Hyperspectrum-based intelligent monitoring method and system for rice false smut

Abstract

The invention relates to the technical field of agricultural disease monitoring, in particular to an intelligent monitoring method and system for rice false smut based on hyperspectrum. The method comprises the following steps: acquiring point spectrum data of rice ears from a rice jointing and ear growing period to a mature period through a ground portable surface feature spectrometer; the unmanned aerial vehicle flies at the low altitude of 5 m and the high altitude of 30-50 m to collect hyperspectral image data, meteorological data are recorded synchronously, and image geographical registration is achieved through ground control points arranged at the interval of 50 m. According to the invention, aiming at the core limitation of the prior art, a multi-source hyperspectral data collaborative acquisition system of'ground object spectrometer-low-altitude unmanned aerial vehicle-high-altitude unmanned aerial vehicle 'is constructed, point-level precise spectrums, single-plant-level high-resolution images and area-level wide coverage data are integrated, multi-scale data coverage from a single plant to a whole domain is realized, and the multi-scale data coverage is realized. The contradiction between coverage and precision of a single data source in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of agricultural disease monitoring technology, specifically to a method and system for intelligent monitoring of rice false smut based on hyperspectral imaging. Background Technology

[0002] Rice false smut is a significant fungal disease affecting rice yield and quality. Its disease cycle encompasses three stages: latent infection (infectious disease without showing symptoms), manifest disease (manifestation of typical symptoms), and spread (expansion and spread of lesions). This places extremely high demands on the timeliness, accuracy, and multi-scale collaborative capabilities of monitoring technologies. Among existing monitoring technologies, manual surveys rely on visual identification, which is extremely inefficient and unable to capture latent diseases. While satellite remote sensing can achieve wide-area coverage, its long revisit cycle and low spatial resolution make it difficult to identify diseases at the individual plant level, resulting in insufficient accuracy in monitoring latent diseases. Ground image recognition requires manual feature extraction, has poor adaptability to different varieties and growth stages, and limited generalization ability. Although single hyperspectral monitoring can acquire detailed spectral information, it suffers from high data redundancy, low processing efficiency, and cannot integrate temporal dynamic features, making it difficult to track the disease's development from latent to manifest stages. These technologies either have bottlenecks in the accuracy of identifying latent diseases or are limited in their ability to coordinate and integrate multi-scale data and monitor time-series dynamics. As a result, they cannot achieve full-cycle, high-precision, and wide-coverage intelligent monitoring of rice false smut from early latent infection to later spread, which has become a technical problem restricting the development of current precision agriculture disease monitoring. Summary of the Invention

[0003] This disclosure proposes a method and system for intelligent monitoring of rice false smut based on hyperspectral imaging, aiming to overcome at least one of the defects in the existing technology.

[0004] To achieve the above objectives, the technical solution disclosed in this invention is as follows: According to one aspect of this disclosure, a method for intelligent monitoring of rice false smut based on hyperspectral imaging is provided, comprising the following steps: Point spectral data of rice panicles from the jointing and heading stage to maturity were collected using a ground-based portable ground object spectrometer; hyperspectral image data were collected by drones flying at a low altitude of 5m and a high altitude of 30-50m, respectively, while meteorological data were recorded simultaneously, and image georegistration was achieved through ground control points set up at 50m intervals. Atmospheric correction was performed on low-altitude hyperspectral images using dark target subtraction, and atmospheric correction was performed on high-altitude hyperspectral images using the FLAASH model. Geometric correction was performed based on the ground control points using quadratic polynomial interpolation. Images were stitched together using the SIFT algorithm, and the spectral curves were smoothed and denoised using Savitzky-Golay filtering. A genetic algorithm is used to coarsely select hyperspectral bands to obtain 30-40 candidate bands. Then, a continuous projection algorithm is used to select key bands from the candidate bands to obtain 10-15 key bands. The key band data is input into a CNN module to extract spatial features. The CNN module includes three convolutional layers. The spatial features are input into a GRU module to process temporal dynamic features. The GRU module includes two GRU units. Temporal data with a preset interval of several days is input. The spatial features and temporal features are fused through a CBAM module, which includes channel attention and spatial attention. Disease monitoring results are obtained through the classification output module. The regional disease incidence rate is output through the large-scale monitoring model. When the incidence rate is greater than 5%, the small-scale precise model is triggered to locate diseased plants. The future disease spread rate is predicted by combining time series data from 3-5 time points and an early warning map is generated.

[0005] Furthermore, the conditions for the ground-based portable ground object spectrometer to collect spectral data of rice ears are as follows: clear, cloudless weather, sensor probe pointing vertically downwards, 15-20cm away from the rice ears, field of view of 25°, calibration with a standard white board before each collection, and simultaneous recording of the sample's growth period and disease classification.

[0006] Furthermore, the low-altitude flight path of the drone is planned in a zigzag pattern to avoid strong midday sunlight, and each flight covers 0.5-1 acre, acquiring high-resolution hyperspectral images with a pixel size of 5cm×5cm. The high-altitude flight of the drone covers 50-100 acres per day to scan the entire contiguous rice paddy.

[0007] Furthermore, the dark target subtraction is used to remove the atmospheric scattering effect of low-altitude hyperspectral images. The FLAASH model is input with atmospheric profile data obtained from local meteorological stations. After correction, the spectral reflectance error is less than 3%. The root mean square error of the geometric correction is less than 0.5 pixels. The image stitched by the SIFT algorithm has no obvious gaps and the spatial continuity error is less than 1%.

