Road greening project maintenance method and system based on machine learning

Abstract

The invention belongs to the technical field of road maintenance, and provides a road greening project maintenance method and system based on machine learning, and the method comprises the following steps: S100, obtaining a to-be-maintained greening project region map, generating a collection path, and enabling an unmanned plane to fly along the collection path to obtain a plurality of hyperspectral images of the region map; s200, generating a first image according to the plurality of acquired hyperspectral images; s300, marking an insect pest area in the first image and updating the insect pest area into a second image; and S400, training the decision tree model by using the second image to obtain an initialized decision tree model, and training and learning insect pest image features and environmental parameters by introducing the decision tree model, so that key influence factors can be autonomously extracted, intelligent identification and diffusion trend judgment of an insect pest area are realized, and the accuracy of insect pest recognition is improved. And targeted maintenance suggestions are generated according to training results, so that the accuracy and reliability of disease and pest detection are improved, and the method is particularly suitable for expressway scenes with wide green belt distribution and fast environment change.

Description

Technical Field

[0001] The present invention belongs to the technical field of highway maintenance, and in particular relates to a highway greening project maintenance method and system based on machine learning. Background Art

[0002] Highway green belts face numerous challenges during maintenance due to their wide geographical distribution, diverse plant species, and significant climate differences. In particular, pests and diseases frequently occur and are difficult to prevent and control. Traditional maintenance methods rely primarily on manual inspections and empirical judgments, which can lead to delayed responses, missed inspections, misjudgments, and untimely treatment. These can easily lead to the spread of pests and the death of vegetation, increasing maintenance costs while also impacting road safety and the overall landscape. Therefore, a more efficient and intelligent pest and disease control method is urgently needed to accurately identify and dynamically manage the health status of green plants, thereby improving the sustainable maintenance level of highway greening projects. Summary of the Invention

[0003] The purpose of the present invention is to propose a highway greening project maintenance method and system based on machine learning to solve one or more technical problems existing in the prior art and at least provide a beneficial option or create conditions.

[0004] To achieve the above objectives, according to one aspect of the present invention, a highway greening project maintenance method based on machine learning is provided. The highway greening project maintenance method based on machine learning comprises the following steps:

[0005] S100, obtaining a map of the greening project area to be maintained and generating a collection path, and flying a drone along the collection path to obtain multiple hyperspectral images of the area map;

[0006] S200, generating a first image based on the acquired multiple hyperspectral images;

[0007] S300, marking the pest-infested area in the first image and updating it into the second image;

[0008] S400: Initialize the decision tree model obtained by training the decision tree model with the second image.

[0009] Furthermore, the method further includes step S500 of identifying a maintenance location for the second image by initializing a decision tree model.

[0010] Furthermore, the UAV is a fixed-wing UAV or a rotary-wing UAV including a hyperspectral imaging system, and the initial parameters of the UAV are set: the field of view angle of the hyperspectral imaging system (10-30°), the route spacing (10-50m), the flight altitude (15-60m), the flight speed (8-30m / s), the boundary safety distance (3-7m), and the obstacle safety distance (3-7m).

[0011] Further, in S100, the region map to be maintained is obtained, and a collection path is generated, and the specific method of flying the unmanned aerial vehicle along the collection path to obtain multiple hyperspectral images of the region map is: obtaining the region map to be maintained, setting the starting point, the ending point and the obstacle position information of the unmanned aerial vehicle flying in the region map to be maintained, generating the flight path of the unmanned aerial vehicle from the starting point to the ending point, and generating collection points according to fixed distances, and the unmanned aerial vehicle flies according to the path, and the spectral image is collected by the spectral instrument carried at the collection points.

[0012] Further, the unmanned aerial vehicle is a coaxial dual-rotor unmanned aerial vehicle, a micro-rotor unmanned aerial vehicle or a multi-rotor unmanned aerial vehicle carrying a spectral instrument.

[0013] Further, the spectral instrument is a GaiaField-mini spectral instrument, a Specim AFX series hyperspectral camera or an ATH9020 hyperspectral imager.

[0014] Further, the flight path of the unmanned aerial vehicle from the starting point to the ending point is generated by using any one of A* algorithm, K-shortest path algorithm, Dijkstra algorithm, APF algorithm (APF: Artificial Potential Field) and SAA algorithm (SAA: Simulated Annealing Algorithm).

