departments to fully grasp the farmland disaster data, make emergency response quickly, and improve the emergency rescue and management ability of flood.

The method of extracting rice-crayfish paddy fields distribution in the Jianghan Plain based on Sentinel-2 data

WANG Jing, WAN Jun, DENG Huan-huan

2024, 63(8):  194-200.  doi:10.14088/j.cnki.issn0439-8114.2024.08.033

Asbtract ( 109 )   PDF (10647KB) ( 14 )  

References | Related Articles | Metrics

Taking the Jianghan Plain as the study area, and based on the Sentinel-2 MSI L2 data provided by AI Earth cloud platform, on the basis of field sampling samples and visually interpreted samples, the critical time and index threshold to distinguish rice-crayfish paddy fields from other ground types were summarized by analyzing the temporal change of rice-crayfish paddy fields in remote sensing images, so as to construct a decision tree model for rice-crayfish paddy fields extraction. Finally, the spatial distribution of rice-crayfish paddy fields in the Jianghan Plain from 2022 to 2023 was extracted. Finally, the accuracy of the method was evaluated based on the sample data, and the overall accuracy reached 93.25% with Kappa coefficient of 0.842 9, which showed that the method had better extraction results.

Extraction of winter wheat planting area in county regions based on principal component analysis features fused with GF-6 WFV image

ZHANG Meng, XU Jian-peng, ZHOU Lu-yang, WANG Jie, WANG Zhuang, YUE Wei

2024, 63(8):  201-208.  doi:10.14088/j.cnki.issn0439-8114.2024.08.034

Asbtract ( 95 )   PDF (7286KB) ( 10 )  

References | Related Articles | Metrics

In order to obtain the planting information of winter wheat at county level accurately and quickly, Guzhen County of Anhui Province was selected as the research area, aiming at the problems of high cost, low efficiency and complex process of multi-temporal methods. An effective area extraction method based on single temporal GF-6 WFV image principal component analysis and original spectral band normalization fusion was proposed, and K-nearest neighbor algorithm was used for land cover classification. The results showed that the proposed method was superior to the other two benchmark methods of RAW and PDR, and the best effect was achieved when the dimensionality reduction parameter was 3. The overall accuracy and Kappa coefficient were 89.71% and 0.87, respectively. The actual accuracy of the winter wheat extraction area was 98.49%, with a relative error of only 1.51%.

Yield estimation and precision evaluation of dry-fed maize based on WOFOST model and remote sensing data

HOU Chen-lian, ZHANG Wu-ping, WANG Guo-fang, LI Fu-zhong

2024, 63(8):  209-215.  doi:10.14088/j.cnki.issn0439-8114.2024.08.035

Asbtract ( 117 )   PDF (5669KB) ( 24 )  

References | Related Articles | Metrics

Lingqiu County, Jiexiu County, Xi County and Yanhu County of Shanxi Province in the eastern part of the Loess Plateau were selected as the study area. Field observation data from 2005 to 2012 were used to analyze the sensitivity of model parameters by using EFAST method, and the growth parameters of maize were adjusted by trial and error method. On this basis, MCD15A3H remote sensing data was fused, leaf area index （LAI） data, which was taken as the coupling variable was assimilated into the calibrated WOFOST model using SUBPLEX algorithm, and the growth and development process of maize in each region was simulated again. The results showed that the calibrated WOFOST model had better simulation results for growth period and yield. The average error between simulated value and measured value in the growth period was less than 3 days, the correlation coefficient （r） between simulated value and measured value of yield was 0.80, and the root mean square error （RMSE） was 956 kg/hm². After assimilating remote sensing data with WOFOST model, the r of simulated and measured yield increased from 0.80 to 0.91, and the RMSE decreased from 956 kg/hm² to 660 kg/hm².

