A Deep Learning-Based Comprehensive Growth Environment Detection Method

Abstract

This invention provides a comprehensive monitoring method for crop growth environment based on deep learning, belonging to the field of environmental monitoring technology. The method includes: collecting crop growth disaster information in a target area using a pre-configured device, performing data preprocessing and image annotation to obtain farmland images, environmental and soil data, and remote sensing data; performing image analysis on the farmland images, data analysis on the environmental and soil data, and data parsing on the remote sensing data; and obtaining pest and disease detection results, soil drought classification results, air quality analysis results, and spatiotemporal analysis results based on the image analysis results, data analysis results, and data parsing results, and then summarizing and outputting the data. By integrating advanced sensing technology, autonomous navigation technology, and artificial intelligence algorithms, comprehensive, efficient, and accurate monitoring of the farmland environment is achieved, providing strong technical support for the sustainable development of modern agriculture.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring technology, and in particular to a comprehensive monitoring method for growth environments based on deep learning. Background Technology

[0002] With the rapid development of modern agricultural technology and the complex situation of global climate change, agricultural production is entering a new era full of challenges. Pests and diseases, as two major intractable problems, are increasingly becoming key bottlenecks restricting the improvement of crop yields and the assurance of quality. The rampant spread of pests and diseases not only directly reduces crop yields, but also causes long-term damage to the agricultural ecosystem by spreading pathogens and viruses; while waterlogging disasters, due to their suddenness and unpredictability, often cause devastating damage to farmland in a short period of time, seriously affecting the stability and sustainability of agricultural production.

[0003] Faced with this severe situation, traditional pest and disease monitoring methods are proving inadequate. While traditional manual monitoring methods can identify pests and diseases and assess waterlogging disasters to some extent, they are limited by human perception and reaction speed, often failing to detect and take timely measures in the early stages of pests and diseases or before waterlogging disasters occur. With the expansion of agricultural planting scale and the increase in intensification, the demand for monitoring the farmland environment is also growing. Traditional monitoring methods can no longer meet the requirements of modern agriculture for efficient, accurate, and comprehensive monitoring. While existing automated monitoring methods and equipment have improved monitoring efficiency, their accuracy still needs improvement, especially for pests and diseases that are highly concealed and have inconspicuous symptoms, as well as the complex and ever-changing environment of waterlogging disasters, where they struggle to effectively distinguish between them.

[0004] Therefore, this invention proposes a comprehensive detection method for growth environment based on deep learning. Summary of the Invention

[0005] This invention provides a comprehensive monitoring method for farmland growth environments based on deep learning. By integrating advanced sensing technology, autonomous navigation technology, and artificial intelligence algorithms, it achieves comprehensive, efficient, and accurate monitoring of the farmland environment, providing strong technical support for the sustainable development of modern agriculture. This invention not only fills a market gap but also promotes the development of intelligent and automated agriculture. It enables the automatic detection of farmland pests, diseases, and waterlogging-related growth environments, thereby helping farmland managers make better decisions.

[0006] This invention proposes a comprehensive detection method for growth environment based on deep learning, comprising: Step 1: Collect crop growth disaster information in the target area based on the pre-configured device, and perform data preprocessing and image annotation to obtain farmland images, environmental and soil data, and remote sensing data; Step 2: Perform image analysis on the farmland image, data analysis on the environmental and soil data, and data parsing on the remote sensing data; Step 3: Based on the image analysis results, data analysis results, and data parsing results, obtain the pest and disease detection results, soil drought classification results, air quality analysis results, and spatiotemporal analysis results, and summarize and output the data.

[0007] Preferably, information on crop growth disasters in the target area is collected based on a pre-configured device, including: Based on the soil detector, the soil interior of the target area is detected to obtain the first key data, which includes: soil moisture content, heavy metal content, pesticide residue concentration, soil pH value and organic matter content. Initial images of the soil surface and crop growth in the target area were captured using a high-speed linear array camera. The target area is scanned using a three-dimensional laser scanning device to construct a three-dimensional point cloud model of the target area; An infrared thermal imaging camera captures infrared radiation from the target object in the target area to generate a temperature distribution image. The target object in the target area is located based on the BeiDou positioning device; The crop growth disaster information includes: first key data, initial image, three-dimensional point cloud model, temperature distribution image and location information.