[0015] Further, the starting point is the position (generally the current position) of the unmanned aerial vehicle taking off in the region to be monitored, the ending point is the terminal position of the unmanned aerial vehicle flying in the region to be monitored, and the obstacle position information is the position of the obstacle in the region to be maintained, and the unmanned aerial vehicle needs to keep a preset obstacle safety distance (2-5m) from the obstacle position.

[0016] Further, in S200, the specific method of generating the first image according to the obtained multiple hyperspectral images is: splicing all the obtained multispectral images in the order of collection time to generate the first image.

[0017] Further, the splicing process includes the steps of: geometric correction, image preprocessing, image registration and image fusion.

[0018] The current common pest monitoring method mainly relies on manual patrol and fixed-point image acquisition, supplemented by periodic spraying or pruning measures. However, in practical application, such means have many limitations: on the one hand, due to the dependence on fixed cameras or manual collection for image acquisition, the data update delay and information feedback lag problems exist due to the limitations of the number of devices, network transmission and image processing capacity, making it difficult to realize real-time monitoring and rapid response of pests; on the other hand, the spread of pests in green belts is affected by factors such as traffic disturbance, wind changes and vegetation structure differences, showing complex characteristics such as unstable transmission path and uneven spread speed, resulting in inaccurate identification of high-incidence areas of pests. In addition, the current system lacks the ability to predict and dynamically analyze the development trend of pests, resulting in untimely, mismatched and insufficient intervention measures, ultimately affecting maintenance effectiveness and increasing operation and maintenance costs. To solve the above problems, the present application proposes the following method: by identifying the pest area and calculating the diffusion coefficient of all pest areas, the specific geographic location of the pest is accurately predicted, and the intelligent maintenance level of the high-speed green engineering is improved:

[0019] Further, in S300, the specific method for marking the pest area in the first image and updating the second image is: marking multiple pest areas PA in the first image by RX algorithm, taking i as the serial number of the pest area, PA i represents the i-th pest area, and all pest areas form a set INP; using an edge algorithm, the first image is divided into multiple sub-regions RAj, taking j as the serial number of the sub-region, RAj represents the j-th sub-region;

[0020] The geometric center of all pest areas is denoted as P(i), and the geometric center of all sub-regions is denoted as R(i). All pest areas in the set INP are traversed, the distance between the geometric center P(i) of all pest areas and the geometric center R(i) of all projections is calculated as PR(j), and all projections with PR(j) less than PR_mean are added to the list Listi, where PR_mean is the average distance between the point P(i) and all points R(i).

[0021] Further, the pest area is an area composed of pest pixels, or the internal area of an edge composed of pest pixels; wherein the pest pixel is a pixel in the spectral image whose visible light reflectance is higher than the average visible light reflectance of each pixel in the spectral image, or the pest pixel is a pixel in the spectral image whose near-infrared band spectral reflectance is lower than the average near-infrared band spectral reflectance of each pixel in the spectral image.

[0022] Within the value range of k, all sub-regions in Listi are traversed in turn, and the pest area PA iThe diffusion coefficient Diff in each sub-region in Listi is calculated in the following way: the average pixel value in each sub-region is calculated in sequence, denoted as GreenK_Mean, the maximum pixel value in the kth sub-region is denoted as GreenK_Max, the minimum pixel value is denoted as GreenK_Min, and the pixel value of the edge point of the sub-region is denoted as Green_edg(k, r), wherein K represents the serial number of the sub-region contained in Listi, r represents the serial number of the edge point on the edge, and Green_edg(k, r) represents the pixel value of the rth edge point on the edge of the kth sub-region. The diffusion coefficient Diff is calculated by the formula: The diffusion coefficient Diff is calculated, wherein q(k) is the distance between the pixel point with the maximum pixel value and the pixel point with the minimum pixel value in the kth sub-region, and PR(k) represents the distance between the geometric center of the pest area PA i and the geometric center of the sub-region.

[0023] If Diff is greater than Green_Mean, wherein Green_Mean is the average value of the pixels in all sub-regions contained in Listi, it is considered that the current pest area has occurred diffusion phenomenon, and the current pest area needs to be corrected, otherwise the current pest area has not occurred diffusion phenomenon, and the current pest area does not need to be corrected.

