Wheat stripe rust remote sensing monitoring method based on optimized hyperspectral vegetation index

Abstract

The present application belongs to the technical field of plant disease monitoring, and more particularly relates to a wheat stripe rust remote sensing monitoring method based on optimized hyperspectral vegetation index. The method comprises: test design and sample collection, multi-scale spectral data acquisition, vegetation index screening and optimization, PLSR monitoring model construction, model verification and disease spatial inversion, and result output. The sensitive vegetation index is screened through a test field and optimized, and then the model adaptability is verified in a real field, so that the problems of significant decrease in precision when the existing wheat stripe rust hyperspectral monitoring technology is transplanted from a controllable test field to a real field, inability to cope with complex interference such as soil background heterogeneity, environmental fluctuation and agronomic condition difference, etc. are solved.

Description

Technical Field

[0001] This invention belongs to the technical field of plant disease monitoring, and more specifically, relates to a remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index. Background Technology

[0002] Wheat stripe rust is one of the most destructive airborne fungal diseases affecting wheat production globally, causing approximately 10-70% of global yield losses annually. Traditional disease monitoring relies primarily on manual field surveys, which suffer from low efficiency, high subjectivity, and limited coverage. Against this backdrop, developing rapid, accurate, and non-destructive wheat stripe rust monitoring technologies has become a current hot topic in the field of agricultural remote sensing. This invention relates to the field of agricultural remote sensing monitoring technology, specifically a remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation indices. Its core principle is to utilize the "image-spectrum integration" characteristic of hyperspectral remote sensing to capture the subtle spectral responses (400-2500 nm) of wheat after stripe rust infection, achieving accurate and rapid monitoring of wheat stripe rust at the field scale. This effectively solves the problems of untimely and low-precision stripe rust monitoring in agricultural production, providing technical support for disease control decision-making.

[0003] The closest prior art to this invention includes two main aspects: First, a hyperspectral monitoring method based on disease-specific vegetation indices (Guo, A. 2020; Ahmad, W. 2025). This method identifies characteristic spectral changes after wheat stripe rust infection (such as weakened chlorophyll absorption around 680nm and a blue shift at 700-750nm), and proposes disease-specific vegetation indices such as RI, MCARI, and YRSI. These indices are used to establish disease monitoring models, which have achieved good monitoring results in controlled-environment experimental fields. Second, a monitoring method combining a drone hyperspectral platform with machine learning (Cross, JF2024; Yashu; Kukreja, V.2025; Wang, Y.2024). This method leverages the centimeter-level spatial resolution advantage of UAV hyperspectral systems to acquire hyperspectral data of wheat canopies in the field. It introduces machine learning algorithms such as random forest (RF) to construct a classification model. In the experimental field environment, the accuracy rate of stripe rust identification can reach 92% (Haboudane, D.2002), which promotes the development of precise disease monitoring at the field scale.

[0004] However, the aforementioned background technologies share a common drawback: when transplanted from controlled experimental fields to real field environments, monitoring accuracy generally decreases by 30-50%. The core reason is that real fields present complex interference factors such as soil background heterogeneity, fluctuating meteorological conditions, and differences in wheat varieties. Existing technologies lack targeted and systematic solutions, making it difficult to adapt to the complex environment of real fields and limiting their practical application value. Summary of the Invention

[0005] The present invention aims to overcome at least one of the defects of the prior art and provide a remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index, so as to solve the problems of significant decrease in accuracy when the existing hyperspectral monitoring technology for wheat stripe rust is transplanted from controlled experimental fields to real fields, and the inability to cope with complex interferences such as soil background heterogeneity, environmental fluctuations and differences in agronomic conditions.

[0006] The detailed technical solution of this invention is as follows: A remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index, the method comprising: S1. In the experimental field and the real field, samples were collected using the five-point sampling method. The severity of disease on each leaf was recorded, and the disease index DI of each sample was quantified. S2. Ground-based portable hyperspectral instruments were used to collect ground-based hyperspectral data in the experimental field, and drone hyperspectral image data were collected in the real field using drone platforms equipped with hyperspectral imagers. S3. Based on the correlation analysis between the spectral data obtained from the experimental field and DI, original vegetation indices sensitive to wheat stripe rust were screened, and a stripe rust-specific vegetation index YRVI was designed. The calculation formula is as follows: YRVI = (R720 - R550) / (R720 + R670) R720, R550, and R670 represent the reflectivity of the 720nm, 550nm, and 670nm wavelength bands, respectively. Furthermore, taking the core bands of each original vegetation index and YRVI as the center, the band range is extended by ±2-5nm according to the resolution of the UAV hyperspectral sensor. The average reflectance of all bands within this extended band range is calculated and substituted into the original index calculation formula to generate the corresponding optimized vegetation index. S4. Construct a PLSR model based on experimental field data. Use the optimized vegetation index obtained in step S3 as the input variable. Use leave-one-out cross-validation to determine the number of latent variables in the model. Train and optimize the model by dividing the training set and validation set through stratified sampling. Validate the trained model using independent samples from real fields and spatial inversion results. S5. Apply the validated PLSR model to the UAV hyperspectral image of the field, calculate and optimize the vegetation index pixel by pixel and input it into the PLSR model to predict the spatial distribution map of the disease index DI, classify the disease level according to the DI value, and output the quantitative monitoring results and spatial distribution map.

[0007] According to a preferred embodiment of the present invention, in step S2, when the UAV platform collects UAV hyperspectral image data, the platform parameters must meet the following requirements: band range 400-1000nm, number of bands 300-350, spectral resolution 2.0-2.5nm, spatial resolution ≤10cm / pixel; the UAV flight altitude must meet 50-60m to ensure full coverage of the target area.

