Wheat sharp eyespot early recognition method based on hyperspectral imaging technology

Abstract

The invention provides a wheat sharp eyespot early recognition method based on a hyperspectral imaging technology, and belongs to the field of crop disease recognition, and the method comprises the following steps: respectively carrying out hyperspectral data collection on overground part tissues of different leaves, different leaf sheaths and stem bases of a wheat plant in a seedling stage to be detected, and obtaining hyperspectral images of each part; the method comprises the following steps: determining an optimal region of interest as an overground part of a basal stem, then performing hyperspectral image processing and region-of-interest marking on the overground part, extracting all corresponding spectral data, then performing characteristic wavelength screening, extracting characteristic wavelengths representing wheat sharp eyespot early-stage hazard characteristics, and constructing a wheat sharp eyespot early-stage identification model based on the characteristic wavelengths; and inputting the characteristic wavelength data of the wheat plant to be detected into the model, and outputting the health status of the wheat plant. The early-stage rapid and accurate recognition of the wheat sharp eyespot is realized, the problem that the timeliness of early-stage large-area recognition of the wheat sharp eyespot is seriously delayed is solved, the method is adaptive to key prevention and control requirements in the seedling stage, and a technical support is provided for early prevention and treatment of diseases.

Description

Technical Field

[0001] This invention relates to the field of crop disease identification technology, and in particular to an early identification method for wheat sheath blight based on hyperspectral imaging technology. Background Technology

[0002] Hyperspectral imaging technology, as an advanced spectral analysis method, has been widely applied to the non-destructive detection of crop diseases due to its superior image-spectrum fusion capabilities. The development of this technology mainly revolves around the extraction of spectral features, model optimization, and the expansion of its application scale. For example, researchers have successfully achieved early identification of diseases such as rice sheath blight and wheat powdery mildew by acquiring hyperspectral data from different growth stages of rice and wheat and using machine learning algorithms. In the detection of crop diseases such as cucumber downy mildew and Acanthopanax senticosus black spot, this technology has also achieved high-precision disease diagnosis by constructing targeted identification models, providing efficient technical support for crop disease control. However, most current research still focuses on mining the spectral features of the plant canopy. For soil-borne diseases with hidden disease sites and inconspicuous early symptoms, especially diseases with unique infection patterns such as wheat sheath blight, a mature detection system has not yet been formed.

[0003] With the popularization of drone technology and hyperspectral imaging technology, the efficiency and coverage of disease detection are constantly improving. Drones equipped with hyperspectral imagers can scan large areas of farmland in a short time and transmit data back in real time, providing a more efficient solution for disease detection. In recent years, with the advancement of machine learning and data mining technologies, researchers are increasingly inclined to build efficient identification models, improving the accuracy of disease identification by integrating multiple feature wavelengths and optimizing algorithms. However, for the detection of wheat sheath blight, existing technologies still face many bottlenecks: on the one hand, manual detection is labor-intensive, inefficient, and easily affected by subjective factors; molecular detection has high accuracy, but it is time-consuming, costly, and inefficient, both of which can easily lead to delays in wheat sheath blight control. On the other hand, wheat sheath blight is a major soil-borne disease in wheat-corn double-cropping areas under conservation tillage conditions, with a wide range of damage and great difficulty in control. In severe cases, it can lead to yield losses of up to 70% or more, while the existing conventional control measures of "turning, mixing, and spraying" still have the problem of severe disease occurrence and low control effect.

[0004] Furthermore, wheat sheath blight initially affects the seedling stage, with the pathogen first infecting the coleoptile or young roots and then rapidly spreading to the base of the above-ground stem. Initially, there are no visible symptoms at the base of the above-ground stem, and changes in leaf canopy indicators such as chlorophyll and water content are minimal. Within 7–20 days after emergence, the spectral difference in the leaf canopy is only 3%–5%, far below the model's recognition threshold, resulting in a significant delay in early identification. Existing methods generally have an accuracy rate of less than 35% for identifying seedlings (disease index < 8%), making them unsuitable for the critical control window. In addition, hyperspectral data contains hundreds of bands, requiring multiple dimensionality reduction and feature selection processes in existing methods, leading to lengthy data processing times. Moreover, ground-based detection equipment is mostly laboratory-grade, bulky, complex to operate, and inefficient for mobile field testing, failing to meet the rapid detection needs of large-scale production.

[0005] Therefore, there is an urgent need for an efficient identification method that can adapt to the disease characteristics of wheat sheath blight, accurately capture early disease signals, and be suitable for large-scale planting areas, so as to achieve early, rapid, and accurate detection of wheat sheath blight. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide an early identification method for wheat sheath blight based on hyperspectral imaging technology. This method takes the aboveground tissue 1-4 cm away from the soil surface at the base of the wheat stem as the research object, directly acquires its hyperspectral image, accurately captures the spectral data of early infection of sheath blight fungus in the seedling stage, and realizes rapid and accurate identification of wheat sheath blight in the early stage. This method is suitable for the key control window period in the seedling stage, solves the problem of serious time lag in the early large-area identification of wheat sheath blight, supports early and precise control, thereby reducing pesticide use, ensuring the sustainable development of green agriculture, and laying a technical foundation for stabilizing wheat yield and improving agricultural management.

