Nondestructive detection device for plant chlorophyll based on multispectral imaging and machine learning

Abstract

The application relates to the technical field of plant detection, in particular to a plant chlorophyll nondestructive detection device based on multispectral imaging and machine learning, wherein the multispectral image acquisition module comprises a back-illumination CMOS sensor, an 8-channel narrowband filter wheel and an LED array light supplementing system; the data processing module sequentially normalizes and calculates received original spectral images and positions a circular detection area of a leaf, and extracts a reflectivity spectral feature vector; the model prediction module is internally provided with a trained support vector machine regression model, takes the reflectivity spectral feature vector as input, and outputs a chlorophyll content value of the circular detection area; and the display output module is used for receiving the chlorophyll content value and corresponding coordinates, and generating a chlorophyll content pseudo-color space distribution graph of the whole leaf through an interpolation algorithm. Thus, the problems of destructive detection, poor equipment portability, limited detection precision, weak model generalization, inability to realize micro-area quantification and spatial visualized imaging and the like in the prior art are solved.

Description

Technical Field

[0001] This application relates to the field of plant detection technology, and in particular to a non-destructive detection device for plant chlorophyll based on multispectral imaging and machine learning. Background Technology

[0002] Chlorophyll is the core functional pigment for photosynthesis in plants. Its content and spatial distribution characteristics are key phenotypic indicators for assessing crop growth and nutritional status, stress resistance, and varietal superiority. It has extremely high application value in agricultural research and production fields such as precise diagnosis of crop nutrition, screening of superior varieties, and research on the physiological mechanisms of plant stress. Accurately and efficiently obtaining information on the absolute content and spatial distribution of chlorophyll in plant leaves is an important prerequisite for realizing dynamic monitoring of crop growth and precise field management.

[0003] Currently, the mainstream chlorophyll detection methods in the industry mainly include three categories: spectrophotometry, high-performance liquid chromatography (HPLC), and portable SPAD chlorophyll analyzers. Traditional spectrophotometry requires destroying plant leaf samples and relying on various chemical reagents for extraction and detection, resulting in cumbersome procedures, long detection cycles, and an inability to achieve continuous dynamic monitoring of living plants. HPLC offers high detection accuracy, but its equipment purchase and maintenance costs are high, and its sample pretreatment procedures are complex and highly specialized, making it unsuitable for large-scale, high-throughput field testing scenarios. Portable SPAD chlorophyll analyzers can achieve rapid detection of living plants in the field, but they can only output dimensionless relative detection values. The results are easily affected by parameters such as leaf thickness, moisture content, and texture, resulting in limited stability and accuracy. Furthermore, they can only collect data from a single point on the leaf, failing to characterize the spatial heterogeneity of chlorophyll distribution on the leaf surface, thus failing to meet the needs of refined detection.

[0004] In recent years, multispectral and hyperspectral imaging technologies have been gradually applied to the field of plant phenotypic parameter detection due to their advantages of non-destructive and rapid detection, effectively making up for some of the shortcomings of traditional detection methods. However, existing technical solutions still have many core defects: most of the current mainstream detection equipment are large-scale laboratory hyperspectral imaging systems, which are bulky, have poor portability, and are expensive to manufacture, making them unsuitable for in-situ, mobile, and high-throughput detection scenarios in the field; at the same time, the analysis and calculation of existing spectral detection data highly depend on complex physical mechanism models or fixed empirical statistical models, resulting in poor algorithm adaptability and low computational efficiency.

[0005] In addition, existing chlorophyll spectral prediction models mostly employ traditional algorithms such as partial least squares regression (PLSR) and simple vegetation index linear fitting. These models have weak generalization ability, and their detection accuracy varies significantly across crop varieties and growth cycles, resulting in insufficient versatility. More importantly, current technologies can only achieve a rough calculation of the overall average chlorophyll content of leaves, failing to perform precise quantitative detection of chlorophyll content in millimeter-level micro-areas of leaves. They also cannot generate visualized chlorophyll spatial distribution maps, making it difficult to achieve refined and intuitive analysis of chlorophyll distribution characteristics in plant leaves. Consequently, they cannot meet the high-end demands of modern agricultural precision breeding, stress micro-area physiological research, and refined field diagnosis. Summary of the Invention

[0006] This application provides a non-destructive detection device for plant chlorophyll based on multispectral imaging and machine learning, in order to solve the problems of destructive detection, poor equipment portability, limited detection accuracy, weak model generalization, and inability to achieve micro-area quantitative and spatial visualization imaging in the prior art.

[0007] The first aspect of this application provides a non-destructive testing device for plant chlorophyll based on multispectral imaging and machine learning, comprising: a multispectral image acquisition module, a data processing module, a model prediction module, and a display output module; wherein, The multispectral image acquisition module is connected to the data processing module. The multispectral image acquisition module includes a back-illuminated CMOS sensor, an 8-channel narrowband filter wheel, and an LED array supplementary lighting system. The data processing module is connected to the model prediction module. It sequentially performs dark current correction, reflectance normalization calculation based on black / white reference plate, and locates the circular detection area of ​​the leaf to extract the reflectance spectral feature vector on the received raw spectral image. The model prediction module is connected to the display output module. The model prediction module has a built-in trained support vector machine regression model. It takes the reflectance spectral feature vector as input and outputs the chlorophyll content value of the circular detection area. The display output module is used to receive the chlorophyll content values ​​and corresponding coordinates of multiple circular detection areas output by the model prediction module, and generate a pseudo-color spatial distribution map of the chlorophyll content of the whole leaf through an interpolation algorithm.

