A multi-modal data fusion rice disease and pest diagnosis and trend prediction method

Abstract

The present application relates to the technical field of wisdom agriculture, and more particularly to a rice disease and pest diagnosis and trend prediction method based on multi-modal data fusion. Real-time multi-modal data of a specific rice variety in the current growth period is obtained, and the data is analyzed in combination with benchmark data of the variety retrieved from a pre-built knowledge base. Phenotypic features are extracted from images and compared with standard features of the variety to generate a first abnormal feature vector. Environmental data is compared with environmental response thresholds of the variety to generate a second abnormal feature vector. A dynamic weight distributor is used to fuse the two vectors according to the growth period and stress degree to form a comprehensive disease feature vector. The vector is then input into a decision logic network based on agronomic knowledge for reasoning to obtain a preliminary diagnosis result. Finally, yield risk is evaluated in combination with a historical yield database, and the development of diseases and pests is deduced based on future environmental trends to generate an integrated diagnosis and trend prediction report. The present application realizes precise diagnosis and forward-looking warning specific to the variety.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, and in particular to a method for diagnosing and predicting trends of rice diseases and pests through multimodal data fusion. Background Technology

[0002] Currently, the diagnosis of rice diseases and pests mainly relies on visual field surveys by agricultural technicians. This method is inefficient, subjective, and difficult to implement for large-scale, precise monitoring. With the development of agricultural information technology, some auxiliary diagnostic methods based on single-type sensor data, such as visible light images or specific environmental sensors, have emerged. However, these methods often overlook the inherent differences in morphology, spectrum, and stress resistance among different rice varieties. Using uniform diagnostic standards can easily lead to misjudgments. Furthermore, they lack cross-modal collaborative analysis and dynamic weighting of phenotypic abnormalities reflected in images and environmental stress factors, resulting in one-sided and delayed diagnostic results that cannot effectively support precise management and forward-looking early warning for specific varieties.

[0003] Furthermore, existing technological solutions have significant limitations in terms of knowledge-driven and predictive capabilities. Most methods only involve simple analysis of real-time collected data and threshold alerts, failing to deeply integrate domain knowledge and systematically correlate real-time data with variety-specific historical growth benchmarks, disease occurrence environmental thresholds, and historical yield impact records. In addition, existing solutions typically only provide a judgment of the current state, lacking the ability to predict trends based on multimodal information fusion and future environmental extrapolation. They cannot assess the potential development paths of pests and diseases and their possible impact on final yield, making it difficult to meet the needs of modern precision agriculture for risk prediction and decision support.

[0004] Therefore, there is an urgent need for an intelligent diagnostic method that can fully consider variety differences, achieve deep integration of multimodal information, and possess knowledge-driven and trend prediction capabilities. Summary of the Invention

[0005] To achieve the above objectives, this application provides the following technical solution: A diagnostic and trend prediction method based on multimodal data fusion, characterized in that the method includes: S1. Obtain real-time multimodal data of the target rice plant during its current growth stage. Based on the pre-determined identifier of the target rice variety, retrieve and call up the corresponding variety baseline data from the pre-built rice variety knowledge base. S2, input the visible light image and near-infrared image from the real-time multimodal data into the pre-configured image feature extraction module to extract real-time image features; S3, based on the standard morphological spectral features in the variety benchmark data, the extracted real-time image features are compared with the standard morphological spectral features of the corresponding standard growth period, the deviation information of the image features is calculated, and a first abnormal feature vector is generated based on the deviation information. S4, compare the environmental data with the preset environmental response threshold corresponding to the standard growth period in the variety benchmark data, calculate the weighted stress index for different environmental stress factors based on the comparison results, and generate a second abnormal feature vector based on the weighted stress index. S5, invoke the dynamic weight allocator, determine the weight allocation strategy for fusing the first abnormal feature vector and the second abnormal feature vector according to the current reproductive period and the stress level indicated by the weighted stress index, and fuse the first abnormal feature vector and the second abnormal feature vector according to the weight allocation strategy to form a comprehensive disease feature vector; S6, input the comprehensive disease feature vector into the decision logic network constructed based on agronomic knowledge, perform logical reasoning based on the input vector, and output preliminary disease and pest diagnosis results; S7. Access the historical yield database to obtain the yield impact coefficients of different rice varieties under historical pest and disease occurrence conditions. Based on the identification of the target rice variety and the preliminary pest and disease diagnosis results, match the corresponding historical impact coefficients from the historical yield database. S8. Combining the historical impact coefficients of the matching and the preliminary disease and pest diagnosis results, assess the potential impact of the currently diagnosed diseases and pests on the final yield of the target rice plants, obtain a yield risk forecast, and generate an integrated diagnosis and trend prediction report by combining the extrapolation of environmental stress trends in the future preset time period.

[0006] Furthermore, the image feature extraction module, the dynamic weight allocator, the decision logic network, the historical yield database, and the rice variety knowledge base together constitute the diagnosis and prediction system; The method also includes a system initialization step: Before performing the diagnosis on the target rice variety for the first time, the corresponding benchmark data of the variety is loaded from the rice variety knowledge base to the local cache according to the identifier of the target rice variety, and the feature extraction parameters corresponding to the standard morphological spectral features in the image feature extraction module are configured.