[0007] To achieve the above objectives, the present invention provides the following solution: a method for early identification of wheat sheath blight based on hyperspectral imaging technology, comprising the following steps: S1. To address the disease characteristics of wheat sheath blight, which initially damages the coleoptile or young root of wheat, then rapidly spreads to the base of the aboveground stem, and in the middle and late stages damages the leaf sheath and stem, hyperspectral data were collected from different leaves, leaf sheaths, and aboveground tissues of wheat seedlings during the one-leaf to four-leaf stages. This yielded hyperspectral images of different leaves, leaf sheaths, and aboveground tissues of the stem base of wheat plants. S2. Based on the obtained hyperspectral images, the correlation between spectral signals from different parts and infection by *Rhizoctonia solani* was analyzed, and the optimal region of interest for *Rhizoctonia solani* was determined to be the aboveground part of the wheat plant stem base. S3. After determining the optimal region of interest, the hyperspectral image of the aboveground tissue at the base of the wheat plant stem is processed and the region of interest is marked. All spectral data corresponding to the region of interest are extracted and preprocessed. S4. Screen the preprocessed spectral data for characteristic wavelengths and extract characteristic wavelengths used to characterize the early damage characteristics of wheat sheath blight caused by infection with Rhizoctonia graminearum or Rhizoctonia solani. S5. Construct an early identification model for wheat sheath blight based on the extracted feature wavelengths and train it. S6. Input the characteristic wavelength data of the wheat plant to be tested into the trained recognition model, output the health status of the wheat plant, and complete the early recognition of wheat sheath blight.

[0008] Compared with the prior art, the present invention discloses at least the following technical effects: (1) This invention provides an early identification method for wheat sheath blight based on hyperspectral imaging technology. By deeply integrating hyperspectral imaging technology with spectral data processing software and machine learning models, a standardized early identification process for diseases is constructed, which effectively solves the problem of serious time delay in the early identification of large-scale wheat sheath blight. This method greatly reduces labor costs and improves detection efficiency. It abandons the traditional model of relying on manual field surveys of each plant. Data is collected by using a portable hyperspectral imager or a drone equipped with a hyperspectral imager. Data correction, preprocessing and feature wavelength extraction are completed by combining professional software such as ENVI and The Unscrambler X. Subsequently, the identification results are automatically output through the trained model. There is no need for manual identification of diseased plants or search for diseased areas. It is more suitable for large-scale field detection scenarios and significantly improves detection efficiency.

[0009] (2) The method provided by this invention has high accuracy and strong stability, avoiding the accuracy fluctuation problem caused by subjective experience and ambient light in traditional manual detection. Noise interference is eliminated by four preprocessing methods: Savitzky-Golay smoothing, SG first derivative, SNV, and MSC. Characteristic wavelengths are screened by principal component analysis (PCA)-loading coefficient method and / or continuous projection algorithm (SPA). Finally, an early identification model for wheat sheath blight is constructed based on 100 characteristic wavelengths. Linear discriminant analysis (LDA) or support vector machine (SVM) is used as the identification algorithm, thereby achieving an early identification accuracy of no less than 90%.

[0010] (3) The method provided by the present invention achieves identification by capturing early biochemical changes caused by diseases, and data can be collected and detected at the one-leaf, two-leaf, three-leaf and four-leaf stages of wheat. While achieving early warning, it breaks the limitations of traditional technology that relies on canopy spectral signals and has a delayed timeliness of early large-area identification. It can achieve accurate identification when there are no visible symptoms on the ground, providing timely guidance for targeted drug application in the seedling stage, effectively blocking the overwintering reproduction chain of pathogens, and greatly improving the disease control effect.

