Intelligent monitoring system for prevention and control of diseases and insect pests for bean planting

Abstract

The invention relates to an intelligent monitoring system for prevention and control of diseases and insect pests for bean planting, in particular to the field of intelligent monitoring of prevention and control of agricultural diseases and insect pests, and can realize extremely early precise recognition and early warning of diseases and insect pests of bean crops. The weak lesion spectral features and the continuous deterioration trend of the plant can be sensitively captured in a stage which is difficult to perceive by naked eyes, real diseases and environmental interference are effectively distinguished, the system can automatically locate a suspicious lesion area by using a weak supervised learning mechanism and only needing image-level labeling, the dependence on expensive pixel-level labeling data is greatly reduced, and the accuracy of the system is improved. Finally, a high-confidence early warning signal is output through multi-source information collaborative decision, a scientific basis is provided for accurate and green prevention and control, blind pesticide application is effectively avoided, and crop health and yield are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring for the prevention and control of agricultural pests and diseases, and more specifically, to an intelligent monitoring system for the prevention and control of pests and diseases in legume cultivation. Background Technology

[0002] Legumes are highly susceptible to various pests and diseases during their growth, posing a serious threat to yield and quality. Currently, large-scale soybean fields often use fixed poles or mobile monitoring equipment to continuously collect image data of leaves, stems, and pods through visible light cameras to achieve early identification and warning of pests and diseases. However, the field imaging environment is extremely challenging: natural lighting conditions vary drastically, including strong midday light, early morning twilight, and cloudy shadows; overlapping leaves within the crop canopy can obscure the view, and dew, dust, or soil easily adheres to the plant surface; in addition, early symptoms of pests and diseases often manifest as tiny chlorotic spots, weak insect holes, or indistinct mycelia, whose visual characteristics are very similar to normal physiological changes, mechanical damage, or pesticide residues, together forming a complex and ever-changing image background. These internal and external factors significantly increase the difficulty of extracting target information, placing extremely high demands on the reliability and robustness of intelligent monitoring systems.

[0003] In existing technologies, agricultural pest and disease image recognition largely relies on publicly available datasets constructed under standardized conditions, such as PlantVillage, and employs convolutional neural networks for classification or color and texture-based threshold segmentation methods for region extraction. While these methods show good recognition performance for mid-to-late-stage diseases, they have significant limitations when dealing with early-stage and subtle pest and disease features. On the one hand, early lesions or pest areas are highly similar to background noise in terms of pixel-level and texture-level features, resulting in an extremely low signal-to-noise ratio, making it difficult for traditional image processing methods to achieve stable segmentation. On the other hand, in laboratory environments... The trained models lack generalization ability when faced with complex real-world field scenarios, exhibiting significant performance degradation, i.e., a prominent domain adaptation problem. Furthermore, consumer-grade cameras, widely adopted to control costs, have inherent limitations in dynamic range and color fidelity, further weakening the ability to express useful information in images. Therefore, current technologies have not effectively solved the problem of enhancing and extracting weak pest and disease features and achieving high-precision recognition in complex environments. There is an urgent need to develop an image data processing method that can effectively enhance target signals in noisy backgrounds and improve the domain adaptation ability of models, so as to achieve truly intelligent early monitoring and warning of pests and diseases. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing an intelligent monitoring system for pest and disease control in legume cultivation. By integrating multispectral and temporal information, it can keenly capture the faint spectral characteristics of plant lesions and the trend of continuous deterioration at stages that are difficult to detect with the naked eye, effectively distinguishing between real diseases and environmental interference. The system utilizes a weakly supervised learning mechanism, requiring only image-level annotation to automatically locate suspicious lesion areas, significantly reducing the reliance on expensive pixel-level annotation data. Finally, through multi-source information collaborative decision-making, it outputs high-confidence early warning signals, providing a scientific basis for precise and green prevention and control, effectively avoiding blind pesticide application, and ensuring crop health and yield, thereby solving the problems mentioned in the background art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: Specifically, it includes: a data acquisition and registration module, a spectral feature enhancement module, a temporal change detection module, a weakly supervised screening module, and a multi-source decision-making and early warning module, wherein... Data acquisition and registration module: used to control the RGB image sensor and multispectral image sensor to simultaneously acquire image data of soybean plants, and to use a feature point matching algorithm to register the RGB image and multispectral image, and output the registered multimodal image data; Spectral feature enhancement module: Connected to the data acquisition and registration module, it is used to receive the registered multimodal image data and calculate the band reflectance of each pixel in the multispectral image based on the preset disease sensitivity spectral index (DSI) calculation rules to generate a DSI feature image that highlights the spectral features of early lesions. Temporal change detection module: connected to the spectral feature enhancement module, used to acquire DSI feature image sequences of the same observation target over multiple consecutive days in chronological order, and to analyze the trend and statistical significance of the DSI value change of each pixel in the DSI feature image sequence using the Mann-Kendall trend test method. Pixels with a significantly deteriorating trend and a change magnitude exceeding a preset ecological threshold are clustered into a set of candidate regions of interest (ROIs) with significant changes. Weakly supervised screening module: connected to the temporal change detection module, it takes the set of candidate regions of significant change (ROI) and the image-level labels of their respective DSI feature images as input, uses a multi-instance learning model based on attention mechanism for training and inference, calculates the attention weight of each ROI, and filters out suspected lesion regions with high confidence based on the principle that the weight value is higher than the preset confidence threshold. Multi-source decision-making and early warning module: Connected to the weakly supervised screening module, the spectral feature enhancement module and the temporal change detection module, it is used to backtrack from the spectral feature enhancement module to obtain the registered multimodal image data corresponding to the high-confidence suspected lesion area, obtain the temporal change trend features of the DSI value of the area from the temporal change detection module, construct a multi-dimensional feature vector and use a classifier to perform multi-source information fusion decision-making, and finally generate and output early pest and disease warning signals; In a preferred embodiment, the data acquisition and registration module uses a feature point matching algorithm to register the RGB image and the multispectral image as follows: The system controls RGB and multispectral image sensors to receive synchronous trigger signals to acquire RGB and multispectral images of soybean plants. Lens distortion correction and color balance processing are performed on the RGB images, while radiometric calibration correction is applied to the multispectral images to convert their raw digital values ​​into reflectance values. The preprocessed RGB images and a near-infrared band image selected from the multispectral images are then input into a shared-weight Siamese convolutional neural network to extract a first high-dimensional feature vector from the RGB images and a second high-dimensional feature vector from the near-infrared band images. The interaction between the first and second high-dimensional feature vectors is then calculated. The system performs cross-correlation analysis to identify extreme points in the cross-correlation spectrum, thereby establishing a preliminary set of original feature matching point pairs. A cyclic consistency constraint based on cosine similarity is then applied to refine this set, eliminating mismatched pairs and obtaining a precise set of feature matching points. Based on this precise set, the optimal perspective transformation matrix from the multispectral image coordinate system to the RGB image coordinate system is estimated. The multispectral image is then resampled using this optimal perspective transformation matrix to generate a registered multispectral image that is spatially aligned with the RGB image. Finally, the RGB image and the registered multispectral image are packaged and output as multimodal image data with a unified georeferenced reference.