[0008] According to a preferred embodiment of the present invention, in step S3, the original vegetation index includes the Normalized Difference Vegetation Index (NDVI), the Structure Insensitive Pigment Index (SIPI), the Photochemical Reflectance Index (PRI), the Plant Senescence Reflectance Index (PSRI), and the Modified Simple Ratio Index (MSR).

[0009] According to a preferred embodiment of the present invention, in step S3, the method for determining the extended band range is as follows: by analyzing the correlation between the experimental field spectral data and DI, 440-600nm and 665-895nm are determined as the sensitive band ranges for stripe rust; the range is extended with the core bands of the original vegetation index and YRVI as the center, so that the extended band range covers the sensitive band ranges. The optimized vegetation indices include NDVIO, SIPIO, PRIO, PSRIO, MSRO, and YRVIO, and their calculation formulas are as follows: NDVIO = (R830±x_avg - R675±x_avg) / (R830±x_avg + R675±x_avg); SIPIO = (R800±x_avg - R445±x_avg) / (R800±x_avg + R680±x_avg); PRIO = (R570±x_avg - R531±x_avg) / (R570±x_avg + R531±x_avg); PSRIO = (R680±x_avg - R500±x_avg) / R750±x_avg; MSRO = ((R800±x_avg / R670±x_avg) - 1) / sprt((R800±x_avg / R670±x_avg) + 1); YRVIO = (R720±x_avg - R550±x_avg) / (R720±x_avg + R670±x_avg); Where Rxx±x_avg represents the average reflectance of all bands within the band range after being extended ±x nm with xx nm as the center, and the value of x ranges from 2 to 5 nm.

[0010] Preferably according to the present invention, in step S4, before constructing the PLSR model, the input variables are preprocessed by Z-score normalization, and the data outliers are removed by the box plot method.

[0011] Preferably according to the present invention, in step S4, the leave-one-out cross-validation method is used to iterate 800 - 1200 times, and the number of latent variables when the root mean square error RMSE of the leave-one-out cross-validation is the smallest is selected, and the optimal number of latent variables of the PLSR model is determined to be 6.

[0012] Preferably according to the present invention, in step S4, the model verification includes experimental field verification and large field verification; For the experimental field verification, the leave-one-out cross-validation is adopted. Each time, 1 sample is selected as the verification set, and the remaining samples are used as the training set. Repeat the number of times consistent with the sample size of the experimental field, and calculate the average R² and RMSE; The large field verification is divided into two stages: First, 30 - 50 ground measured DI samples in the large field and their corresponding optimized vegetation indices are used for independent sample quantitative verification, and the determination coefficient R² and the root mean square error RMSE are calculated; Subsequently, the model is applied to the entire large field hyperspectral image for disease spatial inversion, and 10 - 15 representative regions are selected for field verification to verify the coincidence degree between the spatial inversion result and the actual field situation.

[0013] Preferably according to the present invention, in step S5, the disease grades are classified as: healthy: DI ≤ 5%, mild infection: 5% < DI ≤ 20%, moderate infection: 20% < DI ≤ 50%, severe infection: DI > 50%.

[0014] Compared with the prior art, the beneficial effects of the present invention are: (1) The present invention designs a novel stripe rust specific vegetation index YRVI, which integrates the core spectral responses at different stages of the disease, simultaneously reflects the degree of red edge migration, captures the enhanced scattering signal after the formation of uredinia, and characterizes the degree of chlorophyll loss. The three work together to achieve multi-dimensional disease feature capture, which is more specific than a single-dimensional index.

[0015] (2) Based on the analysis of the sensitive bands in the experimental field, according to the sensor resolution, the core bands of the original index are extended by ±2 - 5 nm, and the average reflectance within the extended bands is calculated to generate an optimized index, which effectively filters out spectral noise, separates the soil and disease signals. The accuracy of the optimized vegetation index in the experimental field is significantly improved, and the cross-scale monitoring stability is significantly enhanced.

[0016] (3) The present invention designs a cross-scale framework of “experimental field screening and optimization + field verification and application”. The experimental field screens and optimizes the sensitivity index, and the field verifies the adaptability of the model, forming a complete technology link. This solves the pain point of traditional technology being “good in the laboratory but ineffective in the field”, realizes stable technology transfer from experimental field to commercial field, and provides support for the practical application of hyperspectral technology.

[0017] (4) The present invention adopts the partial least squares regression (PLSR) model. By stratified sampling, the distribution of disease levels in the training set and the validation set is consistent. Combined with leave-one-out cross-validation (LOOCV) iterations of 800-1200 times, 3-7 optimal latent variables are determined. The model is robust and can reduce overfitting. The prediction error (RMSE) of wheat stripe rust DI is ≤0.16, and the monitoring results are reliable. Attached Figure Description

[0018] Figure 1 This is a flowchart of the remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index as described in this invention.

[0019] Figure 2 This is a comparison diagram of the spectral characteristics of healthy and diseased wheat in an embodiment of the present invention.

[0020] Figure 3 This is a correlation analysis diagram between the optimization index YRVIO and the disease index DI in an embodiment of the present invention.

[0021] Figure 4 This is a spatial inversion result diagram of wheat stripe rust in an embodiment of the present invention. Detailed Implementation

[0022] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0023] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of this disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure pertains.

[0024] This invention provides a cross-scale remote sensing monitoring scheme for wheat stripe rust based on optimized hyperspectral vegetation index. The core of this scheme is to optimize the vegetation index through collaborative research between experimental fields and large fields, and combine it with a UAV hyperspectral system and a partial least squares regression (PLSR) model to achieve accurate monitoring of wheat stripe rust in complex field environments.

[0025] The overall technical route of the present invention is: experimental design and sample collection → multi-scale spectral data acquisition → vegetation index screening and optimization → PLSR monitoring model construction → model verification and disease spatial inversion → result output. By screening and optimizing sensitive vegetation indices in experimental fields (controlled environments), and then verifying the model adaptability in large fields (complex environments), the technical transformation from experimental fields to real large field plots is ultimately achieved, solving the problems of low efficiency of traditional monitoring methods and easy interference of spectral signals.