[0008] Furthermore, the fitness function of the genetic algorithm is the accuracy of rice blast disease identification, and the candidate bands include 550nm, 680nm, 850nm and 950nm bands that are sensitive to the disease. The core logic of the continuous projection algorithm is to calculate the projection vector between bands and select the band combination with the lowest redundancy and the most information to reduce the dimensionality of the hyperspectral data.

[0009] Furthermore, the CNN module's three convolutional layers are followed by max pooling with a 2×2 pooling kernel and a stride of 2, outputting a 64-dimensional spatial feature vector to capture subtle spectral changes caused by latent diseases; the GRU module includes two GRU units and one 64-dimensional fully connected layer, taking into input temporal hyperspectral features at 3-day intervals and outputting a 64-dimensional temporal feature vector to capture the dynamic trend of disease changes from latent to overt.

[0010] Furthermore, the channel attention of the CBAM module performs weight calibration on the spectral feature channels of the convolution output through global average pooling and fully connected layers to enhance the response to the disease-sensitive bands. The spatial attention focuses on the spectral changes at key time nodes in the disease development process through a 1D convolution kernel of size 7.

[0011] Furthermore, the two fully connected layers of the classification output module are 64-dimensional to 32-dimensional, with a Dropout rate of 0.2 used in between to prevent overfitting. The output layer is 32-dimensional to 3-dimensional, and uses the Softmax activation function to output the three-class probability distribution. The training strategy is a batch size of 32, 3400 training rounds, early stopping, patience=10, and the dataset is divided into training and test sets in a 7:3 ratio.

[0012] Furthermore, the large-scale monitoring model inputs 10-15 key bands of 30-50m high-altitude hyperspectral data from the UAV and outputs the regional disease incidence rate and disease index, with a coverage efficiency of 50 mu / hour. The small-scale precision model inputs 10-15 key bands of 5m low-altitude hyperspectral data from the UAV and outputs the single-plant disease level of 0-3.

[0013] According to another aspect of this disclosure, a hyperspectral-based intelligent monitoring system for rice false smut is provided to implement the hyperspectral-based intelligent monitoring method for rice false smut as described above, comprising: The multi-source hyperspectral data acquisition module is used to collect point spectral data of rice panicles from the jointing and heading stage to the maturity stage using a ground-based portable ground object spectrometer, and to collect hyperspectral image data by flying a drone at a low altitude of 5m and a high altitude of 30-50m respectively, while simultaneously recording meteorological data, and to achieve image georegistration through ground control points deployed at 50m intervals. The hyperspectral data preprocessing module is used to perform atmospheric correction on low-altitude hyperspectral images using dark target subtraction, atmospheric correction on high-altitude hyperspectral images using the FLAASH model, geometric correction based on the ground control points using quadratic polynomial interpolation, image stitching using the SIFT algorithm, and smoothing and denoising the spectral curves using Savitzky-Golay filtering with a window size of 5 and a polynomial order of 2. The intelligent feature extraction and dimensionality reduction module is used to coarsely select hyperspectral bands using a genetic algorithm to obtain 30-40 candidate bands, and then to select key bands from the candidate bands using a continuous projection algorithm. The projection dimension of the continuous projection algorithm is in the range of 5-15, with an error threshold of 0.01, to obtain 10-15 key bands. An integrated deep learning model processing module is used to input the key band data into a CNN module to extract spatial features, input the spatial features into a GRU module to process temporal dynamic features, fuse the spatial features and temporal features through a CBAM module, and obtain the disease monitoring results through a classification output module. The multi-scale model application module is used to output the regional disease incidence rate through a large-scale monitoring model. When the incidence rate is greater than 5%, a small-scale precise model is triggered to locate diseased plants. Combined with time series data from 3-5 time points, the spread rate of the disease is predicted and an early warning map is generated.

[0014] The beneficial effects of this invention are: This invention addresses the core limitations of existing technologies by constructing a multi-source hyperspectral data collaborative acquisition system consisting of a ground object spectrometer, a low-altitude UAV, and a high-altitude UAV. This system integrates point-level precise spectra, single-tree-level high-resolution images, and area-level wide-coverage data, achieving multi-scale data coverage from single trees to the entire region. This resolves the contradiction between coverage and accuracy inherent in single data sources in existing technologies.

[0015] Specifically, by using a hybrid dimensionality reduction strategy combining genetic algorithms and continuous projection algorithms, sensitive bands for rice blast disease were accurately selected, reducing data redundancy, improving data processing efficiency, and providing high-information-density input for subsequent model training.

[0016] Furthermore, by using the CNN-GRU-CBAM ensemble model, which integrates the ability of CNN to extract spectral spatial features, the ability of GRU to capture the temporal dynamics of diseases, and the attention mechanism of CBAM to enhance key features, the model effectively identifies subtle spectral changes caused by latent diseases and the dynamic development trend of diseases from latent to manifest, thus breaking through the limitations of existing models in the fusion of spatial and temporal features.