[0011] (4) The method provided by this invention is easy to promote and apply in practice. The operation details of the whole process and the selection of software and equipment are clear and standardized, which reduces the operation threshold. It can be used with a portable hyperspectral instrument to adapt to field sampling and laboratory analysis. Combined with a drone equipped with a hyperspectral imager, it can realize large-scale rapid scanning in the field, and can quickly complete large-scale field screening, which helps to prevent and treat wheat sheath blight in the early stage, and provides technical support for food production security and sustainable development of green agriculture. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart of a method for early identification of wheat sheath blight based on hyperspectral imaging technology according to the present invention; Figure 2 This invention provides a comparison diagram of region-of-interest markers and spectral reflectance curves of the first leaf of healthy and diseased wheat; wherein, Figure 2 In the diagram, A represents the region of interest (ROI) for healthy wheat. Figure 2 In the diagram, B represents the region of interest (ROI) for diseased wheat. Figure 2 C in the figure is a comparison of the spectral reflectance curves of the first leaf of wheat; Figure 3 This invention provides a comparison diagram of region-of-interest markers and spectral reflectance curves of second leaves for healthy and diseased wheat; wherein, Figure 3 In the diagram, A represents the region of interest (ROI) for healthy wheat. Figure 3 In the diagram, B represents the region of interest (ROI) for diseased wheat. Figure 3 C in the figure is a comparison of the spectral reflectance curves of the second leaf of wheat; Figure 4 This invention provides a comparison of region-of-interest markers for healthy and diseased wheat, as well as spectral reflectance curves of wheat stem bases; wherein, Figure 4 In the diagram, A represents the region of interest (ROI) for healthy wheat. Figure 4 In the diagram, B represents the region of interest (ROI) of the infected wheat. Figure 4 C in the figure is a comparison of the spectral reflectance curves of the wheat stem base; Figure 5 This invention provides a comparison of the reflectance of all stem base samples after pretreatment using different methods during the seedling stage; wherein, Figure 5 In the image, A represents a comparison of the reflectance of all stem base samples after SG smoothing. Figure 5 In the image, B represents a comparison of the reflectance of all stem base samples after processing with the first derivative of SG. Figure 5 In the image, C represents a comparison of the reflectance of all stem base samples after SNV treatment. Figure 5 In the figure, D represents a comparison of the reflectance of all stem base samples after MSC treatment; Figure 6 This is a PCA clustering diagram of all stem base samples after pretreatment using different methods during the seedling stage, as provided by this invention; wherein, Figure 6 In the diagram, A represents the PCA clustering diagram of all stem base samples in the original spectral raw data. Figure 6 In the diagram, B represents the PCA clustering diagram of all stem base samples after SG smoothing. Figure 6 In the diagram, C represents the PCA clustering diagram of all stem base samples after the first derivative of SG. Figure 6 In the diagram, D represents the PCA clustering diagram of all stem base samples after SNV processing. Figure 6 In the diagram, E represents the PCA clustering diagram of all stem base samples after MSC treatment; Figure 7 Load factor diagrams preprocessed by different methods provided by the present invention; wherein, Figure 7 In the diagram, A represents the loading coefficients of the original spectral raw data. Figure 7 In the diagram, B represents the load factor after SG smoothing. Figure 7 C in the diagram represents the load factor after processing with the first derivative of SG. Figure 7 In the diagram, D represents the load factor after SNV processing. Figure 7 In the diagram, E represents the load factor after MSC processing; Figure 8 The image provided by this invention shows a comparison of characteristic wavelength extraction and corresponding spectral reflectance curves under SPA transform; wherein... Figure 8 (a) in the figure shows the relationship between the number of eigenvectors and the RMSE value under the SPA transform. Figure 8 (b) in the figure is a comparison of the original spectral data and the spectral reflectance curves after the characteristic wavelengths were extracted by SPA; Figure 9 This invention provides a comparison chart of characteristic wavelength extraction and corresponding spectral reflectance curves of spectral data after SG smoothing and SPA transformation; wherein, Figure 9(a) in the figure shows the relationship between the number of eigenvectors and the RMSE value of the spectral data after SG smoothing and SPA transformation. Figure 9 (b) in the figure is a comparison of the spectral reflectance curves after SG smoothing and SPA extraction of characteristic wavelengths; Figure 10 This invention provides a comparison chart of characteristic wavelength extraction and corresponding spectral reflectance curves of spectral data after SG first-order derivative processing and SPA transformation; wherein, Figure 10 (a) in the figure shows the relationship between the number of eigenvectors and the RMSE value of the spectral data after SPA transformation following the first derivative of SG. Figure 10 (b) in the figure is a comparison of the spectral reflectance curves after the first derivative of SG is processed and the characteristic wavelengths are extracted by SPA. Figure 11 This invention provides a comparison chart of characteristic wavelength extraction and corresponding spectral reflectance curves of SNV-processed spectral data after SPA transformation; wherein, Figure 11 (a) in the figure shows the relationship between the number of eigenvectors and the RMSE value of the spectral data after SNV processing and SPA transformation. Figure 11 (b) is a comparison of the spectral reflectance curves of the characteristic wavelengths extracted by SPA after SNV processing; Figure 12 This invention provides a comparison chart of characteristic wavelength extraction and corresponding spectral reflectance curves of MSC-processed spectral data after SPA transformation; wherein, Figure 12 (a) in the figure shows the relationship between the number of eigenvectors and the RMSE value of the spectral data after SPA transformation following MSC processing. Figure 12 (b) in the figure is a comparison of the spectral reflectance curves of the characteristic wavelengths extracted by SPA after MSC treatment; Figure 13 This is a region of interest (ROI) marker map of wheat seedlings in an experimental field provided in Application Example 1 of the present invention; wherein, Figure 13 In the diagram, A represents the region of interest (ROI) for diseased wheat. Figure 13 In the diagram, B represents the region of interest (ROI) for healthy wheat. Figure 13 C in the figure is a comparison of the spectral reflectance curves of the aboveground tissues at the base of the wheat stem; Figure 14 This is a region of interest (ROI) marker map of mature wheat plants in an experimental field provided in Application Example 2 of the present invention; wherein, Figure 14 In the diagram, A represents the region of interest (ROI) for healthy wheat. Figure 14 In the diagram, B represents the region of interest (ROI) of the infected wheat. Figure 14 C in the figure is a comparison of the spectral reflectance curves of the aboveground tissues at the base of the wheat stem. Detailed Implementation

[0014] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0015] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] like Figure 1 As shown, this embodiment provides a method for early identification of wheat sheath blight based on hyperspectral imaging technology, including the following steps: S1. To address the disease characteristics of wheat sheath blight, which initially damages the coleoptile or young root, then rapidly spreads to the base of the aboveground stem, and later damages the leaf sheath and stem, hyperspectral data were collected from different leaves, leaf sheaths, and aboveground tissues of wheat seedlings during the one-leaf to four-leaf stages. This yielded hyperspectral images of different leaves, leaf sheaths, and aboveground tissues of the stem base.

[0017] In this embodiment, before acquiring hyperspectral data, the following steps are also included: First, disease samples were prepared using the wheat grain inoculum method to simulate disease occurrence under natural conditions. Plump, appropriately sized wheat seeds were selected, disinfected by soaking in 75% alcohol for 1 minute, and then rinsed three times with sterile water. The seeds were placed in petri dishes lined with moist filter paper and cultured at a constant temperature of 25℃ for 3 days. Using artificial inoculation, 5-10 wheat grain inoculum were evenly scattered per hole. After covering with soil, wheat with 5mm coleoptiles were transplanted into 5×10 cm seedling trays containing a 1:1 ratio of nutrient soil and sand, and watered using a bottom irrigation method.