[0024] The beneficial effects of the above steps are: by calculating the diffusion coefficient Diff of the pest area in each sub-region in the image, the pixel statistical characteristics and edge point deviation in the region are considered, the gray gradient distribution and spatial geometric relationship of the pest area are comprehensively considered, the diffusion of the pest is accurately judged, and the diffusion is no longer judged by relying on a single threshold or visual connectivity. Instead, a diffusion formula containing the maximum pixel difference, edge offset and center distance weight is introduced, which effectively improves the sensitivity and stability of diffusion recognition while maintaining the calculation efficiency. Especially in the scene of highway green belt and the like which is greatly affected by traffic disturbance and light change, the method can significantly improve the accuracy of pest dynamic diffusion recognition, reduce the misjudgment rate, and provide a scientific and quantitative decision basis for subsequent pest area correction, thereby optimizing the green maintenance strategy and improving the timeliness and accuracy of pest control.

[0025] Further, the specific correction method is: within the value range of k, the diffusion coefficient Diff of the pest area in each sub-region in Listi is calculated in sequence k , wherein k = 1, 2, 3…m, m represents the number of sub-regions contained in Listi, all sub-regions contained in Listi are rearranged in the order of Diff from small to large, and the diffusion coefficient Diff of the sub-region in the updated Listi is iteratively updated. kIf satisfied, Diff k -Diff k-1 <Diff k+1 -Diff k and Diff k >Diff mean where Diff mean represents the average value of the diffusion coefficient Diff of the sub-region in Listi, Diff k The corresponding sub-region is recorded as a strong influence sub-region,

[0026] The pest area PA i is traversed in sequence, the pixel value on the edge is recorded as PL, the minimum value in the pixel value is selected as the starting point ST, the maximum value in the pixel value is selected as the end point ED, and the PA i is searched from the ST along the minimum path to the ED, the size of the pixel value PL on the edge is taken as s and q as the serial number of the pixel value on the edge, if satisfied and the pixel point corresponding to PLs is updated as the starting point ST; the PA i is searched from the ED along the minimum path to the ST, the size of the pixel value PL on the edge is taken as s as the serial number of the pixel value on the edge, if satisfied and the pixel point corresponding to PLs is updated as the end point ED, and z represents the number of pixel points on the edge. i

[0027] The position of the pixel point corresponding to ST is recorded as point A, the position of the pixel point corresponding to ED is recorded as point B, the minimum curve segment between point A and point B is recorded as the deletion segment and is deleted, the pixel value on the edge of the strong influence sub-region is sequentially recorded, the point corresponding to the maximum pixel value is recorded as point C, the point corresponding to the minimum pixel value is recorded as point D, the minimum curve segment between point C and point D is recorded as the replacement segment, the replacement segment is copied and is scaled to the point A and the point B to modify the pest area PA i .

[0028] According to the above method, all pest areas in the first image are modified, and the first image is updated as the second image.

[0029] ​The beneficial effect of the above steps is that the above steps combine the quantitative judgment of the diffusion trend, which can effectively make up for the time delay and lag response problems existing in the current pest monitoring means, by constructing the path retrieval logic of the pest edge pixels, accurately extracting the extreme points of the boundary pixels in the pest area, and using the statistical difference between the minimum path and the edge pixel sequence, the real diffusion boundary of the pest area is dynamically judged, in addition, by executing the boundary line segment replacement and interpolation repair strategy in the strong influence sub-area, the automatic deviation correction and structured reconstruction of the pest area boundary are realized, so as to generate a corrected area graph that is more in line with the actual pest transmission characteristics, the above method not only improves the timeliness and accuracy of pest area identification, but also enhances the adaptability of the system to the dynamic diffusion process of the pest, provides a reliable basis for subsequent precise maintenance and decision-making of greenery, and significantly improves the intelligent and automatic level of pest control of greenery.

[0030] Further, in S400, the specific method for training the second image to obtain the initialized decision tree model of the decision tree model is:

[0031] Each pest area in the second image is divided into a training set and a validation set, specifically: each pest area is marked as 1 as a positive sample, and the gray pest area with a diffusion coefficient greater than the average value of the pixels of the second image is selected as a negative sample and marked as 0, thereby completing the labeling of the samples; the labeled positive samples and negative samples constitute the training sample set; the training sample set is divided into a training set and a validation set according to a ratio of 4:1, the training set is used to train the decision tree model, and the validation set is used to verify the prediction performance of the trained decision tree model.