[0017] Furthermore, through a multi-scale monitoring application mode, high-altitude data is used to achieve regional-level early warning of disease incidence, triggering low-altitude data for precise diagnosis at the individual plant level. Combined with time-series data, the spread rate of the disease is predicted, forming a full-cycle monitoring system of "early warning-inspection-diagnosis-prediction," providing an early decision-making window for precise disease control. The synergistic effect of the technical means in this invention overcomes the limitations of existing technologies in terms of the accuracy of latent disease identification, multi-scale data fusion capabilities, and time-series dynamic monitoring. It achieves full-cycle, high-precision, and wide-coverage intelligent monitoring of rice false smut from early latent infection to later spread, meeting the high requirements of precision agriculture for disease monitoring and promoting the transformation of disease monitoring from "post-event remediation" to "pre-event early warning."

[0018] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the following describes the preferred embodiments of the present invention in detail with reference to the accompanying drawings. Attached Figure Description

[0019] Figure 1 This is a flowchart of the intelligent monitoring method for rice false smut based on hyperspectral imaging according to the present invention; Figure 2 This is a general technical roadmap of the present invention; Figure 3 This is a low-altitude hyperspectral image of the UAV of the present invention; Figure 4 This is a feature selection diagram of the GA-SPA of the present invention; Figure 5 This is a structural diagram of the composite deep learning model of the present invention; Figure 6 This is a precise disease diagnosis diagram for single plants according to the present invention. Detailed Implementation

[0020] 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, not all, of the embodiments of the present invention. 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.

[0021] The term "comprising," and any variations thereof, used in the specification and claims of this application, is intended to cover a non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or apparatus. Furthermore, the use of "and / or" in the specification and claims indicates at least one of the connected objects, such as A and / or B, indicating the inclusion of A alone, B alone, or both A and B.

[0022] In embodiments of the present invention, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in embodiments of the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0023] The present invention provides the following preferred embodiments: Example 1: To address the issues of insufficient multi-source data synergy, hyperspectral feature redundancy, and inaccurate temporal dynamic capture in rice false smut monitoring, this example provides the steps of a hyperspectral-based intelligent monitoring method for rice false smut, such as... Figure 1 As shown, through multi-scale data acquisition, precise preprocessing, hierarchical feature dimensionality reduction, and integrated deep learning models, the entire lifecycle monitoring of diseases from early latent infection to later spread can be achieved.

[0024] During the data acquisition phase, focusing on the critical period from latent infection to manifest disease in rice false smut, the monitoring window was selected from the jointing and panicle elongation stage to maturity. Point spectral data of rice panicles were collected using a portable ground-based spectrometer. During operation, the sensor probe was positioned vertically downwards towards the rice panicle, maintaining a distance of 15 to 20 centimeters to avoid spectral saturation due to excessive proximity or stray light interference from excessive distance. 50 to 70 replicates were collected for each sample type, with spectral curves collected three times for each replicate, and the average value was taken to ensure data stability. Simultaneously, hyperspectral image data was collected using a drone flying at low altitudes of 5 meters and high altitudes of 30 to 50 meters. At low altitudes, a slower flight speed and higher directional and lateral overlap were selected to acquire high-resolution images to meet the needs of single-plant-level diagnosis. At high altitudes, flight parameters were adjusted to cover a larger area to meet the needs of area-level monitoring. Meteorological data such as temperature, humidity, and wind speed were recorded simultaneously during the flight, providing basic parameters for subsequent atmospheric correction. To ensure spatial consistency of multi-source data, georegistration of UAV hyperspectral images is achieved by deploying ground control points at 50-meter intervals, so that each pixel in the image can correspond to the actual location on the ground.

[0025] Furthermore, the acquired hyperspectral data underwent precise preprocessing. For low-altitude hyperspectral images, due to the low flight altitude and close proximity to the ground, atmospheric scattering effects were relatively minor. Dark target subtraction was used for atmospheric correction, selecting shadowed areas without vegetation as dark targets and calculating their spectral reflectance to remove spectral distortion caused by atmospheric scattering. For high-altitude hyperspectral images, due to the high flight altitude and long atmospheric paths, atmospheric effects were more complex. The FLAASH model was used for atmospheric correction, inputting synchronously recorded meteorological data and atmospheric profile information to accurately correct for atmospheric absorption and scattering effects in high-altitude images. Based on ground control points spaced 50 meters apart, quadratic polynomial interpolation was used for geometric correction of both high- and low-altitude hyperspectral images. By fitting the coordinate relationship between image pixels and ground control points, spatial distortion caused by UAV attitude changes or terrain undulations was eliminated, ensuring the accuracy of image spatial positioning. Subsequently, the SIFT algorithm was used to stitch together overlapping areas of adjacent images. By extracting and matching scale-invariant feature points in the images, a complete regional image was formed, ensuring spatial continuity. To remove instrument noise and environmental interference (such as spectral fluctuations caused by leaf shaking), Savitzky-Golay filtering was used to smooth the spectral curves. By selecting an appropriate window size and polynomial order, the influence of noise on subsequent analysis was reduced while preserving spectral characteristics (such as chlorophyll absorption peaks and water absorption peaks).