[0026] The following further illustrates the wheat stripe rust remote sensing monitoring method based on optimized hyperspectral vegetation indices of the present invention in combination with specific embodiments.

[0027] Example 1 Refer Figure 1 , this example provides a remote sensing monitoring method for wheat stripe rust in large fields based on optimized hyperspectral vegetation indices, and the method includes: (I) Experimental design and sample collection 1. Selection of experimental sites Experimental fields (model training and index screening areas): The selected site needs to meet the temperate monsoon climate, with uniform soil texture and no obvious terrain undulations. An agricultural research base or a standardized experimental field can be selected.

[0028] Large fields (model verification and application areas): The selected site is a large-scale wheat planting area with a natural disease occurrence background, and the soil type and wheat planting density are the same as those in the local main production areas.

[0029] 2. Wheat varieties and planting conditions Variety selection: It is preferred to select the local main cultivated wheat varieties with a medium-susceptible or susceptible level of resistance to stripe rust (to facilitate the formation of a gradient disease level after natural disease occurrence or artificial inoculation).

[0030] Planting conditions: Follow the local conventional agricultural management measures, with a sowing density of 200 - 300 kg / hm², and consistent fertilization and irrigation conditions to avoid spectral interference caused by management differences.

[0031] 3. Disease treatment and sample collection (1) Disease induction in experimental fields The artificial inoculation method is used to simulate natural disease occurrence. The inoculation pathogen source is the pathogen of wheat stripe rust, and the inoculation time is from the jointing stage to the booting stage of wheat.

[0032] Inoculation method: It is carried out in accordance with the agricultural industry standard NY / T 1443.1 - 2007. The spore suspension is evenly sprayed on the wheat leaves. After inoculation, the relative humidity in the field is maintained at 80% - 90% and the temperature is 15 - 20 °C for 3 - 5 days to promote spore germination.

[0033] (2) Sample collection and quantification of disease levels Sampling time: Select the wheat grain-filling period (peak period of disease occurrence), collect samples 2-3 times, with an interval of 7-10 days between each collection, and collect samples from 10:30 to 14:00 (Beijing time). Meteorological conditions must be cloudless and wind force ≤ level 3.

[0034] Sampling method: The five-point sampling method was adopted. 5-8 sampling points were set up in the experimental field and 10-15 sampling points were set up in the field. A 1m² quadrat was selected at each sampling point. 20 wheat plants were randomly selected in the quadrat, and the severity of disease on the leaves of each plant was recorded.

[0035] Disease severity quantification: According to the national standard GB / T 15795-2011, the severity of diseases is divided into 9 levels (0%, 1%, 10%, 20%, 30%, 45%, 60%, 80%, 100%), and the disease index (DI) is calculated using the following formula:

[0036] Where x is the disease gradient level value, n is the highest gradient value (9), and f is the number of leaves corresponding to each gradient.

[0037] Sample size range: 60-80 samples in the experimental field and 150-160 samples in the field to ensure even distribution of samples for each disease level (healthy, mild, moderate, severe).

[0038] (II) Acquisition of multi-scale hyperspectral data 1. Ground hyperspectral data acquisition in the experimental field Data acquisition equipment: A ground-based portable hyperspectral analyzer (such as the ASD FieldSpec series) is used. The core parameters must meet the following requirements: band range of 350-2500nm, spectral resolution of 3-10nm (350-1000nm range) and 8-15nm (1001-2500nm range), and sampling interval of 1.0-1.5nm.

[0039] Data Acquisition Method: The sensor probe was positioned vertically downwards, 0.5-1.0 m away from the wheat canopy, ensuring complete canopy coverage without soil exposure. Whiteboard calibration was performed before acquisition, and recalibrated after every 10 samples. Ten spectra were acquired for each sample, with 20 dark current calibrations performed simultaneously. The average value was used as the final spectral data. Effective spectral band selection: The 2501-2500 nm band, which exhibits significant noise, was removed, retaining the effective 350-1800 nm band.

[0040] 2. Daejeon UAV hyperspectral data acquisition Data acquisition equipment: UAV platform (such as DJI M600, Matrice 300 RTK and other multi-rotor drones), equipped with a hyperspectral imager (such as Resonon Pika L, Headwall Nano-Hyperspec, etc.), the core parameters must meet: band range 400-1000nm, number of bands 300-350, spectral resolution 2.0-2.5nm, spatial resolution ≤10cm / pixel.

[0041] Data collection parameters: flight altitude 50-60m, flight speed 5-8m / s, flight path overlap rate 75%, lateral overlap rate 75%, ensuring full coverage of the target area. Data collection time is consistent with the ground spectrum (10:30-14:00), and meteorological conditions are cloudless and wind force ≤3 to avoid spectral distortion caused by changes in illumination.

[0042] Data preprocessing: Radiometric calibration was performed using the sensor's accompanying software to eliminate sensor response differences. Geometric correction, image registration, and data format conversion were completed using software such as ArcGIS and ENVI, without additional band filtering.

