Analysis system and analysis method for fruit tree pruning scheme

Abstract

The invention discloses an analysis system and method for a fruit tree pruning scheme, and the system comprises an image collection module, a processing unit electrically connected with the image collection module, a display module electrically connected with the processing unit, and a sensor module electrically connected with the processing unit, and the processing unit is electrically connected with a communication module. The communication module is electrically connected with the control module, and a positioning assembly is arranged in the control module. According to the analysis system and the analysis method for the fruit tree pruning scheme, the image of the fruit tree can be collected and analyzed through a corresponding algorithm, so that an accurate pruning method is given, the analysis of the fruit tree pruning scheme is realized, and the cost of fruit growers is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of image analysis, and in particular relates to an analysis system and an analysis method for a fruit tree pruning scheme. Background Art

[0002] Fruit tree pruning is a very important management measure in fruit tree planting. The pruning plan will directly affect the production and yield of fruit trees. The analysis of fruit tree pruning plans is an important task to promote the development of fruit tree planting, which requires accurate and effective analysis of fruit tree pruning plans. The pruning methods for fruit trees in different periods are usually different. The pruning period and pruning focus are different in different seasons. For example, in winter pruning (dormant pruning), the pruning time is after the fruit trees fall leaves and before buds sprout. The pruning focus is on adjusting the tree shape, thinning large branches, renewing fruiting branches, controlling tree height, etc. In summer (growing season pruning), the pruning time is from May to July. The pruning focus is on removing buds, pinching, thinning out overcrowded branches, etc. In addition, the pruning methods are different for trees of different ages.

[0003] In the prior art, pruning of fruit trees is often performed by pruning technicians based on manual experience, which makes it difficult to achieve precise pruning. In addition, it is difficult for inexperienced fruit farmers to complete the pruning work. This also leads to high training costs for excellent fruit farmers and difficulty in changing jobs, resulting in the loss of young talents, over-reliance on experience, and personalized pruning work. Therefore, there is a need for a method that can reduce the cost of training fruit farmers and achieve precise pruning. The present invention solves this technical problem. Summary of the Invention

[0004] The present invention provides an analysis system and method for fruit tree pruning plans, which can collect images of fruit trees and analyze the images through corresponding algorithms, thereby providing accurate pruning methods, realizing the analysis of fruit tree pruning plans, and reducing the cost of fruit cultivation for farmers.

[0005] A fruit tree pruning plan analysis system includes an image acquisition module, a processing unit electrically connected to the image acquisition module, a display module electrically connected to the processing unit, and a sensor module electrically connected to the processing unit. The processing unit is electrically connected to a communication module, the communication module is electrically connected to a control module, and a positioning component is provided in the control module.

[0006] The image acquisition module is used to acquire images of fruit trees, the display module is used to display pruning plans, the positioning component is used to recognize gestures, the control module is used to control the acquisition of images according to gestures, and the sensor module is used to acquire information about the growth environment of fruit trees.

[0007] Furthermore, the control module includes a smart watch, the smart watch is connected to the positioning component through a communication protocol, and the processor of the smart watch is connected to the processing unit through signals.

[0008] A method for analyzing a fruit tree pruning plan, based on the above-mentioned fruit tree pruning plan analysis system, comprises the following steps:

[0009] S1. Aim the image acquisition module at the fruit tree to be photographed, and control the image acquisition module to collect images of the fruit tree through set gestures. The collected image data is uploaded to the cloud through the processing unit;

[0010] S2, preprocessing the obtained image;

[0011] S3. Analyze the problems of fruit trees based on the color of their leaves and fruits, classify the diseases of the fruit trees through the pre-trained model, and provide corresponding solutions;

[0012] S4. Construct a three-dimensional model of the fruit tree, analyze the condition of the fruit tree through the model, and simulate the results of different pruning methods;

[0013] S5. Select the best pruning solution based on the simulation results;

[0014] S6. Optimize the image recognition method to improve the accuracy of image recognition;

[0015] S7. Using image recognition technology to obtain spatial distribution information of fruits for thinning out problematic fruits;

[0016] S8. Send the pruning plan to the processing unit and display it through the display module.

[0017] Furthermore, step S2 includes the following steps:

[0018] S21. Use the cv2.imread() function to load the captured image.

[0019] S22, extracting the key area in the image, and manually defining the coordinates of the upper left corner and lower right corner of the area to be extracted;

[0020] S23. Use the extracted_region = image[y1:y2, x1:x2] code to extract the image of the key area;

[0021] S24. Calculate a color histogram of the image to describe the color distribution of the image;

[0022] S25. Calculate multi-dimensional features of objects in the image to describe the shape of the objects.

