Tobacco quality evaluation system based on unmanned aerial vehicle multispectral camera recognition

Abstract

The invention discloses an unmanned aerial vehicle multispectral camera recognition tobacco quality evaluation system, and relates to the technical field of tobacco leaf grading. The data acquisition module is used for synchronously acquiring spectrum, temperature and three-dimensional structure information of a tobacco canopy, and processing and fusing the spectrum, temperature and three-dimensional structure information to generate a multi-dimensional data cube; the feature extraction module is used for extracting a classic vegetation index according to the multi-dimensional data cube, and constructing to obtain a space-spectrum feature; and the inversion decision module is used for obtaining a total sugar predicted value, a nicotine predicted value and a total nitrogen content predicted value through the spatial-spectral characteristics, and determining a corresponding quality grade through a set grade knowledge base. According to the method, data are collected at different time points, a time sequence data cube is constructed, and dynamic growth information such as the growth speed and the yellowing rate of tobacco is analyzed, so that final quality evaluation is performed, a decision basis of the whole growth period can be provided for accurate operation, and process management is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of tobacco grading, in particular to a UAV multi-spectral camera recognition tobacco quality evaluation system. BACKGROUND

[0002] Tobacco is an important economic crop, and its quality is directly related to the product quality and market value of the cigarette industry. Traditional tobacco quality evaluation mainly relies on manual field observation and laboratory physical and chemical analysis. Field evaluation is usually based on visual indicators such as leaf color, size, pest and disease conditions, which is highly subjective and inefficient, and difficult to cover large planting areas. Laboratory analysis can provide more accurate chemical composition data (such as total sugar, total nitrogen, nicotine content, etc.), but the sampling process is destructive and the detection period is long, which cannot realize dynamic and real-time monitoring of the whole growth process of tobacco. With the development of large-scale and intensive tobacco planting, traditional methods have been difficult to meet the needs of modern agriculture for precision and efficient management.

[0003] The Chinese invention patent application with publication number CN114757936A discloses an automatic tobacco grading system based on image and infrared spectrum chemical component analysis, including the following steps: sequentially collecting tobacco information by industrial cameras and infrared spectrometers; extracting tobacco data features according to image and infrared spectrum data, extracting shape, color, chroma, length, and damage features from tobacco front images, extracting leaf structure features and identity features from tobacco back images, and extracting chemical component features, tobacco maturity, and oil content from infrared spectrum analysis; inputting the extracted data features into a multi-layer perception machine, calculating the probability of each grade that the tobacco belongs to, and taking the highest probability corresponding grade as the predicted grade of the tobacco. This invention combines image and infrared spectrum to analyze the appearance quality and internal quality of tobacco, improving the tobacco feature extraction capability and grading accuracy.

[0004] The above technical solution can objectively and accurately detect the internal chemical components that determine the core quality of tobacco, improving the accuracy and scientificity of grading. However, it only detects the chemical components of single tobacco leaves, which cannot quickly obtain the quality distribution map of the whole tobacco field, nor can it guide precise agricultural operations. SUMMARY

[0005] The present application aims to provide a UAV multi-spectral camera recognition tobacco quality evaluation system to solve the problems raised in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solution: a UAV multi-spectral camera recognition tobacco quality evaluation system, comprising: The data acquisition module synchronously acquires the spectrum, temperature and three-dimensional structure information of the tobacco canopy, and processes and fuses the spectrum, temperature and three-dimensional structure information to generate a multi-dimensional data cube. The feature extraction module extracts a classic vegetation index according to the multi-dimensional data cube to construct a spatial-spectral feature. The inversion decision module obtains total sugar, nicotine and total nitrogen content prediction values through the spatial-spectral feature, and determines the corresponding quality grade through a set rating knowledge base, including: SC1: Chemometric inversion: training a machine learning regression model through a sample data set to obtain a chemometric inversion model, and inputting the spatial-spectral feature into the chemometric inversion model to output the corresponding total sugar content, nicotine content and total nitrogen content; SC2: Data comparison: determining the sugar-nicotine ratio according to the total sugar content and nicotine content, comparing the total sugar content, nicotine content, total nitrogen content and sugar-nicotine ratio with the set rating knowledge base, and determining the corresponding quality grade according to the comparison result; SC3: Data display: displaying different quality grades through different colors to obtain a quality grade distribution map.