[0032] Further, the decision tree model is a GBDT model.

[0033] The training set and the validation set are used to train the decision tree model to obtain a pre-trained decision tree model, wherein the method for training the decision tree model using the training set and the validation set is: extracting the features of the training set and inputting the decision tree model, using a grid search method to optimize the hyperparameters in the decision tree model, retraining the decision tree model according to the optimized hyperparameters, and obtaining a trained initialized decision tree model.

[0034] Further, in S500, the specific method for identifying the maintenance site in the to-be-maintained green engineering area graph by using the initialized decision tree model is:

[0035] The unmanned aerial vehicle carrying the spectrometer obtains the hyperspectral image of the to-be-maintained green engineering area map, and identifies multiple pest areas through the RX algorithm, and uses the initialized decision tree model to screen the positive sample area in the multiple pest areas, if the positive sample area can be screened, it is judged that the to-be-maintained green engineering area map exists a diffusion pest area, and the geometric center of all diffusion pest areas is taken as the maintenance site, and all maintenance sites are generated by the curve fitting method to generate the pesticide application path.

[0036] Beneficial effects: the present application can realize accurate positioning and dynamic diffusion prediction of pest areas by fusing hyperspectral image analysis and environmental factor modeling, effectively breaking through the limitations of existing highway green pest monitoring methods relying on manual experience, lagging response and limited coverage, the system can autonomously extract key influencing factors by introducing the decision tree model to train and learn pest image features and environmental parameters, realize intelligent identification and diffusion trend judgment of pest areas, and generate targeted maintenance suggestions according to the training results, compared with the traditional static image recognition method, the present application uses the high interpretability and strong generalization ability of the decision tree classification mechanism, improves the accuracy and reliability of pest detection, and is especially suitable for highway scenes with wide green belt distribution and fast environmental change, significantly enhances the intelligent level and practical value of the pest monitoring system.

[0037] The present application also provides a highway green engineering maintenance system based on machine learning, which comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to run in the following system units:

[0038] A path generation unit is used to obtain a to-be-maintained green engineering area map and generate a collection path;

[0039] An image collection unit is used to control the unmanned aerial vehicle to fly along the collection path to obtain multiple hyperspectral images of the area map;

[0040] An image splicing unit is used to generate a first image according to the obtained multiple hyperspectral images;

[0041] A pest marking unit is used to mark the pest area in the first image;

[0042] A coefficient calculation unit is used to calculate the diffusion coefficient of all marked pest areas;

[0043] An image correction unit is used to correct the first image to obtain a second image according to the calculated diffusion coefficient;

[0044] A model training unit is used to train a decision tree model to obtain an initialized decision tree model.

[0045] The method has the advantages that the method can realize accurate positioning and dynamic diffusion prediction of the pest and disease area by fusing hyperspectral image analysis and environmental factor modeling, effectively breaks through the limitations of the existing highway greening pest and disease monitoring means, such as dependence on manual experience, lagging response and limited coverage, and the system can autonomously extract key influencing factors to realize intelligent identification and diffusion trend judgment of the pest and disease area, and generate targeted maintenance suggestions according to the training results, compared with the traditional static image recognition mode, the method uses the high interpretability and strong generalization ability of the decision tree classification mechanism to improve the accuracy and reliability of pest and disease detection, and is especially suitable for the highway scene with wide distribution of green belts and rapid environmental change, and significantly enhances the intelligent level and practical value of the pest and disease monitoring system. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 Fig. 1 shows a flowchart of a highway greening engineering maintenance method based on machine learning;

[0047] Figure 2 Fig. 2 shows a system structure diagram of a highway greening engineering maintenance system based on machine learning. DETAILED DESCRIPTION

[0048] Embodiments of the present application will be described in detail below, examples of which are shown in the drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0049] Example 1:

[0050] Figure 1 Fig. 1 shows a flowchart of a highway greening engineering maintenance method based on machine learning.

[0051] With reference to Figure 1 The present application provides a highway greening engineering maintenance method based on machine learning, which comprises the following steps:

[0052] S100, acquiring a to-be-maintained greening engineering area map and generating a collection path, and flying a UAV along the collection path to acquire multiple hyperspectral images of the area map;

[0053] S200, generating a first image according to the acquired multiple hyperspectral images;

[0054] S300, marking a pest and disease area in the first image and updating it to a second image;

[0055] S400, training the second image to obtain an initialized decision tree model.