[0026] To address the high-dimensional redundancy problem of hyperspectral data, a hierarchical feature reduction strategy is adopted. First, a genetic algorithm is used for coarse selection: a population of a certain size is initialized, with each individual representing a subset of bands using binary encoding. During iteration, individuals are optimized through selection, crossover, and mutation operations. Classification performance is used as the fitness index to select 30 to 40 candidate bands, initially removing irrelevant or redundant bands. Then, a continuous projection algorithm is used to refine the selection from the candidate bands. By minimizing redundancy between bands through continuous projection, the band with the largest projection vector is selected, ultimately yielding 10 to 15 key bands. Sensitive spectral features related to rice false smut (such as spectral bands corresponding to changes in chlorophyll content and moisture content caused by the disease) are retained. It is important to understand that the genetic algorithm excels at global search in high-dimensional data, while the continuous projection algorithm excels at reducing redundancy between bands. The combination of the two achieves effective dimensionality reduction from high to low dimensions, providing efficient feature input for subsequent model processing.

[0027] During the model building phase, key band data is input into a CNN module to extract spatial features. This module consists of three convolutional layers. The first convolutional layer uses a small kernel to extract simple spatial features from the spectral data (such as the basic shape of the spectral curve). The second convolutional layer uses a larger kernel to extract more complex features (such as the combination of spectral peaks and valleys). The third convolutional layer extracts high-level spatial features (such as the spatial distribution pattern of spectral features). After each convolutional layer, a pooling layer is used to reduce the feature dimensionality and retain key information. The extracted spatial features are then input into a GRU module to process temporal dynamic features. This module consists of two GRU units, which take temporal data at preset intervals (such as spectral data collected weekly) as input. Through gating mechanisms (update gate, reset gate), it captures the temporal trend of disease development (such as the gradual change in spectral reflectance during the process from latent infection to apparent disease). Subsequently, spatial and temporal features are fused using the CBAM module: first, spatial features are weighted using a channel attention mechanism to calculate the importance of each band and assign corresponding weights, highlighting the features of disease-sensitive bands; then, the channel-weighted features are processed using a spatial attention mechanism to calculate the importance of each spatial location, highlighting the spatial features of the lesion area, thus achieving effective fusion of spatial and temporal features.

[0028] Furthermore, the fused features are processed through a classification output module to obtain disease monitoring results (e.g., healthy plants, latently infected plants, mildly diseased plants, and severely diseased plants). High-altitude hyperspectral data is input into a large-scale monitoring model to output the regional disease incidence rate (e.g., the disease incidence rate in a 100-mu (approximately 6.7 hectares) paddy field). When the incidence rate exceeds 5%, a low-altitude UAV is triggered to collect high-resolution hyperspectral images, which are then input into a small-scale precision model to locate diseased plants (e.g., the disease severity of a rice panicle). Combining time-series data from 3 to 5 time points, the GRU module predicts the future disease spread rate (e.g., changes in the disease incidence rate over the next 7 days) and generates an early warning map (marking disease spread areas), providing decision support for precise disease control. In essence, large-scale monitoring is used for overall early warning, small-scale precision models are used for locating diseased plants in key areas, and time-series prediction is used for early intervention. These three elements form a complete monitoring system from the overall to the local, from the current situation to the future.

[0029] This embodiment ensures the comprehensiveness and spatial consistency of data through multi-scale data acquisition, improves data quality through precise preprocessing, removes redundancy and retains sensitive features through hierarchical feature dimensionality reduction, and achieves effective fusion of spatial and temporal features through the integration of deep learning models. Ultimately, it realizes full-cycle monitoring of rice false smut from its early stage to its spread, providing technical support for disease prevention and control.

[0030] Example 2: In response to the problems of existing technologies such as "difficulty in identifying latent diseases, poor multi-scale data fusion, insufficient time-series dynamic monitoring, and low model efficiency", this example aims to provide an intelligent monitoring method based on hyperspectral multi-source data fusion and deep learning to achieve full-cycle, high-precision, and wide-coverage monitoring of rice false smut from "early latent to mid-term manifestation to late-term spread", providing early warning support 5-7 days in advance for precise disease prevention and control.

[0031] refer to Figures 2 to 6 As shown, the overall technical approach of this embodiment constructs a full-cycle monitoring system for rice false smut through five major stages: "multi-source hyperspectral data acquisition → precise preprocessing → intelligent feature extraction → deep learning modeling → multi-scale monitoring application". The specific technical steps are as follows: 1. Multi-source hyperspectral data acquisition (1) Ground point spectral acquisition (ground object spectrometer) Equipment parameters: Ground object spectrometer, spectral range 350-2500nm, spectral resolution 3nm (350-1000nm), 10nm (1000-2500nm), sampling interval 1nm.

[0032] Data collection subjects: rice panicles from the jointing and panicle elongation stage to maturity stage. 50-70 replicates were collected for each type of sample, and the spectral curves of each replicate were collected 3 times and the average value was taken.

[0033] Data collection conditions: Clear, cloudless weather (10:00-14:00), sensor probe pointing vertically downwards, 15-20cm away from the rice ear, field of view 25°, calibrated with a standard white board before each data collection.

[0034] Data labeling: The growth period and disease level of the samples were recorded synchronously (classified according to GB / T15790-2017, level 0: healthy; level 1: latent infection; level 2: mild disease; level 3: severe disease).

[0035] (2) Low-altitude hyperspectral imaging by UAV (precise monitoring in a small area) Equipment configuration: DJI Matrice 300RTK drone, equipped with a hyperspectral camera with a spectral range of 400-1000nm, a spectral resolution of 5nm, a spatial resolution of 0.003m (flight altitude of 5m), and a frame rate of 15fps.