[0043] (III) Vegetation Index Screening and Optimization 1. Screening based on original vegetation indices First, five core indices related to the severity of wheat stripe rust were selected from existing validated crop disease monitoring indices, covering different stages of the disease (early to late stages) and different physiological responses (pigmentation changes to structural damage), as shown in Table 1: Table 1 Original Vegetation Index Table

[0044] Screening criteria: The selected index must be verified by literature and have a correlation coefficient ≥0.5 with the severity of wheat stripe rust. It can be replaced with other disease-related indices with equivalent response characteristics according to actual monitoring needs. Secondly, because existing disease-related indices (such as YRSI, RI, and MCARI) only focus on a single physiological response (e.g., YRSI focuses on spore characteristics, and RI focuses on chlorophyll degradation), their band combinations are limited and they do not simultaneously integrate multi-dimensional spectral signals of red-edge migration, chlorophyll loss, and spore mass formation. Their anti-interference ability is limited. However, the disease process of wheat stripe rust is accompanied by a chain of physiological changes: "chlorophyll degradation → red-edge blue shift → spore mass formation." A single-dimensional index cannot fully capture the disease characteristics, which leads to susceptibility to soil and light interference in the field environment, resulting in a significant decrease in accuracy. Therefore, based on the correlation analysis between the 350-1149nm spectrum of the experimental field and the disease index DI, three stripe rust-specific sensitive bands were selected: the spore heap scattering characteristic band at 550nm, the chlorophyll strong absorption band at 670nm, and the red-edge core response band at 720nm. Based on this, the YRVI (Yellow Rust Vegetation Index) was designed, and the calculation formula is as follows: YRVI = (R720 - R550) / (R720 + R670) Among them, R720 reflects the degree of red edge migration (the disease causes the red edge to shift to blue, and the reflectance of R720 decreases), R550 captures the enhanced scattering signal after the formation of spore heaps, and R670 characterizes the degree of chlorophyll loss. The three work together to achieve multi-dimensional disease feature capture.

[0045] 2. Vegetation Index Optimization Methods To address disturbances such as soil background heterogeneity and light fluctuations in the field environment, the above six original indices were optimized based on the analysis results of sensitive bands in the experimental field. The core steps are as follows: Sensitive band determination: Through correlation analysis between spectral data of 350-1149nm from the experimental field and DI, the 440-600nm chlorophyll absorption region and the 665-895nm leaf structure scattering region were determined to be the sensitive band range for stripe rust.

[0046] Core band extension: Centered on the core band of the original index, extend the range by ±2-5nm (adjusted according to the resolution of the UAV hyperspectral sensor; for example, if the sensor resolution is 2.1nm, ±4.2nm can be used, corresponding to 2 band widths), covering the sensitive band range.

[0047] Optimized index calculation: Calculate the average reflectance of all bands within the extended band range, substitute it into the original index formula, and generate the optimized vegetation index, as shown in the following formula: NDVIO: (R830±4.2_avg - R675±4.2_avg) / (R830±4.2_avg + R675±4.2_avg) SIPIO: (R800±4.2_avg - R445±4.2_avg) / (R800±4.2_avg + R680±4.2_avg) PRIO: (R570±4.2_avg - R531±4.2_avg) / (R570±4.2_avg + R531±4.2_avg) PSRIO: (R680±4.2_avg - R500±4.2_avg) / (R750±4.2_avg ) MSRO: ((R800±4.2_avg) / (R670±4.2_avg)-1) / sprt((R800±4.2_avg) / (R670±4.2_avg) + 1) YRVIO=(R720±4.2_avg - R550±4.2_avg) / (R720±4.2_avg + R670±4.2_avg)) Where Rxx±x_avg represents the average reflectance of all bands within the range of x nm above and below the band xx nm.

[0048] Optimize parameter range: The extended band width can be adjusted according to the sensor's spectral resolution. For a resolution of 1-2nm, use ±2-3nm; for a resolution of 2-3nm, use ±3-5nm. The core is to ensure that the extended band can cover the sensitive range and effectively reduce noise.

[0049] (iv) Construction and validation of PLSR monitoring model 1. Model Input and Data Preprocessing Input variables: Calculation results of 6 optimized vegetation indices (NDVIO, SIPIO, PRIO, PSRIO, MSRO, YRVIO).

[0050] Output variable: Wheat stripe rust DI (0-100).

[0051] Data preprocessing: Z-score standardization was used to normalize the input variables and eliminate differences in units; outliers were removed using box plots to ensure model stability.

[0052] 2. PLSR Model Parameter Settings Number of latent variables: Based on the numerical distribution of YRVIO (range 0.2-0.8) and the disease response pattern, the number of latent variables was determined by leave-one-out cross-validation (LOOCV). The optimal number of latent variables was determined to be 6, which is more in line with the characteristic dimensions of YRVIO than the 5 latent variables of traditional index fitting.

[0053] Dataset partitioning: Stratified sampling was used to divide the experimental field data into training and validation sets in a 7:3 ratio, while the field data was used directly as an independent validation set to ensure that the disease severity distribution of each dataset was consistent.

[0054] Number of iterations: The model is trained for 800-1200 iterations to ensure model convergence.

[0055] 3. Model Validation Methods Verification metrics: The coefficient of determination R² and the root mean square error RMSE are used as the core verification metrics. The closer R² is to 1 and the closer RMSE is to 0, the better the model performance. Among them, for the experimental field model, R² ≥ 0.55 and RMSE ≤ 0.15, and for the large field model, R² ≥ 0.50 and RMSE ≤ 0.16 are the qualified standards.

[0056] Verification method: 1. Experimental field verification: LOOCV is adopted. Each time, 1 sample is selected as the verification set, and the remaining samples are used as the training set. Repeat 60 - 80 times (the same as the sample size of the experimental field), and calculate the average R² and RMSE.

[0057] 2. Large field verification: Large field verification is divided into two progressive stages: First is the independent sample quantification verification: Select 30 - 50 ground measured DI samples, extract the optimized vegetation indices corresponding to the UAV hyperspectral data, such as YRVIO, substitute them into the model trained in the experimental field to obtain the predicted DI, and verify the prediction accuracy of the model at a single point in the large field by calculating the coefficient of determination R² and the root mean square error RMSE. Secondly is the spatial inversion and field verification: After the model passes the quantification verification, apply it to the entire large field UAV hyperspectral data, inversely retrieve the spatial distribution map of wheat stripe rust pixel by pixel; then select 10 - 15 representative regions including different disease grades and boundary regions, carry out field verification with GPS positioning, confirm the coincidence degree between the inversion result and the actual disease occurrence in the field, and complete the spatial accuracy verification of the whole field. Only when both stages of verification pass, the large field verification is considered to pass completely, and finally the trained PLSR model is obtained.