[0007] Further, the multi-dimensional data cube is generated, including: SA1: Image acquisition: acquiring spectral data, temperature data and three-dimensional structure data of a tobacco field in real time through a drone provided with a multispectral camera, a thermal infrared camera and a laser radar camera; SA2: Data processing: processing the spectral data, temperature data and three-dimensional structure data through radiation calibration, atmospheric correction and image registration to obtain preprocessed sensor data, and dividing the preprocessed sensor data into ground points and non-ground points through a point cloud processing algorithm, obtaining a digital elevation map according to the ground points, and obtaining a digital surface map according to the ground points and non-ground points; SA3: Data fusion: combining and processing multispectral data images, canopy height images and temperature data images through a created multi-band raster file to obtain a multi-dimensional data cube.

[0008] Further, the multispectral camera is provided with five optical sensors and filters, and each optical sensor receives corresponding narrow-band light, including blue band, green band, red band, red edge band and near-infrared band.

[0009] Further, according to the ground elevation corresponding to each pixel point in the digital elevation map and the object top elevation corresponding to each pixel point in the digital surface model, the object itself net height corresponding to each pixel point is determined, and a canopy height map is obtained.

[0010] Further, the spatial-spectral feature is constructed, including: SB1: vegetation index extraction: according to the red light reflectivity, red edge light reflectivity and near-infrared light reflectivity in the multi-dimensional data cube, the corresponding normalized difference vegetation index, normalized difference red edge index and optimized soil-adjusted vegetation index are obtained; SB2: spatial-spectral feature construction: the normalized difference vegetation index, the normalized difference red edge index, the optimized soil-adjusted vegetation index and the feature data in the multi-dimensional data cube are combined through the multi-band raster file to construct a final feature cube.

[0011] Further, the final feature cube includes a feature cube with 10 channels, and the multi-dimensional data cube includes a feature cube with 7 channels.

[0012] Further, according to the final feature cube at different time points, a time series data cube is constructed, and a corresponding decision result is determined, including: W1: curve extraction: according to the set remote sensing monitoring time window, the spectral data at each time point is collected and obtained, and the normalized difference vegetation index and the normalized difference red edge index at each time point are determined, and the time series feature curve is obtained according to the normalized difference vegetation index and the normalized difference red edge index; W2: analysis and decision: the growth speed, growth intensity, growth cycle, yellowing rate and whole period vigor are determined through the time series feature curve, a feature matrix is constructed, and the feature matrix is taken as the input of a machine learning model to output the probability of different quality grades.

[0013] Further, the total sugar content, nicotine content and total nitrogen content of tobacco leaves in the tobacco field to be monitored are obtained by a chemical analysis method, and the sampling point feature vector is determined from the spatial-spectral feature according to the sampling point GPS coordinates of the tobacco leaves, and the sample data set is determined according to the total sugar content, nicotine content, total nitrogen content and sampling point feature vector.