[0018] A blank control group was also set up, using wheat seeds from the same batch treated with sterile water and cultivated according to the same procedure as healthy samples. Each treatment was repeated three times, with 100 wheat plants cultivated in each replicate to ensure sufficient and statistically significant sample size, providing a stable database of healthy and diseased samples for subsequent data collection and model training.

[0019] Furthermore, the specific implementation process of step S1 above is as follows: In this embodiment, the key window period and core region for hyperspectral data acquisition were first identified. A hyperspectral imager was used to acquire spectral images of the first, second, and third leaves, the leaf sheaths of the first and second leaves, and the aboveground tissues at the stem base during the three-leaf stage of wheat. The acquired spectral images were then corrected to obtain hyperspectral images of different leaves, leaf sheaths, and aboveground tissues at the stem base of the wheat plant. The acquisition period covered the one-leaf to four-leaf stage of wheat, which is the critical period for the initial infection of the sheath blight pathogen (Rhizoctonia graminearum or Rhizoctonia solani), allowing for precise capture of early microphysiological changes in the leaf sheaths and stem base. The core region for acquisition was the aboveground leaf sheath of the first true leaf, i.e., the early core infection area of ​​sheath blight, mainly the aboveground tissues at the stem base 1-4 cm from the soil surface. This area is the initial site of sheath blight infection, and the spectral signal can most directly reflect the early characteristics of the disease.

[0020] Secondly, select a suitable hyperspectral imaging system and parameters. The acquisition equipment can be a portable hyperspectral imager (HG101) or a drone-mounted hyperspectral imager (such as Parrot Sequoia). These two devices are suitable for fine-scale laboratory testing and field-scale population testing, respectively. The original coverage band of the hyperspectral imager is 396–1008 nm. After acquisition, noise bands need to be removed, and the final effective spectral range is 400–1001 nm. This range can completely cover the spectral response characteristics of leaf sheaths and stem base aboveground tissues caused by *Sheath blight* infection, laying the foundation for subsequent characteristic wavelength selection.

[0021] Secondly, the data collection environment and operating procedures should be set according to the equipment type. If a portable hyperspectral imager is used, it should be conducted in an indoor darkroom environment. After cleaning and drying the wheat plants, place them on the platform directly below the camera. Scan the first, second, and third leaves of healthy and diseased wheat, the leaf sheaths of the first and second leaves, and the core area of ​​the aboveground leaf sheath of the first true leaf. Subsequent spectral difference analysis revealed that, except for the aboveground leaf sheath of the first true leaf, the reflectance of other parts of healthy and diseased plants showed no significant differentiation. The image of this core area was ultimately selected for data analysis. If a hyperspectral imager mounted on a drone is used, the data collection time should be selected between 10:00 and 14:00, in sunny, windless, and cloudless weather. Before the flight, the flight route should be planned in advance using professional software to ensure coverage of all detection areas in the field. Reflectance correction should be performed using a standard white board before takeoff, and the drone should automatically return to base after the flight to ensure the accuracy and consistency of data collection at the population scale.

[0022] S2. Based on the obtained hyperspectral images, the correlation between spectral signals from different parts and infection by *Rhizoctonia solani* was analyzed, and the optimal region of interest for *Rhizoctonia solani* was determined to be the aboveground part of the wheat plant stem base.

[0023] Furthermore, the specific implementation process of step S2 above is as follows: Based on the above, this invention also compared the differences in spectral reflectance of different parts of wheat seedlings, and found that: Figure 2 A, B and Figure 3 As shown in A and B, spectral reflectance was collected from the same parts of healthy and diseased wheat plants, and the results are shown in Figure 1. Figure 2 C and Figure 3 As shown in C, within the 400–1001 nm wavelength range, the reflectance of the first, second, and third leaves of wheat plants was indistinguishable between healthy and diseased plants; furthermore, the reflectance of the leaf sheaths of the second and third leaves of diseased plants also showed no significant difference. However… Figure 4 A and B were collected from the aboveground tissues at the base of the stems of healthy and diseased wheat plants, respectively. The obtained spectral reflectance is as follows: Figure 4 As shown in C, the spectral reflectance of the aboveground tissues at the base of the wheat stem differs significantly in the wavelength ranges of 400–696 nm and 700–900 nm. Therefore, the optimal region of interest for early identification of wheat sheath blight is determined to be the aboveground part of the stem base, i.e., the aboveground leaf sheath of the first leaf.

[0024] S3. After determining the optimal region of interest, the hyperspectral image of the aboveground tissue at the base of the wheat plant stem is processed and the region of interest is marked. All spectral data corresponding to the region of interest are extracted and preprocessed.

[0025] Furthermore, the specific implementation process of step S3 above is as follows: Based on the above analysis, the optimal region of interest can be determined as the aboveground part of the wheat plant stem base. Subsequently, this embodiment first performs radiometric correction on the original hyperspectral image of the aboveground tissue of the wheat plant stem base to eliminate spectral errors caused by factors such as ambient light variations and equipment malfunctions. The correction employs the whiteboard correction method, executed using ENVI 5.3.1 software. Using the standard reflectance of the whiteboard as a reference, the image grayscale values ​​are converted into true reflectance data to ensure the reliability and comparability of the spectral data.