[0008] Optionally, the center wavelengths of the 8-channel narrowband filter wheel are configured to be 450nm, 550nm, 680nm and 800nm, respectively, and the bandwidth of each channel filter is 10nm±2nm; the spectral response range of the back-illuminated CMOS sensor is 400-900nm, and the resolution is not less than 5 million pixels; the illumination uniformity of the LED array supplementary lighting system at a working distance of 30cm is not less than 90%, and the illuminance adjustment range is 500-5000 lux.

[0009] Optionally, the support vector machine regression model uses radial basis functions as kernel functions; The support vector machine regression model outputs a predicted value after Yeo-Johnson power transformation, and the predicted value is calculated through a decision function, which is constructed based on the support vector spectral features and kernel function. The model prediction module applies a Yeo-Johnson inverse transform to the predicted value to obtain the predicted chlorophyll content of the circular detection area.

[0010] Optionally, the reflectance spectral feature vector includes reflectance values ​​at center wavelengths of 450 nm, 680 nm, and 800 nm, as well as normalized vegetation index, ratio vegetation index, and red edge position parameters calculated based on the reflectance values.

[0011] Optionally, when the data processing module locates the circular detection area of ​​the blade using the Canny edge detection and Hough circle transform algorithm, the radius of the detection circle is set to 1.25mm ± 0.05mm, and the positioning accuracy is no greater than 0.05mm.

[0012] Optionally, the support vector machine regression model uses 5-fold cross-validation during the training phase, and the dataset is divided into training and validation sets in an 8:2 ratio; the coefficient of determination R² of the trained model is not less than 0.92, and the root mean square error is not greater than 0.35 mg / g.

[0013] Optionally, the multispectral image acquisition module further includes a PID closed-loop control unit, which is used to automatically adjust the driving current of the LED array supplementary lighting system according to the feedback value of the ambient light sensor in order to maintain a constant target illuminance.

[0014] Optionally, it also includes a calibration module, which automatically triggers the black / white reference board calibration process and updates the correction parameters in the data processing module each time the device is powered on or when the ambient light intensity changes beyond a preset threshold.

[0015] Optionally, the display output module generates the pseudo-color spatial distribution map using a bilinear interpolation algorithm, and uploads the chlorophyll content value and spatial distribution map to the agricultural Internet of Things management platform via a communication interface, for use in generating variable fertilization prescription maps or crop stress classification reports.

[0016] A second aspect of this application provides a method for visual non-destructive testing of plant chlorophyll, comprising the following steps: S1. Control the multispectral image acquisition module to acquire multispectral images of plant leaves; S2. Use the data processing module to perform reflectance correction and circular detection area localization on the image, and extract the reflectance spectral feature vector; S3. Input the reflectance spectral feature vector into the trained SVM regression model in the model prediction module to calculate the chlorophyll content value. S4. Visualize the chlorophyll content value as a whole leaf distribution map through the display output module and upload it to the agricultural IoT management platform.

[0017] Therefore, this application has at least the following beneficial effects: (1) This application uses multispectral imaging technology, which does not require grinding, cutting and chemical reagents. The in-situ non-destructive determination of leaves is achieved through the multispectral image acquisition module. It does not damage plant samples and is pollution-free. The samples after detection can be reused for subsequent growth monitoring or other phenotypic analysis, which overcomes the defects of traditional spectrophotometry and high performance liquid chromatography, which require destructive sampling, are cumbersome to operate and have a long cycle. (2) This application uses the Canny edge detection and Hough circle transform algorithm to accurately locate a circular detection area with a diameter of 2.5 mm (radius 1.25 mm ± 0.05 mm) on the leaf, with a positioning accuracy of ≤ 0.05 mm. It can achieve independent detection of multiple micro-areas on a single leaf. Compared with the traditional SPAD chlorophyll meter, which can only output the detection mode of a single point or average value of the leaf, this application breaks through the limitation of coarse averaging of leaf chlorophyll content. It achieves the first millimeter-level precise quantitative analysis of chlorophyll content in plant leaves, providing a new technical means for studying the nutritional heterogeneity and stress response micro-area differences within the leaf. (3) This application is based on the support vector machine regression framework, and uses kernel functions to complete nonlinear spatial mapping. It follows the principle of minimizing structural risk, adapts to the characteristics of small sample data, and avoids the problem of model overfitting. The radial basis kernel function is selected, which has the characteristics of good stability and strong nonlinear approximation ability. It can accurately adapt to the leaf spectral data characteristics of different varieties and growth stages, and consolidate the foundation of cross-species prediction algorithm. In view of the non-normal distribution problem of chlorophyll content data, Yeo-Johnson power transformation is introduced to correct the data distribution, balance the training weights of different content intervals, improve the model fitting bias problem, and the detection accuracy can reach 0.1 mg / g after the inverse transformation restores the data. Based on the model decision function, adaptive prediction is achieved by weighting the support vectors and the spectral similarity of the test samples. It can flexibly adapt to various types of sample data and finally achieve performance indicators of determination coefficient R²≥0.92 and root mean square error RMSE≤0.35mg / g. It has both excellent detection accuracy and cross-species generalization ability, and overcomes the shortcomings of traditional algorithms in terms of insufficient universality and limited detection accuracy. It meets the needs of high-throughput in-situ fine detection of chlorophyll in multiple species in the field. (4) This application integrates a multispectral image acquisition module, a data processing module, a model prediction module, and a display output module into a handheld portable housing. It has a built-in LED array supplementary lighting system and a PID closed-loop control unit, which can realize real-time detection in various environments such as fields, greenhouses, and laboratories, and is not affected by changes in ambient light. The detection data can be seamlessly connected to the agricultural Internet of Things management platform through communication interfaces (such as USB, WiFi, Bluetooth) to generate variable fertilizer prescription maps or crop stress classification reports, meeting the high-end needs of modern agriculture for precision breeding, refined field diagnosis, and high-throughput phenotypic analysis; (5) This application is equipped with a calibration module that automatically triggers the black / white reference board calibration process each time the device is powered on or when the ambient light intensity changes beyond a preset threshold. At the same time, a PID closed-loop control unit is set up to automatically adjust the driving current of the LED array supplementary lighting system according to the feedback value of the ambient light sensor to maintain a constant target illuminance. The dual mechanism ensures the consistency and repeatability of the device under different ambient light conditions.