[0014] Compared with the prior art, the beneficial effects of the present application are: Firstly, the present application collects data at different time points to construct a time series data cube, so as to analyze the dynamic growth vigor information such as the growth speed and the yellowing rate of tobacco, thereby performing final quality evaluation, and further providing a decision basis for precise operation throughout the whole growth period, realizing process management; Secondly, the present application can overcome the shortcomings of strong subjectivity and low efficiency of traditional manual field observation by remote sensing monitoring by the unmanned aerial vehicle, and can quickly and non-destructively obtain detailed data of the whole tobacco field, so as to realize full coverage evaluation and improve the evaluation efficiency. Thirdly, the present application can accurately predict the key internal chemical components that determine the core quality of tobacco by machine learning model through the use of multi-dimensional information such as spectrum, temperature and three-dimensional structure, so that the evaluation result is more scientific. Fourthly, the present application directly and powerfully supports the precise harvesting, zoned processing and final commercial pricing of tobacco by intuitively displaying the quality grades of different areas of the tobacco field with different colors, so that the user can quickly grasp the overall quality difference. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a system block diagram of the tobacco quality evaluation system in the present application; Figure 2 is a flowchart of the tobacco quality evaluation system in the present application; Figure 3 is a flowchart of the acquisition of the multi-dimensional data cube in the present application; Figure 4 is a flowchart of the determination of the quality grade in the present application; Figure 5 is a total sugar content distribution map in the present application; Figure 6 is a tobacco quality grade distribution map in the present application. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0017] The existing tobacco quality identification and evaluation method mostly relies on an ordinary RGB camera for aerial photography. Since only the color and shape of the tobacco leaf can be obtained, the key chemical components such as total sugar, nicotine and total nitrogen that determine the tobacco rating and industrial usability cannot be accurately detected, so that the evaluation result of the tobacco leaf only stays on the surface level and cannot truly reflect the core quality of the tobacco leaf. The technical scheme of the present application synchronously collects the spectral, temperature and three-dimensional structure information of the tobacco canopy by the unmanned aerial vehicle equipped with a multispectral camera, a thermal infrared camera and a laser radar, and fuses to generate a multidimensional data cube. Meanwhile, the corresponding classical vegetation index is extracted from the multidimensional data cube, and a spectral-spatial feature is constructed. Meanwhile, the spectral-spatial feature is taken as the input of the machine learning model to inversely predict the key chemical components of the tobacco such as the total sugar, nicotine and total nitrogen content according to the spectral-spatial feature, and determine the quality grade through the set rating knowledge base to generate a quality grade distribution map, so that the rapid, non-destructive and accurate quality evaluation of the large-area tobacco field is realized.

[0018] Embodiment 1 Reference Figures 1-6 The present embodiment provides a kind of unmanned aerial vehicle multispectral camera identification tobacco quality evaluation system, which includes data acquisition module, feature extraction module and inversion decision module. Wherein data acquisition module is used to be acquired by the unmanned aerial vehicle with multiple sensors, the spectral, temperature and three-dimensional structure information of tobacco canopy. Feature extraction module is used to extract the corresponding high-dimensional feature vector from the multidimensional data cube of spectral, temperature and three-dimensional structure information. Inversion decision module is used to identify high-dimensional feature vector by machine learning regression model, and obtain the total sugar, nicotine and total nitrogen content prediction value of each pixel point, while determining the corresponding tobacco quality grade according to the content prediction value of each parameter.

[0019] In the present embodiment, data acquisition module is used to collect the spectral, temperature and three-dimensional structure information of tobacco canopy by the unmanned aerial vehicle equipped with multispectral camera, thermal infrared camera and laser radar camera, and process and fuse the collected spectral, temperature and three-dimensional structure information to generate corresponding multidimensional data cube. Specifically as follows: Step SA1: image acquisition. That is, multispectral camera, thermal infrared camera and laser radar camera are deployed on the same unmanned aerial vehicle to collect spectral data, temperature data and three-dimensional structure data of tobacco field in real time during monitoring tobacco field by unmanned aerial vehicle.

[0020] In the embodiment, the flight monitoring route of the unmanned aerial vehicle is set according to the area range of the tobacco field to be monitored. That is, the unmanned aerial vehicle is used to monitor the tobacco field to be monitored in real time according to the flight monitoring route. Meanwhile, in the monitoring process, the GNSS / PTK positioning module is used to record the three-dimensional coordinates of the multispectral camera center point corresponding to the sensor data, and the IMU module is used to record the pitch angle, roll angle and yaw angle of the multispectral camera corresponding to the sensor data. That is, the unmanned aerial vehicle in the embodiment is also provided with a GNSS / PTK positioning module and an IMU module to collect the geographic position information and attitude information of each image. That is, the metadata corresponding to each sensor data collected includes but is not limited to GNSS position (such as longitude, latitude and altitude), IMU attitude (such as pitch angle, roll angle and yaw angle), time stamp and image data.

[0021] Further, the unmanned aerial vehicle in the embodiment is provided with a multispectral camera, a thermal infrared camera and a laser radar, and in the process of collecting sensor data, the GNSS / PTK positioning module records the three-dimensional coordinates of the multispectral camera center point corresponding to the data, and the IMU module records the pitch angle, roll angle and yaw angle of the multispectral camera corresponding to the data. That is, the unmanned aerial vehicle in the embodiment is also provided with a GNSS / PTK positioning module and an IMU module to collect the geographic position information and attitude information of each image. That is, the metadata corresponding to each sensor data collected includes but is not limited to GNSS position (such as longitude, latitude and altitude), IMU attitude (such as pitch angle, roll angle and yaw angle), time stamp and image data.