[0006] In a preferred embodiment, the specific process of calculating the band reflectance of each pixel in the multispectral image based on the preset disease sensitivity spectral index (DSI) calculation rules in the spectral feature enhancement module is as follows: The reflectance values ​​of each pixel in different bands are extracted from the registered multimodal image data to form the reflectance vector of that pixel. The first derivative of the reflectance vector in multiple consecutive bands is calculated to highlight the subtle changes in spectral morphology. Based on the prior knowledge of the spectral response of early legume diseases, two characteristic differential bands are selected, and the relative divergence of the first derivative of the pixel in these two characteristic differential bands is calculated. The relative divergence is obtained by calculating the ratio of the absolute value of the difference between the two differential values ​​to their sum plus a small constant. The calculated relative divergence value is input into an S-shaped nonlinear enhancement function for transformation. This function can significantly stretch the input value in the middle region and compress the output range to between zero and one. Finally, the output value of the function is determined as the disease sensitivity spectral index value of that pixel.

[0007] In a preferred embodiment, generating a DSI feature image that highlights the spectral characteristics of early lesions specifically includes: Traverse each pixel in the multispectral image and calculate its corresponding disease sensitivity spectral index value according to the calculation rules. Rearrange the disease sensitivity spectral index values ​​calculated for all pixels according to their original spatial coordinates to form a two-dimensional grayscale image with the same spatial size as the original multispectral image. This two-dimensional grayscale image is the DSI feature image.

[0008] In a preferred embodiment, the specific operation of the temporal change detection module in analyzing the trend and statistical significance of the DSI value change of each pixel in the DSI feature image sequence using the Mann-Kendall trend test method is as follows: For each pixel location in the DSI feature image sequence, its DSI values ​​at different time points are extracted to form the DSI time series of that pixel. The Mann-Kendall statistic for this DSI time series is calculated. This statistic is obtained by summing the sign function values ​​of the differences between the values ​​at each subsequent time point and the values ​​at the previous time point in the sequence. The Mann-Kendall statistic is standardized, and the standard normal distribution table is consulted based on the standardized result to obtain the significance probability value used to measure whether the trend of the DSI value change of that pixel is statistically significant. At the same time, the slope of the DSI time series of that pixel is calculated using the Thiel-Sen estimator. This slope is the median of the slope values ​​for all possible points in the sequence and is used to describe the magnitude of the change in the DSI value per unit time.

[0009] In a preferred embodiment, the specific process of clustering pixels whose change trend is significantly deteriorating and whose change magnitude exceeds a preset ecological threshold into candidate Regions of Interest (ROIs) of significant change is as follows: The calculated saliency probability value of each pixel is compared with a preset statistical significance threshold, and the calculated slope value of each pixel is compared with a preset ecological change magnitude threshold. Only when the saliency probability value of a pixel is less than or equal to the statistical significance threshold and its slope value is greater than or equal to the ecological change magnitude threshold is the pixel marked as a significantly deteriorated pixel. A binary mask image is generated for all marked significantly deteriorated pixels. Morphological closing operation is performed on the binary mask image to connect neighboring pixels, and a connected component labeling algorithm is used to identify all connected pixel regions. Each connected region is defined and output as a significant change candidate region (ROI). The set of all significant change candidate ROIs is denoted as the significant change candidate ROI set. The significant change candidate ROI set and its spatial location information are output to the weakly supervised screening module.

[0010] In a preferred embodiment, the specific process of training and inference using a multi-instance learning model based on an attention mechanism in the weakly supervised screening module is as follows: A set of candidate regions of significant change (ROIs) and their associated DSI feature images are constructed as a bag in a multi-instance learning task. The overall image-level label of the DSI feature image is the label of the bag, and each candidate ROI contained in the image is an instance in the bag. Image patches corresponding to each instance are cropped from the DSI feature image, and a feature encoding network is used to extract features from each instance, converting each region image into a high-dimensional feature vector. An attention network is applied to the high-dimensional feature vectors of all instances. This network calculates the importance of each instance in determining the label of the entire bag through nonlinear transformation, i.e., the attention weight. Finally, a normalized exponential function is used to make the sum of the attention weights of all instances equal to one. The high-dimensional feature vectors of all instances are weighted and summed according to their corresponding attention weights to obtain a comprehensive feature vector representing the entire bag. Finally, this comprehensive feature vector is input into a classifier to calculate the predicted probability that the bag is a disease bag. The parameters of the feature encoding network, attention network, and classifier are jointly optimized by minimizing the difference between the predicted probability and the true image-level label.

[0011] In a preferred embodiment, the specific operation of calculating the attention weight of each ROI and filtering out high-confidence suspected lesion areas based on the principle that the weight value is higher than a preset confidence threshold is as follows: In the inference phase after model training, the set of candidate regions of significant change to be screened and their corresponding DSI feature images are input into the multi-instance learning model. After forward propagation, the attention weight value calculated by the attention network for each instance in the model is obtained. The attention weight value is compared with a pre-set confidence threshold, and instances with attention weights greater than or equal to the confidence threshold are screened out. The original candidate regions of significant change corresponding to these screened instances are determined as high-confidence suspected lesion regions.