The relationship between vegetation cover change and its influencing factors in Xinjiang based on machine learning

MA Nan, CAI Zhao-zhao, BAI Tao

2024, 63(8):  216-222.  doi:10.14088/j.cnki.issn0439-8114.2024.08.036

Asbtract ( 120 )   PDF (6586KB) ( 33 )  

References | Related Articles | Metrics

Taking the relationship between vegetation cover change and its influencing factors in Xinjiang as the research object, by comparing multiple linear regression, random forest regression, XGBoost and support vector machine regression, the model with the highest accuracy was selected, and 15 influencing factors were reorganized and analyzed according to the attribute importance degree calculated by the optimal model. The effects of 15 factors including air temperature, precipitation, radiation amount, potential evaporation, longitude, dimension, elevation, landform type, slope, slope orientation, human impact index, runoff, soil type, soil moisture, and vegetation type on vegetation cover change were explored. The results showed that XGBoost model had the highest prediction accuracy for normalized vegetation index （NDVI）, followed by random forest regression. In the study area, the most influential factors on NDVI were soil moisture, runoff, vegetation type, longitude, potential evaporation, air temperature, radiation amount, landform type and precipitation. In terms of the types of influencing factors, climatic conditions had the greatest influence on NDVI, followed by soil characteristics, and the topographic and geomorphic factors were the lowest compared with the first two.

Soil salinization monitoring model based on remote sensing derivative processing and optimal spectral index

TANG Zi-ru, WU Tong, TAN Shi-lin, YUE Sheng-ru

2024, 63(8):  223-230.  doi:10.14088/j.cnki.issn0439-8114.2024.08.037

Asbtract ( 105 )   PDF (19263KB) ( 14 )  

References | Related Articles | Metrics

Using Landsat-8 remote sensing data, the correlation of band reflectance, 2D and 3D indices with soil conductivity was analyzed based on three treatments： Raw spectra, first-order derivatives and second-order derivatives. The optimal spectral index was selected as the input parameter of the neural network algorithm, and the soil salinization prediction model was constructed based on MATLAB. The results showed that the 2D and 3D spectral indices had a higher correlation with the conductivity than the original spectra, and the overall correlation between the 2D and 3D indices constructed after the second-order derivative treatment and soil conductivity was better than that of the first-order derivative treatment and the original spectra. The accuracy of the model constructed by choosing B1 to B7 as the input parameters of the neural network algorithm under the original spectra was optimal, the correlation coefficients of the training set, validation set, test set and the whole were 0.732 4, 0.716 4, 0.444 5, 0.691 9, respectively, and the constructed model had high prediction accuracy when the soil conductivity was around 1 000 μS/cm.

Dynamic inversion of plant height of soybean and maize under composite planting system

WANG Huan-chen, ZHANG Wu-ping, LI Fu-zhong

2024, 63(8):  231-235.  doi:10.14088/j.cnki.issn0439-8114.2024.08.038

Asbtract ( 101 )   PDF (4437KB) ( 17 )  

References | Related Articles | Metrics

In order to accurately estimate the plant height of soybean and maize under composite planting system, the digital terrain model （DTM） generated by the mean interpolation method and linear interpolation method and the digital surface model （DSM） generated by UAV image were used for processing, respectively. The canopy height model （CHM） of soybean and maize under composite planting mode was obtained. The linear regression method combined with the determination coefficient （R²） and root mean square error （RMSE） was used to analyze the accuracy of the model, and the interpolation method corresponding to the plant height model with higher accuracy was selected, and the dynamic inversion of plant height of soybean and maize was realized by using the model. The results showed that the DTM obtained by linear interpolation had a better CHM estimation effect after processing, and the model’s R² was 0.85 and RMSE was 7.05 cm. It was indicated that in the absence of complete DTM data, it was feasible to construct the canopy height model by using small-area DTM generated by interpolation and DSM generated by UAV image, and realize dynamic inversion of soybean and maize plant height under composite planting mode.

Extraction of rice from remote sensing images based on deep heterogeneous transfer learning

QIU Ru-qiong, HE Li-hua, LI Meng-fan

2024, 63(8):  236-242.  doi:10.14088/j.cnki.issn0439-8114.2024.08.039

Asbtract ( 103 )   PDF (7465KB) ( 9 )  

References | Related Articles | Metrics

In order to achieve high-quality construction and reuse of rice extraction models from remote sensing images based on heterogeneous with only unlabeled samples in the target domain, a deep heterogeneous feature transfer learning model based on temporal and spatial constraints was constructed. Firstly, unlabeled sample groups in the source domains and target domains were constructed based on spatial location, and their deep features were extracted; secondly, in order to reduce the negative transfer impact of features and realize precise transfer of heterogeneous features, a heterogeneous feature transferring model was constructed by using a composite loss function including corresponding sample feature conversion loss, corresponding sample feature regular loss, and sample reconstruction loss; finally, in order to improve the accuracy of classification, a semi-supervised classification model was established, and HingLoss was introduced to eliminate the impact of wrong pseudo labels. The results showed that the research method could realize sample feature transfer between images at different resolutions. Compared with the case without feature transfer, the accuracy rate was improved by 27.68 percentage points, and the F1 score was improved by 17.3 percentage points.