[0022] It is worth noting that the multispectral camera in the embodiment is provided with five independent optical sensors and filters, and in the process of collecting image data, each optical sensor only receives corresponding narrow-band light. Specifically, the multispectral camera in the embodiment collects five different waveband gray scale images, including but not limited to blue waveband (pixel value is 85), green waveband (pixel value is 105), red waveband (pixel value is 62), red edge waveband (pixel value is 215) and near-infrared waveband (pixel value is 780).

[0023] Step SA2: data processing. That is, according to the sensor data collected in step SA1, the original digital quantization value of the sensor data is converted into the corresponding ground reflectivity through radiation calibration, the influence of atmospheric molecules and sols is removed through an atmospheric transmission model, and the images obtained by different wavebands and different sensors in the same area are pixel-level aligned based on the positioning and attitude data to form a data cube containing multi-dimensional information for each pixel. That is, after the obtained sensor data is subjected to radiation calibration, atmospheric correction and image registration processing, the preprocessed sensor data is obtained.

[0024] Furthermore, point cloud processing algorithms such as LASTools and TerraSolid are used to classify the preprocessed point cloud data into ground points and non-ground points. Simultaneously, based on all the identified ground points, interpolation algorithms are used to obtain the corresponding digital elevation maps. Based on the identified ground and non-ground points, interpolation algorithms are used to obtain the corresponding digital surface maps.

[0025] In this embodiment, the net height of the object corresponding to each pixel is determined based on the ground elevation corresponding to each pixel in the digital elevation map and the top elevation of the object corresponding to each pixel in the digital surface map. Specifically: in: The net height of the object itself. The elevation of the top of the object. This refers to the ground elevation.

[0026] To elaborate further, the corresponding canopy height map is obtained based on the net height of the object itself corresponding to each pixel. In other words, each pixel in the canopy height map corresponds to the net height of the object itself.

[0027] Step SA3: Data Fusion. This involves using GIS software to unify the spatial reference and resolution of the acquired multispectral data images, canopy height images, and temperature data images, placing them in the same geographic coordinate system and projection method, thus ensuring completely consistent pixel size and row and column counts.

[0028] In this embodiment, a multi-band raster file is created using selected software such as ENVI or QGIS. The created multi-band raster file includes seven bands, which are used to combine the processed multispectral data image, canopy height image, and temperature data image to obtain the corresponding multidimensional data cube.

[0029] Furthermore, the multispectral data image in this embodiment includes blue light reflectance, green light reflectance, red light reflectance, red-edge light reflectance, and near-infrared light reflectance; the canopy height image includes the net height of the object itself, i.e., the canopy height; and the temperature data image includes the canopy temperature. In other words, blue light reflectance, green light reflectance, red light reflectance, red-edge light reflectance, near-infrared light reflectance, canopy height, and canopy temperature are stored in different bands of a multi-band raster file, thereby obtaining the corresponding multidimensional data cube.

[0030] In the embodiment, the feature extraction module is configured to extract the corresponding classic vegetation index from the multi-dimensional data cube in the data acquisition module, and construct the corresponding spatial-spectral feature according to the extracted classic vegetation index. Specifically, Step SB1: vegetation index extraction. That is, according to the red light reflectivity, red edge light reflectivity and near-infrared light reflectivity in the multi-dimensional data cube, the corresponding normalized difference vegetation index, normalized difference red edge index and optimized soil-adjusted vegetation index are obtained, specifically: wherein: is the normalized difference vegetation index, is the near-infrared light reflectivity, is the red light reflectivity, is the normalized difference red edge index, is the red edge light reflectivity, is the optimized soil-adjusted vegetation index.

[0031] In the process of specific implementation, the red light reflectivity is 3.5%, the red edge light reflectivity is 8.75%, and the near-infrared light reflectivity is 45%. The corresponding normalized difference vegetation index is 0.86, the corresponding normalized difference red edge index is 0.67, and the corresponding optimized soil-adjusted vegetation index is 0.75.