[0018] This solves the problems of destructive detection, poor equipment portability, limited detection accuracy, weak model generalization, and inability to achieve micro-area quantitative and spatial visualization imaging in existing technologies.

[0019] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0020] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a block diagram of a non-destructive testing device for plant chlorophyll based on multispectral imaging and machine learning, provided according to an embodiment of this application. Figure 2 This is a flowchart of a non-destructive detection method for plant chlorophyll based on multispectral imaging and machine learning, according to an embodiment of this application. Detailed Implementation

[0021] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0022] The following description, with reference to the accompanying drawings, describes a plant chlorophyll non-destructive testing device based on multispectral imaging and machine learning, according to embodiments of this application. Addressing the limited detection accuracy mentioned in the background section, this application provides a plant chlorophyll non-destructive testing device based on multispectral imaging and machine learning. This device comprises a multispectral image acquisition module, a data processing module, a model prediction module, and a display output module. The multispectral image acquisition module is equipped with a back-illuminated CMOS sensor, an 8-channel narrowband filter wheel, and an LED array supplementary lighting system. It can acquire high-quality original multispectral images of leaves. Relying on a matching PID closed-loop control and automatic black-and-white board calibration mechanism, it can adaptively adjust the supplementary lighting intensity and complete reflectivity correction according to changes in ambient light, effectively avoiding interference from external light fluctuations and ensuring the stability and repeatability of image acquisition in different field, greenhouse, and laboratory settings. This application utilizes multispectral imaging technology to achieve in-situ non-destructive testing of leaves, eliminating the need for sample grinding, slicing, and chemical reagent assistance. The testing process is non-damaging and pollution-free, and the samples can be repeatedly used for subsequent growth monitoring and phenotypic analysis. This completely overcomes the shortcomings of traditional spectrophotometry and high-performance liquid chromatography, which involve destructive sampling, cumbersome operations, and long testing cycles. The data processing module sequentially performs dark current correction and black-and-white reference plate reflectance normalization preprocessing on the original spectral images. Combined with Canny edge detection and Hough circle transform algorithms, it accurately locates circular micro-areas on the leaves with a diameter of 2.5 mm and a positioning accuracy of ≤0.05 mm. Multiple micro-areas can be independently detected on a single leaf, overcoming the technical limitations of traditional SPAD chlorophyll meters, which can only obtain single-point or overall average values ​​and cannot characterize the spatial heterogeneity of chlorophyll distribution in leaves. This achieves precise positioning and feature extraction of chlorophyll content in millimeter-level micro-areas of leaves. The model prediction module incorporates a support vector machine regression machine learning model trained and optimized with multi-species samples. It uses preprocessed micro-area reflectance spectral feature vectors as input to complete intelligent predictions. Based on the SVR framework, the model adapts to small agricultural samples and nonlinear regression scenarios by relying on the structural risk minimization criterion, effectively suppressing overfitting. It is equipped with a radial basis kernel function with excellent stability and nonlinear approximation performance, adaptively adapting to the spectral data distribution characteristics of different species and different growth stages. At the same time, it introduces Yeo-Johnson power transform to correct the non-normal distribution defects of chlorophyll content data, balances the training weights between high and low content intervals, eliminates the fitting bias problem of traditional models, and combines an adaptive weighted decision function to break away from the constraints of fixed empirical formulas. Finally, it achieves an ultra-high detection accuracy of 0.1 mg / g. The overall performance of the model reaches R²≥0.92 and RMSE≤0.35 mg / g, and it has a very strong cross-species and cross-growth stage generalization prediction capability. The display output module can receive multiple micro-area chlorophyll content detection values ​​and corresponding spatial coordinates, and reconstruct and generate a pseudo-color spatial distribution map of chlorophyll content in the whole leaf through interpolation algorithms, truly realizing the micro-area quantitative detection and spatial visualization of chlorophyll content.Meanwhile, this application highly integrates each functional module into a handheld portable structure, making the device lightweight, easy to carry, and adaptable to in-situ high-throughput detection in multiple scenarios. Moreover, the detection data can be accessed to the agricultural IoT platform via communication methods such as USB, WiFi, and Bluetooth, supporting precision agricultural applications such as precision fertilization prescription mapping and crop stress classification.

[0023] The following description, with reference to the accompanying drawings, describes a non-destructive testing device for plant chlorophyll based on multispectral imaging and machine learning, according to an embodiment of this application.