[0058] (V) Disease spatial inversion and result output 1. Spatial inversion process Extract the optimized vegetation indices from the UAV hyperspectral image pixel by pixel, substitute them into the trained PLSR model, predict the DI value pixel by pixel, and divide the disease grades according to the DI range: healthy (DI ≤ 5%), mildly infected (5% < DI ≤ 20%), moderately infected (20% < DI ≤ 50%), severely infected (DI > 50%), and generate the disease spatial distribution map.

[0059] 2. Derived results obtained by further processing, statistics or integration using the predicted DI values output by the model: Quantitative results: Output the model performance metrics R² and RMSE of each optimized index, and the recognition accuracy of different disease grades. The R² and RMSE are obtained by statistical fitting of the "predicted DI value" output by the model and the ground "measured DI value", and calculated through formulas, which are used to evaluate the model accuracy. The accuracy rate of different disease levels is determined by classifying the levels into healthy, mild, moderate, and severe based on the range of predicted DI values, and by statistically analyzing the percentage of samples whose predicted levels match the actual levels.

[0060] Spatial results: The PLSR model was applied to the whole field UAV hyperspectral imagery, and the predicted DI values ​​were output pixel by pixel. TIFF format raster maps were generated according to the grade coding (green / yellow-green / yellow / red) to mark the spatial range and area proportion of each disease grade. Application Results: Generates disease monitoring reports, integrates the above quantitative and spatial results, and combines them with agricultural plant protection and control standards to extract disease severity, key control areas, and pesticide / management recommendations, forming a directly applicable report that includes disease occurrence severity, core disease areas, and control recommendations, supporting the application of intelligent agricultural pest and disease monitoring and early warning systems.

[0061] Example 2 This embodiment applies the wheat stripe rust remote sensing monitoring method based on optimized hyperspectral vegetation index described in Embodiment 1 to Luyuan 502, specifically including: 1. Experimental materials and equipment Wheat variety: Luyuan 502 (moderately susceptible to stripe rust) Experimental field: Shandong Academy of Agricultural Sciences Experimental Base (116°58′43″E, 36°59′1″N), area 300m², soil texture loam, pH 7.2, sown in October 2022, density 270kg / hm².

[0062] Datian: A real wheat field in Heze City, Shandong Province (116°3′12″E, 35°27′53″N), covering an area of ​​3000m², with natural disease, wheat variety Luyuan 502, and planting conditions consistent with the main producing areas in the region.

[0063] Instruments and equipment: Ground-based spectrometer: ASD FieldSpec (band 350-2500nm, spectral resolution 3nm / 8nm, sampling interval 1.4nm / 1.1nm).

[0064] UAV platform: DJI M600 multi-rotor drone, equipped with Resonon Pika L hyperspectral imager (400-1000nm bands, 332 bands, spectral resolution 2.1nm).

[0065] Auxiliary equipment: GPS locator (accuracy ±1m), whiteboard (reflectivity 99%), Python 3.9.12 (scikit-learn 1.2.2), Excel 365v2308, ArcGIS 10.5.

[0066] Reagents and standards: Wheat stripe rust pathogen spore suspension (concentration 3×10⁻⁶) 5 (number of cells / mL), conforming to the NY / T1443.1-2007 inoculation standard; disease surveys follow GB / T 15795-2011.

[0067] 2. Experimental Procedure (1) Disease induction and sample collection in experimental fields Artificial inoculation was carried out on April 10, 2022, by spraying the spore suspension evenly. After inoculation, the field humidity was maintained at 85% and the temperature at 18℃ for 4 consecutive days.

[0068] (2) Sample collection from experimental fields and field sites and quantification of disease severity The five-point sampling method was used, with 6 sampling points set up. Each sampling point was a 1m² quadrat with 20 plants per quadrat. The disease severity was recorded, and the disease intensity (DI) was calculated. A total of 68 experimental field samples were obtained (20 healthy, 17 mild, 16 moderate, and 15 severe).

[0069] (3) Multi-scale hyperspectral data acquisition Experimental field: Ground-based spectrometers were used to collect near-ground canopy hyperspectral data of wheat stripe rust. Spectral data were collected on May 6, 16, and 23, 2022 (10:30-14:00, cloudless, wind force 2). Whiteboard calibration: Calibrate once for every 10 samples collected, scan each sample 10 times, and perform 20 simultaneous dark current calibrations, taking the average value.

[0070] Datian: Using a UAV platform to collect hyperspectral data of wheat stripe rust canopy; UAV flight to take place on May 19, 2022 (10:30-12:00, clear skies, wind force 1): Flight parameters: Altitude 50m, speed 6m / s, flight path overlap 80%, lateral overlap 60%.

[0071] Data preprocessing: SBGcenter 3.5.498 radiometric calibration -- AirlineDivision 1.7 geometric correction -- ArcGIS 10.5 image registration and format conversion.

[0072] Ground validation: Five-point sampling method, 15 sampling points, 1m² per quadrat, 20 plants / quadrat, DI calculated, a total of 155 field samples were obtained (35 healthy, 40 mild, 50 moderate, and 30 severe).

[0073] (4) Vegetation index screening Five indices associated with the severity of wheat stripe rust were screened: NDVI, SIPI, PRI, PSRI, and MSR. Then, based on the correlation analysis between the 350-1149nm spectrum of the experimental field and the disease index DI, three stripe rust-specific sensitive bands were selected: the spore heap scattering characteristic band is 550nm, the chlorophyll strong absorption band is 670nm, and the red edge core response band is 720nm. Based on this, the YRVI (Yellow Rust Vegetation Index) was designed. The YRVI index integrates the core spectral responses of different stages of disease and is more specific than single-dimensional indices. The R² in the experimental field reached 0.6325, and the R² in the field reached 0.5917. The cross-scale accuracy decreased by only 6.5%, and the resistance to soil and light interference was improved by 12%-15% compared with NDVI.