[0032] Step SB2: spatial-spectral feature construction. That is, the multi-dimensional data cube in the data acquisition module is combined with the normalized difference vegetation index, normalized difference red edge index and optimized soil-adjusted vegetation index obtained in step SB1 to construct the corresponding ultimate feature cube. Specifically, a multi-band raster file is created through a selected software such as ENVI or QGIS, and the multi-band raster file includes 10 bands, i.e., a feature cube containing 10 channels is constructed.

[0033] Further, in the multi-band raster file containing 10 bands, the blue light reflectivity, green light reflectivity, red light reflectivity, red edge light reflectivity, near-infrared light reflectivity, normalized difference vegetation index, normalized difference red edge index, optimized soil-adjusted vegetation index, canopy height and canopy temperature are sequentially saved wave by wave, thereby obtaining the corresponding ultimate feature cube.

[0034] In the embodiment, the inversion decision module is configured to identify the high-dimensional feature vector by a machine learning regression model to obtain corresponding total sugar prediction value, nicotine prediction value and total nitrogen content prediction value, generate corresponding chemical composition spatial distribution thematic map according to the obtained total sugar prediction value, nicotine prediction value and total nitrogen content prediction value, and perform comprehensive calculation on all chemical composition prediction values by setting a rating knowledge base, and determine the corresponding quality grade according to the obtained chemical composition comprehensive result. Specifically as follows: Step SC1: Chemical metrology inversion. That is, ground sampling is performed in the tobacco field to be monitored, and tobacco leaves in the tobacco field to be monitored are randomly sampled, and the sampled tobacco leaves are subjected to chemical analysis to determine the total sugar content, nicotine content and total nitrogen content. Further, according to the GPS coordinates of the sampling points, the feature vectors corresponding to the GPS coordinates of the sampling points are extracted from the ultimate feature cube of the feature extraction module. That is, the feature vectors corresponding to the GPS coordinates of the sampling points, the determined total sugar content, nicotine content and total nitrogen content are the sample data set.

[0035] In the embodiment, the machine learning regression model (such as a convolutional neural network model) is trained according to the obtained sample data set to obtain a corresponding chemical metrology inversion model. That is, the feature vectors corresponding to the GPS coordinates of the sampling points are taken as the input of the machine learning regression model, and the determined total sugar content, nicotine content and total nitrogen content are taken as the output of the machine learning regression model, so as to obtain the corresponding chemical metrology inversion model. Further, the feature data in the ultimate feature cube is taken as the input of the chemical metrology inversion model, and the total sugar content, nicotine content and total nitrogen content corresponding to each ultimate feature cube are obtained. Reference Figure 5 , wherein the minimum value of the total sugar content is 16.11%, the maximum value is 23.46%, and the average value is 19.43%.

[0036] Step SC2: Data comparison. That is, according to the total sugar content and nicotine content obtained in step SC1, the size of the sugar-nicotine ratio is determined, specifically as follows: , wherein: is the sugar-nicotine ratio, is the total sugar content, is the nicotine content.

[0037] Further, the total sugar content, nicotine content, total nitrogen content and sugar-nicotine ratio obtained in step SC1 are compared with the data in the set rating knowledge base, and the corresponding quality grade is determined according to the comparison result. It should be noted that the rating knowledge base in this embodiment is specifically set according to the tobacco purchase standard or agricultural specification digitalization, so it can be specifically set according to the actual data requirement, so only an example is given in this embodiment.

[0038] In the process of specific implementation, the quality grades corresponding to the rating knowledge base in this embodiment include superior product, good product, medium product and poor product. The rating range corresponding to the superior product is that the total sugar content is greater than 18%, the nicotine content is between 1.5% and 2.5%, and the sugar-nicotine ratio is between 8 and 12; the rating range corresponding to the good product is that the total sugar content is greater than 15%, the nicotine content is between 1% and 3%, and the sugar-nicotine ratio is between 6 and 15; the rating range corresponding to the medium product is that the total sugar content is greater than 12%, the nicotine content is between 0.5% and 3.5%, and the sugar-nicotine ratio is between 4 and 17; otherwise, it does not satisfy the rating range of the superior product, the good product and the medium product, that is, it is the poor product.