[0036] Flight parameters: flight altitude 5m, flight speed 3m / s, forward overlap 80%, lateral overlap 60%, flight path planning adopts a zigzag pattern to avoid strong midday sunlight (suspended from 11:00 to 13:00).

[0037] Collection scope: Targeting the "key monitoring area" (high disease incidence area) in paddy fields, each frame covers 0.5-1 acre to acquire high-resolution hyperspectral images (pixel size 5cm×5cm), which can identify spectral anomalies in individual rice panicles.

[0038] (3) High-altitude hyperspectral imaging by UAVs (wide-area coverage monitoring) Equipment configuration: DJI Matrice 300RTK drone, equipped with a hyperspectral camera with a spectral range of 400-1000nm, a spectral resolution of 8nm, a spatial resolution of 0.02m (flight altitude of 50m), and a frame rate of 15fps.

[0039] Flight parameters: flight altitude 30-50m, flight speed 8m / s, forward overlap rate 70%, lateral overlap rate 50%, can cover 50-100 mu per day, suitable for full-area scanning of contiguous rice fields.

[0040] Data synchronization: Meteorological data (temperature, humidity, wind speed) are recorded before each flight, and image georegistration is achieved after the flight through ground control points (GCPs, deployed at 50m intervals).

[0041] 2. Precise preprocessing of hyperspectral data (1) Atmospheric correction Method selection: For low-altitude images (5m), Dark Object Subtraction (DOS) was used to remove the effects of atmospheric scattering; for high-altitude images (30-50m), the FLAASH atmospheric correction model (Fast Line-of-Sight Atmospheric Analysis of Spectral Hypercubes) was used, with atmospheric profile data (obtained from local weather stations) as input.

[0042] Correction effect: The spectral reflectance error after correction is less than 3%, eliminating the interference of water vapor absorption band.

[0043] (2) Geometric correction and image stitching Geometric correction: Based on ground control points (GCP), a quadratic polynomial interpolation method is used, and the root mean square error (RMSE) of the corrected image is <0.5 pixels.

[0044] Image stitching: The SIFT feature matching algorithm is used to merge the overlapping areas of adjacent images. The stitched image has no obvious gaps and the spatial continuity error is <1%.

[0045] (3) Spectral curve smoothing and noise reduction Smoothing algorithm: Savitzky-Golay filter (window size 5, polynomial order 2) is used to remove instrument noise and environmental interference, such as spectral fluctuations caused by blade shaking.

[0046] Denoising effect: The signal-to-noise ratio (SNR) of the processed spectral curve is improved from 20dB to over 45dB, and the characteristic peaks (such as 550nm, 680nm, and 850nm) are clearer.

[0047] 3. Intelligent Feature Extraction and Dimensionality Reduction (GA-SPA Hybrid Algorithm) To address the issues of "high dimensionality and redundancy" in hyperspectral data, a hybrid dimensionality reduction strategy is proposed, combining a genetic algorithm (GA) for coarse selection with a continuous projection algorithm (SPA) for fine selection. The steps are as follows: (1) GA coarse selection: Initialize the population: During the initialization phase, a population containing... A population of individuals. Each individual A feature subset is represented, consisting of multiple bands from the hyperspectral data. Individuals can be represented using binary encoding or other methods, with each bit indicating whether a particular band is selected. The initial population is established by randomly generating feature subsets.

[0048]

[0049] in, It is the first Individual.

[0050] Fitness calculation: For each individual Calculate its fitness Fitness reflects the performance of a subset of features in a specific task (such as classification or regression). Assumptions Represents an individual The performance metrics (such as classification accuracy, error rate, etc.) and the fitness calculation formula are as follows:

[0051] in, Individual The performance evaluation value is calculated with the sum of the fitness of all individuals in the denominator, and is used for normalization.

[0052] Selection: Selection is based on an individual's fitness value; individuals with higher fitness have a greater probability of being included in the next generation. Selection methods can include roulette wheel selection and tournament selection. Selection Probability Given by the following formula:

[0053] in, Individual The probability of being selected ensures that individuals with higher fitness are more likely to participate in subsequent crossover and mutation operations.

[0054] Crossover: Crossover simulates the natural process of gene recombination by selecting two parent individuals. and Then they exchange some of their genes to generate new individuals. Crossover operations come in various forms, such as single-point crossover, double-point crossover, and uniform crossover. The formula for crossover operations is:

[0055] Mutation operation: The mutation operation increases population diversity and prevents early convergence by randomly altering the genes of individuals. The mutation probability is set to... If the generated random number is less than the probability, the individual is mutated. The mutated individual is... :

[0056] Mutation operations generate new solutions by altering parts of an individual's genes.

[0057] Fitness update: For each newly generated individual Its fitness value needs to be recalculated. To assess its performance in the current task, so as to provide a basis for selection in the next round.

[0058] Population renewal: A new generation of individuals replaces those with higher fitness, forming a new population. The renewed population includes individuals resulting from selection, crossover, and mutation, and then proceeds to the next round of genetic operations.

[0059] Termination condition: If the optimal individual in the population... The set threshold is reached, or the algorithm executes its maximum number of iterations. The algorithm terminates if the condition is met.

[0060] Output the optimal solution: After the algorithm terminates, output the individual with the highest fitness. As the final solution, this solution corresponds to the optimal feature subset, which is used for subsequent intelligent monitoring of rice false smut.