Intelligent Aquaculture

Optimization of multistory pig building environment precise control

HAN Qi-zheng, WANG Yi-ning, SUN Xiao-tong, HU Pei-ao, YIN Bao-quan

2024, 63(8):  243-246.  doi:10.14088/j.cnki.issn0439-8114.2024.08.040

Asbtract ( 98 )   PDF (3700KB) ( 19 )  

References | Related Articles | Metrics

Multistory pig building could make full use of land resources, and had the advantages of improving production efficiency per unit area and reducing environmental pollution area. However, due to the different environmental requirements of different floors in the multistory pig building, it was difficult to accurately control the environment in the multistory pig building. In order to study the precise control scheme of multistory pig building environment, ANSYS FLUENT software was used to optimize and improve the airflow field distribution of multistory pig building based on computational fluid dynamics （CFD）. Based on the standard k-ε turbulence model, the airflow field in the multistory pig building was simulated, and the ventilation position in the piggery was adjusted and optimized. The results showed that the optimized design scheme could make the airflow field in the piggery more uniform and avoid large environmental deviation in the same piggery. Finally, based on the optimized design scheme, PLC was used to design and realize the precise control of the environment in the multistory pig building.

Intelligent Agricultural Machinery

Research on path optimization of unmanned agricultural machinery based on hybrid ant colony algorithm

YANG Hu-jia, ZHANG Ya-jun, WANG Peng-jie, WANG Dong, WANG Ya-ping

2024, 63(8):  247-251.  doi:10.14088/j.cnki.issn0439-8114.2024.08.041

Asbtract ( 130 )   PDF (9431KB) ( 10 )  

References | Related Articles | Metrics

In addressing the challenges of slow iteration speed and low path safety in the optimization process of unmanned agricultural machinery path planning under complex environments in smart agriculture, a hybrid ant colony algorithm was proposed, integrating artificial potential fields, quantum behavior and a B-spline-based smoothing strategy. This method introduced artificial potential fields in the early iterations to address the issues of slow iteration speed and balance global optimality. In the mid-term of path optimization, quantum behavior was incorporated to enhance the algorithm’s capability to obtain high-quality solutions by adjusting the information density threshold, improving algorithm state selection probabilities, and avoiding local optima. In the later stages of iteration, the B-spline-based smoothing strategy was integrated to optimize the optimal path and enhance the obstacle avoidance capability of unmanned agricultural machinery. Simulation experiment results demonstrated that the unmanned agricultural machinery based on the hybrid ant colony algorithm showed significantly improved path optimization ability in complex environments. The response speed of path optimization was increased by 73 times, and the distance was reduced by over 11.8% after path optimization.

Intelligent Monitoring

Design on electronic nose for detecting the freshness of Larimichthys polyactis based on sensor array

HUANG Can-can, CHEN Ya-long, CHEN Hai-gen

2024, 63(8):  252-256.  doi:10.14088/j.cnki.issn0439-8114.2024.08.042

Asbtract ( 102 )   PDF (2808KB) ( 20 )  

References | Related Articles | Metrics

To develop an electronic nose that can be used to determine the freshness of agricultural products, taking the Larimichthys polyactis as the research object, the corruption test was carried out under constant conditions, and the data were collected. The sensor response value was taken as the independent variable and the total volatile basic nitrogen （TVB-N） value of Larimichthys polyactis was taken as the dependent variable. The multiple linear regression, partial least squares and BPNN were used to establish a prediction model of Larimichthys polyactis meat quality grade, and the performance of the model was analyzed by comparing the correlation coefficient R and the average error percentage RE-mean of the three models for TVB-N value prediction. The results showed that the electronic nose could distinguish the fresh rotten samples of Larimichthys polyactis by using the prediction model. It could be seen that the prediction algorithm of BP neural network could achieve the best prediction for the samples, and the performance of the multiple linear regression model and the least square method was poor.