[0012] In a preferred embodiment, the specific operation of constructing the multi-dimensional feature vector in the multi-source decision-making and early warning module is as follows: First, based on the spatial location information of high-confidence suspected lesion regions received from the weakly supervised screening module, the registered multimodal image data corresponding to this region is retrieved from the spectral feature enhancement module. Then, the temporal variation trend features of the DSI value of this region are retrieved from the temporal change detection module, including the significance probability value of the Mann-Kendall trend test and the Thiel-Sen estimation slope. Subsequently, the texture feature vector of the high-confidence suspected lesion region in the RGB band is extracted from the retrieved registered multimodal image data. The spectral index feature vector of the high-confidence suspected lesion region in the multispectral band is calculated from the retrieved registered multimodal image data. The temporal variation trend features of the DSI value of this region obtained from the temporal change detection module are directly used to construct the temporal change feature vector. The texture feature vector, spectral index feature vector, and temporal change feature vector are concatenated, and the concatenated vector is standardized to finally generate a joint multimodal feature vector representing the multi-source information of this region.

[0013] In a preferred embodiment, the specific process of using a classifier for multi-source information fusion decision-making is as follows: A classification model based on Bayesian inference is used for decision-making. This model takes a joint multimodal feature vector as input and calculates the posterior probability that the input feature vector belongs to each predefined disease category. The calculation of the posterior probability depends on the prior probability of the category obtained from historical data and the class conditional probability estimated from the training data. Finally, the model selects the disease category with the largest posterior probability value as the final decision output and uses this posterior probability value as the confidence level of this decision. The decision result, confidence level, and spatial location information of the area are packaged to generate a structured early disease and pest warning signal. This warning signal is output to a display terminal, control center, or precision application equipment to guide subsequent prevention and control actions.

[0014] The beneficial effects of this invention are as follows: This system can achieve very early and accurate identification and early warning of diseases and pests in legume crops. By integrating multispectral and temporal information, it can keenly capture the weak spectral characteristics of plant lesions and the trend of continuous deterioration at a stage that is difficult to detect with the naked eye, effectively distinguishing between real diseases and environmental interference. The system utilizes a weakly supervised learning mechanism, which can automatically locate suspicious lesion areas with only image-level annotation, greatly reducing the dependence on expensive pixel-level annotation data. Finally, through multi-source information collaborative decision-making, it outputs high-confidence early warning signals, providing a scientific basis for precise and green prevention and control, effectively avoiding blind application of pesticides, and ensuring crop health and yield. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of the system structure of the present invention. Detailed Implementation

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

[0017] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0019] Example 1 This embodiment provides, for example Figure 1-2 The system, shown, is an intelligent monitoring system for pest and disease control in legume cultivation. Specifically, it includes: a data acquisition and registration module, a spectral feature enhancement module, a temporal change detection module, a weakly supervised screening module, and a multi-source decision-making and early warning module. Data acquisition and registration module: used to control the RGB image sensor and multispectral image sensor to simultaneously acquire image data of soybean plants, and to use a feature point matching algorithm to register the RGB image and multispectral image, and output the registered multimodal image data; Spectral feature enhancement module: Connected to the data acquisition and registration module, it is used to receive the registered multimodal image data and calculate the band reflectance of each pixel in the multispectral image based on the preset disease sensitivity spectral index (DSI) calculation rules to generate a DSI feature image that highlights the spectral features of early lesions. Time-series change detection module: connected to the spectral feature enhancement module, used for time-series detection. ( This indicates a time index, specifically the point in time. Represents the first point in time (e.g., day one). Represents the last point in time (e.g., the last time point) sky), This indicates the total length of the time series, i.e., the total number of images at different time points (e.g., if observations were conducted continuously for 10 days, then T=10). It acquires a sequence of DSI feature images of the same observed target over multiple consecutive days. (A large set containing data from time...) arrive All DSI images were analyzed, and the Mann-Kendall trend test was used to analyze the trend and statistical significance of the DSI value change of each pixel in the DSI feature image sequence. Pixels with a significantly deteriorating trend and a change magnitude exceeding the preset ecological threshold were clustered into a set of candidate regions of interest (ROIs) with significant changes. Weakly supervised screening module: connected to the temporal change detection module, it takes the set of candidate regions of significant change (ROI) and the image-level labels of their respective DSI feature images as input, uses a multi-instance learning model based on attention mechanism for training and inference, calculates the attention weight of each ROI, and filters out suspected lesion regions with high confidence based on the principle that the weight value is higher than the preset confidence threshold. Multi-source decision-making and early warning module: Connected to the weakly supervised screening module, spectral feature enhancement module and temporal change detection module, it is used to backtrack from the spectral feature enhancement module to obtain the registered multimodal image data corresponding to the high-confidence suspected lesion area, obtain the temporal change trend features of the DSI value of the area from the temporal change detection module, construct a multi-dimensional feature vector and use a classifier to perform multi-source information fusion decision, and finally generate and output early pest and disease warning signals.

[0020] In this embodiment, it should be specifically explained that in the data acquisition and registration module, the feature point matching algorithm is used to register the RGB image and the multispectral image as follows: The RGB and multispectral image sensors are controlled to receive synchronous trigger signals to acquire RGB and multispectral images of the bean plant, ensuring that they capture the same scene within a millisecond time difference. This operation aims to minimize inter-frame motion blur caused by plant swaying (such as wind). (subscript) This indicates that the image is a color image composed of the three primary color channels: red, green, and blue. Lens distortion correction and color balance processing are performed, in which... Represents the pixel coordinates of the image, and also for multispectral images. ( Indicates spectral band, It is a variable used to specify which band the image is from, for example: Represents the blue light band. Represents the green light band. Represents the red light band. Represents the red-edge band. Representing the near-infrared band, The two-dimensional coordinates of pixels in a multispectral sensor image are similar to... However, due to the different sensors, the initial coordinate systems are inconsistent. Radiometric calibration correction is then performed to retrieve their original digital values. Convert to reflectance value ( Reflectivity, a physical quantity obtained after radiometric calibration of raw sensor data, represents the ability of a ground feature to reflect sunlight. It has real physical meaning and can be used to compare data from different times and different sensors. Indicates spectral band, Similarly, (representing the two-dimensional coordinates of pixels in a multispectral sensor), the preprocessed RGB image Near-infrared band images selected from multispectral images (subscript) It is an abbreviation for Near-Infrared, specifically referring to the near-infrared band. This band was chosen because its texture features are similar to those of RGB images, which facilitates matching. The data are then input into a shared-weight Siamese convolutional neural network to extract the first high-dimensional feature vector of the RGB image. and the second high-dimensional feature vector of near-infrared band images Calculate the first high-dimensional eigenvector. With the second high-dimensional feature vector Cross-correlation spectrum between (Its value is calculated using cosine similarity; the closer the value is to 1, the greater the probability that the two points are in the same spatial location.) Extreme points are then searched in the cross-correlation spectrum to initially establish a set of original feature-matching point pairs. ( Represents a set The Middle The coordinates of the matching points on the RGB image. Represents a set The Middle (The coordinates of each matching point on the multispectral image) are used to apply a cosine similarity-based cycle consistency constraint to the original set of feature matching point pairs. The process involves purification to remove mismatched point pairs and obtain a set of accurate feature matching point pairs. (subscript) This indicates that the set is "refined or purified," meaning it was eliminated using an algorithm (such as cycle consistency constraints). (High-quality matching pairs obtained after mismatches in the image), and based on the set of accurate feature matching point pairs, estimate the optimal perspective transformation matrix from the multispectral image coordinate system to the RGB image coordinate system. (This matrix defines the spatial transformation relationship (including rotation, translation, scaling, and shearing) between the multispectral image coordinate system and the RGB image coordinate system, used to "distort" and align the multispectral image to the RGB image, utilizing the optimal perspective transformation matrix.) For multispectral images Resampling is performed to generate an RGB image. Spatially perfectly aligned registered multispectral images (superscript) It is an abbreviation for "registered," meaning "already registered." These are its coordinates. Note that the coordinates here have changed to a coordinate system consistent with the RGB image, meaning that the image is now aligned with... (fully aligned in space), ultimately converting the RGB image and registered multispectral images Packaged and output as multimodal image data with unified georeferenced .