[0039] Step SC3: data display. That is, according to the rating result corresponding to each ultimate feature cube obtained in step SC2, the quality grade corresponding to each ultimate feature cube is determined, and different quality grades are displayed through different colors. That is, according to the quality grade marked by different colors, the final detection result of the tobacco field to be monitored is obtained, that is, the corresponding quality grade distribution diagram.

[0040] Reference Figure 6 It can be seen that the pixel distribution of the poor product is 2 pixels, and the proportion is 0%; the pixel distribution of the medium product is 4854 pixels, and the proportion is 48.5%; the pixel distribution of the good product is 5144 pixels, and the proportion is 51.4%.

[0041] Embodiment 2 The embodiment provides a kind of unmanned aerial vehicle multispectral camera recognition tobacco quality evaluation system, its specific implementation method same as embodiment 1, its difference lies in, according to the ultimate feature cube obtained in feature extraction module, each feature data corresponding to different time points is determined from it, time series data cube is constructed, corresponding time series feature curve is generated, and corresponding decision result is determined. The present application is illustrated by combining the specific implementation of the embodiment.

[0042] In this embodiment, the time series data cube is constructed by the ultimate feature cube corresponding to different time points, the corresponding time series feature curve is generated, and the corresponding decision result is determined. Specifically as follows: Step W1: Curve extraction. That is, through the set remote sensing monitoring time window, the same flight parameters are used to collect spectral data in the same area range under the same meteorological conditions by the same unmanned aerial vehicle, including blue light reflectivity, green light reflectivity, red light reflectivity, red edge light reflectivity and near-infrared light reflectivity. At the same time, the spectral data collected at different times are spatiotemporally aligned at the pixel level to obtain the spatiotemporally aligned spectral data.

[0043] Further, according to the spatiotemporally aligned spectral data obtained at different times, that is, according to the red light reflectivity, red edge light reflectivity and near-infrared light reflectivity corresponding to each time point in the remote sensing monitoring time window, the normalized difference vegetation index and the normalized difference red edge index corresponding to each time point are obtained.

[0044] In this embodiment, according to the normalized difference vegetation index and the normalized difference red edge index corresponding to each time point in the remote sensing monitoring time window, the normalized difference vegetation index curve and the normalized difference red edge index curve in the remote sensing monitoring time window are obtained, that is, the corresponding time sequence characteristic curve is obtained.

[0045] Step W2: Analysis and decision. That is, according to the time sequence characteristic curve obtained in step W1, the corresponding rising slope, peak height, peak occurrence time, falling slope and curve area are determined, that is, the corresponding growth speed, growth intensity, growth period, yellowing rate and whole-period growth potential are extracted. At the same time, the extracted growth speed, growth intensity, growth period, yellowing rate and whole-period growth potential are combined to construct the corresponding feature matrix.

[0046] Further, the obtained feature matrix is taken as the input of the constructed machine learning model (such as a random forest model or a gradient boosting machine model), and the probability of obtaining the corresponding different quality grades is output.

[0047] In the process of specific implementation, the growth speed is 0.015 / day, the growth intensity is 0.75, the growth period is 55 days, the yellowing rate is -0.01 / day, and the whole-period growth potential is 45.2. Therefore, the corresponding feature matrix is [0.015, 0.75, 55, -0.01, 45.2]. That is, the feature matrix [0.015, 0.75, 55, -0.01, 45.2] is taken as the input of the machine learning model, and the probability of obtaining the corresponding different quality grades is output, that is, the probability of the quality grade being “excellent” is 80%.

[0048] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the embodiments disclosed except insofar as recited in the claims.

Claims

1. A multispectral camera-based tobacco quality assessment system for unmanned aerial vehicles (UAVs), characterized in that, Including: Data acquisition module: synchronously acquires the spectral, temperature and three-dimensional structural information of the tobacco canopy, and processes and fuses the spectral, temperature and three-dimensional structural information to generate a multi-dimensional data cube; Feature extraction module: Based on the multi-dimensional data cube, extract classic vegetation indices and construct spatial-spectral features; Inversion Decision Module: Using the aforementioned spatial-spectral features, it obtains predicted values ​​for total sugar, nicotine, and total nitrogen content. Simultaneously, through a set rating knowledge base, it determines the corresponding quality level, including: SC1: Chemometric Inversion: A chemometric inversion model is obtained by training a machine learning regression model using a sample dataset. The spatial-spectral features are used as the input to the chemometric inversion model, and the corresponding total sugar content, nicotine content, and total nitrogen content are output. SC2: Data Comparison: Based on the total sugar content and nicotine content, determine the sugar-to-nicotine ratio, and compare the total sugar content, nicotine content, total nitrogen content, and sugar-to-nicotine ratio with the established rating knowledge base. Based on the comparison results, determine the corresponding quality grade. SC3: Data shows that different quality levels are displayed using different colors to obtain a quality level distribution map.