[0074] (5) Optimization of vegetation index Sensitive band identification: Through correlation analysis between experimental field spectral data and DI, 440-600nm and 665-895nm were identified as sensitive bands; Figure 2 The heatmap shows the correlation between the measured wheat canopy spectral reflectance and DI in the experimental field. It indicates the correlation between the severity of wheat stripe rust infection and the corresponding canopy spectral reflectance data, with orange-red indicating high correlation and green indicating low correlation. Figure 2 As can be seen, strong correlations were observed in the wavelength ranges of 440-600nm and 665-895nm. Therefore, these regions were identified as sensitive spectral bands for the detection of wheat stripe rust.

[0075] Optimization parameters: Core band extension ±4.2nm (corresponding to 2 band widths of the Pika L sensor), calculate the average reflectivity of the extended band, and generate 6 optimization indices: NDVIO: (R830±4.2_avg - R675±4.2_avg) / (R830±4.2_avg + R675±4.2_avg) SIPIO: (R800±4.2_avg - R445±4.2_avg) / (R800±4.2_avg + R680±4.2_avg) PRIO: (R570±4.2_avg - R531±4.2_avg) / (R570±4.2_avg + R531±4.2_avg) PSRIO: (R680±4.2_avg - R500±4.2_avg) / (R750±4.2_avg ) MSRO: ((R800±4.2_avg) / (R670±4.2_avg)-1) / sprt((R800±4.2_avg) / (R670±4.2_avg) + 1) YRVIO=(R720±4.2_avg - R550±4.2_avg) / (R720±4.2_avg + R670±4.2_avg)); Where Rxx±x_avg represents the average reflectance of all bands within the range of x nm above and below the core band.

[0076] (6) PLSR model construction and validation Data preprocessing: Z-score standardization to optimize the index, and box plot method to remove two outliers (one from the experimental field and one from the main field).

[0077] Model parameters: Stratified sampling was used to divide the training set (156 samples) and the validation set (67 samples), LOOCV was iterated 1200 times, and 6 latent variables were determined.

[0078] Validation methods: LOOCV validation was used in the experimental field, and 30 independent samples were used in the field validation.

[0079] 3. Experimental Results and Analysis Test field performance: YRVIO's R²=0.6118 and RMSE=0.1404 are better than other optimization indices (NDVIO R²=0.5144, PRIO R²=0.5907).

[0080] Daejeon performance: Figure 3 To validate the results using field data, the horizontal axis represents the predicted DI value, and the vertical axis represents the measured DI value; the scatter plots represent 155 field samples; the blue dashed line is the fitted regression line, labeled with the fitted equation (y=6.4698x-276.47), coefficient of determination (R²=0.5713), and root mean square error (RMSE=0.1512). As shown in the figure, YRVIO has R²=0.5713 and RMSE=0.1512. Validation with 30 independent samples indicates that the accuracy rate for identifying heavily infected areas (DI>50%) is 85.2%, which is 15-25 percentage points higher than traditional UAV monitoring methods (accuracy rate 60-70%).

[0081] Spatial inversion: Figure 4This is the spatial inversion result map of wheat stripe rust. In the upper part of the figure is the visible light image taken by the drone, and in the lower part is the YRVIO spatial inversion result map of wheat stripe rust, which uses four-color coding: green = healthy (DI ≤ 5%), yellow-green = mildly infected (5% < DI ≤ 20%), yellow = moderately infected (20% < DI ≤ 50%), red = severely infected (DI > 50%); the results of the YRVIO spatial inversion result map of wheat stripe rust are in 83.8% agreement with the ground measured disease occurrence areas, and the healthy areas in the figure have the highest matching degree with the visible light image taken by the drone.

[0082] Advantage verification: (1)The monitoring accuracy is significantly improved, and the cross-scale stability is better than traditional technologies a. Technical principle The optimized vegetation index is designed by "core band ± 2 - 5nm extended average", which not only retains the disease-sensitive spectral characteristics but also reduces the single-band noise; the PLSR model combines stratified sampling and LOOCV verification to reduce the risk of overfitting and ensure the generalization ability of the model.

[0083] b. Experimental data and comparison Accuracy in the experimental field: For the optimized optimal index YRVIO, R² = 0.6118 and RMSE = 0.1404. Compared with traditional disease indices (such as R² ≈ 0.55 for RI and R² ≈ 0.53 for MCARI), the accuracy is improved by 11% - 15%; Accuracy in the large field: In the large field (n = 155), for the optimized YRVIO, R² = 0.5713 and RMSE = 0.1512. The accuracy only drops by 6.6% compared with that in the experimental field, while the accuracy of traditional technologies generally drops by 30 - 50% when transplanted from the experimental field to the large field; Identification of severely infected areas: The recognition accuracy of YRVIO for severely infected areas (DI > 50%) reaches 85.2%, while the accuracy of traditional UAV monitoring methods is 60 - 70%, which is 15 - 25 percentage points higher than traditional UAV monitoring methods.

[0084] c. Effect conclusion High-precision and stable monitoring from the experimental field to commercial large fields has been achieved, solving the key problem of "sharp decline in scale transfer accuracy" of traditional technologies.

[0085] (2)The anti-interference ability is enhanced, adapting to complex field environments a. Technical principle Soil background interference: Through the extended average of sensitive bands (440 - 600nm, 665 - 895nm), the spectral signals of soil and crop diseases are separated, reducing the interference of soil signals (accounting for up to 40%) under low coverage; Illumination fluctuation interference: UAV hyperspectral data is preprocessed through standardization (radio calibration, geometric correction), and combined with the band integration characteristics of the optimization index, the influence of illumination fluctuation within 15% is suppressed; Variety-specific interference: Cross-scale samples (Luyuan 502 in the experimental field and the local main cultivated varieties in the field) verified that the optimization index focuses on disease-specific physiological signals, such as the destruction of the spore wax layer and chlorophyll degradation, and weakens the influence of differences in the health spectrum among varieties.