[0021] In this embodiment, it is specifically necessary to explain the process by which the spectral feature enhancement module calculates the band reflectance of each pixel in the multispectral image based on the preset calculation rules of the disease sensitivity spectral index (DSI): From the registered multimodal image data Extract the reflectance values ​​of each pixel in different bands to construct the reflectance vector of that pixel, as expressed by: ; in, This represents the reflectance vector of that pixel. The arrow on the symbol indicates that this is a vector, not a single numerical value. This means that the vector corresponds to the spatial coordinates located at The reflectance vector represents the reflectance value of a specific pixel in different spectral bands, serving as its unique spectral "fingerprint". This represents the reflectance value at a specific wavelength and location, and the subscript indicates... This indicates the first band (e.g., the blue band). This indicates that the reflectivity value belongs to spatial coordinates. The pixels, therefore Indicates "in spatial location" Above, the reflectivity value received by the sensor in band 1. Indicates the transpose symbol. This indicates the total number of bands possessed by a multispectral camera; for example, a camera with 5 bands. The first-order derivative of the reflectance vector over multiple consecutive bands is calculated to eliminate background noise and highlight subtle changes in spectral morphology. The formula for the first-order derivative is: ; in, Indicates in the middle section The first-order differential value calculated at this point represents the rate of change of the spectral curve near that wavelength band. It can effectively eliminate background noise caused by changes in illumination and highlight subtle changes in spectral morphology. Indicates in the middle section The reflectance value measured on the next adjacent band, Indicates in the middle section The reflectance value measured in the previous adjacent band, Indicates band The wavelength interval (in nanometers) between them is used to normalize the differential values, making them comparable. The two-dimensional spatial coordinates of the pixel specify the location in the image where the calculation is performed. Based on prior knowledge of the spectral response of early legume diseases, two characteristic differential bands are selected, and the relative divergence of the first-order differential values ​​of the pixel on these two characteristic differential bands is calculated. This relative divergence is obtained by calculating the ratio of the absolute value of the difference between the two differential values ​​to their sum plus a small constant. The formula for relative divergence is: ; in, This represents the calculated relative divergence, a dimensionless value used to quantify the degree of difference between two characteristic differential bands. This indicates the first selected characteristic differential band. The first derivative value on, This indicates the selection of the second characteristic differential band. The first derivative value on, This represents a very small constant (e.g., 0.001), whose main function is to prevent the denominator from being zero and to ensure the numerical stability of the formula calculations. The absolute value sign is used to ensure that the numerator is always positive, amplifying the difference between the two differential values. The calculated relative divergence value is input into an S-shaped nonlinear enhancement function for transformation. This function can significantly stretch the input value in the intermediate region and compress the output range to between zero and one. Finally, the output value of the function is determined as the disease sensitivity spectral index value of that pixel. The formula for the disease sensitivity spectral index value is: ; in, This represents the final calculated disease sensitivity spectral index value, ranging from 0 to 1. A higher value indicates a higher probability that the pixel represents an early-stage disease. This represents the natural constant, approximately equal to 2.71828. This represents the gain factor (a positive real number), which controls the steepness of the Sigmoid function curve. The larger the value, the steeper the curve, and the more sensitive it is to changes in the RD value within the middle range. The offset parameter (a real number) determines the location of the center point of the Sigmoid function, that is, it controls the vicinity of which RD value the DSI value begins to change drastically from 0 to 1; Generating DSI feature images that highlight the spectral characteristics of early lesions specifically includes: Traverse each pixel in the multispectral image and calculate its corresponding disease sensitivity spectral index value according to the calculation rules. The disease sensitivity spectral index values ​​calculated from all pixels are rearranged according to their original spatial coordinates to form a two-dimensional grayscale image with the same spatial dimensions as the original multispectral image. This two-dimensional grayscale image is the DSI feature image. (i.e., in coordinates) The grayscale value of the pixel at position [position] is [value]. This value is a floating-point number, ranging from (0,1). The closer the value is to 1, the higher the confidence that there is early disease at that point. The gray value of each pixel represents the confidence or significance of the presence of early disease symptoms at that spatial location. Finally, the generated DSI feature image is output to the time series change detection module for subsequent analysis.