[0056] Further, the method further comprises a step S500 of identifying the maintenance site from the second image by initializing the decision tree model.

[0057] Further, the unmanned aerial vehicle is a fixed-wing unmanned aerial vehicle or a rotor unmanned aerial vehicle comprising a hyperspectral imaging system, and the initial parameters of the unmanned aerial vehicle are set as a field angle of the hyperspectral imaging system of 16°, a flight height of 30 m, a flight speed of 15 m, a route interval of 30 m, a boundary safety distance of 5 m, and an obstacle safety distance of 5 m.

[0058] Further, in S100, the specific method of obtaining the greenery engineering area to be maintained and generating a collection path, and flying the unmanned aerial vehicle along the collection path to obtain a plurality of hyperspectral images of the area map is as follows: obtaining the greenery engineering area to be maintained, setting the starting point, the ending point and the obstacle position information of the unmanned aerial vehicle flying in the greenery engineering area to be maintained, generating a route path of the unmanned aerial vehicle from the starting point to the ending point, and generating collection points at fixed distances, and the unmanned aerial vehicle flies along the path and collects spectral images at the collection points by the carried spectrometer.

[0059] Further, the unmanned aerial vehicle is a coaxial dual-rotor unmanned aerial vehicle carrying a spectrometer.

[0060] Further, the spectrometer is a GaiaField-mini spectrum.

[0061] Further, the route path of the unmanned aerial vehicle from the starting point to the ending point is generated by using an A* algorithm to generate a collection path of the unmanned aerial vehicle from the starting point to the ending point.

[0062] Further, the starting point is the position where the unmanned aerial vehicle takes off in the area to be monitored, the ending point is the position where the unmanned aerial vehicle flies in the area to be monitored, and the obstacle position information is the position of the obstacle in the greenery engineering area to be maintained, and the unmanned aerial vehicle needs to maintain a preset obstacle safety distance (5 m) from the obstacle position.

[0063] Further, in S200, the specific method of generating the first image according to the obtained plurality of hyperspectral images is as follows: splicing all the multispectral images in the order of collection time to generate the first image.

[0064] Further, the splicing process comprises the steps of: geometric correction, image preprocessing, image registration and image fusion.

[0065] Further, in S300, the specific method of marking the pest area in the first image and updating the second image is as follows: marking a plurality of pest areas PA in the first image by RX algorithm, taking i as the serial number of the pest area, and PA iRepresents the i-th pest area, and all pest areas constitute a set INP; use the edge algorithm to divide the first image into multiple sub-areas RA j , with j as the sub-region number, RA j represents the jth sub-region;

[0066] Let the geometric center of all pest-infested areas be P(i), and the geometric center of all sub-areas be R(i). Traverse all pest-infested areas in the set INP and calculate the distance PR(j) from the geometric center P(i) of all pest-infested areas to the geometric center R(i) of all projections. Add all projections whose PR(j) is less than PR_mean to the list Listi, where PR_mean is the average distance between point P(i) and all points R(i).

[0067] Furthermore, the pest area is an area composed of pest and disease pixel points, or an internal area of ​​an edge composed of pest and disease pixel points; wherein the pest and disease pixel points are corresponding pixel points in the spectral image whose visible light reflectance is higher than the average value of the visible light reflectance of each pixel point in the spectral image, or the pest and disease pixel points are corresponding pixel points in the spectral image whose near-infrared band spectral reflectance is lower than the average value of the near-infrared band spectral reflectance of each pixel point in the spectral image.