[0086] b. Experimental Data and Comparison Soil disturbance scenario: In low vegetation cover areas with a coverage of <60%, YRVIO's R² = 0.52, an improvement of 36.8% compared to the original YRVI's R² = 0.38; Illumination fluctuation scenario: The RMSE fluctuation of DI retrieved from spectral data during different acquisition periods (10:30-14:00) is 0.01-0.03, which is much lower than the traditional index (RMSE fluctuation 0.08-0.12).

[0087] c. Conclusion of Results This invention effectively resists complex field interferences such as soil, light, and variety, solves the pain point of "poor environmental adaptability" of traditional technologies, and ensures the reliability of monitoring in actual production scenarios.

[0088] (3) The monitoring efficiency and practicality have been greatly improved, making it suitable for agricultural production needs. a. Technical principles The UAV hyperspectral platform (flying at an altitude of 50-60m) achieves centimeter-level spatial resolution, and a single flight can cover 1000-5000m² of field, improving data acquisition efficiency by 50-100 times compared to manual surveys. The standardized process (data preprocessing → index calculation → model inversion → result output) can be automated using Python and Excel, eliminating the need for complex professional operations and lowering the application threshold.

[0089] b. Experimental Data and Comparison Manual survey: It takes 2 people 4 hours to survey a 1000m² field, with a survey coverage of 30%, and is highly subjective; The method of this invention: One person operates a UAV for 30 minutes, covering the entire field, and the consistency between the inversion results and the ground-measured DI is over 80%.

[0090] c. Conclusion of Results This invention significantly improves monitoring efficiency and reduces labor costs, while achieving full field coverage and objective quantification. It solves the problems of "inefficiency, subjectivity, and limited coverage" of traditional manual monitoring and meets the actual needs of precise prevention and control in smart agriculture.

[0091] (4) Summary of core effects This invention directly achieves the technical effects of "stable accuracy, strong anti-interference, high efficiency, and easy expansion" through three major technical features: vegetation index optimization, cross-scale framework construction, and UAV and PLSR synergy. Its core advantages are reflected in: In terms of accuracy: the accuracy reduction in cross-scale monitoring was only 6.6%, and the accuracy rate in heavily infected areas was 85.2%, far exceeding that of traditional technologies; In terms of environmental adaptability: It effectively suppresses disturbances from soil, light, and varieties, and is suitable for complex field environments; In terms of practicality: the monitoring efficiency is improved by more than 50 times, the operation is simple, and it can be directly applied to agricultural production; In terms of scalability: the methodology framework can be transferred to various crop diseases, and its application value is wide-ranging.

[0092] This invention provides key technical support for the transition of hyperspectral technology from the laboratory to actual production, promotes precise control of wheat stripe rust, and helps ensure food security.

[0093] Example 3 This embodiment also provides a remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index applied to Luyuan 502. The difference from Embodiment 2 is that the core band extension width is adjusted to ±3nm (corresponding to 1.4 band widths of the Pika L sensor). The final experimental results and analysis are as follows: Test field performance: YRVIO's R²=0.5824, RMSE=0.1536, a decrease of 4.8% compared to Example 2 (R²=0.6118).

[0094] Performance in the field: YRVIO has an R² of 0.5347, an RMSE of 0.1628, and an accuracy of 79.6% in identifying heavily infected areas.

[0095] Conclusion: The extended band width is too narrow (±3nm), failing to fully cover the sensitive band, resulting in insufficient noise immunity and performance slightly below optimal conditions.

[0096] Example 4 This embodiment also provides a remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index applied to Luyuan 502. The difference from Embodiment 2 is that the core band extension width is adjusted to ±5nm (corresponding to 2.4 band widths of the Pika L sensor). The final experimental results and analysis are as follows: Test field performance: YRVIO's R²=0.5917, RMSE=0.1483, a decrease of 3.3% compared to Example 1 (R²=0.6118).

[0097] Performance in the field: YRVIO has an R² of 0.5482, an RMSE of 0.1581, and an accuracy of 81.3% in identifying heavily infected areas.

[0098] Conclusion: The excessively wide extended band width (±5nm) introduces noise in insensitive bands, resulting in a slight decrease in model accuracy. It is still better than the traditional exponential model, but not as good as the optimal extended band width (±4.2nm).

[0099] Example 5 This embodiment also provides a remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index applied to Luyuan 502. The difference from Embodiment 2 is that the UAV flight altitude is adjusted to 40m, while other flight parameters remain unchanged. After data preprocessing, the spatial resolution of the UAV image is improved to 8cm / pixel (10cm / pixel in Embodiment 2). The final experimental results and analysis are as follows: Performance in the field: YRVIO has an R² of 0.5628, an RMSE of 0.1547, and an accuracy of 83.7% in identifying heavily infected areas.

[0100] Spatial inversion: The boundary identification of the disease area is more refined, but the amount of data increases by 30% and the preprocessing time is extended by 25%.

[0101] Conclusion: Lowering the flight altitude can improve spatial resolution, but it increases data processing costs. The monitoring accuracy is close to that of the optimal altitude (50m), and it can be considered as an alternative.

[0102] Example 6 This embodiment also provides a remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index applied to Luyuan 502 wheat variety. The difference from Embodiment 2 is that the sample size is reduced to: Experimental field sample: 50 (15 healthy, 12 mild, 13 moderate, and 10 severe); Daegu sample: 120 (28 healthy, 32 mild, 40 moderate, and 20 severe); The model parameters were changed to: 119 samples in the training set, 51 samples in the validation set, 1000 LOOCV iterations, and 5 latent variables. The final experimental results and analysis are as follows: Test field performance: YRVIO's R²=0.5532, RMSE=0.1641, a decrease of 9.6% compared to Example 1 (R²=0.6118).