[0008] Preferably, data preprocessing and image annotation are performed on crop growth disaster information, including: The first key data is cleaned, integrated, transformed, and reduced to obtain the first processed data. The initial image is annotated at both the image level and the example level to obtain an annotated image; From the first processed data, labeled images, 3D point cloud models, temperature distribution images, and positioning information, farmland images, environmental and soil data, and remote sensing data are obtained.

[0009] Preferably, the process of detecting the interior of the soil in the target area using a soil detector includes: The target area is divided into units, and the estimated planting information and edge shape of each target unit are obtained. Based on the planting density and edge shape of the estimated planting information, a first number of detection positions for the corresponding target unit is determined, and these positions are uniformly set in the corresponding target unit according to the first number. Where N1 represents the corresponding first quantity; This indicates the planting density of the corresponding target unit; Indicates the unit measurement density of the corresponding target cell; This indicates the perimeter of the target cell. Indicates the unit measurement perimeter of the corresponding target unit; This represents the cell area of ​​the corresponding target cell; This indicates that the minimum radius is calculated for the element edge shape of the corresponding target element. The area of ​​the circle obtained by dividing it; This indicates a constant value greater than 3. The detection results of the soil detector at the detection location of the corresponding target unit are statistically analyzed.

[0010] Preferably, the first key data is cleaned, including: Construct a detection matrix for the detection results of the target unit, wherein the rows of the detection matrix are the values ​​of the same element at different detection positions, and the columns are the values ​​of different elements at the same detection position; Calculate the first feature coefficient of each row vector in the detection matrix, and simultaneously calculate the second feature coefficient of each column vector in the detection matrix; in, This represents the first eigenvalue of the i-th row vector; Represents the values ​​of all elements in the i-th row vector. The variance; ln denotes the logarithmic function sign; Represents the values ​​of all elements in the i-th row vector. The average value; Represents the values ​​of all elements in the i-th row vector. The maximum value; Represents the values ​​of all elements in the i-th row vector. The minimum value; in, This represents the second characteristic coefficient of the j-th column vector; This represents the value of the i-th element in the j-th column vector. The value of the kth element The relevant functions; n1 represents the number of elements in the j-th column vector; Represents all columns in the j-th column vector. The variance; This represents the average value obtained by standardizing and averaging the values ​​of the elements in the j-th column vector. The detection period for each target unit is determined, and the stability of the corresponding detection period is obtained based on the first feature coefficient and the second feature coefficient under each detection period. Then, a weight coefficient is set for the corresponding target unit. Based on the weighting coefficients and application scenarios, the cleaning rules for the corresponding target units are extracted from the coefficient-scenario-cleaning reference table, and the detection results of the target units are cleaned according to the cleaning rules.

[0011] Preferably, the stability of the corresponding detection period is obtained based on the first feature coefficient and the second feature coefficient for each detection period, and then a weighting coefficient is set for the corresponding target unit, including: in, This indicates the stability during the h-th detection cycle; Indicates the detection period h based on The variance; Indicates the detection period h based on The variance; Indicates the symbol for the exponential function; This indicates all data in the h-th detection period. The maximum value in; This indicates all data in the h-th detection period. The minimum value in; This indicates all data in the h-th detection period. The average value in; This indicates all data in the h-th detection period. The maximum value in; This indicates all data in the h-th detection period. The minimum value in; This indicates all data in the h-th detection period. The average value in; in, These are the weighting coefficients for the corresponding target units; For the target unit, the stability of all detection cycles over Tn is... The average stable value; The sum of the average stable values ​​of all target units; The variance of the average stable value of all target units; For all stability under the corresponding target element The variance.

[0012] Preferably, the initial image is annotated at both the image level and the example level to obtain an annotated image, including: Obtain the edge information set and background information set of each state image under different annotation categories, and extract information pairs consistent with the number of states from the edge information set and background information set to construct the annotation standard for each state image; By merging all annotation standards under the same annotation category, the content annotation mechanism and border annotation mechanism of the corresponding annotation category are obtained; The initial image is annotated at the image level according to the content annotation mechanism and at the example level according to the border annotation mechanism.