2. The UAV multispectral camera recognition tobacco quality assessment system according to claim 1, characterized in that, Generate a multi-dimensional data cube, including: SA1: Image Acquisition: Real-time acquisition of spectral data, temperature data, and three-dimensional structural data of tobacco fields using drones equipped with multispectral cameras, thermal infrared cameras, and lidar cameras; SA2: Data Processing: The spectral data, temperature data, and three-dimensional structural data are processed through radiometric calibration, atmospheric correction, and image registration to obtain preprocessed sensor data. The preprocessed sensor data is then divided into ground points and non-ground points using a point cloud processing algorithm. A digital elevation map is obtained based on the ground points, and a digital surface map is obtained based on the ground points and non-ground points. SA3: Data Fusion: By combining multispectral data images, canopy height images, and temperature data images through the creation of multiband raster files, a multidimensional data cube is obtained.

3. The UAV multispectral camera recognition tobacco quality assessment system according to claim 2, characterized in that, The multispectral camera is equipped with five optical sensors and filters, and each optical sensor receives corresponding narrow band light, including blue band, green band, red band, red edge band and near-infrared band.

4. The UAV multispectral camera recognition tobacco quality assessment system according to claim 2, characterized in that, Based on the ground elevation corresponding to each pixel in the digital elevation map and the top elevation of the object corresponding to each pixel in the digital surface map, the net height of the object itself corresponding to each pixel is determined, and the canopy height map is obtained.

5. The UAV multispectral camera recognition tobacco quality assessment system according to claim 1, characterized in that, The constructed spatial-spectral features include: SB1: Vegetation index extraction: Based on the red light reflectance, red edge light reflectance and near-infrared light reflectance in the multi-dimensional data cube, the corresponding normalized differential vegetation index, normalized differential red edge index and optimized soil-regulated vegetation index are obtained. SB2: Spatial-spectral feature construction: By combining the feature data in the normalized differential vegetation index, normalized differential red edge index, optimized soil-regulated vegetation index and multi-dimensional data cube through multi-band raster files, the final feature cube is constructed.

6. The UAV multispectral camera recognition tobacco quality assessment system according to claim 5, characterized in that, The ultimate feature cube includes a feature cube with 10 channels, and the multi-dimensional data cube includes a feature cube with 7 channels.

7. The UAV multispectral camera recognition tobacco quality assessment system according to claim 5, characterized in that, Based on the ultimate feature cubes at different time points, a time-series data cube is constructed to determine the corresponding decision results, including: W1: Curve extraction: According to the set remote sensing monitoring time window, collect spectral data at each time point, and determine the normalized difference vegetation index and normalized difference red edge index at each time point. At the same time, obtain the time series characteristic curve based on the normalized difference vegetation index and normalized difference red edge index. W2: Analysis and Decision Making: Based on the time-series characteristic curves, the growth rate, growth intensity, growth cycle, yellowing rate, and overall growth status are determined, and a feature matrix is ​​constructed. At the same time, the feature matrix is ​​used as the input of a machine learning model, and the output is the probability of obtaining the corresponding different quality levels.

8. The UAV multispectral camera recognition tobacco quality assessment system according to claim 1, characterized in that, Chemical analysis was used to obtain the total sugar content, nicotine content, and total nitrogen content of tobacco leaves in the tobacco field to be monitored. Based on the GPS coordinates of the sampling points of the tobacco leaves, the feature vector of the sampling points was determined from the spatial-spectral characteristics. Based on the total sugar content, nicotine content, total nitrogen content, and the feature vector of the sampling points, the sample dataset was determined.

Images