[0023] Furthermore, step S3 includes the following steps:

[0024] S31. Collect photos of fruit trees at different stages, including leaves, branches, and fruits, and use annotation tools to mark any problems with the fruit trees in the images. The annotated dataset is then divided for model training.

[0025] S32. The pre-processed image is input to the model. The model analyzes the leaves, branches, and fruits in the image and compares the corresponding colors with the data in the model to classify the different problems existing in the fruit trees.

[0026] S33. Connect the model to the database of related industries, use natural language processing technology to identify the problems existing in fruit trees, and use relevant knowledge and information to generate solutions.

[0027] Furthermore, step S4 includes the following steps:

[0028] S41. Deploy several sensors in the orchard to collect various data on the fruit tree growth environment in real time, as well as the fruit tree growth data over the years. Combine the two types of data to analyze the growth trends of the fruit trees and establish a fruit tree growth model.

[0029] S42. Use a depth camera to collect data, then process the collected point cloud data and extract feature points. By matching features between adjacent frames, determine the relative position changes of sensors in the orchard at different times. Based on the position of the sensors and the collected point cloud data, gradually build a three-dimensional map of the fruit trees and determine the shape of the fruit trees;

[0030] S43. Using reinforcement learning algorithms, simulate the long-term effects of different pruning strategies on the yield and quality of fruit trees.

[0031] Furthermore, the step S43 includes the following steps:

[0032] S431. Select the DQN network, input the initialization state of the fruit tree, and train the model;

[0033] S432, using the trained model to simulate the growth of fruit trees over a long period of time, and conducting in-depth analysis of the simulation results;

[0034] S433. Compare the effects of different pruning strategies on yield and quality in the long term, and predict the subsequent growth of fruit trees.

[0035] Furthermore, step S6 includes the following steps:

[0036] S71. Using image recognition technology to obtain spatial distribution information of the fruit and calculate the density of the fruit;

[0037] S72. Regularly measure fruit growth data including fruit size and weight;

[0038] S73, determining a fruit density threshold based on the variety, age, and vigor of the fruit trees and the management level of the orchard, comparing the fruit density data with the threshold, identifying areas where the fruit density exceeds the threshold, and marking these areas as overcrowded areas;

[0039] S74. In overcrowded areas, further determine the location of fruits that need to be thinned based on fruit growth data, and give priority to thinning out fruits that are small, poorly developed, or have diseases and insect pests.

[0040] Furthermore, step S6 includes the following steps:

[0041] S61. In the Yolo algorithm network Googlenet, the convolution kernel is positioned as 1*1 in size and placed directly in the center of the image to reduce image background prediction, thereby reducing the confidence=0 situation and reducing the error of the loss function.

[0042] S62, in the inception structure, auxiliary prediction is performed on the four branches around the center position to improve accuracy and efficiency;

[0043] S63. Use the convolution kernel dimensionality reduction method in the inception structure to reduce calculations.

[0044] Furthermore, the loss function is optimized in the following way:

[0045] ;

[0046] Where: N is the total number of all predicted boxes; Represents the indicator function. If the i-th prediction box is responsible for detecting the target, then =1; if the target is not included, then =0; is the prediction confidence of the i-th prediction box, which ranges from [0,1] and represents the probability that the prediction box contains the target.

[0047] The technical effects of the present invention are as follows:

[0048] (1) The present invention can collect images of fruit trees, identify diseases therein, analyze relevant data such as branches and fruits, and provide corresponding pruning strategies, thereby realizing the function of automatically analyzing pruning plans. Compared with traditional methods, it reduces the cost of cultivating fruit farmers and makes pruning plans more scientific, accurate, and practical.

[0049] (2) The pre-trained model in the present invention can identify the colors of objects such as leaves and fruits through image recognition, thereby analyzing the problems existing in fruit trees and classifying fruit trees according to different problems, so as to facilitate the subsequent provision of corresponding solutions based on industry databases and natural language processing technology, so that fruit farmers can more quickly understand the problems existing in fruit trees and help solve the problems in a timely manner, thereby realizing the function of automatic disease analysis;

[0050] (3) The present invention can construct a three-dimensional model of the fruit tree, which helps to understand the shape of the fruit tree and facilitate the subsequent provision of pruning strategies. It can also predict the subsequent growth of the fruit tree through the fruit tree growth model, which is conducive to judging the impact of different pruning strategies and helping to obtain the best pruning strategy, thereby realizing automatic analysis of pruning plans and making fruit tree pruning more intelligent and accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a structural block diagram of the present invention.