[0103] Performance in the field: YRVIO has an R² of 0.5129, an RMSE of 0.1734, and an accuracy of 78.5% in identifying heavily infected areas.

[0104] Conclusion: The reduction in sample size leads to a decrease in the model's generalization ability and accuracy, but it still meets the needs of field monitoring (R²≥0.5) and is suitable for scenarios where sample collection is difficult.

[0105] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solutions of the present invention, and are not intended to limit the specific implementation of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the claims of the present invention should be included within the protection scope of the claims of the present invention.

Claims

1. A remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index, characterized in that, The method includes: S1. In the experimental field and the real field, samples were collected using the five-point sampling method. The severity of disease on each leaf was recorded, and the disease index DI of each sample was quantified. S2. Ground-based portable hyperspectral instruments were used to collect ground-based hyperspectral data in the experimental field, and drone hyperspectral image data were collected in the real field using drone platforms equipped with hyperspectral imagers. S3. Based on the correlation analysis between the spectral data obtained from the experimental field and DI, original vegetation indices sensitive to wheat stripe rust were screened, and a stripe rust-specific vegetation index YRVI was designed. The calculation formula is as follows: YRVI = (R720 - R550) / (R720 + R670); R720, R550, and R670 represent the reflectivity of the 720nm, 550nm, and 670nm wavelength bands, respectively. Furthermore, taking the core bands of each original vegetation index and YRVI as the center, the band range is extended by ±2-5nm according to the resolution of the UAV hyperspectral sensor. The average reflectance of all bands within this extended band range is calculated and substituted into the original index calculation formula to generate the corresponding optimized vegetation index. S4. Construct a PLSR model based on experimental field data. Use the optimized vegetation index obtained in step S3 as the input variable. Use leave-one-out cross-validation to determine the number of latent variables in the model. Train and optimize the model by dividing the training set and validation set through stratified sampling. Validate the trained model using independent samples from real fields and spatial inversion results. S5. Apply the validated PLSR model to the UAV hyperspectral image of the field, calculate and optimize the vegetation index pixel by pixel and input it into the PLSR model to predict the spatial distribution map of the disease index DI, classify the disease level according to the DI value, and output the quantitative monitoring results and spatial distribution map.

2. The remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index according to claim 1, characterized in that, In step S2, when the UAV platform collects hyperspectral image data, the platform parameters must meet the following requirements: band range 400-1000nm, number of bands 300-350, spectral resolution 2.0-2.5nm, spatial resolution ≤10cm / pixel; the UAV flight altitude must meet 50-60m to ensure full coverage of the target area.

3. The remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index according to claim 1, characterized in that, In step S3, the original vegetation indices include the Normalized Difference Vegetation Index (NDVI), the Structure Insensitive Pigment Index (SIPI), the Photochemical Reflectance Index (PRI), the Plant Senescence Reflectance Index (PSRI), and the Modified Simple Ratio Index (MSR).

4. The remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index according to claim 3, characterized in that, In step S3, the method for determining the extended band range is as follows: by analyzing the correlation between the experimental field spectral data and DI, 440-600nm and 665-895nm are determined as the sensitive band ranges for stripe rust; the range is extended with the core bands of the original vegetation index and YRVI as the center, so that the extended band range covers the sensitive band ranges. The optimized vegetation indices include NDVIO, SIPIO, PRIO, PSRIO, MSRO, and YRVIO, and their calculation formulas are as follows: NDVIO = (R830±x_avg - R675±x_avg) / (R830±x_avg + R675±x_avg); SIPIO = (R800±x_avg - R445±x_avg) / (R800±x_avg + R680±x_avg); PRIO = (R570±x_avg - R531±x_avg) / (R570±x_avg + R531±x_avg); PSRIO = (R680±x_avg - R500±x_avg) / R750±x_avg; MSRO = ((R800±x_avg / R670±x_avg) - 1) / sprt((R800±x_avg / R670±x_avg) + 1); YRVIO = (R720±x_avg - R550±x_avg) / (R720±x_avg + R670±x_avg); Where, Rxx±x_avg represents the average reflectance of all bands within the wavelength range expanded by ±x nm centered at xx nm, and the value range of x is 2 - 5 nm.

5. The remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index according to claim 1, characterized in that, In step S4, before constructing the PLSR model, the input variables are preprocessed by Z - score standardization, and the box - plot method is used to remove data outliers.

6. The remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index according to claim 1, characterized in that, In step S4, the leave - one - out cross - validation method is used to iterate 800 - 1200 times, and the number of latent variables is selected when the root mean square error RMSE of the leave - one - out cross - validation is the smallest. The optimal number of latent variables of the PLSR model is determined to be 6.

7. The remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index according to claim 1, characterized in that, The model validation includes experimental field validation and large - field validation; The experimental field validation uses the leave - one - out cross - validation. Each time, 1 sample is selected as the validation set, and the remaining samples are used as the training set. Repeat the number of times consistent with the sample size of the experimental field, and calculate the average R² and RMSE; The large - field validation is divided into two stages: First, 30 - 50 ground - measured DI samples in the large - field and their corresponding optimized vegetation indices are used for independent sample quantitative validation, and the determination coefficient R² and the root mean square error RMSE are calculated. Subsequently, the model is applied to the entire large - field hyperspectral image for disease spatial inversion, and 10 - 15 representative regions are selected for on - site verification to verify the coincidence degree between the spatial inversion result and the actual field situation.

8. The remote sensing monitoring method for wheat stripe rust based on optimized hyperspectral vegetation index according to claim 3, characterized in that, In step S5, the disease levels are classified as: healthy: DI≤5%, mild infection: 5% < DI≤20%, moderate infection: 20% < DI≤50%, severe infection: DI > 50%.

Images