[0024] Specifically, Figure 1 This is a block diagram illustrating a non-destructive testing device for plant chlorophyll based on multispectral imaging and machine learning, provided as an embodiment of this application.

[0025] like Figure 1 As shown, the plant chlorophyll non-destructive testing device 10 based on multispectral imaging and machine learning includes: a multispectral image acquisition module 100, a data processing module 200, a model prediction module 300, and a display output module 400.

[0026] The multispectral image acquisition module 100 is connected to the data processing module 200. The multispectral image acquisition module 100 includes a back-illuminated CMOS sensor, an 8-channel narrowband filter wheel, and an LED array supplementary lighting system. The data processing module 200 is connected to the model prediction module 300. It sequentially performs dark current correction, reflectance normalization calculation based on black / white reference plate, and locates the circular detection area of ​​the leaf to extract the reflectance spectral feature vector. The model prediction module 300 is connected to the display output module 400. The model prediction module 300 has a built-in trained support vector machine regression model. It takes the reflectance spectral feature vector as input and outputs the chlorophyll content value of the circular detection area. The display output module 400 is used to receive the chlorophyll content values ​​and corresponding coordinates of multiple circular detection areas output by the model prediction module 300, and generate a pseudo-color spatial distribution map of the chlorophyll content of the whole leaf through an interpolation algorithm.

[0027] It is understood that the embodiments of this application acquire high signal-to-noise ratio raw spectral images through a multispectral image acquisition module based on a back-illuminated CMOS sensor and a multi-channel narrowband filter wheel. The data processing module performs dark current correction, reflectivity normalization, and region localization to effectively remove ambient light interference and accurately extract leaf spectral features. Then, a pre-trained support vector machine regression model is used to quickly predict the chlorophyll content in the circular detection area. Finally, the display output module generates a pseudo-color spatial distribution map of the chlorophyll content of the whole leaf through interpolation. This achieves non-contact, high-throughput, and visualized high-precision detection of chlorophyll content in plant leaves, overcoming the problems of insufficient representativeness of traditional point measurements and low efficiency of destructive detection.

[0028] It should be noted that, under laboratory conditions, plant leaf samples of broad representativeness were collected, covering leaf materials from different species (such as rice, corn, wheat, soybean, etc.), different growth stages (seedling stage, jointing stage, heading stage, etc.), and different nutritional states (normal, nitrogen-deficient, potassium-deficient, etc.). The chlorophyll content at a specific location (a circular area with a diameter of 2.5 mm) on each sample was accurately determined using standard physicochemical methods (such as spectrophotometry or high-performance liquid chromatography) as a standard value, which was recorded as the measured value.

[0029] Simultaneously, the multispectral image acquisition module of this application is used to acquire reflectance spectral images of the corresponding leaf regions, with the acquisition wavelength range covering characteristic bands such as 450nm, 550nm, 680nm, and 800nm. The acquired spectral images are then preprocessed sequentially using dark current correction, reflectance normalization, and denoising, and spectral feature vectors related to chlorophyll sensitivity are extracted, including but not limited to: characteristic band reflectance values, normalized difference vegetation index (NDVI), ratio vegetation index (RVI), and red-edge position parameters.

[0030] Then, the support vector machine regression algorithm is used, with the spectral feature vectors extracted above as input and the chlorophyll content value measured by standard physicochemical methods as output. The model is trained by 5-fold cross-validation and 8:2 training and validation set partitioning. The non-normal distribution of chlorophyll content data is corrected by Yeo-Johnson power transform, and finally the optimized chlorophyll content prediction model is obtained.

[0031] Finally, the optimized model is solidified and integrated into the model prediction module. During actual detection, the device automatically executes the above preprocessing and feature extraction steps, inputs the extracted feature vectors into the integrated model, and can instantly calculate the chlorophyll content value of the circular detection area.

[0032] In this embodiment, the center wavelengths of the 8-channel narrowband filter wheel are 450nm, 550nm, 680nm and 800nm, respectively, and the bandwidth of each channel filter is 10nm±2nm; the spectral response range of the back-illuminated CMOS sensor is 400-900nm, and the resolution is not less than 5 million pixels; the illumination uniformity of the LED array supplementary lighting system at a working distance of 30cm is not less than 90%, and the illuminance adjustment range is 500-5000 lux.

[0033] It is understood that, by rationally configuring the center wavelength of the 8-channel narrowband filter wheel, this embodiment of the application can specifically capture the strong absorption peak of chlorophyll in the visible light region and the abrupt reflection characteristics in the near-infrared region, thereby providing sensitive spectral basis for biochemical parameter inversion. Combined with a narrowband bandwidth design of 10nm±2nm, it effectively suppresses environmental stray light interference and improves spectral resolution. Simultaneously, the back-illuminated CMOS sensor, with its wide-spectral high-sensitivity response in the 400-900nm range and a spatial resolution of no less than 5 million pixels, ensures the capture capability of weak spectral signals and clear imaging of leaf details. Furthermore, the LED array supplemental lighting system achieves no less than 90% illumination uniformity and an adjustable illuminance range of 500-5000 lux at a specific working distance, eliminating reflectivity errors caused by uneven illumination and ensuring the stability of sample illumination. The coordinated operation of the above hardware modules ensures high accuracy and high consistency of the input data for the subsequent chlorophyll content inversion model from the source of optical signal acquisition.