[0026] Secondly, regions of interest (ROIs) at the base of the stem are marked. In the radiometrically corrected hyperspectral image, a rectangular area of ​​50×50 pixels is marked as the ROI for the aboveground tissue at the base of the stem, which is 1-4 cm from the soil surface. This size ensures that enough information about the stem base tissue is included while avoiding interference from surrounding soil, leaves and other irrelevant areas.

[0027] Finally, all spectral data of the region of interest are extracted. Specifically, the spectral data of all pixels in the marked region are extracted, and the average spectral data of the base of a single wheat stem is obtained by averaging. This data can reflect the spectral characteristics of the stem base and provide standardized input for subsequent preprocessing and feature extraction.

[0028] In addition, this step involves preprocessing the spectral data, including: First, importing all the obtained spectral data from the stem base into The Unscrambler X 10.4 software, a professional tool for spectral data processing capable of efficiently executing various preprocessing algorithms. Second, a preprocessing strategy is employed to eliminate noise interference. The preprocessing methods used include Savitzky-Golay (SG) smoothing, Savitzky-Golay (SG) first derivative, Standard Normal Transform (SNV), and Multivariate Scattering Correction (MSC). SG smoothing filters out high-frequency random noise while preserving spectral profile features; SG first derivative enhances subtle spectral differences, highlighting early spectral changes in disease; SNV eliminates scattering differences caused by sample particle size and surface condition; and MSC corrects multivariate scattering effects in the spectral data, improving spectral consistency. By using these four methods, interference factors such as environmental noise, soil scattering, and equipment system noise are effectively eliminated, resulting in high-quality standardized spectral data.

[0029] S4. Select characteristic wavelengths from the preprocessed spectral data to characterize the early damage characteristics of wheat sheath blight caused by infection with Rhizoctonia graminearum or Rhizoctonia solani.

[0030] Furthermore, the specific implementation process of step S4 above is as follows: This embodiment employs Principal Component Analysis (PCA) with loading coefficients for preliminary screening of characteristic wavelengths. PCA analysis is performed on both the raw and preprocessed spectral data. PCA1 and PCA2 from the principal component score plots are selected for sample clustering to initially distinguish between healthy and diseased samples. By calculating the contribution rate of the principal components, principal components with a cumulative contribution rate greater than 90% are selected, as these principal components contain the core information of the spectral data. Subsequently, the loading curves of the selected principal components are analyzed, and the wavelengths corresponding to their peaks or troughs are extracted as characteristic wavelengths, denoted as Raw1, SG smoothing1, SG first derivative1, SNV1, and MSC1 according to the preprocessing methods.

[0031] Specifically, by Figure 5 It can be seen that the original spectral data of the wheat plant stem base (i.e., the above-ground leaf sheath of the first leaf) during the seedling stage were subjected to SG smoothing, SG first-order derivative processing, standard normal variable transformation (SNV) processing, and multivariate scattering correction (MSC) processing, respectively. Figure 5As shown in A, SG smoothing makes the original spectrum smoother, but still retains its original shape; as... Figure 5 As shown in B, the first derivative of SG significantly increases the number of its detailed eigenvalues; for example... Figure 5 As shown in C and D, SNV and MSC processing significantly enhance the matching degree of different sample spectra in a specific band range.

[0032] To further clarify the spectral differences between healthy and diseased samples after different pretreatments, principal component analysis was first performed on the spectral data of healthy and diseased samples after different pretreatments, followed by PCA cluster analysis. The results are as follows: Figure 6 As shown.

[0033] The cumulative contribution rate of the first principal component (PC1) and the second principal component (PC2) of the original spectral data was 88.7%. After SG smoothing preprocessing, the cumulative contribution rate of PC1 and PC2 was 89.0%. After SG first derivative preprocessing, the cumulative contribution rate of PC1 and PC2 was 33.9%. After SNV preprocessing, the cumulative contribution rate of PC1 and PC2 was 88.7%. After MSC preprocessing, the cumulative contribution rate of PC1 and PC2 was 89.0%. The data indicate that PC1 and PC2 can represent the vast majority of the original data information, meaning that PC1 and PC2 can effectively represent the data information of wheat samples with sheath blight and healthy wheat samples. Specifically, after PCA cluster analysis, the original spectra (RAW) and SG smoothed spectra could basically distinguish between wheat samples with sheath blight and healthy wheat samples, but some samples remained confused and difficult to distinguish. However, after PCA cluster analysis, the classification boundary between diseased and healthy wheat samples was clearer after SG first derivative, SNV, and MSC preprocessing.

[0034] like Figure 6 As shown in A and B, after the original raw and SG smoothing processes, the sample spectral data, after PCA transformation, can distinguish most healthy wheat samples from diseased wheat samples, but a small number of samples are still difficult to distinguish.

[0035] like Figure 6 As shown in C, D, and E, after processing with the first derivative of SG, SNV, and MSC, PCA classification results are significantly better and the boundaries are clearer. SNV and MSC processing, in particular, yield the best results. This demonstrates that the spectral data clearly contains two categories of samples, thus allowing for the differentiation between diseased and healthy wheat samples.

[0036] Based on the above, Figure 7 These are the spectral characteristic wavelengths obtained using Principal Component Analysis (PCA) with loading coefficients. The peaks and troughs of each principal component loading contain a significant amount of useful information and are therefore selected as spectral characteristic wavelengths.