[0022] In this embodiment, it is specifically necessary to explain the following operation in the time-series change detection module: the Mann-Kendall trend test method is used to analyze the trend and statistical significance of the DSI value change of each pixel in the DSI feature image sequence. For each pixel location in the DSI feature image sequence The DSI values ​​at different time points are extracted to construct the DSI time series of that pixel. ( This represents the DSI time series of a pixel, recording a fixed location. The DSI values ​​of the pixels at all time points are the data foundation for all subsequent analyses. (Representing the DSI value at time point 1), the Mann-Kendall statistic (SR) of this DSI time series is calculated. This statistic is obtained by summing the sign function values ​​of the differences between all subsequent time point values ​​and previous time point values ​​in the series. The formula for the Mann-Kendall SR statistic is: ; in, This represents the Mann-Kendall SR statistic, an integer whose sign indicates the overall direction of the trend (positive for upward, negative for downward), and whose absolute value reflects the degree of consistency of the trend. This represents a sign function, a mathematical function that takes any real number as input and outputs only three possibilities: 1. When the input is a positive number, the output is 1; 2. When the input is negative, the output is -1; 3. When the input is zero, the output is 0; Used to determine whether the value between two points increases or decreases. Indicates the total length of the time series. Indicates the first DSI values ​​at each time point, superscript Indicates the first in chronological order The value of the pixel in the [number]th observation is [value]. The disease sensitivity spectral index obtained by calculation. Indicates the first DSI values ​​at each time point, superscript Indicates the first in chronological order After each observation, the Mann-Kendall statistic is standardized, and the expression is: ; in, This represents the standard deviation of the Mann-Kendall SR statistic, which is based on the total length of the time series. A theoretical value is calculated and used to standardize the S-statistic. Its calculation formula is: , The Z-statistic is represented by the standardized Z-statistic, which is obtained by subtracting the sign (+1 or -1) from the Mann-Kendall SR statistic and then dividing by its standard deviation. This operation transforms the Mann-Kendall SR statistic into a statistic that approximately follows a standard normal distribution, facilitating significance testing. Based on the standardized result, the standard normal distribution table is consulted to obtain the significance probability value used to measure whether the DSI value change trend of this pixel is statistically significant. (according to The p-value is obtained from the standard normal distribution table. It represents the probability that the currently observed trend is caused by random factors (rather than the true trend). The smaller the p-value (e.g., p < 0.05), the more significant the trend. Simultaneously, the slope of the DSI time series for this pixel is calculated using the Thiel-Sen estimator. This slope is the median of the slope values ​​for all possible points in the series, used to describe the magnitude of the DSI value change per unit time. The formula for the slope of the DSI time series for this pixel is: ; in, This represents the slope of the DSI time series of that pixel, i.e., at the pixel coordinates. At this point, the average rate of change of the DSI value over time is calculated by the Thiel-Sen estimator and is insensitive to outliers. The median function is a statistical function that represents the middle value of a sorted set of values. It's used to calculate the slope of a sequence of values, avoiding the influence of extreme values. Indicates the first DSI values ​​at each time point, superscript Indicates the first in chronological order The value of the pixel in the [number]th observation is [value]. The disease sensitivity spectral index obtained by calculation. Indicates the first DSI values ​​at each time point, superscript Indicates the first in chronological order Second observation, This represents a point-in-time index, a sequential numbered index, and is defined as follows: That is, time point At the point of time Before, The denominator represents the time interval, the time difference between two points in time (usually in "days"); it is used to calculate the change per unit of time. The specific process for clustering pixels with a significantly deteriorating trend and a change magnitude exceeding a preset ecological threshold into candidate Regions of Interest (ROIs) of significant change is as follows: The calculated saliency probability value of each pixel is compared with a preset statistical significance threshold. (Its value is set to 0.05) and compared, and the slope value of each calculated pixel is compared with a preset ecological change amplitude threshold. (Its value is set to) The comparison is performed only if the significance probability value of a pixel is less than or equal to the statistical significance threshold. And its slope value is greater than or equal to the ecological change magnitude threshold ( Only when the pixel is marked as significantly degraded will a binary mask image be generated from all marked significantly degraded pixels. (An image of the same size as the original image, with pixel values ​​of only 0 or 1, where a value of 1 represents that pixel) The binarized mask image was labeled as "significantly degraded pixels". Morphological closing operations are performed to connect neighboring pixels, and a connected component labeling algorithm (such as the Two-Pass algorithm based on disjoint-set data structure) is used to identify all connected pixel regions. Each connected region is defined and output as a candidate region of significant change (ROI). The set of all candidate ROIs of significant change is denoted as the set of candidate ROIs of significant change. (subscript) This represents each region in the enumerated set (e.g., It's the first area. It is the second area), therefore, This represents a set containing all candidate regions, and outputs a set of candidate regions (ROIs) with significant changes and their spatial location information to the weakly supervised screening module.