Monitoring and diagnosis of potassium nutrition in Ipomoea batatas leaves based on spectral reflectivity

LU Yan-jun, WANG Xu-wei, HU Ji-jie, CHEN Shao-jie, CHEN Yu, LYU Zun-fu

2024, 63(8):  257-261.  doi:10.14088/j.cnki.issn0439-8114.2024.08.043

Asbtract ( 96 )   PDF (5291KB) ( 27 )  

References | Related Articles | Metrics

Two Ipomoea batatas varieties, Shangshu 19 and Xinxiang, were used as experimental materials. By setting different gradient potassium treatments to determine the spectral reflectance of leaves, Ipomoea batatas leaves potassium content and potassium nutrient index prediction models were constructed based on the ratio vegetation index （RVI）. The results showed that the linear model constructed by RVI and potassium content in leaves showed that RVI （R_{1 598 nm}, R_{1 771 nm}） had a high prediction accuracy for potassium content in Ipomoea batatas leaves,the regression equation was y=58.601 0x-58.446（R²=0.741 4, RMSE=0.83）,using validation data to test the linear model, the model showed good predictive ability for potassium content in Ipomoea batatas leaves under different potassium fertilizer levels （R²=0.732 4, RMSE=0.85）;the linear model constructed by RVI and potassium nutrition index indicated that RVI （R_{700 nm}, R_{1 385 nm}） had a high prediction accuracy for the potassium nutrition index of Ipomoea batatas leaves,the regression equation was y=6.032 9x-0.833 （R²=0.768 8, RMSE=0.15）,using validation data to test the linear model, the model showed good predictive ability for the potassium nutrient index of Ipomoea batatas leaves under different potassium fertilizer levels （R²=0.639 5,RMSE=0.20）;the use of RVI could effectively monitor and diagnose potassium nutrition in Ipomoea batatas.

Forest pest monitoring and prevention based on UAV image detection

QIU Ya-lin, LIU Xiang-long, HE Xiao-jun, ZHAO Qing-long, JIA Cun-fang

2024, 63(8):  262-266.  doi:10.14088/j.cnki.issn0439-8114.2024.08.044

Asbtract ( 98 )   PDF (4157KB) ( 46 )  

References | Related Articles | Metrics

In order to solve the problem of low efficiency and poor effect of existing pest control methods, which required a lot of manpower and material resources, the research built a forest pest detection framework based on deep learning, which transferred the feature information extracted from the shallow network to the deep network, and made lightweight improvements to the model through pruning and batch normalization folding. The results showed that, when each model tended to be stable during training, the average accuracy of the improved YOLOv4 model reached 97.38%, and compared with the original YOLOv4 model, the computing cost and storage requirements were reduced by 17.81 percent points and 23.38%, respectively. The average detection accuracy was 12.75 percent points higher than before.

Carbon flux prediction in farmland ecosystem based on hydrometeorological factors

WU Cheng-qiu, CAO Zhao-dan, ZHAO Xiao-er, WU Hong-yu, DENG Ke

2024, 63(8):  267-280.  doi:10.14088/j.cnki.issn0439-8114.2024.08.045

Asbtract ( 114 )   PDF (34347KB) ( 22 )  

References | Related Articles | Metrics

Using the carbon and water fluxes and hydrometeorological data observed by the flux tower of Yucheng Station of China Flux Observation Network （ChinaFLUX） in the lower reaches of the Yellow River Basin, the main controlling environmental factors affecting the CO₂ exchange capacity of the farmland ecosystem were determined based on the feature importance method. A machine learning model for carbon flux prediction was constructed based on all environmental factors and master environmental factors, and the mean square error （MSE）, mean absolute error （MAE） and coefficient of determination （R²） were used to evaluate the model prediction performance of the test set. The results showed that, the main environmental driving factors affecting carbon flux in Yucheng agro-ecosystems were net radiation, soil temperature, vapor pressure deficit and soil water content. Compared with single models, the ensemble models had better learning and prediction performances in the testing set. Among the single models, MLPRegressor model could better predict NEE with R² of 0.830, MSE of 3.113 and MAE of 1.283. Among the ensemble models, XGBRegressor model had better prediction performance with R² of 0.845, MSE of 2.838 and MAE of 1.149. The machine learning models using the main four environmental driving factors had the same prediction performances as the models using all environmental factors.