[0013] Preferably, the fusion process is performed based on all annotation standards under the same annotation category, including: Based on the crop growth attributes of the target area, priority weights are assigned to the labeling standards under the corresponding labeling categories; Based on the weight settings and annotation standards, the content annotation mechanism that outputs the corresponding annotation category is sequentially input into the fusion model. The outer minimum border is expanded according to the annotation shape that matches the annotation standard, and the boundary of the content annotation mechanism is adjusted to obtain the border annotation mechanism.

[0014] This invention provides a robot for applying any of the deep learning-based comprehensive detection methods for growth environments.

[0015] Compared with the prior art, the beneficial effects of this application are as follows: By integrating advanced sensing technology, autonomous navigation technology, and artificial intelligence algorithms, this invention achieves comprehensive, efficient, and precise monitoring of the farmland environment, providing strong technical support for the sustainable development of modern agriculture. This invention not only fills a market gap but also promotes the development of intelligent and automated agriculture. It enables automatic detection of farmland pests, diseases, and waterlogging-related environmental conditions, thereby helping farmland managers make better decisions.

[0016] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a comprehensive detection method for growth environment in an embodiment of the present invention; Figure 2 This is a structural diagram of the robot in an embodiment of the present invention; Figure 3 This is a flowchart of the integrated pest and disease detection process in farmland, as described in this embodiment of the invention. Detailed Implementation

[0019] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] This invention proposes a comprehensive detection method for the growth environment based on deep learning, such as... Figure 1 As shown, it includes: Step 1: Collect crop growth disaster information in the target area based on the pre-configured device, and perform data preprocessing and image annotation to obtain farmland images, environmental and soil data, and remote sensing data; Step 2: Perform image analysis on the farmland image, data analysis on the environmental and soil data, and data parsing on the remote sensing data; Step 3: Based on the image analysis results, data analysis results, and data parsing results, obtain the pest and disease detection results, soil drought classification results, air quality analysis results, and spatiotemporal analysis results, and summarize and output the data.

[0021] Preferably, information on crop growth disasters in the target area is collected based on a pre-configured device, including: Based on the soil detector, the soil interior of the target area is detected to obtain the first key data, which includes: soil moisture content, heavy metal content, pesticide residue concentration, soil pH value and organic matter content. Initial images of the soil surface and crop growth in the target area were captured using a high-speed linear array camera. The target area is scanned using a three-dimensional laser scanning device to construct a three-dimensional point cloud model of the target area; An infrared thermal imaging camera captures infrared radiation from the target object in the target area to generate a temperature distribution image. The target object in the target area is located based on the BeiDou positioning device; The crop growth disaster information includes: first key data, initial image, three-dimensional point cloud model, temperature distribution image and location information.

[0022] Preferably, data preprocessing and image annotation are performed on crop growth disaster information, including: The first key data is cleaned, integrated, transformed, and reduced to obtain the first processed data. The initial image is annotated at both the image level and the example level to obtain an annotated image; From the first processed data, labeled images, 3D point cloud models, temperature distribution images, and positioning information, farmland images, environmental and soil data, and remote sensing data are obtained.

[0023] In this embodiment, the soil detector can penetrate deep into the soil to accurately obtain various key data of the soil sample. This data provides a scientific basis for assessing soil health and guiding precision fertilization and irrigation.

[0024] The high-speed line scan camera, equipped with a high-resolution high-speed line scan camera, can capture subtle changes in soil surface and crop growth, generating high-definition RGB images.

[0025] 3D Laser Scanning System: Utilizing advanced laser scanning technology, this robot can rapidly construct 3D point cloud models of farmland and its crops. These models not only accurately reflect the undulations of the terrain but also precisely measure growth parameters such as crop height and canopy structure, providing rich 3D data support for spatial layout optimization and crop growth model construction.