[0052] Figure 2 It is the workflow diagram of the present invention.

[0053] Figure 3 Schematic diagram of the convolutional neural network process in the present invention. DETAILED DESCRIPTION

[0054] The technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments and drawings.

[0055] See also Figure 1 A fruit tree pruning plan analysis system includes an image acquisition module, a processing unit electrically connected to the image acquisition module, a display module electrically connected to the processing unit, and a sensor module electrically connected to the processing unit, wherein the processing unit is electrically connected to a communication module, the communication module is electrically connected to a control module, and a positioning component is provided in the control module;

[0056] The image acquisition module is used to acquire images of fruit trees, the display module is used to display pruning plans, the positioning component is used to recognize gestures, the control module is used to control the acquisition of images according to gestures, and the sensor module is used to acquire information about the growth environment of fruit trees.

[0057] In this embodiment, the image acquisition module and the display module are both arranged on the AR glasses, wherein the image acquisition module is a high-resolution camera (main camera) that supports dynamic focus and ambient light adaptation to ensure clear capture of problem scenes. The display module can adopt a split design (such as MIPI interface connection), or refer to NVIDIA's backlight-free digital holographic technology to achieve lightweight and natural superposition of virtual information, and can also add a SLAM camera for subsequent three-dimensional environment modeling. The processing unit adopts a high-performance embedded processor (such as FPGA or ARM architecture) to support local preliminary image processing and real-time data transmission. The processing unit can be connected to the communication module, specifically to the cloud server via a Wi-Fi or 5G module to ensure high-speed data transmission.

[0058] Preferably, an infrared camera and a visible light camera can be embedded in the edge of the optical waveguide lens, and dual-band data acquisition can be achieved through optical waveguide spectrometry technology. The infrared camera is used for environmental depth perception (such as nighttime or low-light scenes), and the visible light camera is used for real-time image rendering.

[0059] Optionally, this solution can adopt a split design, for example, the processing unit, communication module and other components are connected in the host, and the image acquisition module and display module are both set on the AR glasses, thereby reducing the burden on fruit farmers wearing AR glasses. In the split design, the connection between modules adopts the MIPI interface solution, which can reduce the complexity of cables.

[0060] Furthermore, the control module includes a smart watch, the smart watch is connected to the positioning component through a communication protocol, and the processor of the smart watch is connected to the processing unit through signals.

[0061] Specifically, this embodiment installs a BMX160 sensor on a wristwatch. The BMX160 sensor has dedicated pins for connecting to the wristwatch's I2C bus. During wristwatch design, the BMX160 sensor can be directly integrated into the internal space and connected to other circuit components of the wristwatch via pins on the circuit board to achieve signal transmission and power supply. The wristwatch's operating system or related applications interact with the BMX160 through the I2C communication protocol. When sensor data needs to be obtained, the wristwatch's processor sends a read command to the BMX160, which is transmitted to the sensor via the I2C bus. After receiving the command, the BMX160 returns data collected by the watch's accelerometer, gyroscope, geomagnetic sensor, etc. to the wristwatch processor via the SDA data line. The wristwatch processor then transmits the millimeter-wave gesture signal to the processing unit, thereby realizing the gesture recognition function.

[0062] See also Figure 2 A method for analyzing a fruit tree pruning plan, based on the above-mentioned fruit tree pruning plan analysis system, comprises the following steps:

[0063] S1. Aim the image acquisition module at the fruit tree to be photographed and control it through set gestures to capture images of the fruit tree. The collected image data is uploaded to the cloud through the processing unit. Specifically, an API interface is built to upload the image to the cloud big data platform, and an encryption protocol (such as HTTPS) is used to ensure data security.

[0064] S2, preprocessing the obtained image;

[0065] S3. Analyze the problems of fruit trees based on the color of their leaves and fruits, classify the diseases of the fruit trees through the pre-trained model, and provide corresponding solutions;

[0066] S4. Construct a three-dimensional model of the fruit tree, analyze the condition of the fruit tree through the model, and simulate the results of different pruning methods;

[0067] S5. Select the best pruning solution based on the simulation results;

[0068] S6. Optimize the image recognition method to improve the accuracy of image recognition;

[0069] S7. Using image recognition technology to obtain spatial distribution information of fruits for thinning out problematic fruits;

[0070] S8. Send the pruning plan to the processing unit and display it through the display module.