Digital Village

The development level differences and influencing factors of 57 Taobao villages in Jiangxi Province

WANG Fang, LI Han, DING Zhi-wei

2024, 63(8):  281-288.  doi:10.14088/j.cnki.issn0439-8114.2024.08.046

Asbtract ( 107 )   PDF (8791KB) ( 11 )  

References | Related Articles | Metrics

The spatio-temporal characteristics of Taobao village development level were evaluated, and its influencing factors were explained using the entropy weight-TOPSIS method and spatial analysis technology. The results showed that, from the perspective of spatial development level, the overall development in the northern and southern regions of Jiangxi Province was better than that in the central region. High-value and low-value areas of development level in the southern region showed in a “staggered” manner, and medium-level areas were widely distributed, presenting an overall pattern of development radiating outwards from Nanchang City and Ganzhou City as the core. From the perspective of time, in the slow star stage （2013 to 2016）, the development of Taobao villages started slowly and gradually expanded from Hongxing Village, forming four Taobao villages; in the steady growth phase （2017 to 2018）, the number of Taobao villages increased steadily with an annual increment of four; in the recent explosive growth phase（2019 to 2021）, it rapidly increased to 57 Taobao villages. Although the spatial distribution of the nuclear density of the number and development level of Taobao village showed an uneven distribution of “scattered in the whole, relatively concentrated in the local area”, a high-density value area in the shape of a triangle had formed in the north of Jiangxi Province. In the initial stage, spatial diffusion of Taobao villages was dominated by “point-like” independent development; in the mid-stage of diffusion, it relied on professional driving, developing in a “point-line” mode with the point as the core and radiating outwards; in the late stage, it mainly developed through common development and government support, weaving and expanding in a “line-net” mode. In terms of influencing factors, the development of Taobao villages was the result of a combination of factors, with economic fundamentals, transport levels and platform support being the main influencing factors, and service levels, internet levels and information communication having a positive impact on the development of Taobao villages.

The theoretical logic, evaluation and strategy of the digital transformation of rural governance： A case study of Sichuan Province

MA Chen, LI Jin, LI Ze-xin

2024, 63(8):  289-295.  doi:10.14088/j.cnki.issn0439-8114.2024.08.047

Asbtract ( 102 )   PDF (1591KB) ( 35 )  

References | Related Articles | Metrics

Based on the perspective of digitalization and centering on the core elements and characteristics of digitalization, the theoretical logic of digitalization to break down the barriers of rural governance and improve the effectiveness of rural governance was expounded. On this basis, a digital evaluation index system of rural governance covering 4 secondary indicators and 14 tertiary indicators was constructed. Taking Sichuan Province as an example, the digitalization level of rural governance in Sichuan Province was evaluated. According to calculations, the comprehensive implementation level of digital rural governance in Sichuan Province was 71.73%, which was in the primary stage of digital development of rural governance. From a dimensional perspective, the highest level of digital implementation of rural party building was achieved, reaching 74.94%. From a regional perspective, the degree of digitalization in rural governance in the Chengdu Plain Economic Zone was 74.81%, leading the digital development of rural governance in Sichuan Province. On this basis, optimization and upgrading strategies were proposed from the aspects of optimizing and upgrading rural information infrastructure, exploring digital application scenarios according to local conditions, emphasizing multiple participation to stimulate the internal driving force of rural governance, and conducting pilot demonstrations of rural digital governance, so as to promote the digital development of rural governance in Sichuan Province.

Digital economy and agricultural green development： Theoretical mechanism and empirical test

MA Li, TANG Mi

2024, 63(8):  296-302.  doi:10.14088/j.cnki.issn0439-8114.2024.08.048

Asbtract ( 122 )   PDF (1518KB) ( 52 )  

References | Related Articles | Metrics

Based on the panel data of 30 provinces （municipalities and autonomous regions） in China from 2013 to 2022, the level of agricultural green development was comprehensively measured, and an empirical model was constructed to study the heterogeneity, intermediary effect and threshold effect of digital economy on agricultural green development. The results showed that digital economy had a promoting effect on the green development of agriculture, and there was regional heterogeneity; the digital economy could indirectly improve the level of agricultural green development through green technology innovation; the impact of digital economy on green development of agriculture had a double threshold effect in the eastern region. Therefore, it was necessary to strengthen the construction of agricultural digital infrastructure, narrow the differences in the development of digital economy in various regions, promote the green development of agriculture with the support of technology, and improve the awareness of the governance of the agricultural ecological environment.