[0023] In this embodiment, it is specifically necessary to explain the training and inference process of the multi-instance learning model based on the attention mechanism in the weakly supervised screening module as follows: Set of candidate regions (ROIs) that will significantly change and its associated DSI feature image Constructed as a package for a multi-instance learning task, where the overall image-level label of the DSI feature image is... (subscript) It's an abbreviation for "image," emphasizing that this is a label for the entire image, not a label for a specific region within the image. Furthermore, It is a binary scalar, serving as a supervisory signal in weakly supervised learning. This indicates that there is a disease in the image. The label for this packet is the region of health (indicating the image is healthy), and each candidate region of significant change contained in the image is considered a potential candidate region. This is an instance within the package, derived from the DSI feature image. Image patches corresponding to each instance are cropped from the image, and a feature encoding network (such as a convolutional neural network) is used to extract features from each instance, converting each region image into a high-dimensional feature vector. ( Indicates the first The feature vector of the nth instance, that is, representing the nth instance The high-dimensional numerical vectors obtained after the candidate region image patches are processed by the feature encoding network are abstract representations of the spectral-spatial features of the region and form the basis for subsequent calculation of attention weights. An attention network is applied to the high-dimensional feature vectors of all instances. This network calculates the importance of each instance in determining the overall packet label through nonlinear transformations, i.e., the attention weight. Finally, a normalized exponential function is used to make the sum of the attention weights of all instances equal to one, as expressed in the following expression: ; in, This represents the weight matrix in an attention network. It is a learnable parameter matrix used to weight the input feature vector. The first linear transformation maps it to a new feature space, aiming to better extract features that help distinguish the importance of instances. This represents the weight vector in the attention network. It is a learnable parameter vector that acts on the features after tanh activation. Essentially, it is a query vector used to perform a dot product with the transformed instance features to calculate the relevance score of that instance to the current learning task. This represents the hyperbolic tangent activation function, a nonlinear function that introduces nonlinearity into the transformation process, enabling the model to learn more complex patterns while compressing the output value to the range of (-1, 1) to prevent excessively large values. Indicates the first The unnormalized attention score for each instance, which is a scalar reflecting the original importance of that instance. Indicates the first The normalized attention weight for each instance is a scalar between 0 and 1, with the sum of all weights being 1. It represents the probability or contribution of the model in considering that instance to be a true lesion. The higher the value, the more confident the model is that the area is a true lesion. The lower the value, the more likely the area is a false positive (such as shadows, dust, or abnormalities in healthy tissue). The transpose symbol indicates that a vector or matrix is ​​transposed, for example... This indicates that the column vector The vector is transposed to a row vector for dot product operations. `exp` represents the natural exponential function, which maps the input to a positive range. This is a standard component of the Softmax function, ensuring that the output weights are positive. Indicates the first The feature vector of each instance The total number of instances refers to the total number of candidate regions (ROIs) in an image, expressed as a summation operation in the denominator. This is calculated for all K instances. The index representing the summation operation, and the loop variable used for summing the denominator. Iterating from 1 to K represents iterating through each instance. The same numerator calculation is performed on all instances, and the high-dimensional feature vectors of all instances are weighted and summed according to their corresponding attention weights, thereby fusing them to obtain a comprehensive feature vector representing the entire package. The expression is: ; in, The aggregate representation of the package is a vector that is the feature vector of all instances (candidate regions). According to their respective attention weights The weighted summation yields a comprehensive feature vector that represents the overall features of the entire image (packet). The symbol represents the summation operation, which sums the values ​​from the first instance to the Kth instance (a total of K instances). Indicates the first Normalized attention weights for each instance, Indicates the first The feature vectors of each instance are then combined and input into a classifier to calculate the predicted probability that the package is a diseased package. (It is a scalar (a number between 0 and 1), which is the aggregate representation vector of the package.) The output obtained after inputting into a classifier (such as a sigmoid classifier) ​​represents the probability that the model predicts the presence of a disease in the image. For example, =0.85 indicates that the model has an 85% confidence level in identifying lesions in the image, and this is achieved by minimizing the predicted probability relative to the true image-level label. (This is a binary scalar (taking values ​​of 0 or 1). These are manually labeled image-level tags, where 1 represents that the image is "sick" and 0 represents "healthy." It doesn't tell you where the lesions are, but only whether the whole image is sick.) The differences between these tags are used to jointly optimize the parameters of the feature encoding network, attention network, and classifier. The specific steps for calculating the attention weight of each ROI and filtering out high-confidence suspected lesion areas based on weight values ​​higher than a preset confidence threshold are as follows: During the inference phase after model training, the set of candidate regions of significant change (ROIs) to be screened and their respective DSI feature images are input into the multi-instance learning model. After forward propagation, the attention weight values ​​calculated by the attention network for each instance in the model are obtained. The attention weight value is then compared with a pre-set confidence threshold. ,For example ) are compared, and instances whose attention weight is greater than or equal to the confidence threshold are selected. These selected instances correspond to the original significant change candidate regions. The regions identified as high-confidence suspected lesions are then output to subsequent modules along with their spatial location information in the original image.

[0024] In this embodiment, it is necessary to specifically explain the operation of constructing multi-dimensional feature vectors in the multi-source decision-making and early warning module as follows: First, based on the high-confidence suspected lesion areas received from the weakly supervised screening module... The spatial location information is used to retrieve the registered multimodal image data corresponding to the region from the spectral feature enhancement module. This data block contains RGB information and multispectral band information, and retrieves the time-series change trend characteristics of the DSI values ​​in this region from the time-series change detection module, including the significance probability value of the Mann-Kendall trend test. And Thiel-Sen estimated the slope Subsequently, the registered multimodal image data obtained from the retrospective process... In the process, texture feature vectors of high-confidence suspected lesion regions in the RGB band are extracted. Registered multimodal image data obtained by backtracking from the spectral feature enhancement module. In this study, the spectral index eigenvectors of high-confidence suspected lesion regions in the multispectral bands were calculated. The time-series change trend characteristics of the DSI values ​​in this region obtained from the time-series change detection module will be directly used to construct the time-series change feature vector. The texture feature vector, spectral index feature vector, and temporal variation feature vector are concatenated, and the concatenated vector is standardized to generate a joint multimodal feature vector that represents the multi-source information of the region. The expression is: ; in, This represents the joint multimodal feature vector, a high-order, high-dimensional vector. Its unique feature lies in its simultaneous integration of spatial (texture), spectral, and temporal information, providing unprecedentedly comprehensive data for subsequent classifiers to make more accurate judgments. The texture feature vector representing the RGB band, i.e., the features extracted from the RGB image, typically contains information describing the visual texture of the diseased area. Examples include feature values ​​calculated using algorithms such as Local Binary Pattern (LBP) and Gray-Level Co-occurrence Matrix (GLCM). These features can capture changes in leaf surface roughness and pattern caused by lesions. The spectral index feature vector represents features extracted from multispectral images. It typically contains a series of vegetation indices (such as NDVI, EVI, PRI, etc.) calculated from the reflectance of different spectral bands. These indices can sensitively reflect the biochemical parameters of vegetation (such as chlorophyll content and water stress) and are core spectral evidence for detecting early diseases. This represents the feature vector of time-series changes, i.e., the features extracted from the time-series image, and usually includes significance probability values ​​obtained from the Mann-Kendall trend test. The value) and the slope of change obtained from the Thiel-Sen estimator ( These characteristics quantify the region's continued deterioration trend and statistical significance over a historical period. This represents the RGB feature processing function, which is a processing function whose input is the original RGB texture feature vector. The output is a new vector after processing. The purpose of processing is usually standardization (to eliminate the influence of dimensions) or dimensionality reduction (to extract core information and reduce the amount of computation). This represents the spectral feature processing function, which is a processing function whose input is the original spectral feature vector. The output is a new vector after processing, and the purpose of the processing is the same as... Similarly, ensuring that spectral features can be effectively integrated with other features on the same scale, This represents a time-series feature processing function. It is a processing function whose input is the original time-series variation feature vector. The output is the processed new vector. This represents the vector concatenation operator; The specific process of using a classifier for multi-source information fusion decision-making is as follows: A classification model based on Bayesian inference is used for decision-making, which employs a joint multimodal feature vector. As input, the posterior probability of the input feature vector belonging to each predefined disease category is calculated. The calculation of the posterior probability depends on the prior probability of the category obtained from historical data and the class-conditional probability estimated from the training data, as shown in the formula: ; in, Represents the joint multimodal feature vector. The category label is a variable representing the possible categories to which a sample might belong. In this application, its value comes from a predefined set of categories. ,For example It could be "health," "rust," "mold," etc. This represents a predefined set of categories, containing all possible category labels, for example... ={health, rust, mold} This represents the prior probability, that is, the probability that a sample belongs to a class without any data evidence. The initial probability, for example, if 80% of the bean plants in the field are healthy, then P(healthy) = 0.8. This probability is usually obtained from historical data. This represents the class conditional probability (likelihood), which is the probability given that a sample belongs to a given class. Under these conditions, joint multimodal eigenvectors were observed. The probability of the joint multimodal feature vector is measured. Category The degree of "support" is usually estimated from the training data. The evidence (marginal probability) is represented by the joint multimodal feature vector. The total probability of occurrence (regardless of category) is, in practice, a normalization constant that ensures the sum of all posterior probabilities is 1, and is therefore often ignored in comparisons. Representing the posterior probability, this is the core output of Bayesian inference, indicating the probability of the observed multimodal feature vectors. Under these conditions, the sample belongs to category The probability is what we ultimately want to calculate. This indicates finding the parameter corresponding to the maximum value. The predicted category label is obtained by... The operation finds the class that maximizes the posterior probability, which is the class the model believes the sample most likely belongs to. Ultimately, the model selects the posterior probability value. The largest category of disease As the final decision output, this posterior probability value is used. As a measure of confidence in this decision, the decision outcome will be... Confidence level The system packages the spatial location information of the area to generate a structured early warning signal for pests and diseases. This warning signal is then output to a display terminal, control center, or precision application equipment to guide subsequent prevention and control actions.