[0034] In this embodiment of the application, the support vector machine regression model uses the radial basis function as the kernel function; The support vector machine regression model outputs a predicted value after Yeo-Johnson power transformation, and the predicted value is calculated through a decision function, which is constructed based on the spectral features of support vectors and a kernel function. The model prediction module applies an inverse Yeo-Johnson transform to the predicted values ​​to obtain the predicted chlorophyll content of the circular detection area.

[0035] It should be noted that the support vector machine regression model uses radial basis functions as the kernel function and: ; Where γ = 0.1, and the penalty coefficient C = 10; The predicted value is calculated using a decision function, where the decision function is: ; in, The number of support vectors, , For Lagrange multipliers, For support vector spectral features, Let be the spectral eigenvector of the reflectance to be predicted, and b be the bias term. For kernel functions; The model prediction module applies an inverse Yeo-Johnson transform to the predicted values ​​to obtain the predicted chlorophyll content of the circular detection area, using the following formula: ; in, λ is the Yeo-Johnson transform parameter, determined by maximum likelihood estimation during the training phase; λ is the kernel function parameter, and the output range of the predicted chlorophyll content value y is 0.1 mg / g ≤ y ≤ 10 mg / g.

[0036] In this embodiment, the reflectance spectral feature vector includes reflectance values ​​at center wavelengths of 450 nm, 680 nm, and 800 nm, as well as the normalized vegetation index, ratio vegetation index, and red edge position parameters calculated based on the reflectance values.

[0037] It should be noted that the inputs to the chlorophyll content prediction model are: The reflectance values ​​at center wavelengths of 450nm, 680nm and 800nm ​​are R450, R680 and R800, respectively.

[0038] Normalized Difference Vegetation Index (NDVI): NDVI = (R800−R680) / (R800+R680) Ratio vegetation index: RVI = R800 / R680 Improved red-edge normalized difference index: mNDVI_rededge = (R800− R700) / (R800 + R700− 2·R670) The red edge position λ_red edge (unit: nm) = 700 + 40 × [(R_700 − R_670) / (R_730 − R_670)] The first derivative of reflectivity R'(λ) = [R(λ+Δλ) − R(λ−Δλ)] / (2·Δλ), Where Δλ = 10nm, and the value of λ ranges from 450 to 900nm.

[0039] In this embodiment of the application, when the data processing module locates the circular detection area of ​​the blade using the Canny edge detection and Hough circle transform algorithm, the radius of the detection circle is set to 1.25mm ± 0.05mm, and the positioning accuracy is no greater than 0.05mm.

[0040] It is understood that the embodiments of this application achieve millimeter-level precise positioning of the detection area on the blade surface through the aforementioned precise circle detection parameter settings. Specifically, the Canny edge detection algorithm first uses a 5×5 Gaussian filter kernel to smooth the reflectivity grayscale image to suppress high-frequency noise introduced during image acquisition; then, it calculates the gradient magnitude and direction of each pixel in the image, and performs double-threshold edge detection by setting a low threshold of 50 and a high threshold of 150, effectively extracting the complete contour boundary of the blade while suppressing the interference of false edges. Based on the obtained blade contour, the Hough circle transform algorithm searches for a circular detection region within the contour, with the following parameter configurations: accumulator resolution to image spatial resolution ratio dp=1, minimum distance between circle centers 5mm, Canny edge detection high threshold param1=120, circle center detection threshold param2=80, and detection radius ranging from 1.20mm to 1.30mm. The above parameter combination ensures that the algorithm can stably detect the required circular area on blades of different sizes and shapes, and multiple non-overlapping detection circles can be located on the same blade. The radius of the detection circle is strictly controlled within the range of 1.25mm±0.05mm, and the center positioning accuracy is better than 0.05mm.

[0041] The embodiments of this application effectively avoid interference from non-representative areas such as leaf veins, lesions, and leaf margin damage on spectral measurements, ensuring that the spectral reflectance data within each detection circle truly reflects the optical characteristics of the leaf mesophyll tissue. Simultaneously, the joint positioning of multiple detection circles allows for the acquisition of chlorophyll content data from dozens of independent measurement points on a single leaf, providing sufficiently dense sampling points for the subsequent generation of spatial distribution maps. This overcomes the limitations of traditional SPAD chlorophyll meters, which can only acquire single-point or whole-leaf average values, laying a data foundation for refined studies of nutrient heterogeneity within leaves and micro-regional differences in stress response.

[0042] In the embodiments of this application, the support vector machine regression model adopts 5-fold cross-validation during the training phase, and the dataset is divided into training set and validation set in an 8:2 ratio; the coefficient of determination R² of the trained model is not less than 0.92, and the root mean square error is not greater than 0.35mg / g.

[0043] Specifically, the implementation of 5-fold cross-validation is as follows: all labeled sample data is randomly divided into 5 subsets. Four subsets are selected sequentially as the training set, and the remaining subset as the validation set. This training and validation process is repeated 5 times. The model's performance on the validation set is recorded each time, and the average of the 5 results is taken as the model's overall evaluation metric. This strategy effectively utilizes limited labeled sample data, avoids evaluation bias caused by a single division of the training and validation sets, and more realistically reflects the model's predictive performance on unknown data. It is particularly suitable for scenarios in agriculture where sample collection costs are high and labeled data is limited.