[0071] Furthermore, step S2 includes the following steps:

[0072] S21. Use the cv2.imread() function to load the captured image.

[0073] S22, extracting the key area in the image, and manually defining the coordinates of the upper left corner and lower right corner of the area to be extracted;

[0074] S23. Use the extracted_region = image[y1:y2, x1:x2] code to extract the image of the key area;

[0075] S24. Calculate a color histogram of the image to describe the color distribution of the image;

[0076] S25. Calculate multi-dimensional features of objects in the image to describe the shape of the objects.

[0077] Preferably, automatic capture can also be used. Specifically, a real-time image capture program can be developed, using the OpenCV VideoCapture class to capture image frames from the camera. Image clarity can be assessed by calculating the Laplace variance of the image. Different focus settings can be tried during capture, and the image with the highest clarity can be selected. The Laplace variance formula is: σ² = N¹∑i = 1N(Li − μ)². The above is prior art and will not be elaborated on here.

[0078] Regarding step S25, the calculation of the multi-dimensional features in this solution includes the following:

[0079] 1. Calculation of basic geometric features:

[0080] The pixel set of the object in the image is represented as S. For each pixel, its coordinate is represented as (x, y), the pixel value (foreground) is 1, and the background is 0.

[0081] For each object in the image, calculate the total number of pixels contained in each object, that is, the area of ​​the object, represented by A:

[0082] ;

[0083] Calculate each pixel in the object using the above method to obtain the area of ​​the object. Then calculate the perimeter of the object, which can be calculated using the following method:

[0084] Method 1: 4-neighborhood or 8-neighborhood counting method:

[0085] 4-neighborhood perimeter: Count the 4 adjacent edges between the edge pixels of the object and the background pixels, i.e. the top, bottom, left, and right sides;

[0086] 8-neighborhood perimeter: Counts the number of 8-adjacent edges between the pixels in the edge area of ​​the object and the background pixels, including not only the four sides of the top, bottom, left, and right, but also the diagonal.

[0087] Method 2: Contour tracing method:

[0088] In this solution, the Euclidean distance between two adjacent pixels can be calculated by taking the difference in pixel coordinates of the edge area of ​​the object, i.e., the change. The Euclidean distances between all adjacent pixels are then accumulated to obtain the perimeter of the object. This is calculated using the following formula:

[0089] ;

[0090] in, Represents the coordinates of the i-th pixel on the edge area (contour) of the object, n represents the total number of pixels in the edge area, and the total perimeter of the object contour is obtained by connecting two adjacent pixels in the edge area end to end.

[0091] After obtaining the perimeter, in order to determine the position of the object in the image, it is necessary to calculate the geometric center of the object so that it can be used as the reference coordinate of the object:

[0092] ; ;

[0093] in, and The horizontal and vertical coordinates belong to the geometric center respectively.

[0094] 2. Shape description features:

[0095] Since the density of the fruit and other data need to be calculated later, the shape of the objects in the image needs to be calculated.

[0096] 1. Circularity calculation: Since fruits are relatively close to circles, we first calculate the circularity of the fruit. This measures the degree of similarity between an object and a circle. The value range is [0, 1]. The larger the value, the closer the object is to a circle. The calculation formula is as follows:

[0097] ;

[0098] Wherein, A represents the area of ​​the object, and P represents the perimeter of the object. The calculation formulas for A and P have been given in detail in the above content and will not be described in detail here.

[0099] 2. To distinguish branches from fruits, we need to calculate the aspect ratio of the objects in the image, that is, the ratio of the length to the width of the circumscribed rectangle of the object:

[0100] ;

[0101] Where w represents the length of the circumscribed rectangle, h represents the width of the circumscribed rectangle, 、 Respectively represent the coordinate sets of the object contour points.

[0102] 3. Calculate the ratio of the object area to the area of ​​the circumscribed rectangle to reflect the degree of fit of the object to the rectangle:

[0103] ;

[0104] Where A represents the actual area of ​​the object, and Respectively represent the maximum and minimum values ​​of the horizontal coordinates of the pixels in all edge areas of the object, and They respectively represent the maximum and minimum values ​​of the vertical coordinates of the pixels in all edge areas of the object.

[0105] 4. Calculate the ellipticity of the object, specifically the ratio of the major axis to the minor axis of the ellipse:

[0106] ;

[0107] Among them, a represents the length of the major axis of the ellipse, b represents the length of the minor axis of the ellipse, and They represent the eigenvalues ​​of the object covariance matrix respectively.