[0037] like Figure 7 As shown in A, B, C, D, and E, the original spectral data, SG smoothing, SG first derivative, SNV, and MSC selected the first 5 principal components. The characteristic wavelengths selected by the loading coefficient method were 7 (545.68nm, 552.40nm, 569.26nm, etc.), 8 (545.68nm, 552.40nm, 569.26nm, etc.), 11 (545.68nm, 552.40nm, 557.45nm, etc.), 9 (552.40nm, 562.50nm, 569.26nm, etc.), and 7 (552.40nm, 562.50nm, 569.26nm, etc.).

[0038] Furthermore, this embodiment can also optimize and select feature wavelengths using the Continuous Projection (SPA) algorithm. The SPA transform is applied to the aforementioned batch of spectral data. This algorithm, while minimizing information redundancy, can select the optimal feature wavelengths by calculating the number of feature vectors and their corresponding root mean square error (RMSE). Preprocessing methods include Raw2, SG smoothing2, SG first derivative2, SNV2, and MSC2.

[0039] Specifically, by performing the SPA transformation on the original spectral data, the number of eigenvectors and their corresponding RMSE values ​​can be obtained, such as... Figure 8 As shown in (a) and (b), when the RMSE value is 0.2111, the results show that 7 characteristic wavelengths can be extracted.

[0040] By performing SPA transformation on the spectral data after SG smoothing, the number of eigenvectors and their corresponding RMSE values ​​can be obtained, such as... Figure 9 As shown in (a) and (b), when the RMSE value is 0.19928, the results show that 13 characteristic wavelengths can be extracted.

[0041] By performing the SPA transform on the spectral data after taking the first derivative of SG, the number of eigenvectors and their corresponding RMSE values ​​can be obtained, such as... Figure 10 As shown in (a) and (b), when the RMSE value is 0.24408, the results show that 15 characteristic wavelengths can be extracted.

[0042] By performing SPA transformation on the spectral data after SNV processing, the number of eigenvectors and their corresponding RMSE values ​​can be obtained, such as... Figure 11 As shown in (a) and (b), when the RMSE value is 0.20018, the results show that 10 characteristic wavelengths can be extracted.

[0043] By performing SPA transformation on the spectral data processed by MSC, the number of eigenvectors and their corresponding RMSE values ​​can be obtained, such as... Figure 12 As shown in (a) and (b), when the RMSE value is 0.19895, the results show that 13 characteristic wavelengths can be extracted.

[0044] Based on the above-mentioned characteristic wavelength screening, the spectral characteristic wavelengths selected in this embodiment using principal component analysis (PCA)-loading factor method and continuous projection algorithm (SPA) are shown in Table 1.

[0045] Table 1. Wavelengths for identifying wheat sheath blight

[0046] As can be seen from Table 1, the characteristic wavelengths selected by the two methods are somewhat similar, with a total of 100 spectral characteristic wavelengths.

[0047] S5. Construct an early identification model for wheat sheath blight based on the extracted feature wavelengths and train it.

[0048] Furthermore, the specific implementation process of step S5 above is as follows: First, the sample set is divided. The spectral data corresponding to the 10 feature band sets (Raw1-SNV1, Raw2-SNV2) obtained in step S4 are randomly divided into training and test sets in a 7:3 ratio. 70% is used as the training set for model building, and 30% is used as the test set for model performance validation. During the division process, it is ensured that the sample ratio of healthy plants (label "1") and diseased plants (label "2") in the training and test sets is consistent with the overall sample to avoid data bias that could lead to a decrease in the model's generalization ability.

[0049] Secondly, a suitable recognition algorithm is selected and a model is constructed. This embodiment uses either Linear Discriminant Analysis (LDA) or Support Vector Machine (SVM) as the recognition algorithm. Both algorithms have mature applications in crop disease spectral identification and can effectively handle high-dimensional spectral data. During training, key parameters are optimized: the kernel function type and penalty coefficient are optimized for the SVM model, and the discrimination threshold is optimized for the LDA model to ensure optimal model fitting.

[0050] Finally, model performance is evaluated. Test set data is input into the trained model, and the model's recognition accuracy is calculated. When the accuracy is not lower than 90%, the model is saved as the final early identification model for wheat sheath blight.

[0051] The above content specifically includes using principal component analysis (PCA) with loading coefficients and / or continuous projection algorithm (SPA) to screen characteristic wavelengths, and performing discriminant analysis on healthy and diseased wheat plants based on these wavelengths, including: (1) For the spectral characteristic wavelengths selected by principal component analysis (PCA)-loading coefficient method, linear discriminant analysis (LDA) and support vector machine (SVM) were used to establish classification models of healthy wheat and diseased wheat plants, respectively. The discriminant performance is shown in Table 2.

[0052] As can be seen from Table 2, the correct classification accuracy of the LDA model on both the training and test sets is above 94%; among them, the model based on MSC data performs the best, with a correct classification accuracy of 100% on both the training and test sets.

[0053] Table 2 Performance of LDA and SVM models based on characteristic wavelengths using the PCA load factor method

[0054] (2) For the feature wavelengths selected by the continuous projection algorithm (SPA), the classification models of healthy wheat and diseased wheat were established by linear discriminant analysis (LDA) and support vector machine (SVM), respectively. The discriminant performance is shown in Table 3.

[0055] Table 3 Performance of LDA and SVM models based on SPA characteristic wavelengths

[0056] Table 3 shows that the correct classification rate of the LDA model on both the training and test sets is over 90%, with the model based on RAW data performing the best, achieving a correct classification rate of 98% on both the training and test sets.