[0044] The specific operation of dividing the dataset into training and validation sets in an 8:2 ratio is as follows: After parameter tuning through 5-fold cross-validation, all sample data are randomly divided into training and validation sets in an 8:2 ratio. The training set is used for the final training of the model, and the validation set is used to evaluate the final generalization performance of the model. The 8:2 division ratio achieves a good balance between the sufficiency of model training and the reliability of validation results—80% of the training data ensures that the model can fully learn the nonlinear mapping relationship between spectral features and chlorophyll content, while the 20% of the validation data provides a sufficient sample size to statistically reliably evaluate the model performance.

[0045] In this embodiment, the multispectral image acquisition module further includes a PID closed-loop control unit, which is used to automatically adjust the driving current of the LED array supplementary lighting system according to the feedback value of the ambient light sensor in order to maintain a constant target illuminance.

[0046] It is understood that, by adding a PID closed-loop control unit to the multispectral image acquisition module, this embodiment of the application utilizes an ambient light sensor to monitor the current lighting conditions in real time and feeds them back to the control terminal, dynamically and automatically adjusting the driving current of the LED array supplementary lighting system. This effectively counteracts the noise impact caused by fluctuations in external ambient light, ensuring that the target illuminance remains constant at the working distance. This guarantees the consistency of exposure for each frame of the original multispectral image from the source, significantly reducing the reflectance calculation error caused by light drift. This improves the accuracy and repeatability of the spectral feature vector input to the data processing module, ultimately resulting in higher reliability and quantification accuracy for the chlorophyll content value output by the model prediction module and the whole-leaf pseudo-color distribution map generated by the display output module.

[0047] In this embodiment of the application, a calibration module is included, which is used to automatically trigger the black / white reference board calibration process and update the correction parameters in the data processing module each time the device is powered on or the ambient light intensity changes beyond a preset threshold.

[0048] It is understood that, by setting up an independent calibration module, this application embodiment can monitor changes in ambient light intensity in real time during each device startup initialization phase or during operation. Once light fluctuations are detected to exceed a preset threshold, the black / white reference board calibration process is automatically triggered. This dynamically updates the dark current correction coefficient and reflectance normalization reference parameter within the data processing module, promptly eliminating the response drift of the sensor itself and the influence of uncontrollable external ambient light on the absolute reflectance value. This ensures that even in variable optical environments, the extracted reflectance spectral feature vector maintains a high degree of standardization and comparability, preventing chlorophyll content prediction deviations due to calibration failures, and enhancing the robustness and measurement reproducibility of the entire detection system when deployed in different time periods and locations.

[0049] In this embodiment, the display output module generates the pseudo-color spatial distribution map using a bilinear interpolation algorithm, and uploads the chlorophyll content value and spatial distribution map to the agricultural Internet of Things management platform through a communication interface, which is used to generate variable fertilization prescription maps or crop stress classification reports.

[0050] Understandably, this application's embodiments utilize a bilinear interpolation algorithm to densify and reconstruct chlorophyll content data from discrete circular detection areas, effectively eliminating random errors and sampling sparsity caused by single-point detection. This achieves an intuitive transformation from discrete numerical values ​​to a continuous, high-resolution pseudo-color spatial distribution map of whole-leaf chlorophyll content, facilitating users' rapid identification of spatial heterogeneity features such as leaf edge yellowing or central chlorophyll deficiency. Simultaneously, the quantified chlorophyll data and distribution map are uploaded in real-time to the agricultural IoT management platform via a communication interface, establishing a data link. This allows the detection results to directly support the accurate generation of variable fertilizer prescription maps or graded diagnostic reports of crop stress, thereby transforming simple plant phenotypic detection into practical smart agronomic decision-making support, improving the timeliness and automation level of precision field management.

[0051] A non-destructive testing device for plant chlorophyll based on multispectral imaging and machine learning, as proposed in this application, includes a multispectral image acquisition module, a data processing module, a model prediction module, and a display output module. The device accurately locates leaf micro-regions with a diameter of 2.5 mm using Canny edge detection and Hough circle transform, extracting reflectance spectral feature vectors. An integrated SVM regression model, with a radial basis function kernel function as its core, incorporates Yeo-Johnson power transform to correct data distribution, achieving an output accuracy of 0.1 mg / g, R²≥0.92, and RMSE≤0.35 mg / g. An interpolation algorithm generates a pseudo-color spatial distribution map of whole-leaf chlorophyll. The entire device features a handheld portable design, supports access to agricultural IoT platforms, and achieves millimeter-level micro-region quantification of leaf chlorophyll content, high-precision cross-species prediction, and spatial visualization detection.

[0052] This application also provides a method for visual non-destructive testing of plant chlorophyll, such as... Figure 2 As shown, it includes the following steps: S1. Control the multispectral image acquisition module to acquire multispectral images of plant leaves.

[0053] Specifically, the user aligns the handheld end of the device with the leaf to be tested, maintaining a standard working distance of 30cm between the leaf surface and the lens. Under the coordination of the PID closed-loop control unit, the multispectral image acquisition module automatically adjusts the drive current of the LED array supplementary lighting system based on real-time feedback from the ambient light sensor, ensuring the total illuminance on the leaf surface remains within the preset target illuminance range (typically 2000 lux), with illumination uniformity of no less than 90%. Subsequently, the 8-channel narrowband filter wheel sequentially switches to filter channels with center wavelengths of 450nm, 550nm, 680nm, and 800nm. The back-illuminated CMOS sensor acquires the reflectance spectrum image of the leaf in each channel, obtaining a set of raw multispectral images containing four characteristic bands. The entire acquisition process does not require damaging the leaf sample or using chemical reagents, achieving truly in-situ non-destructive testing.