[0108] 5. Perform Fourier transform on the object contour to extract its frequency domain features for shape classification:

[0109] , k=0, 1, ..., N-1;

[0110] Where N represents the total number of pixels on the contour, Represents the coordinates of the nth pixel point on the contour. Low-frequency coefficients correspond to smaller k values. Retaining low-frequency coefficients can reconstruct the shape contour.

[0111] Furthermore, step S3 includes the following steps:

[0112] S31. Collect photos of fruit trees at various stages, including leaves, branches, and fruits, and use annotation tools to annotate the problems with the fruit trees in the images. The annotated dataset is then divided into two parts for model training. The pre-trained model is fine-tuned using the prepared dataset to enable it to identify leaf colors and corresponding tree problems.

[0113] S32. The pre-processed image is input to the model. The model analyzes the leaves, branches, and fruits in the image and compares the corresponding colors with the data in the model to classify the different problems existing in the fruit trees.

[0114] Use the fine-tuned model to analyze new leaf images, classify tree issues by identifying leaf color and comparing it with data in the platform model;

[0115] S33. Connect the model to the database of related industries, use natural language processing technology to identify the problems existing in fruit trees, and use relevant knowledge and information to generate solutions.

[0116] For example, infrared detection can reveal abnormal leaf temperatures (such as localized temperature increases caused by aphid infestations), while visible light analysis can reveal leaf spots or fruit mold, thereby identifying problems with fruit trees. Ideally, simple tasks (such as preliminary classification) can be processed locally on the host, while complex analysis can be offloaded to the cloud (e.g., edge storage and central platform collaboration).

[0117] Furthermore, step S4 includes the following steps:

[0118] S41. Deploy several sensors in the orchard to collect real-time data on the fruit tree growth environment, including light intensity, temperature, humidity, and soil nutrient content. Sensors are also used to collect historical growth data on the fruit trees, including fruit yield, fruit size, pest and disease occurrence, branch length and diameter, and pruning records. These two types of data are combined to analyze the growth trends of the fruit trees and establish a fruit tree growth model.

[0119] In view of the above content, this embodiment is implemented through the following process:

[0120] 1. Data collection and storage:

[0121] First, data is collected, including environmental data of the orchard and growth data of the fruit trees. The environmental data is collected in real time through sensors, including temperature and humidity, light, soil moisture / pH / nutrients, and pest and disease data; the growth data includes leaf indicators, fruit size, phenological period (which can be recorded manually or through image recognition), and historical yield / farming records (historical data).

[0122] Sensors need to be deployed near fruit trees and transmit data via LoRa / 4G. Real-time data is stored in a time-series database (such as InfluxDB), while historical data is stored in a relational database.

[0123] 2. Analyze the growth trend of fruit trees:

[0124] First, it is necessary to clean outliers in the data. For missing or abnormal data, interpolation can be used to fill in the missing data and generate derived indicators, such as effective accumulated temperature and water stress index.

[0125] Then, the relationship between environmental data and growth data is identified. For example, temperature affects fruit expansion, and based on historical records, the seasonal changes in the growth of fruit trees and fruits are analyzed to predict the growth of fruit trees and fruits in the short term. In this embodiment, the ARIMA model is used.

[0126] 3. Establish a fruit tree growth model

[0127] First, it is necessary to systematically sort out the physiological processes of fruit trees, such as photosynthesis, respiration, nutrient distribution, and other physiological processes, establish differential equations to describe the flow of matter and energy, and use this as a physical model;

[0128] Afterwards, machine learning algorithms, such as neural networks and random forests, are used to process the heterogeneous features of the collected environmental data and historical data through random forests. The LSTM neural network is used to predict the time series growth of fruit trees, and cross-validation is used to prevent model overfitting.

[0129] In this embodiment, the physical model is used as the main framework, and the neural network model is used to correct the residual term, thereby improving the accuracy of the prediction.

[0130] Then, machine learning modeling is performed, specifically using the following methods:

[0131] First, we need to define the goal, which can be to predict fruit tree yield, predict when the fruit tree will reach maturity, or predict the risk of pests and diseases. In this example, the goal is to predict fruit tree yield. Then, we perform feature screening to select features related to fruit tree yield. Specifically, we need to screen time features and agronomic features:

[0132] Time features: Use sliding windows to calculate the average temperature over a period of time;

[0133] Agronomic characteristics: Confirm the critical value of accumulated temperature during flower bud differentiation, fertilizer application amount, etc., and record them in a table.

[0134] The model is then trained. XGBoost is used to process tabular data, such as fertilizer application rates, and LSTM+Attention is used to process real-time sensor data, i.e., time series data. Parameters are then optimized through methods such as cross-validation.