[0025] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0027] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0028] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0029] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0030] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0031] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent monitoring system for the prevention and control of diseases and pests in legume cultivation, characterized by, Specifically comprising: The data acquisition and registration module, the spectral feature enhancement module, the time series change detection module, the weakly supervised screening module, and the multi-source decision warning module, wherein The data acquisition and registration module is used to control the RGB image sensor and the multispectral image sensor to synchronously collect image data of the bean plant, and uses a feature point matching algorithm to register the RGB image and the multispectral image, and outputs the registered multi-modal image data; The spectral feature enhancement module is connected to the data acquisition and registration module, and is used to receive the registered multi-modal image data, and calculate the band reflectivity of each pixel in the multispectral image based on a preset disease-sensitive spectral index (DSI) calculation rule, to generate a DSI feature image highlighting the spectral features of early lesions; The time series change detection module is connected to the spectral feature enhancement module, and is used to obtain a sequence of DSI feature images of the same observation target for consecutive days in time sequence, and uses the Mann-Kendall trend test method to analyze the DSI value change trend and statistical significance of each pixel in the DSI feature image sequence, and clusters the pixels with a significant deterioration trend and a change amplitude exceeding a preset ecological threshold into a set of significant change candidate regions (ROIs); The weakly supervised screening module is connected to the time series change detection module, and is used to take the set of significant change candidate regions (ROIs) and the image-level labels of the DSI feature images to which they belong as inputs, and uses a multi-instance learning model based on an attention mechanism to train and infer, calculate the attention weight of each ROI, and screen out high-confidence suspected lesion regions according to the principle that the weight value is higher than a preset confidence threshold; The multi-source decision warning module is connected to the weakly supervised screening module, the spectral feature enhancement module, and the time series change detection module, and is used to obtain the registered multi-modal image data corresponding to the high-confidence suspected lesion region from the spectral feature enhancement module, and obtain the DSI value time series change trend feature of the region from the time series change detection module, construct a multi-dimensional feature vector, and use a classifier to make a multi-source information fusion decision, and finally generate and output an early disease and pest warning signal.

2. The intelligent monitoring system for the prevention and control of diseases and pests for legume planting according to claim 1, characterized in that: In the data acquisition and registration module, the feature point matching algorithm used to register the RGB image and the multispectral image is specifically: The RGB image sensor and the multispectral image sensor receive a synchronous trigger signal to collect an RGB image and a multispectral image of a bean plant, the RGB image is subjected to lens distortion correction and color balance processing, and the multispectral image is subjected to radiometric calibration correction to convert the original digital value of the multispectral image into a reflectance value, the preprocessed RGB image and an image of a near-infrared wave band selected from the multispectral image are input into a convolutional neural network with shared weights to extract a first high-dimensional feature vector of the RGB image and a second high-dimensional feature vector of the image of the near-infrared wave band, the cross-correlation spectrum between the first high-dimensional feature vector and the second high-dimensional feature vector is calculated, and extreme points in the cross-correlation spectrum are searched to preliminarily establish a set of original feature matching point pairs, the set of original feature matching point pairs is purified by using a cyclic consistency constraint based on cosine similarity to eliminate false matching point pairs and obtain a set of accurate feature matching point pairs, an optimal perspective transformation matrix from a multispectral image coordinate system to an RGB image coordinate system is estimated according to the set of accurate feature matching point pairs, the multispectral image is resampled by using the optimal perspective transformation matrix to generate a registered multispectral image that is completely aligned with the RGB image space, and finally the RGB image and the registered multispectral image are packaged as output as multi-modal image data with unified geographic reference.

3. The intelligent monitoring system for the prevention and control of diseases and pests for legume planting according to claim 2, characterized in that: In the spectral feature enhancement module, the specific process of operating the band reflectance of each pixel in the multispectral image based on the preset calculation rule of the disease sensitive spectral index DSI is as follows: The reflectance values of each pixel in the registered multi-modal image data in different wave bands are extracted to form a reflectance vector of the pixel, the first derivative values of the reflectance vector in multiple continuous wave bands are calculated to highlight the subtle changes in spectral morphology, two characteristic differential wave bands are selected according to the prior knowledge of the spectral response of early diseases of legumes, and the relative dispersion of the first derivative values of the pixel in the two characteristic differential wave bands is calculated, the relative dispersion is obtained by calculating the ratio of the absolute value of the difference between the two differential values to the sum of the two differential values plus a small constant, the calculated relative dispersion value is input into an S-shaped nonlinear enhancement function for transformation, the function can significantly stretch the input values in the middle region and compress the output range to between zero and one, and finally the output value of the function is determined as the disease sensitive spectral index value of the pixel.