[0061] Objective function: Using the accuracy of rice blast disease identification (based on SVM model) as the fitness function, candidate bands sensitive to the disease are selected.

[0062] Parameter settings: population size 50, number of iterations 30, crossover probability 0.8, mutation probability 0.05, wavelength range 400-1000nm.

[0063] Output: Select 30-40 candidate bands from 200 original bands (with a focus on retaining bands such as 550nm (chlorophyll absorption), 680nm (photosynthesis sensitivity), 850nm (biomass association), and 950nm (disease stress response).

[0064] (2) SPA Selection: Input and initialization steps: Input data is initialized by extracting the location features of each sample in the hyperspectral data. Specifically, the features of each sample are represented by its band feature values ​​in the hyperspectral data. First, based on the characteristics of the samples, initial hyperspectral bands are selected. As input data for dimensionality reduction, and defining the number of selected bands.

[0065] Initialize the number of hyperspectral bands: In the initial stage, set the number of spectral bands to 1 and extract the initial hyperspectral bands. eigenvectors The purpose of this step is to provide initial spectral feature information for the subsequent dimensionality reduction process.

[0066]

[0067] Calculate the projection matrix: During the dimensionality reduction process, the projection matrix for each hyperspectral band is calculated based on the transmitted hyperspectral data. Therefore, we first define the band feature matrix. Then calculate from the initial hyperspectral band Projection matrix to projection space Its formula is:

[0068] Calculating projection data: When calculating projection data, the projection matrix is ​​first used... Calculate the initial hyperspectral band Projection data of the remaining hyperspectral bands And combine these projection data with their corresponding spectral data sets. Use them together for further feature extraction:

[0069] Band selection based on projection data: Based on the projection data calculated in the previous step, select the hyperspectral band with the highest projection weight from the projection data. The number of projection data selected... A new set of spectral bands was determined. This is used for subsequent analysis. Its calculation method is as follows:

[0070] Output optimized projection data: Finally, by comparing the selected optimal spectral band data with the initial hyperspectral data... The data is then fused to obtain an optimized projection dataset. This fusion step ensures the validity of the data and provides high-quality feature data for subsequent intelligent monitoring tasks.

[0071] Core logic: By calculating the projection vectors between bands, select the band combination with the lowest redundancy and the highest information content to avoid the "curse of dimensionality".

[0072] Parameter settings: Projection dimension range 5-15, error threshold 0.01.

[0073] Output: 10-15 key bands are further selected from the candidate bands, resulting in a reduction of more than 90% in data dimensionality and a significant reduction in model training time.

[0074] (3) Validation: After dimensionality reduction, the discrimination capability of key bands (calculated by Mahalanobis distance) is improved by 40%. Compared with the original data, the accuracy of rice blast disease identification only decreases by 2%-3%, but the processing efficiency is improved by 80%.

[0075] 4. Integrated Deep Learning Models (CNN-GRU-CBAM) A three-order integrated model is constructed, consisting of "spatial feature extraction - temporal dynamic modeling - key area focusing," with the structure as follows: Figure 5 As shown: (1) CNN module (spatial feature extraction): Structure: 3 convolutional layers (3×3 kernel size, number of kernels: 16, 32, 64 respectively; stride=1, padding=1, each layer is followed by batch normalization (BatchNorm) and ReLU activation function, and downsampling is performed using max pooling (pooling kernel 2×2, stride=2).

[0076] Function: Extract local features and spectral response patterns from hyperspectral curves, focusing on capturing subtle spectral changes caused by latent diseases (such as changes in absorption valleys in specific bands, abnormal reflectance, etc.).

[0077] Output: 64-dimensional feature vectors.

[0078] (2) GRU module (time-series dynamic modeling): Structure: 2 layers of GRU units (64 hidden layers, ReLU activation function) + fully connected layer (64 dimensions).

[0079] Function: Input hyperspectral characteristics at different time points (3-day intervals) to capture the dynamic change trend of diseases from "latent to manifest" (such as the continuous decrease of spectral reflectance in the 850nm band).

[0080] Output: 64-dimensional temporal feature vector.

[0081] (3) CBAM module (attention mechanism focusing): Channel attention: Through global average pooling and fully connected layers, the spectral feature channels of the convolution output are weighted and calibrated to enhance the response to disease-sensitive bands.

[0082] Spatial attention: By using one-dimensional convolution operations (convolution kernel size = 7), we focus on the spectral changes at key time points in the disease development process.

[0083] Output: Feature vectors that fuse spectral features, temporal dynamics, and attention weighting.

[0084] (4) Classification output module: Fully connected layer: Two-layer structure (64→32 dimensions), with Dropout (rate=0.2) used in between to prevent overfitting.

[0085] Output layer: 32→3 (outputs a three-class probability distribution), using the Softmax activation function.

[0086] Optimizer: SGD (learning rate 0.0025, decay rate 0.0001).

[0087] Loss function: Cross-entropy loss function (suitable for multi-class classification tasks).

[0088] Training strategy: Batch size 32, training epochs 3400, early stopping (patience=10) to avoid overfitting; retain the model with the lowest validation loss, and divide the dataset into training and test sets in a 7:3 ratio.

[0089] 5. Multi-scale model training and application (1) Model partitioning: Small-scale precision model: Input UAV low-altitude (5m) hyperspectral data (10-15 key bands), output single-plant level disease level (0-3), with an identification accuracy of over 90%, suitable for key area monitoring combining "point-area" approach.