[0135] Finally, data from different years were reserved as validation sets, and indicators such as RMSE / R² were used to test the accuracy of the model. The model was then revised based on the actual measurement data of the orchard.

[0136] 4. Application and Optimization

[0137] The data collected by the sensors is transmitted to the model in real time, and the data is updated dynamically. It can also be optimized based on the actual situation. When the probability of pests and diseases is predicted to be high, an early warning can be triggered, so that fruit farmers can be informed of the situation immediately.

[0138] And when the output target of the model is fruit tree yield, the output results of the model may include fertilization and irrigation suggestions, the source of which may be historical data related to fruit tree yield.

[0139] When the model is initially established, the preliminary plan can be tested on a small scale in an actual orchard to observe the growth of the fruit trees after pruning. After continuous monitoring and updating, the best pruning plan can be obtained;

[0140] S42. Use a depth camera to collect data, then process the collected point cloud data and extract feature points. By matching features between adjacent frames, determine the relative position changes of sensors in the orchard at different times. Based on the position of the sensors and the collected point cloud data, gradually build a three-dimensional map of the fruit trees and determine the shape of the fruit trees;

[0141] S43. Using reinforcement learning algorithms, simulate the long-term effects of different pruning strategies on the yield and quality of fruit trees.

[0142] Furthermore, step S43 includes the following steps:

[0143] S431. Select the DQN network, input the initialization state of the fruit tree, and train the model;

[0144] The initialization state includes the current state of the fruit tree, such as branch density, fruit density, fruit size, pest and disease conditions, etc.

[0145] S432, using the trained model to simulate the growth of fruit trees over a long period of time, and conducting in-depth analysis of the simulation results;

[0146] S433. Compare the effects of different pruning strategies on yield and quality in the long term, and predict the subsequent growth of fruit trees.

[0147] Furthermore, step S6 includes the following steps:

[0148] S71. Using image recognition technology to obtain spatial distribution information of the fruit and calculate the density of the fruit;

[0149] S72. Regularly measure fruit growth data including fruit size and weight;

[0150] S73, determining a fruit density threshold based on the variety, age, and vigor of the fruit trees and the management level of the orchard, comparing the fruit density data with the threshold, identifying areas where the fruit density exceeds the threshold, and marking these areas as overcrowded areas;

[0151] S74. In overcrowded areas, further determine the location of fruits that need to be thinned based on fruit growth data, and give priority to thinning out fruits that are small, poorly developed, or have diseases and insect pests.

[0152] For step S73, this solution is implemented in the following manner:

[0153] 1. Establish the density threshold calculation formula:

[0154] ;

[0155] in, Indicates the variety coefficient, and different coefficients are assigned to different varieties of fruits, for example, the coefficient of apple = 1.0, the coefficient of citrus = 0.8, and the coefficient of pear = 1.2;

[0156] Represents the tree age correction function, the specific calculation method will be described in detail below;

[0157] Indicates the tree vigor correction coefficient, that is, the correction coefficient of the fruit tree growth. For example, a fruit tree with poor growth is a weak tree with a coefficient of 0.8, a fruit tree with normal growth is a mediocre tree with a coefficient of 1.0, and a fruit tree with vigorous growth is a vigorous tree with a coefficient of 1.2.

[0158] Indicates the management level correction coefficient. For fruit trees with poor management level, it is recorded as extensive, with a coefficient of 0.9; for fruit trees with normal management level, it is recorded as general, with a coefficient of 1.0; for fruit trees with good management level, it is recorded as fine, with a coefficient of 1.1;

[0159] It represents the standard density base, that is, the ratio of the number of fruits to the projected area of ​​the canopy per square meter.

[0160] 2. For the tree age correction function, this embodiment calculates it in the following way:

[0161] ;

[0162] 3. For overcrowded areas, this embodiment uses the following method to determine:

[0163] ;

[0164] in, represents the actual fruit density in region i, Indicates the preset density threshold.

[0165] 4. Establish quantitative indicators of overcrowding:

[0166] ;

[0167] in, It can reflect the differences in natural fruit density caused by variety characteristics; Used to reflect the impact of tree age on fruit-bearing ability, so that the predicted results are consistent with the growth laws of fruit trees; Used to correct the difference in fruits caused by the strength of trees. More fruits can be left on the paper strips of vigorous trees, while fewer fruits should be left on weak trees. Used to consider the impact of management level on fruit quality. For example, fruit trees that have been carefully managed grow better, their branches can bear more fruits, and their fruit density can be higher. It represents the standard density base, that is, the ideal fruiting density of the reference variety, which can be determined from literature or experience; The actual regional fruit density can be obtained through remote sensing, drones or manual measurement.