[0026] Infrared thermal imaging camera: Infrared cameras capture the infrared radiation of target objects to generate temperature distribution images. This function is particularly important in monitoring the crop growth environment, as it can promptly detect abnormal temperature areas caused by factors such as pests, diseases, and water stress, providing crucial information for early warning and intervention.

[0027] Beidou Positioning Device: The terminal, equipped with the built-in Beidou satellite positioning system, ensures precise positioning and navigation for the robot in farmland. Whether autonomously planning patrol routes or uploading location and status information in real time, it relies on the support of this high-precision positioning technology. At the same time, the Beidou positioning terminal also provides strong technical support for remote monitoring and dispatching.

[0028] The deployment of the above equipment enables comprehensive and high-precision monitoring of the farmland environment.

[0029] In this embodiment, the purpose of data preprocessing is that, due to the complexity and diversity of the sampling environment, the collected data often suffers from incompleteness, inconsistency, and noise interference. These problems directly affect the accuracy and reliability of subsequent data analysis. Therefore, in order to improve data quality and ensure data integrity and accuracy, data preprocessing is necessary.

[0030] In this embodiment, data cleaning refers to improving the cleanliness of data by identifying and correcting errors, outliers, missing values, etc.

[0031] Data integration involves combining data from different sources and in different formats into a unified dataset for comprehensive analysis. During the integration process, issues such as data redundancy and data conflicts need to be addressed to ensure data consistency and integrity.

[0032] Data transformation: This involves converting data into a form suitable for analysis through methods such as normalization, standardization, and discretization. For example, converting continuous variables into categorical variables, or variables with different dimensions into variables with a unified dimension, to facilitate comparison and analysis.

[0033] Data reduction: Without affecting the data analysis results, data complexity and storage requirements are reduced by methods such as deleting redundant data and reducing data dimensions, thereby improving data processing efficiency.

[0034] In this embodiment, image-level annotation: The labeling content includes categories such as weeds, debris, wild animals, pests and diseases, crop growth cycle status, and drought level.

[0035] Labeling method: Assign one or more category labels to each image to represent the main objects or phenomena present in the image.

[0036] Instance-level annotation: Location annotation: The target area is annotated using bounding boxes. The annotation file stores the coordinates of the four vertices of the bounding box to accurately describe the target's position in the image.

[0037] Category labeling: The category information of the target is recorded in the labeling file that stores the location coordinates, so as to facilitate subsequent classification and identification.

[0038] Pixel-level annotation: For scenarios requiring higher precision recognition, pixel-level annotation can be used. This involves labeling areas such as weeds, debris, pests, and wild animals with different pixel values, where each pixel value directly represents the category of the object at that location. While this method is more time-consuming, it provides richer detail information, helping to improve recognition accuracy.

[0039] Annotation Tool: Labelme, a powerful image annotation tool, was chosen for the annotation work. Labelme supports multiple annotation methods, including rectangular bounding box annotation, polygon annotation, and pixel-level annotation, which can meet the annotation needs of different scenarios. Labelme also provides a user-friendly interface and rich export options, making it convenient for users to view and export annotation results.

[0040] In this embodiment, a manually labeled soil surface dataset is trained using the SegmentAnythingModel (SAM) model. SAM is currently the highest-performing object detection and segmentation model, boasting high accuracy and speed, enabling real-time and rapid object detection. RGB images are input into the SAM model, which uses an FPN network to extract object features. After multiple layers of multi-head attention operations and a segmentation head, the final detection result is obtained, as shown below. Figure 3 As shown, the input image is processed by a pyramid network to extract image patches for embedding. The embedded features are then regularized and input into a multi-head attention layer. The output features are fused with the original features and further regularized. Finally, the results are output through a segmentation head. For weeds, debris, pests, and wild animals, it is necessary to detect the category, location coordinates, and pixel-level labels simultaneously. For drought levels and crop growth cycle status, only image-level category identification is required.