[0068] Within the value range of k, traverse all sub-areas in Listi in turn and calculate the pest area PA by the formula i The diffusion coefficient Diff in each sub-region in Listi is calculated as follows: the average pixel value contained in each sub-region is calculated in turn as GreenK_Mean, the maximum pixel value in the k-th sub-region is recorded as GreenK_Max, the minimum pixel value is recorded as GreenK_Min, and the pixel value of the edge point of the sub-region is recorded as Green_edg(k,r), where K represents the sequence number of the sub-region included in Listi, r represents the sequence number of the edge point on the edge, and Green_edg(k,r) represents the pixel value of the r-th edge point on the edge of the k-th sub-region. By the formula: Calculate the diffusion coefficient Diff, where q(k) is the distance between the pixel point with the maximum pixel value and the pixel point with the minimum pixel value in the kth sub-region, and PR(k) represents the pest area PA i The distance between the geometric center of and the geometric center of the sub-region;

[0069] If Diff is greater than Green_Mean, where Green_Mean is the average value of pixels in all sub-regions included in List1, it is considered that the current pest region has a diffusion phenomenon, and the current pest region needs to be corrected, otherwise the current pest region does not have a diffusion phenomenon, and the current pest region does not need to be corrected;

[0070] Further, the specific correction method is: within the value range of k, the diffusion coefficient Diff of the pest region in each sub-region in Listi is calculated in turn k , where k=1, 2, 3…m, m represents the number of sub-regions included in Listi, all sub-regions included in Listi are rearranged in the order of Diff from small to large, and the diffusion coefficient Diff of the sub-region in the updated Listi is traversed and updated in turn k If Diff k -Diff k-1 <Diff k+1 -Diff k And Diff k >Diff mean , where Diff mean represents the average value of the diffusion coefficient Diff of the sub-region in Listi, then Diff k The corresponding sub-region is recorded as a strong influence sub-region,

[0071] The pixels on the edge of the pest region PA i are recorded as PL, the minimum value in the pixel value is selected as the starting point ST, and the maximum value in the pixel value is selected as the end point ED, and the PA i is retrieved from ST to ED along the minimum path; If the size of the pixel value PL on the edge satisfies and , the pixel point corresponding to PLs is updated as the starting point ST; and the PA i is retrieved from ED to ST along the minimum path; If the size of the pixel value PL on the edge satisfies and , the pixel point corresponding to PLs is updated as the starting point ED, and z represents the number of pixel points on the edge of the pest region PA i ;

[0072] The position of the pixel point corresponding to ST is recorded as point A, the position of the pixel point corresponding to ED is recorded as point B, the minimum curve segment between point A and point B is recorded as a deletion segment and is deleted, the pixel values on the edges of the sub-regions are sequentially influenced, the point corresponding to the maximum pixel value is recorded as point C, the point corresponding to the minimum pixel value is recorded as point D, the minimum curve segment between point C and point D is recorded as a replacement segment, the replacement segment is copied and is scaled to the segment between point A and point B to correct the pest damage area PA i ;

[0073] According to the above method, all pest damage areas in the first image are corrected, and the first image is updated to the second image.

[0074] Each pest damage area in the second image is divided into a training set and a validation set, specifically: each pest damage area is labeled as 1 as a positive sample, and the gray pest damage area with a diffusion coefficient greater than the average value of the pixels of the second image is selected as a negative sample and labeled as 0, thereby completing the labeling of the samples; the labeled positive samples and negative samples constitute a training sample set; the training sample set is divided into a training set and a validation set according to a ratio of 4:1, the training set is used to train the decision tree model, and the validation set is used to verify the prediction performance of the trained decision tree model.

[0075] Further, the decision tree model is a GBDT model.

[0076] The training set and the validation set are used to train the decision tree model to obtain a pre-trained decision tree model, wherein the method for training the decision tree model using the training set and the validation set is: extracting features of the training set and inputting the decision tree model, optimizing hyperparameters in the decision tree model using a grid search method, retraining the decision tree model according to the optimized hyperparameters, and obtaining a trained initialization decision tree model.

[0077] Further, the specific method for identifying the maintenance site in the to-be-maintained green engineering area map by the initialization decision tree model in S500 is:

[0078] A hyperspectral image of the to-be-maintained green engineering area map is obtained by a drone carrying a spectrometer, and a plurality of pest damage areas are identified by an RX algorithm, and the initialization decision tree model is used to screen the positive sample areas in the plurality of pest damage areas, if the positive sample areas can be screened, it is judged that the to-be-maintained green engineering area map has a diffusion pest damage area, the geometric centers of all diffusion pest damage areas are taken as the maintenance sites, and all maintenance sites are generated by curve fitting to generate a pesticide application path.

[0079] In addition, the application also provides an embodiment of a highway green engineering maintenance system based on machine learning, such as Figure 2As shown is a structure diagram of a highway greening engineering maintenance system based on machine learning. The highway greening engineering maintenance system based on machine learning comprises a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above embodiment of the highway greening engineering maintenance system based on machine learning are implemented.