[0168] The calculation can be done in the following way:

[0169] Variety: Apple, ;

[0170] Age: 12 years ;

[0171] Tree vigor: Moderate, ;

[0172] Management level: fine, ;

[0173] Standard density base .

[0174] Based on the above, calculate the density threshold:

[0175] ;

[0176] When the actual density of area A is , calculate the over-density percentage of the area:

[0177] ;

[0178] Conclusion, exceeding the threshold , the density is too dense.

[0179] Therefore, based on the above formula, automatic overcrowded area identification and fruit thinning decision support can be achieved.

[0180] See also Figure 3 , step S6 comprises the following steps:

[0181] S61. In the Yolo algorithm network Googlenet, the convolution kernel is positioned as 1*1 in size and placed directly in the center of the image to reduce image background prediction, thereby reducing the confidence=0 situation and reducing the error of the loss function.

[0182] S62, in the inception structure, auxiliary prediction is performed on the four branches around the center position to improve accuracy and efficiency;

[0183] S63. Use the convolution kernel dimensionality reduction method in the inception structure to reduce calculations.

[0184] In the figure, "Previous Layer" refers to the layer before the current layer, which directly accepts input data and outputs feature maps. "Convolutions" refers to convolution operations, used to extract multi-scale features. "Max Pooling" drives the maximum value in a local area, achieving downsampling and reducing spatial dimensionality. "Fitter concatenation" refers to optimized concatenation, fusing the results of multiple branches through fitter concatenation. "1×1", "3×3", and "5×5" represent the spatial sizes of the convolution kernels, respectively.

[0185] Furthermore, the loss function is optimized in the following way:

[0186] ;

[0187] Where: N is the total number of all predicted boxes; Represents the indicator function. If the i-th prediction box is responsible for detecting the target, then =1; if the target is not included, then =0; is the prediction confidence of the i-th prediction box, which ranges from [0,1] and represents the probability that the prediction box contains the target.

[0188] Preferably, targeted prevention and control plans (such as pesticide ratios and biological control recommendations) can be pushed through the AR interface, the sugar content and hardness of the fruit can be detected using near-infrared technology, and the maturity can be marked through the AR interface, such as color markings: green for unripe, yellow for ready to pick, and red for overripe, to guide fruit farmers to harvest the fruit accurately.

[0189] In the AR interface, redundant branches that need to be pruned can be highlighted to assist novice fruit farmers in learning pruning techniques. It is also necessary to distinguish between seasons, specify different pruning strategies for different tree ages, periods, and seasons, and be precise to a certain branch or location. Specifically, pixel coordinates or a coordinate system based on a geographic information system (GIS) can be used to correspond the location information of the branch with the actual geographical location in the orchard, so as to achieve the effect of being precise to a certain location.

[0190] Intelligent pruning strategies can generate a dynamic pruning rule base: based on the fruit tree pruning knowledge base, combined with real-time environmental data (light, temperature and humidity) and historical growth records (such as fruit yield, disease and pest history), a personalized pruning plan suitable for the fruit tree is generated according to actual conditions.

[0191] The above embodiments are only preferred embodiments of the present invention. Those skilled in the art can derive other embodiments from the above embodiments without creative work. Therefore, this application protects not only the above embodiments, but also the scope consistent with the principles and features of this application.

Claims

1. A fruit tree pruning plan analysis system, characterized in that: The system comprises an image acquisition module, a processing unit electrically connected to the image acquisition module, a display module electrically connected to the processing unit, and a sensor module electrically connected to the processing unit. The processing unit is electrically connected to a communication module, the communication module is electrically connected to a control module, and a positioning component is provided in the control module. The image acquisition module is used to acquire images of fruit trees, the display module is used to display pruning plans, the positioning component is used to recognize gestures, the control module is used to control the acquisition of images according to gestures, and the sensor module is used to acquire information about the growth environment of fruit trees.

2. The fruit tree pruning plan analysis system according to claim 1, characterized in that: The control module includes a smart watch, the smart watch is connected to the positioning component through a communication protocol, and the processor of the smart watch is connected to the processing unit through signals.