[0041] In this embodiment, traditional soil and air analysis instruments have relatively limited functions. This invention proposes a comprehensive environmental detector capable of acquiring and analyzing air and soil data in real time. The comprehensive environmental detector is equipped on a comprehensive inspection robot for crop growth environment during disasters. It enters the soil through a drilling device to collect data and simultaneously analyze multiple air components. The comprehensive soil detector collects and analyzes data every three meters, ensuring the integrity and uniformity of the farmland samples. Furthermore, by combining remote sensing spatiotemporal data analysis, soil drought levels and air quality classifications are determined, thereby identifying waterlogging disasters.

[0042] In this embodiment, the present invention can comprehensively detect crop disasters and growth environment. Therefore, the data is complex and diverse, and the data volume is large. In order to facilitate classification and management, the data is classified and summarized into the background management system, so that relevant management departments can carry out supervision and early warning based on it.

[0043] The beneficial effects of the above technical solution are: by integrating advanced sensing technology, autonomous navigation technology, and artificial intelligence algorithms, it achieves comprehensive, efficient, and accurate monitoring of the farmland environment, providing strong technical support for the sustainable development of modern agriculture. This invention not only fills a market gap but also promotes the development of intelligent and automated agriculture. It enables automatic detection of farmland pests and diseases, as well as the growth environment affected by waterlogging, thereby helping farmland managers make better decisions.

[0044] This invention proposes a comprehensive detection method for growth environment based on deep learning. The method involves detecting the interior of the soil in a target area using a soil detector, and includes: The target area is divided into units, and the estimated planting information and edge shape of each target unit are obtained. Based on the planting density and edge shape of the estimated planting information, a first number of detection positions for the corresponding target unit is determined, and these positions are uniformly set in the corresponding target unit according to the first number. Where N1 represents the corresponding first quantity; This indicates the planting density of the corresponding target unit; Indicates the unit measurement density of the corresponding target cell; This indicates the perimeter of the target cell. Indicates the unit measurement perimeter of the corresponding target unit; This represents the cell area of ​​the corresponding target cell; This indicates that the minimum radius is calculated for the element edge shape of the corresponding target element. The area of ​​the circle obtained by dividing it; This indicates a constant value greater than 3. The detection results of the soil detector at the detection location of the corresponding target unit are statistically analyzed.

[0045] In this embodiment, the unit splitting can be a sequential division of the target area according to a certain area to obtain several target units. The estimated planting information refers to the crops that the unit was originally planned to plant and the planting density, which are preset.

[0046] In this embodiment, the cell edge shape refers to the outline shape of the target cell.

[0047] In this embodiment, the purpose of determining the first quantity is to perform reasonable and reliable detection of the target unit.

[0048] In this embodiment, the area of ​​the circle obtained by dividing by the minimum radius refers to the area of ​​the circle based on the inscribed circle of the corresponding target unit.

[0049] The beneficial effects of the above technical solution are: by dividing the area into units, it is easier to improve the detection accuracy, and the number of detection positions can be determined by planting density and edge shape, which facilitates reasonable and reliable detection of the unit and provides a basis for analyzing the growth environment.

[0050] This invention proposes a comprehensive detection method for growth environment based on deep learning, which includes data cleaning of the first key data, including: Construct a detection matrix for the detection results of the target unit, wherein the rows of the detection matrix are the values ​​of the same element at different detection positions, and the columns are the values ​​of different elements at the same detection position; Calculate the first feature coefficient of each row vector in the detection matrix, and simultaneously calculate the second feature coefficient of each column vector in the detection matrix; in, This represents the first eigenvalue of the i-th row vector; Represents the values ​​of all elements in the i-th row vector. The variance; ln denotes the logarithmic function sign; Represents the values ​​of all elements in the i-th row vector. The average value; Represents the values ​​of all elements in the i-th row vector. The maximum value; Represents the values ​​of all elements in the i-th row vector. The minimum value; in, This represents the second characteristic coefficient of the j-th column vector; This represents the value of the i-th element in the j-th column vector. The value of the kth element The relevant functions; n1 represents the number of elements in the j-th column vector; Represents all columns in the j-th column vector. The variance; This represents the average value obtained by standardizing and averaging the values ​​of the elements in the j-th column vector. The detection period for each target unit is determined, and the stability of the corresponding detection period is obtained based on the first feature coefficient and the second feature coefficient under each detection period. Then, a weight coefficient is set for the corresponding target unit. Based on the weighting coefficients and application scenarios, the cleaning rules for the corresponding target units are extracted from the coefficient-scenario-cleaning reference table, and the detection results of the target units are cleaned according to the cleaning rules.