[0054] S2. Use the data processing module to perform reflectance correction and circular detection area localization on the image, and extract the reflectance spectral feature vector.

[0055] The data processing module performs reflectance correction and circular detection region localization on the image, extracting reflectance spectral feature vectors. First, the module sequentially performs dark current correction and reflectance normalization calculation based on a black / white reference plate on the original multispectral image, converting the original response values ​​into standard reflectance data and eliminating the influence of sensor background noise and ambient light fluctuations on spectral measurements. Then, the Canny edge detection and Hough circle transform algorithms are used to accurately locate the circular detection region on the blade surface: the Canny edge detection algorithm first uses a 5×5 Gaussian filter kernel to smooth and denoise the reflectance grayscale image, setting a low threshold of 50 and a high threshold of 150 to extract the complete blade contour; the Hough circle transform algorithm searches for a circular region with a radius of 1.25mm ± 0.05mm within the contour, with a minimum distance of 5mm between the centers to avoid overlapping detection circles, achieving a localization accuracy better than 0.05mm. After completing the localization of the circular detection area, the average reflectance values ​​of all pixels within each detection circle at four feature bands of 450nm, 550nm, 680nm, and 800nm ​​are extracted to construct a reflectance spectral feature vector. Optionally, the Normalized Difference Vegetation Index (NDVI), Ratio Vegetation Index (RVI), and red-edge position parameters are calculated as supplementary features.

[0056] S3. Input the reflectance spectral feature vector into the trained SVM regression model in the model prediction module to calculate the chlorophyll content value.

[0057] The reflectance spectral feature vector is input into the trained SVM regression model in the model prediction module to calculate the chlorophyll content value. The built-in support vector machine regression model in the model prediction module uses the radial basis function as the kernel function. The model first calculates the kernel function value between the input spectral feature vector and each support vector determined during the training phase, obtaining the spectral similarity between the test sample and each support vector. Then, it calculates the predicted value after Yeo-Johnson power transformation using a decision function. Finally, it applies an inverse Yeo-Johnson transform to the predicted value to restore the absolute chlorophyll content value of the circular detection area. During the training phase, the model employs a 5-fold cross-validation strategy, with the dataset divided into training and validation sets in an 8:2 ratio. The model has a coefficient of determination R² ≥ 0.92, a root mean square error RMSE ≤ 0.35 mg / g, and an output accuracy of 0.1 mg / g. It can quickly calculate the chlorophyll content of multiple micro-region detection points on a single leaf.

[0058] S4. Visualize the chlorophyll content value as a whole leaf distribution map through the display output module and upload it to the agricultural IoT management platform.

[0059] The display output module visualizes chlorophyll content values ​​as a whole-leaf distribution map and uploads it to the agricultural IoT management platform. The display output module receives chlorophyll content values ​​and their spatial coordinates from multiple circular detection areas output by the model prediction module. It then uses a bilinear interpolation algorithm to perform gridded interpolation on the discrete detection point data, generating a continuous pseudo-color spatial distribution map of chlorophyll content covering the entire leaf, visually presenting the heterogeneity of chlorophyll distribution on the leaf surface. Users can annotate the chlorophyll content value of any selected point through the interactive interface and save or export the detection results. Simultaneously, the display output module uploads the chlorophyll content values ​​and spatial distribution map to the agricultural IoT management platform via communication interfaces such as USB, WiFi, or Bluetooth. The platform generates variable fertilization prescription maps or crop stress classification reports based on the received data, providing data support for precision agriculture decision-making.

[0060] In summary, the plant chlorophyll visualization non-destructive detection method provided in this application realizes end-to-end automated processing from the acquisition of the original multispectral image to the output of chlorophyll content value through the organic connection of the four steps S1 to S4. The entire process does not require sample destruction or chemical reagents, is simple to operate, and has a fast detection speed (the detection time for a single leaf does not exceed 5 seconds). The output results have both high precision (0.1 mg / g) and spatial visualization features, and can be widely used in scenarios such as crop nutrition diagnosis, variety selection, stress physiology research, and precision agricultural management.

[0061] Specifically, in this embodiment of the application, the specific operation procedure for a user to perform visual non-destructive testing of plant chlorophyll using the above-mentioned device is as follows: Step 1: Press and hold the power button for more than five seconds to start the handheld multispectral chlorophyll meter and complete the system initialization; Step 2: Adjust the detection parameters by using the function keys and the display screen, and perform a self-test of the device. After confirming that each module is working properly, enter the chlorophyll measurement mode. Step 3: Hold the main body of the device, align the detection window with the surface of the leaf to be tested, maintain a standard working distance of 30cm, and press the measurement button to start the chlorophyll content measurement. The device automatically completes multispectral image acquisition, data processing, and model prediction; Step 4: After the measurement is completed, the display screen will automatically show a pseudo-color spatial distribution map of chlorophyll content of the whole leaf. Users can select different leaf parts through function buttons, and the display screen will simultaneously give the chlorophyll content value of the corresponding single point. The output accuracy can reach 0.1mg / g. Step 5: After the test is completed, the user can connect the device to the computer through a communication interface (such as a USB data cable) to export the test data and spatial distribution map for subsequent analysis or to generate variable fertilization prescription maps.