3. A method for analyzing a fruit tree pruning plan, based on the fruit tree pruning plan analysis system according to claim 1, characterized in that: The following steps are involved: S1. Aim the image acquisition module at the fruit tree to be photographed, and control the image acquisition module to collect fruit tree images through gestures. The collected image data is uploaded to the cloud through the processing unit; S2, preprocessing the obtained image; S3. Analyze the problems of fruit trees based on the color of their leaves and fruits, classify the diseases of the fruit trees through the pre-trained model, and provide corresponding solutions; S4. Construct a three-dimensional model of the fruit tree to analyze the condition of the fruit tree and simulate the results of different pruning methods; S5. Select the best pruning solution based on the simulation results; S6. Optimize the image recognition method to improve the accuracy of image recognition; S7. Using image recognition technology to obtain spatial distribution information of fruits for thinning out problematic fruits; S8. Send the pruning plan to the processing unit and display it through the display module.

4. The method for analyzing a fruit tree pruning plan according to claim 3, wherein: The step S2 comprises the following steps: S21, loading the captured image; S22, extracting the key area in the image, and manually defining the coordinates of the upper left corner and lower right corner of the area to be extracted; S23, extracting the image of the key area; S24. Calculate a color histogram of the image to describe the color distribution of the image; S25. Calculate multi-dimensional features of objects in the image to describe the shape of the objects.

5. The method for analyzing a fruit tree pruning plan according to claim 3, characterized in that: The step S3 comprises the following steps: S31. Collect photos of fruit trees at different stages, including leaves, branches, and fruits, and use annotation tools to mark any problems with the fruit trees in the images. The annotated dataset is then divided for model training. S32. The pre-processed image is input to the model. The model analyzes the leaves, branches, and fruits in the image and compares the corresponding colors with the data in the model to classify the different problems existing in the fruit trees. S33. Connect the model to the database of related industries, use natural language processing technology to identify the problems existing in fruit trees, and use relevant knowledge and information to generate solutions.

6. The method for analyzing a fruit tree pruning plan according to claim 3, characterized in that: The step S4 comprises the following steps: S41. Deploy several sensors in the orchard to collect various data on the fruit tree growth environment in real time, as well as the fruit tree growth data over the years. Combine the two types of data to analyze the growth trends of the fruit trees and establish a fruit tree growth model. S42. Use a depth camera to collect data, then process the collected point cloud data and extract feature points. By matching features between adjacent frames, determine the relative position changes of sensors in the orchard at different times. Based on the position of the sensors and the collected point cloud data, gradually build a three-dimensional map of the fruit trees and determine the shape of the fruit trees; S43. Using reinforcement learning algorithms, simulate the long-term effects of different pruning strategies on the yield and quality of fruit trees.

7. The method for analyzing a fruit tree pruning plan according to claim 6, characterized in that: The step S43 includes the following steps: S431. Select the DQN network, input the initialization state of the fruit tree, and train the model; S432, using the trained model to simulate the growth of fruit trees over a long period of time, and conducting in-depth analysis of the simulation results; S433. Compare the effects of different pruning strategies on yield and quality in the long term, and predict the subsequent growth of fruit trees.

8. The method for analyzing a fruit tree pruning plan according to claim 3, wherein: The step S6 comprises the following steps: S71. Using image recognition technology to obtain spatial distribution information of the fruit and calculate the density of the fruit; S72. Regularly measure fruit growth data including fruit size and weight; S73, determining a fruit density threshold based on the variety, age, and vigor of the fruit trees and the management level of the orchard, comparing the fruit density data with the threshold, identifying areas where the fruit density exceeds the threshold, and marking these areas as overcrowded areas; S74. In overcrowded areas, further determine the location of fruits that need to be thinned based on fruit growth data, and give priority to thinning out fruits that are small, poorly developed, or have diseases and insect pests.

9. The method for analyzing a fruit tree pruning plan according to claim 3, wherein: The step S6 comprises the following steps: S61. In the Yolo algorithm network Googlenet, the convolution kernel is positioned at a size of 1*1 and placed at the center of the image to reduce image background prediction, thereby reducing the confidence=0 situation and reducing the error of the loss function. S62, in the inception structure, auxiliary prediction is performed on the four branches around the center position to improve accuracy and efficiency; S63. Use the convolution kernel dimensionality reduction method in the inception structure to reduce calculations.

10. The method for analyzing a fruit tree pruning plan according to claim 9, characterized in that: The loss function is optimized in the following way: ； Where: N is the total number of all predicted boxes; Represents the indicator function. If the i-th prediction box is responsible for detecting the target, then =1; if the target is not included, then =0; is the prediction confidence of the i-th prediction box, which ranges from [0,1] and indicates the probability that the prediction box contains the target.

Images