[0090] Large-scale monitoring model: Input high-altitude (30-50m) hyperspectral data (10-15 key bands) from UAVs, output regional disease incidence rate (%) and disease index, with a coverage efficiency of 50 acres / hour, suitable for global trend early warning.

[0091] (2) Model fusion application: When the large-scale model detects that the disease incidence rate in a certain area is >5%, the small-scale model is automatically triggered to conduct precise inspections and locate the diseased plants (error <1m).

[0092] By combining time-series data from 3-5 time points, the spread rate of the disease in the next 7 days (such as the daily growth rate of the disease index) can be predicted, and a visual early warning map can be generated.

[0093] (III) Key Technological Innovations 1. Collaborative acquisition of multi-source hyperspectral data: For the first time, three-level data fusion of "ground point spectrum (high precision) - low-altitude image (single tree level) - high-altitude image (global level)" is realized, resolving the contradiction between "precision and coverage".

[0094] 2. GA-SPA hybrid dimensionality reduction algorithm: By combining GA coarse selection and SPA fine selection, it achieves efficient dimensionality reduction of hyperspectral data, taking into account both feature effectiveness and processing efficiency.

[0095] 3. CNN-GRU-CBAM integrated model: It integrates spatial features, temporal dynamics and attention mechanism to overcome the bottleneck of traditional models in low recognition rate of latent diseases and weak temporal prediction ability.

[0096] 4. Multi-scale intelligent monitoring system: Construct a hierarchical application model of "full-area early warning - key inspection - precise diagnosis" to realize full-cycle, closed-loop monitoring of rice blast disease.

[0097] Experimental verification (2025 Heilongjiang Wuchang Rice Major Production Area Trial) 1. Test conditions: Experimental location: Rice planting base in Wuchang City, Heilongjiang Province (100 mu in area, rice variety "Songjing 83").

[0098] Experiment period: August-October 2025 (from rice jointing stage to maturity).

[0099] Comparison methods: traditional satellite remote sensing (Landsat-8), ground image recognition (RGB+SVM), and single CNN model.

[0100] 2. Comparison of core indicators:

[0101] Table 1 Comparison of Core Indicators

[0102] Table 2 Performance Comparison of Low-Altitude Data Models (Test Set)

[0103] Table 3 Performance Comparison of High-Altitude Data Models (Test Set) Summary of technical advantages based on Tables 1 to 3: 1. Improved accuracy: The accuracy of identifying latent diseases is up to 53.8% higher than existing technologies, solving the problem of monitoring diseases that are not obvious; the accuracy of identifying obvious diseases is over 98%, which is close to the accuracy of human visual identification (99%), but the efficiency is increased by 60 times.

[0104] 2. Efficiency optimization: By using GA-SPA dimensionality reduction and CNN-GRU-CBAM model, data processing efficiency is improved by 80%, and a single drone (high altitude) can monitor an area of ​​100 acres per day, which is 200 times that of manual surveys.

[0105] 3. Time-based early warning: Enables dynamic trend prediction of diseases, with an early warning time of up to 7 days, providing ample decision-making window for precision pesticide application (reducing pesticide use by 20%-30%).

[0106] 4. Strong generalization ability: The cross-regional monitoring accuracy rate in major rice-producing areas of different accumulated temperature zones, such as Heilongjiang Province, is >85%, and it is adaptable to different varieties and planting patterns.

[0107] Although the present invention has been specifically described above with reference to preferred embodiments, it should be understood that the present invention is not limited to the embodiments described above. Various modifications and variations can be made by those skilled in the art without departing from the spirit of the present invention, and such modifications and variations should fall within the scope defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring of rice false smut based on hyperspectral imaging, characterized by the following steps: include: Point spectral data of rice panicles from the jointing and heading stage to maturity were collected using a ground-based portable ground object spectrometer; hyperspectral image data were collected by drones flying at a low altitude of 5m and a high altitude of 30-50m, respectively, while meteorological data were recorded simultaneously, and image georegistration was achieved through ground control points set up at 50m intervals. Atmospheric correction was performed on low-altitude hyperspectral images using dark target subtraction, and atmospheric correction was performed on high-altitude hyperspectral images using the FLAASH model. Geometric correction was performed based on the ground control points using quadratic polynomial interpolation. Images were stitched together using the SIFT algorithm, and the spectral curves were smoothed and denoised using Savitzky-Golay filtering. A genetic algorithm is used to coarsely select hyperspectral bands to obtain 30-40 candidate bands. Then, a continuous projection algorithm is used to select key bands from the candidate bands to obtain 10-15 key bands. The key band data is input into a CNN module to extract spatial features. The CNN module includes three convolutional layers. The spatial features are input into a GRU module to process temporal dynamic features. The GRU module includes two GRU units. Temporal data with a preset interval of several days is input. The spatial features and temporal features are fused through a CBAM module, which includes channel attention and spatial attention. Disease monitoring results are obtained through the classification output module. The regional disease incidence rate is output through the large-scale monitoring model. When the incidence rate is greater than 5%, the small-scale precise model is triggered to locate diseased plants. The future disease spread rate is predicted by combining time series data from 3-5 time points and an early warning map is generated.