[0062] According to the embodiments of this application, a non-destructive detection method for plant chlorophyll based on multispectral imaging and machine learning is proposed. By using a chlorophyll content prediction model and multispectral imaging technology, the spatial distribution of plant chlorophyll content can be visualized and predicted. This method can quickly obtain the precise chlorophyll content at one or more points, solving the problem that traditional methods can only obtain a rough value for the overall leaf and cannot characterize the spatial heterogeneity. Multispectral imaging enables non-destructive determination of plant chlorophyll content, successfully overcoming the limitations of traditional methods that damage plant samples due to grinding or cutting. In addition, multispectral imaging measurement can accurately determine chlorophyll content without relying on chemical reagents. The detection process is convenient, pollution-free, and meets green and environmentally friendly requirements.

[0063] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

Claims

1. A non-destructive detection device for plant chlorophyll based on multispectral imaging and machine learning, characterized in that, include: The system comprises a multispectral image acquisition module, a data processing module, a model prediction module, and a display output module; among which, The multispectral image acquisition module is connected to the data processing module. The multispectral image acquisition module includes a back-illuminated CMOS sensor, an 8-channel narrowband filter wheel, and an LED array supplementary lighting system. The data processing module is connected to the model prediction module. It sequentially performs dark current correction, reflectance normalization calculation based on black / white reference plate, and locates the circular detection area of ​​the leaf to extract the reflectance spectral feature vector on the received raw spectral image. The model prediction module is connected to the display output module. The model prediction module has a built-in trained support vector machine regression model. It takes the reflectance spectral feature vector as input and outputs the chlorophyll content value of the circular detection area. The display output module is used to receive the chlorophyll content values ​​and corresponding coordinates of multiple circular detection areas output by the model prediction module, and generate a pseudo-color spatial distribution map of the chlorophyll content of the whole leaf through an interpolation algorithm.

2. The plant chlorophyll non-destructive testing device based on multispectral imaging and machine learning according to claim 1, characterized in that, The 8-channel narrowband filter wheel is configured with center wavelengths of 450nm, 550nm, 680nm and 800nm, respectively, and the bandwidth of each channel filter is 10nm±2nm; the back-illuminated CMOS sensor has a spectral response range of 400-900nm and a resolution of no less than 5 million pixels; the LED array supplementary lighting system has an illumination uniformity of no less than 90% at a working distance of 30cm and an illuminance adjustment range of 500-5000 lux.

3. The plant chlorophyll non-destructive testing device based on multispectral imaging and machine learning according to claim 1, characterized in that, The support vector machine regression model uses radial basis functions as kernel functions; The support vector machine regression model outputs a predicted value after Yeo-Johnson power transformation, and the predicted value is calculated by a decision function, which is constructed based on the support vector spectral features and kernel function. The model prediction module applies a Yeo-Johnson inverse transform to the predicted value to obtain the predicted chlorophyll content of the circular detection area.

4. The plant chlorophyll non-destructive testing device based on multispectral imaging and machine learning according to claim 1, characterized in that, The reflectance spectral feature vector includes reflectance values ​​at center wavelengths of 450 nm, 680 nm, and 800 nm, as well as the normalized vegetation index, ratio vegetation index, and red edge position parameters calculated based on the reflectance values.

5. The plant chlorophyll non-destructive testing device based on multispectral imaging and machine learning according to claim 1, characterized in that, When the data processing module locates the circular detection area of ​​the blade using the Canny edge detection and Hough circle transform algorithm, the radius of the detection circle is set to 1.25mm ± 0.05mm, and the positioning accuracy is no greater than 0.05mm.

6. The plant chlorophyll non-destructive testing device based on multispectral imaging and machine learning according to claim 1, characterized in that, The support vector machine regression model uses 5-fold cross-validation during the training phase, and the dataset is divided into training and validation sets in an 8:2 ratio. The trained model has a determination coefficient R² of no less than 0.92 and a root mean square error of no more than 0.35 mg / g.

7. The plant chlorophyll non-destructive testing device based on multispectral imaging and machine learning according to claim 1, characterized in that, The multispectral image acquisition module also includes a PID closed-loop control unit, which is used to automatically adjust the driving current of the LED array supplementary lighting system according to the feedback value of the ambient light sensor in order to maintain a constant target illuminance.

8. The non-destructive testing device for plant chlorophyll based on multispectral imaging and machine learning according to claim 1, characterized in that, It also includes a calibration module, which automatically triggers the black / white reference board calibration process and updates the correction parameters in the data processing module each time the device is powered on or when the ambient light intensity changes beyond a preset threshold.

9. The non-destructive testing device for plant chlorophyll based on multispectral imaging and machine learning according to claim 1, characterized in that, The display output module generates the pseudo-color spatial distribution map through a bilinear interpolation algorithm, and uploads the chlorophyll content value and spatial distribution map to the agricultural Internet of Things management platform through the communication interface, which is used to generate variable fertilization prescription maps or crop stress classification reports.

10. A method for visual non-destructive detection of plant chlorophyll using the apparatus according to any one of claims 1-9, characterized in that, Includes the following steps: S1. Control the multispectral image acquisition module to acquire multispectral images of plant leaves; S2. Use the data processing module to perform reflectance correction and circular detection area localization on the image, and extract the reflectance spectral feature vector; S3. Input the reflectance spectral feature vector into the trained SVM regression model in the model prediction module to calculate the chlorophyll content value. S4. Visualize the chlorophyll content value as a whole leaf distribution map through the display output module and upload it to the agricultural IoT management platform.

Images