[0057] S6. Input the characteristic wavelength data of the wheat plant to be tested into the trained recognition model, output the health status of the wheat plant, and complete the early recognition of wheat sheath blight.

[0058] Furthermore, the specific implementation process of step S6 above is as follows: This embodiment first performs a preprocessing procedure on the wheat plants to be tested. Hyperspectral images of the aboveground tissues at the base of the wheat stem are acquired according to the aforementioned acquisition specifications. These images are then sequentially processed through radiometric correction and region of interest extraction, spectral preprocessing, and characteristic wavelength screening to obtain the characteristic wavelength data of the wheat plants to be tested, ensuring that the input data format is consistent with the model training data.

[0059] Secondly, disease identification and result output are performed. The characteristic wavelength data of the wheat plants to be tested are imported into the trained early identification model. The model automatically judges the health status of the wheat plants by analyzing the spectral characteristics of the input data and outputs the identification results in the form of labels: "1" indicates that the wheat plants to be tested are healthy, and "2" indicates that the wheat plants to be tested have been infected with sheath blight.

[0060] Finally, early large-scale identification and control guidance are achieved. The model output results quickly clarify the disease status of the wheat plants under test. Data can be collected at the one-leaf, two-leaf, three-leaf, and four-leaf stages of wheat to complete the detection, realize early warning, provide technical support for targeted pesticide application in autumn, effectively block the reproduction chain of pathogens before overwintering, and achieve early and precise control of wheat sheath blight.

[0061] In addition, this embodiment also includes using a drone equipped with a hyperspectral imager to conduct large-scale field scanning, expanding the application scenarios for disease identification.

[0062] Specifically, in this embodiment, a drone equipped with an imaging hyperspectral imager is used to achieve large-scale acquisition of hyperspectral images of wheat fields. Flight routes must be planned in advance to ensure full field coverage. After the flight, the equipment automatically returns to base, improving operational efficiency. Image data acquisition should be conducted between 10:00 and 14:00, under sunny, windless, and low-cloud conditions to avoid insufficient light or airflow interference that could degrade data quality. Before each flight, a reference board is used to perform radiometric calibration on the spectrometer; that is, the reflectance of the spectral calibration board is collected before each flight to ensure consistency of data collected from different batches and regions. The drone flight parameters are set as follows: lens angle 67~88°, flight altitude 10.5~19.0m, flight speed 1.1~2.8m / s, with 80%~90% lateral overlap and 70%~85% lateral overlap to ensure complete image stitching. Markers are placed at ground sampling points to facilitate accurate matching of subsequent images with actual locations. The collected hyperspectral images of the field are processed and analyzed according to steps S2 to S6 to achieve large-scale early screening of wheat sheath blight in hectare-level fields, meeting the disease detection needs of wheat sheath blight in large-scale wheat production scenarios.

[0063] The following specific application examples will further illustrate the above content to verify the recognition effect of the early identification model for wheat sheath blight in the seedling and mature stages.

[0064] Application Example 1 Using the raw feature wavelengths selected by the PCA-loading factor method, classification models for healthy and diseased wheat in an experimental field were established using linear discriminant analysis (LDA) and support vector machine (SVM), respectively. The specific regions of interest for wheat seedlings were marked as follows: Figure 13 As shown in A and B in the table, their discrimination performance is shown in Table 4.

[0065] In addition, such as Figure 13 As shown in C, there is a significant difference in spectral reflectance at the base of the stems of healthy wheat and diseased wheat. In the range of 500-1000 nm, the spectral reflectance at the base of the stems of healthy wheat is significantly higher than that of diseased wheat.

[0066] Table 4 shows that the correct classification accuracy of both the LDA model and the SVM model on the modeling set and the prediction set is above 83%.

[0067] Table 4 Performance of LDA and SVM models for seedling wheat based on raw spectral data (Raw).

[0068] Application Example 2 Using the raw feature wavelengths selected by the PCA-loading factor method, classification models for healthy and diseased wheat were established using both Linear Discriminant Analysis (LDA) and Support Vector Machine (SVM). Specific regions of interest for mature wheat plants were marked as follows: Figure 14 As shown in A and B in the table, their discrimination performance is shown in Table 5.

[0069] In addition, such as Figure 14 As can be seen from C, the spectral curves of healthy and diseased wheat stem bases show roughly the same trend in the 400-590nm and 720-1000nm ranges. However, in the 590-720nm range, the spectral reflectance of the healthy stem base is significantly lower than that of the diseased plant, thus identifying this band as the main characteristic wavelength range for identifying wheat sheath blight in mature plants.

[0070] As can be seen from Table 5, the correct classification recognition rate of both the LDA model and the SVM model on the test set is above 95%.

[0071] Table 5 Performance of LDA and SVM models for mature wheat based on raw spectral data (Raw).

[0072] Therefore, the above-mentioned method for early identification of wheat sheath blight based on hyperspectral imaging technology solves the problem of severe time lag in the early large-area identification of wheat sheath blight. By accurately acquiring spectral data of the stem base through hyperspectral imaging technology, the disease status can be identified quickly and accurately, thus providing technical support for wheat production safety and the sustainable development of green agriculture.