[0051] In this embodiment, the detection matrix = .

[0052] In this embodiment, the detection cycle can be once every 10 days, and the application scenario refers to the type of crop being planted and the climate conditions of the geographical location where the crop is planted.

[0053] In this embodiment, the coefficient-scenario-cleaning comparison table contains cleaning rules matched with weight coefficients under different application scenarios. Different cleaning rules result in different data cleaning results for the detection results. Furthermore, the cleaning rules involved in this database are existing technologies.

[0054] The beneficial effects of the above technical solution are: by performing row and column analysis on the matrix to obtain the corresponding feature coefficients, the stability under different detection cycles can be calculated, providing a basis for matching the data cleaning rules of the unit. Furthermore, by matching the weight coefficients and application scenarios from the reference table, the cleaning rules can be effectively obtained, ensuring the reliability of data cleaning.

[0055] This invention proposes a comprehensive detection method for growth environments based on deep learning. The method obtains the stability of the corresponding detection period based on the first feature coefficient and the second feature coefficient under each detection period, and then sets weight coefficients for the corresponding target units, including: in, This indicates the stability during the h-th detection cycle; Indicates the detection period h based on The variance; Indicates the detection period h based on The variance; Indicates the symbol for the exponential function; This indicates all data in the h-th detection period. The maximum value in; This indicates all data in the h-th detection period. The minimum value in; This indicates all data in the h-th detection period. The average value in; This indicates all data in the h-th detection period. The maximum value in; This indicates all data in the h-th detection period. The minimum value in; This indicates all data in the h-th detection period. The average value in; in, These are the weighting coefficients for the corresponding target units; For the target unit, the stability of all detection cycles over Tn is... The average stable value; The sum of the average stable values ​​of all target units; The variance of the average stable value of all target units; For all stability under the corresponding target element The variance.

[0056] In this embodiment, the value of h is 10.

[0057] The beneficial effects of the above technical solution are: stability is calculated based on the first feature coefficient and the second feature coefficient, and weight coefficients are subsequently calculated based on the stability under different detection cycles, providing a data basis for subsequent matching with the reference table.

[0058] This invention proposes a deep learning-based method for comprehensive detection of growth environments, which performs image-level and instance-level annotation on the initial image to obtain an annotated image, including: Obtain the edge information set and background information set of each state image under different annotation categories, and extract information pairs consistent with the number of states from the edge information set and background information set to construct the annotation standard for each state image; By merging all annotation standards under the same annotation category, the content annotation mechanism and border annotation mechanism of the corresponding annotation category are obtained; The initial image is annotated at the image level according to the content annotation mechanism and at the example level according to the border annotation mechanism.

[0059] In this embodiment, different labeling categories refer to categories such as weeds, debris, pests and diseases, and wild animals. Images under different categories need to be manually labeled to distinguish them based on edge information and background information. The state image is used to reflect the state of objects under the corresponding labeling category in different states. Specifically, it reflects the state of objects corresponding to crops in different growth states.

[0060] In this embodiment, the number of states refers to the number of growth states of crops. The number of states is different for different crops and can be obtained from a crop-quantity lookup table, which contains different crops and the corresponding number of growth states of crops.

[0061] In this embodiment, the information pair refers to the extracted edge information and background information, and the edge information set contains edge information under different state images, and the background information set contains background information under different state images.

[0062] In this embodiment, the annotation standard is to extract the annotation results from the information pairs to obtain the annotation vector for the corresponding state image. All annotation vectors are input into the fusion model to obtain the content annotation mechanism, such as the boundary annotation of adjacent objects (image-level annotation).

[0063] In this embodiment, the edge annotation mechanism refers to selecting objects by rectangular borders (executive-level annotation).

[0064] The beneficial effects of the above technical solution are: by acquiring background and edge information of different categories to construct annotation standards, and then by fusing the two mechanisms, effective annotation of images can be achieved.