[0080] A path generation unit is configured to acquire a greening engineering region to be maintained and generate a collection path.

[0081] An image collection unit is configured to control the UAV to fly along the collection path and acquire a plurality of hyperspectral images of the region.

[0082] An image splicing unit is configured to generate a first image according to the plurality of acquired hyperspectral images.

[0083] A pest marking unit is configured to mark a pest region in the first image.

[0084] A coefficient calculation unit is configured to calculate a diffusion coefficient of all the marked pest regions.

[0085] An image correction unit is configured to correct the first image according to the calculated diffusion coefficient to obtain a second image.

[0086] A model training unit is configured to train a decision tree model according to the second image to obtain an initialized decision tree model.

[0087] The highway greening engineering maintenance system based on machine learning can be run on a desktop computer, a notebook computer, a palm computer, a cloud server, and other computing devices. The highway greening engineering maintenance system based on machine learning can be run on a system that can include, but is not limited to, a processor and a memory. Those skilled in the art can understand that the example is only an example of the highway greening engineering maintenance system based on machine learning and does not constitute a limitation on the highway greening engineering maintenance system based on machine learning. It can include more or fewer components, or combine certain components, or different components, for example, the highway greening engineering maintenance system based on machine learning can also include an input / output device, a network access device, a bus, and the like.

[0088] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware component, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The processor is a control center of the machine learning-based highway greening engineering maintenance system running system, and is connected with each part of the machine learning-based highway greening engineering maintenance system running system through various interfaces and lines.

[0089] The memory can be used to store the computer program and / or module, and the processor realizes various functions of the machine learning-based highway greening engineering maintenance system by running or executing the computer program and / or module stored in the memory, and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store an operating system, at least one application program required by a function (such as a sound playing function, an image playing function, etc.), etc. The data storage area can store data created according to the use of the mobile phone (such as audio data, a phone book, etc.), etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, for example, a hard disk, a memory, a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0090] Although the description of the present application has been quite detailed and particularly described with respect to several embodiments, it is not intended to be limited to any of these details or embodiments or any special embodiment, so as to effectively cover the intended scope of the present application. In addition, the present application is described above in the embodiments that the inventors can foresee, and the purpose is to provide a useful description, and non-essential modifications to the present application that have not yet been foreseen can still represent equivalent modifications of the present application.

Claims

1. A highway greening project maintenance method based on machine learning, characterized in that: The method comprises the following steps: S100, obtaining a map of the greening project area to be maintained and generating a collection path, and flying a drone along the collection path to obtain multiple hyperspectral images of the area map; S200, generating a first image based on the acquired multiple hyperspectral images; S300, marking the pest-infested area in the first image and updating it into the second image; S400: Initialize the decision tree model obtained by training the decision tree model with the second image.

2. A highway greening project maintenance method based on machine learning according to claim 1, characterized in that: In S100, a map of the greening project area to be maintained is obtained and a collection path is generated. The method for flying a drone along the collection path to obtain multiple hyperspectral images of the area map is as follows: a map of the greening project area to be maintained is obtained, the starting point, end point and obstacle position information of the drone's flight in the greening project area to be maintained are set, a flight path of the drone from the starting point to the end point is generated, and collection points are generated at fixed distances. The drone flies along the path and collects spectral images at the collection points through the onboard spectrometer.

3. The highway greening project maintenance method based on machine learning according to claim 2 is characterized in that: In S200 , the method for generating the first image based on the acquired multiple hyperspectral images is: all the acquired multispectral images are spliced ​​in the order of acquisition time to generate the first image.