4. The intelligent monitoring system for the prevention and control of diseases and pests for legume planting according to claim 3, characterized in that: The specific process of generating the DSI feature image highlighting the spectral features of early diseases includes: Each pixel in the multispectral image is traversed, and the disease sensitive spectral index value of each pixel is calculated according to the calculation rule, and the disease sensitive spectral index values of all pixels are rearranged according to their original spatial coordinate positions to form a two-dimensional gray image with the same spatial size as the original multispectral image, and the two-dimensional gray image is the DSI feature image.

5. The intelligent monitoring system for the prevention and control of diseases and pests for legume planting according to claim 4, characterized in that: In the time series change detection module, the specific operation of analyzing the change trend and statistical significance of the DSI value of each pixel in the DSI feature image sequence by using the Mann-Kendall trend test method is as follows: For each pixel position in the DSI feature image sequence, the DSI values at different time points are extracted to form a DSI time sequence of the pixel, a Mann-Kendall statistic of the DSI time sequence is calculated, the statistic is obtained by calculating the sum of the sign function values of the differences between the values at all subsequent time points and the values at the previous time points in the sequence, the Mann-Kendall statistic is normalized, and according to the normalized result, a standard normal distribution table is queried to obtain a significance probability value for measuring whether the change trend of the DSI value of the pixel has statistical significance, at the same time, a Theil-Sen estimator is used to calculate the slope of the DSI time sequence of the pixel, the slope is the median of all possible point slope values in the sequence, which is used to describe the change amplitude of the DSI value per unit time.

6. The intelligent monitoring system for the prevention and control of diseases and pests for legume planting according to claim 5, characterized in that: The specific process of clustering the pixels with significant deterioration and change amplitude exceeding the preset ecological threshold into a significant change candidate region ROI is: The significance probability value of each pixel calculated is compared with a preset statistical significance threshold, and the slope value of each pixel calculated is compared with a preset ecological change amplitude threshold, only when the significance probability value of a certain pixel is less than or equal to the statistical significance threshold and its slope value is greater than or equal to the ecological change amplitude threshold, the pixel is marked as a significant deterioration pixel, a binary mask image is generated for all marked significant deterioration pixels, a morphological closing operation is performed on the binary mask image to connect adjacent pixels, and a connected component labeling algorithm is used to identify all connected pixel regions, each connected region is defined and output as a significant change candidate region ROI, and the set of all significant change candidate regions ROI is called a significant change candidate region ROI set, and the significant change candidate region ROI set and its spatial position information are output to a weak supervision screening module.

7. The intelligent monitoring system for the prevention and control of diseases and pests for legume planting according to claim 6, characterized in that: In the weak supervision screening module, the specific process of training and inference using a multi-instance learning model based on an attention mechanism is: A set of candidate regions of significant change (ROIs) and their associated DSI feature images are constructed as a bag in a multi-instance learning task. The overall image-level label of the DSI feature image is the label of the bag, and each candidate ROI contained in the image is an instance in the bag. Image patches corresponding to each instance are cropped from the DSI feature image, and a feature encoding network is used to extract features from each instance, converting each region image into a high-dimensional feature vector. An attention network is applied to the high-dimensional feature vectors of all instances. This network calculates the importance of each instance in determining the label of the entire bag through nonlinear transformation, i.e., the attention weight. Finally, a normalized exponential function is used to make the sum of the attention weights of all instances equal to one. The high-dimensional feature vectors of all instances are weighted and summed according to their corresponding attention weights to obtain a comprehensive feature vector representing the entire bag. Finally, this comprehensive feature vector is input into a classifier to calculate the predicted probability that the bag is a disease bag. The parameters of the feature encoding network, attention network, and classifier are jointly optimized by minimizing the difference between the predicted probability and the true image-level label.

8. The intelligent monitoring system for the prevention and control of diseases and pests in legume planting according to claim 7, characterized in that: The specific steps for calculating the attention weight of each ROI and filtering out high-confidence suspected lesion areas based on the principle that the weight value is higher than a preset confidence threshold are as follows: In the inference phase after model training, the set of candidate regions of significant change to be screened and their corresponding DSI feature images are input into the multi-instance learning model. After forward propagation, the attention weight value calculated by the attention network for each instance in the model is obtained. The attention weight value is compared with a pre-set confidence threshold, and instances with attention weights greater than or equal to the confidence threshold are screened out. The original candidate regions of significant change corresponding to these screened instances are determined as high-confidence suspected lesion regions.

9. The intelligent monitoring system for the prevention and control of diseases and pests in legume cultivation according to claim 8, characterized in that: In the multi-source decision-making and early warning module, the specific operation for constructing the multi-dimensional feature vector is as follows: First, based on the spatial location information of high-confidence suspected lesion regions received from the weakly supervised screening module, the registered multimodal image data corresponding to this region is retrieved from the spectral feature enhancement module. Then, the temporal variation trend features of the DSI value of this region are retrieved from the temporal change detection module, including the significance probability value of the Mann-Kendall trend test and the Thiel-Sen estimation slope. Subsequently, the texture feature vector of the high-confidence suspected lesion region in the RGB band is extracted from the retrieved registered multimodal image data. The spectral index feature vector of the high-confidence suspected lesion region in the multispectral band is calculated from the retrieved registered multimodal image data. The temporal variation trend features of the DSI value of this region obtained from the temporal change detection module are directly used to construct the temporal change feature vector. The texture feature vector, spectral index feature vector, and temporal change feature vector are concatenated, and the concatenated vector is standardized to finally generate a joint multimodal feature vector representing the multi-source information of this region.

10. The intelligent monitoring system for the prevention and control of diseases and pests for legume planting according to claim 9, characterized in that: The specific process of using a classifier for multi-source information fusion decision-making is as follows: A classification model based on Bayesian inference is used for decision-making. This model takes a joint multimodal feature vector as input and calculates the posterior probability that the input feature vector belongs to each predefined disease category. The calculation of the posterior probability depends on the prior probability of the category obtained from historical data and the class conditional probability estimated from the training data. Finally, the model selects the disease category with the largest posterior probability value as the final decision output and uses this posterior probability value as the confidence level of this decision. The decision result, confidence level, and spatial location information of the area are packaged to generate a structured early disease and pest warning signal. This warning signal is output to a display terminal, control center, or precision application equipment to guide subsequent prevention and control actions.

Images