[0065] This invention proposes a comprehensive detection method for growth environments based on deep learning, which performs fusion processing on all annotation standards under the same annotation category, including: Based on the crop growth attributes of the target area, priority weights are assigned to the labeling standards under the corresponding labeling categories; Based on the weight settings and annotation standards, the content annotation mechanism that outputs the corresponding annotation category is sequentially input into the fusion model. The outer minimum border is expanded according to the annotation shape that matches the annotation standard, and the boundary of the content annotation mechanism is adjusted to obtain the border annotation mechanism.

[0066] In this embodiment, crop growth attributes refer to the importance of protection for crops under different growth states, and the higher the priority of protection, the greater the weight of the corresponding priority.

[0067] In this embodiment, the fusion model is obtained by training the neural network model with samples based on different annotation vectors and weight settings. Therefore, the content annotation mechanism can be obtained, and the shape formed by the annotation boundary is the annotation shape.

[0068] In this embodiment, the expansion of the outer minimum border refers to a rectangle, which facilitates the representation of the object's position.

[0069] The beneficial effects of the above technical solution are: the content annotation mechanism enables the annotation of object boundaries to determine the objects present in the farmland, and the border annotation mechanism expands the annotation shape with the minimum border, which facilitates the effective annotation of the location.

[0070] This invention provides a robot for applying any of the deep learning-based comprehensive detection methods for growth environments, such as... Figure 2 As shown, the robot specifically includes: a camera, a control processing unit, a power supply / motor controller, a processing terminal, a tracked device, a multispectral camera, an infrared camera, a drilling device, a POS stabilizing gimbal, a laser scanner, and a soil detector.

[0071] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A comprehensive detection method for growth environment based on deep learning, characterized in that, include: Step 1: Collect crop growth disaster information in the target area based on the pre-configured device, and perform data preprocessing and image annotation to obtain farmland images, environmental and soil data, and remote sensing data; Step 2: Perform image analysis on the farmland image, data analysis on the environmental and soil data, and data parsing on the remote sensing data; Step 3: Based on the image analysis results, data analysis results, and data parsing results, obtain the pest and disease detection results, soil drought classification results, air quality analysis results, and spatiotemporal analysis results, and summarize and output the data.

2. The deep learning-based comprehensive detection method for growth environment according to claim 1, characterized in that, Information on crop growth disasters in the target area is collected based on pre-configured devices, including: Based on the soil detector, the soil interior of the target area is detected to obtain the first key data, which includes: soil moisture content, heavy metal content, pesticide residue concentration, soil pH value and organic matter content. Initial images of the soil surface and crop growth in the target area were captured using a high-speed linear array camera. The target area is scanned using a three-dimensional laser scanning device to construct a three-dimensional point cloud model of the target area; An infrared thermal imaging camera captures infrared radiation from the target object in the target area to generate a temperature distribution image. The target object in the target area is located based on the BeiDou positioning device; The crop growth disaster information includes: first key data, initial image, three-dimensional point cloud model, temperature distribution image and location information.

3. The deep learning-based comprehensive detection method for growth environment according to claim 2, characterized in that, Data preprocessing and image annotation of crop growth disaster information, including: The first key data is cleaned, integrated, transformed, and reduced to obtain the first processed data. The initial image is annotated at both the image level and the example level to obtain an annotated image; From the first processed data, labeled images, 3D point cloud models, temperature distribution images, and positioning information, farmland images, environmental and soil data, and remote sensing data are obtained.

4. The deep learning-based comprehensive detection method for growth environment according to claim 1, characterized in that, The process of detecting the interior of the soil in a target area using a soil detector includes: The target area is divided into units, and the estimated planting information and edge shape of each target unit are obtained. Based on the planting density and edge shape of the estimated planting information, a first number of detection positions for the corresponding target unit is determined, and these positions are uniformly set in the corresponding target unit according to the first number. Where N1 represents the corresponding first quantity; This indicates the planting density of the corresponding target unit; Indicates the unit measurement density of the corresponding target cell; This indicates the perimeter of the target cell. Indicates the unit measurement perimeter of the corresponding target unit; This represents the cell area of ​​the corresponding target cell; This indicates that the minimum radius is calculated for the element edge shape of the corresponding target element. The area of ​​the circle obtained by dividing it; This indicates a constant value greater than 3. The detection results of the soil detector at the detection location of the corresponding target unit are statistically analyzed.