[0073] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for early identification of wheat sheath blight based on hyperspectral imaging technology, characterized in that, Includes the following steps: S1. To address the disease characteristics of wheat sheath blight, which initially damages the coleoptile or young root of wheat, then rapidly spreads to the base of the aboveground stem, and in the middle and late stages damages the leaf sheath and stem, hyperspectral data were collected from different leaves, leaf sheaths, and aboveground tissues of wheat seedlings during the one-leaf to four-leaf stages. This yielded hyperspectral images of different leaves, leaf sheaths, and aboveground tissues of the stem base of wheat plants. S2. Based on the obtained hyperspectral images, the correlation between spectral signals from different parts and infection by *Rhizoctonia solani* was analyzed, and the optimal region of interest for *Rhizoctonia solani* was determined to be the aboveground part of the wheat plant stem base. S3. After determining the optimal region of interest, the hyperspectral image of the aboveground tissue at the base of the wheat plant stem is processed and the region of interest is marked. All spectral data corresponding to the region of interest are extracted and preprocessed. S4. Screen the preprocessed spectral data for characteristic wavelengths and extract characteristic wavelengths used to characterize the early damage characteristics of wheat sheath blight caused by infection with Rhizoctonia graminearum or Rhizoctonia solani. S5. Construct an early identification model for wheat sheath blight based on the extracted feature wavelengths and train it. S6. Input the characteristic wavelength data of the wheat plant to be tested into the trained recognition model, output the health status of the wheat plant, and complete the early recognition of wheat sheath blight.

2. The method for early identification of wheat sheath blight based on hyperspectral imaging technology according to claim 1, characterized in that, In S1, acquiring the hyperspectral image includes: First, a hyperspectral imager was used to acquire spectral images of the first, second, and third leaves, the leaf sheaths of the first and second leaves, and the above-ground tissues at the base of the stem of wheat plants. Then, the acquired spectral images were corrected to obtain hyperspectral images of different leaves, leaf sheaths, and above-ground tissues at the base of the stem of wheat plants.

3. The method for early identification of wheat sheath blight based on hyperspectral imaging technology according to claim 2, characterized in that, The effective spectral range of the hyperspectral imager is 400~1001nm.

4. The method for early identification of wheat sheath blight based on hyperspectral imaging technology according to claim 1, characterized in that, In S2, the above-ground part of the wheat plant stem base is the area 1-4 cm from the soil surface, and this area is the optimal region of interest for the wheat sheath blight pathogen. The above-ground part of the wheat plant stem base is the above-ground leaf sheath of the first true leaf of the wheat.

5. The method for early identification of wheat sheath blight based on hyperspectral imaging technology according to claim 1, characterized in that, In S3, the hyperspectral image of the aboveground tissue at the base of the wheat plant stem is processed and the region of interest is marked. All spectral data corresponding to the region of interest are extracted and preprocessed, specifically including: The hyperspectral image of the aboveground tissue at the base of the wheat plant stem was radiometrically corrected using the whiteboard correction method to eliminate spectral errors caused by changes in light intensity. Then, the region of interest of the herbaceous blight pathogen was marked in the hyperspectral image of the aboveground tissue of the wheat plant stem base after radiation correction, and the spectral data of the region of interest was extracted to obtain all the spectral data of the corresponding stem base tissue and preprocessed.

6. The method for early identification of wheat sheath blight based on hyperspectral imaging technology according to claim 5, characterized in that, In S3, all spectral data of the stem base tissue were preprocessed using one or more of the following methods: Savitzky-Golay smoothing, Savitzky-Golay first derivative, standard normal transformation (SNV), and multivariate scattering correction (MSC).

7. The method for early identification of wheat sheath blight based on hyperspectral imaging technology according to claim 1, characterized in that, In S4, the preprocessed spectral data is subjected to feature wavelength screening, including: using principal component analysis (PCA) with loading coefficients and / or continuous projection algorithm (SPA) to screen feature wavelengths, with a total of 100 feature wavelengths obtained.

8. The method for early identification of wheat sheath blight based on hyperspectral imaging technology according to claim 7, characterized in that, When using the principal component analysis (PCA)-loading coefficient method for screening, principal components with a contribution rate greater than 90% are selected, and the wavelengths corresponding to their peak or trough positions are extracted. The SPA continuous projection algorithm determines the optimal feature wavelength by calculating the number of feature vectors and the corresponding root mean square error (RMSE) during the screening process.

9. The method for early identification of wheat sheath blight based on hyperspectral imaging technology according to claim 1, characterized in that, In S5, an early identification model for wheat sheath blight is constructed based on the extracted feature wavelengths, and the model is trained, including: The extracted spectral data corresponding to the characteristic wavelengths are randomly divided into training set and test set according to a preset ratio, and the sample ratio of healthy plants and diseased plants in the training set and test set is the same as that of the total sample. The training set is input into either Linear Discriminant Analysis (LDA) or Support Vector Machine (SVM) algorithms to establish an early identification model for wheat sheath blight and train it, optimizing the model parameters during the training process. The test set is input into the trained early identification model of wheat sheath blight, and the model's identification accuracy is calculated to ensure that the early identification accuracy of the model is not less than 90%.

10. The method for early identification of wheat sheath blight based on hyperspectral imaging technology according to claim 1, characterized in that, In S6, the characteristic wavelength data of the wheat plant to be tested are input into the trained recognition model, and the health status of the wheat plant is output, including: The characteristic wavelength data of the wheat plants to be tested are input into the trained early identification model of wheat sheath blight; The wheat sheath blight early identification model will diagnose the health status of wheat plants, mark them as healthy or diseased, and output the early identification results.

Images