4. The highway greening project maintenance method based on machine learning according to claim 3 is characterized in that: In S300, the method for marking the pest areas in the first image and updating the second image is as follows: the first image is marked with a plurality of pest areas PA by the RX algorithm, i is used as the serial number of the pest area, PA i represents the i-th pest area, and all pest areas constitute a set INP; use the edge algorithm to divide the first image into multiple sub-areas RAj, with j as the sub-area sequence number, and RAj represents the j-th sub-area; Let the geometric center of all pest-infested areas be P(i), and the geometric center of all sub-areas be R(i). Traverse all pest-infested areas in the set INP and calculate the distance PR(j) from the geometric center P(i) of all pest-infested areas to the geometric center R(i) of all projections. Add all projections whose PR(j) is less than PR_mean to the list Listi, where PR_mean is the average distance between point P(i) and all points R(i). Within the value range of k, traverse all sub-areas in Listi in turn and calculate the pest area PA by the formula i The diffusion coefficient Diff in each sub-region in Listi is calculated as follows: the average pixel value contained in each sub-region is calculated in turn as GreenK_Mean, the maximum pixel value in the k-th sub-region is recorded as GreenK_Max, the minimum pixel value is recorded as GreenK_Min, and the pixel value of the edge point of the sub-region is recorded as Green_edg(k,r), where K represents the sequence number of the sub-region contained in Listi, r represents the sequence number of the edge point on the edge, Green_edg(k,r) represents the pixel value of the r-th edge point on the edge of the k-th sub-region, and the diffusion coefficient Diff is calculated by the formula; if Diff is greater than Green_Mean, where Green_Mean is the average value of the pixels in all sub-regions contained in List1, it is considered that the current pest area has diffused and needs to be corrected; otherwise, the current pest area has not diffused and does not need to be corrected; The specific correction method is: within the value range of k, calculate the diffusion coefficient Diff of the pest area in each sub-area in Listi in turn k , where k = 1, 2, 3…m, m represents the number of sub-regions contained in Listi, all sub-regions contained in Listi are rearranged in the order of Diff from small to large, and the diffusion coefficients Diff of the sub-regions in Listi after the update are traversed in turn k , if satisfied, Diff k -Diff k-1 <Diff k+1 -Diff k And Diff k >Diff mean , where Diff mean Represents the average diffusion coefficient Diff of the sub-regions in Listi, then Diff k The corresponding sub-region is recorded as a strong influence sub-region. Traverse the pest area PA in turn i The pixel value on the edge is recorded as PL, the minimum pixel value is selected as the starting point ST, the maximum pixel value is selected as the end point ED, and PA is retrieved from ST along the minimum path to ED. i The size of the pixel value PL on the edge, with s and q as the sequence number of the pixel value on the edge, if it satisfies and When PLs is updated to the starting point ST, the pixel point corresponding to PLs is retrieved from ED along the minimum path to ST. i The size of the pixel value PL on the edge is s as the sequence number of the pixel value on the edge. If and When PLs is updated to the starting point ED, z represents PA i The number of pixels on the edge; The position of the pixel corresponding to ST is recorded as point A, the position of the pixel corresponding to ED is recorded as point B, the minimum curve segment between point A and point B is recorded as the deletion segment and deleted, and the pixel values ​​on the edge of the sub-region are strongly affected in turn. The point corresponding to the maximum pixel value is recorded as point C, and the point corresponding to the minimum pixel value is recorded as point D. The minimum curve segment between point C and point D is recorded as the replacement segment, and the replacement segment is copied and scaled proportionally to the corrected insect infestation area PA between point A and point B. i ; According to the above method, all the pest areas in the first image are corrected and the first image is updated to be the second image.

5. The highway greening project maintenance method based on machine learning according to claim 4 is characterized in that: In S400, the method for initializing the decision tree model obtained by training the decision tree model using the second image is as follows: each pest-infested area in the second image is divided into a training set and a validation set, specifically: each pest-infested area is labeled as 1 as a positive sample, and gray pest-infested areas with a diffusion coefficient greater than the average pixel value of the second image are screened out as negative samples and labeled as 0, thereby completing the labeling of samples; the labeled positive samples and negative samples constitute a training sample set; the training sample set is divided into a training set and a validation set in a ratio of 4:1, the training set is used to train the decision tree model, and the validation set is used to verify the prediction performance of the trained decision tree model; The decision tree model is trained using the training set and the validation set to obtain a pre-trained decision tree model, wherein the method of training the decision tree model using the training set and the validation set is: extracting the features of the training set and inputting them into the decision tree model, optimizing the hyperparameters in the decision tree model using the grid search method, and retraining the decision tree model according to the optimized hyperparameters to obtain a trained initialized decision tree model.

6. A highway greening project maintenance system based on machine learning, characterized in that: The highway greening project maintenance system based on machine learning includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the highway greening project maintenance method based on machine learning described in any one of claims 1 to 5 are implemented.

Images