2. The intelligent monitoring method for rice false smut based on hyperspectral imaging as described in claim 1, characterized in that, The conditions for the ground-based portable ground object spectrometer to collect spectral data of rice ears are as follows: clear, cloudless weather, sensor probe pointing vertically downwards, 15-20cm away from the rice ears, field of view 25°, calibration with a standard white board before each collection, and simultaneous recording of the sample's growth period and disease classification.

3. The intelligent monitoring method for rice false smut based on hyperspectral imaging as described in claim 1, characterized in that, The drone's low-altitude flight path is planned in a zigzag pattern to avoid strong midday sunlight. Each flight covers 0.5-1 acre and acquires high-resolution hyperspectral images with a pixel size of 5cm×5cm. The drone's high-altitude flight can cover 50-100 acres per day to scan the entire contiguous rice paddy.

4. The intelligent monitoring method for rice false smut based on hyperspectral imaging as described in claim 1, characterized in that, The dark target subtraction method is used to remove the atmospheric scattering effect of low-altitude hyperspectral images. The FLAASH model is input with atmospheric profile data obtained from local meteorological stations. After correction, the spectral reflectance error is less than 3%. The root mean square error of the geometric correction is less than 0.5 pixels. The image stitched by the SIFT algorithm has no obvious gaps and the spatial continuity error is less than 1%.

5. The intelligent monitoring method for rice false smut based on hyperspectral imaging as described in claim 1, characterized in that, The fitness function of the genetic algorithm is the accuracy of rice blast disease identification. The candidate bands include 550nm, 680nm, 850nm and 950nm bands that are sensitive to the disease. The core logic of the continuous projection algorithm is to calculate the projection vector between bands and select the band combination with the lowest redundancy and the most information to reduce the dimensionality of hyperspectral data.

6. The intelligent monitoring method for rice false smut based on hyperspectral imaging as described in claim 1, characterized in that, The CNN module consists of three convolutional layers, each followed by a max pooling operation with a 2×2 pooling kernel and a stride of 2, outputting a 64-dimensional spatial feature vector to capture subtle spectral changes caused by latent diseases. The GRU module includes two GRU units and one 64-dimensional fully connected layer, taking into input temporal hyperspectral features at 3-day intervals and outputting a 64-dimensional temporal feature vector to capture the dynamic trend of disease changes from latent to overt.

7. The intelligent monitoring method for rice false smut based on hyperspectral imaging as described in claim 1, characterized in that, The channel attention of the CBAM module uses global average pooling and a fully connected layer to weight the spectral feature channels of the convolution output, thereby enhancing the response to the disease-sensitive bands. The spatial attention focuses on the spectral changes at key time points in the disease development process through a 1D convolution kernel of size 7.

8. The intelligent monitoring method for rice false smut based on hyperspectral imaging as described in claim 1, characterized in that, The classification output module has two fully connected layers ranging from 64 to 32 dimensions, with a Dropout rate of 0.2 used in between to prevent overfitting. The output layer is 32 to 3 dimensions and uses the Softmax activation function to output the three-class probability distribution. The training strategy is a batch size of 32, 3400 training rounds, early stopping, and a patience of 10. The dataset is divided into training and test sets in a 7:3 ratio.

9. The intelligent monitoring method for rice false smut based on hyperspectral imaging as described in claim 1, characterized in that, The large-scale monitoring model inputs 10-15 key bands of 30-50m high-altitude hyperspectral data from UAVs and outputs regional disease incidence and disease index, with a coverage efficiency of 50 mu / hour. The small-scale precision model inputs 10-15 key bands of 5m low-altitude hyperspectral data from UAVs and outputs single-plant disease level of 0-3.

10. A hyperspectral-based intelligent monitoring system for rice false smut, used to implement the hyperspectral-based intelligent monitoring method for rice false smut as described in any one of claims 1-9, characterized in that, include: The multi-source hyperspectral data acquisition module is used to collect point spectral data of rice panicles from the jointing and heading stage to the maturity stage using a ground-based portable ground object spectrometer, and to collect hyperspectral image data by flying a drone at a low altitude of 5m and a high altitude of 30-50m respectively, while simultaneously recording meteorological data, and to achieve image georegistration through ground control points deployed at 50m intervals. The hyperspectral data preprocessing module is used to perform atmospheric correction on low-altitude hyperspectral images using dark target subtraction, atmospheric correction on high-altitude hyperspectral images using the FLAASH model, geometric correction based on the ground control points using quadratic polynomial interpolation, image stitching using the SIFT algorithm, and smoothing and denoising the spectral curves using Savitzky-Golay filtering with a window size of 5 and a polynomial order of 2. The intelligent feature extraction and dimensionality reduction module is used to coarsely select hyperspectral bands using a genetic algorithm to obtain 30-40 candidate bands, and then to select key bands from the candidate bands using a continuous projection algorithm. The projection dimension of the continuous projection algorithm is in the range of 5-15, with an error threshold of 0.01, to obtain 10-15 key bands. An integrated deep learning model processing module is used to input the key band data into a CNN module to extract spatial features, input the spatial features into a GRU module to process temporal dynamic features, fuse the spatial features and temporal features through a CBAM module, and obtain the disease monitoring results through a classification output module. The multi-scale model application module is used to output the regional disease incidence rate through a large-scale monitoring model. When the incidence rate is greater than 5%, a small-scale precise model is triggered to locate diseased plants. Combined with time series data from 3-5 time points, the spread rate of the disease is predicted and an early warning map is generated.

Images