5. The deep learning-based comprehensive detection method for growth environment according to claim 4, characterized in that, Data cleaning is performed on the first key data, including: Construct a detection matrix for the detection results of the target unit, wherein the rows of the detection matrix are the values ​​of the same element at different detection positions, and the columns are the values ​​of different elements at the same detection position; Calculate the first feature coefficient of each row vector in the detection matrix, and simultaneously calculate the second feature coefficient of each column vector in the detection matrix; in, This represents the first eigenvalue of the i-th row vector; Represents the values ​​of all elements in the i-th row vector. The variance; ln denotes the logarithmic function sign; Represents the values ​​of all elements in the i-th row vector. The average value; Represents the values ​​of all elements in the i-th row vector. The maximum value; Represents the values ​​of all elements in the i-th row vector. The minimum value; in, This represents the second characteristic coefficient of the j-th column vector; This represents the value of the i-th element in the j-th column vector. The value of the kth element The relevant functions; n1 represents the number of elements in the j-th column vector; Represents all columns in the j-th column vector. The variance; This represents the average value obtained by standardizing and averaging the values ​​of the elements in the j-th column vector. The detection period for each target unit is determined, and the stability of the corresponding detection period is obtained based on the first feature coefficient and the second feature coefficient under each detection period. Then, a weight coefficient is set for the corresponding target unit. Based on the weighting coefficients and application scenarios, the cleaning rules for the corresponding target units are extracted from the coefficient-scenario-cleaning reference table, and the detection results of the target units are cleaned according to the cleaning rules.

6. The deep learning-based comprehensive detection method for growth environment according to claim 5, characterized in that, Based on the first feature coefficient of each detection cycle and the second feature coefficient, the stability of the corresponding detection cycle is obtained, and then weight coefficients are set for the corresponding target units, including: in, This indicates the stability during the h-th detection cycle; Indicates the h-th detection period based on The variance; Indicates the h-th detection period based on The variance; Indicates the symbol for an exponential function; This indicates all data in the h-th detection period. The maximum value in; This indicates all data in the h-th detection period. The minimum value in; This indicates all data in the h-th detection period. The average value in; This indicates all data in the h-th detection period. The maximum value in; This indicates all data in the h-th detection period. The minimum value in; This indicates all data in the h-th detection period. The average value in; in, These are the weighting coefficients for the corresponding target units; For the target unit, the stability of all detection cycles over Tn is... The average stable value; The sum of the average stable values ​​of all target units; The variance of the average stable value of all target units; For all stability under the corresponding target element The variance.

7. The deep learning-based comprehensive detection method for growth environment according to claim 1, characterized in that, The initial image is annotated at both the image level and the application level to obtain an annotated image, including: Obtain the edge information set and background information set of each state image under different annotation categories, and extract information pairs consistent with the number of states from the edge information set and background information set to construct the annotation standard for each state image; By merging all annotation standards under the same annotation category, the content annotation mechanism and border annotation mechanism of the corresponding annotation category are obtained; The initial image is annotated at the image level according to the content annotation mechanism and at the example level according to the border annotation mechanism.

8. The deep learning-based comprehensive detection method for growth environment according to claim 7, characterized in that, The data is merged based on all annotation standards within the same annotation category, including: Based on the crop growth attributes of the target area, priority weights are assigned to the labeling standards under the corresponding labeling categories; Based on the weight settings and annotation standards, the content annotation mechanism that outputs the corresponding annotation category is sequentially input into the fusion model. The outer minimum border is expanded according to the annotation shape that matches the annotation standard, and the boundary of the content annotation mechanism is adjusted to obtain the border annotation mechanism.

9. A robot, characterized in that, Used for applying the deep learning-based comprehensive detection method for growth environment as described in any one of claims 1-8.

Images