[0007] Furthermore, S2 specifically includes the following steps: Spatial registration is performed on the synchronously acquired visible light image and near-infrared image to ensure that pixels at the same spatial location correspond to the same plant part. Based on the registered visible light image, a pre-trained plant segmentation model is used to identify the canopy region of the target rice plant, which is then used as the effective region for subsequent feature analysis. Within the effective area, the visible light image is converted from the original color space to the target color space, and the statistical values ​​of all pixels within the effective area in a specific channel of the target color space are calculated as leaf color features. Within the effective area, a multi-scale texture analysis operator is applied to the grayscale image of the visible light image to calculate the texture descriptor at each scale. The texture descriptors at multiple scales are then concatenated to form a multi-dimensional texture feature vector, which serves as the texture feature. Within the effective area, the pixel intensity value of the corresponding area in the registered near-infrared image is read, and the statistical value of the near-infrared band reflectance within the effective area is calculated as the near-infrared reflectance feature. The leaf color feature, the texture feature vector, and the near-infrared reflectance feature are spliced ​​and normalized in a preset order to form a unified real-time image feature vector for subsequent comparison.

[0008] Furthermore, S3 specifically includes the following steps: Based on the current growth period information, find the standard morphological spectral characteristics corresponding to the standard growth period that is the same as or closest to the current growth period from the loaded variety benchmark data; The real-time image feature vector is subtracted element-wise from the matched standard morphological spectral features to obtain the original deviation vector. Each element in the original deviation vector is divided by the reference value of the corresponding element in the standard morphological spectral features to obtain the relative deviation ratio vector. For each type of feature deviation in the relative deviation ratio vector, a separate anomaly judgment interval is preset to distinguish between normal fluctuations and suspected anomalies. The relative deviation ratio vector is traversed, and each element is mapped to a discrete anomaly level according to the anomaly judgment interval into which its value falls. Using one-hot encoding or numerical encoding, the anomaly level corresponding to each feature is converted into a sub-vector of fixed length. All sub-vectors corresponding to all features are connected in a predetermined order to form the first anomaly feature vector, which reflects the number of features and the anomaly state of the monitored image.

[0009] Furthermore, S4 specifically includes: The temperature data is compared with the suitable temperature range in the environmental response threshold. If the real-time temperature exceeds the suitable temperature range, the temperature stress index is calculated based on the extent of the exceedance and the preset temperature stress coefficient. The humidity data is compared with the suitable humidity range in the environmental response threshold. If the real-time humidity exceeds the suitable humidity range, the humidity stress index is calculated based on the extent of the exceedance and the preset humidity stress coefficient. The weighted stress index is obtained by weighting and summing the temperature stress index and the humidity stress index according to their relative importance in inducing specific diseases.

[0010] Furthermore, S4 also includes: Multiple consecutive weighted stress index intervals are set, each interval corresponding to a preset environmental stress level; The calculated weighted stress index is mapped to the corresponding environmental stress level; The mapped environmental stress level is encoded into a multi-dimensional vector, where each dimension represents a preset disease type or stress category. The value of this dimension represents the environmental stress level or the environmental induced risk level of the disease type derived from the environmental stress level. The resulting vector serves as the second abnormal feature vector.

[0011] Furthermore, the process by which the dynamic weight allocator determines the weight allocation strategy includes: The dynamic weight allocator has a built-in weight configuration table that defines the fusion weight coefficients of the first abnormal feature vector and the second abnormal feature vector under different reproductive periods and different stress levels. The dynamic weight allocator queries the weight configuration table based on the input current reproductive period information and the degree of stress reflected by the second abnormal feature vector or directly indicated by the weighted stress index to obtain a corresponding set of fusion weight coefficients. The process of fusing the first abnormal feature vector and the second abnormal feature vector according to the weight allocation strategy is as follows: Each element of the first abnormal feature vector is multiplied by the corresponding first weight coefficient, and each element of the second abnormal feature vector is multiplied by the corresponding second weight coefficient. The two weighted vectors are then concatenated to form the dimensionally expanded comprehensive disease feature vector.

[0012] Furthermore, the decision logic network is a production rule-based system that performs logical reasoning and outputs preliminary pest and disease diagnosis results, as well as subsequent evaluation and report generation, specifically including the following steps: The decision logic network includes agronomic diagnostic rules, and the condition part of each rule is associated with a specific pattern or value range of the comprehensive disease feature vector; The comprehensive disease feature vector is matched with the condition parts of all rules to filter out all rules whose conditions are met. Each triggered rule outputs one or more candidate pest / disease types and their confidence factors. All candidate pest and disease types output by the triggered rules are integrated, the confidence factors of the same pest and disease type are aggregated and calculated, and the pest and disease type with the highest confidence is selected as the preliminary pest and disease diagnosis result according to the sorting of the aggregated confidence factors. Using the identifier of the target rice variety and the pest and disease types identified in the preliminary pest and disease diagnosis results as the query keys, a search is performed in the historical yield database; The historical yield database is stored in the form of a data structure, in which each record is associated with a variety identifier, a type of pest or disease, and a yield impact coefficient obtained through statistical analysis of historical data; If a perfect match is found, extract its stored production impact coefficient; If no exact match is found, then according to the preset approximate matching rules, find records of similar diseases and pests of the same variety or other varieties of the same diseases and pests, and estimate the historical influence coefficient by interpolation or averaging. Using the historical impact coefficient obtained through matching or estimation as the basic impact factor, the characteristic values ​​representing the severity of pests and diseases in the comprehensive disease feature vector are analyzed, and the ratio of the current severity to the historical typical severity is calculated as the severity adjustment factor. The potential impact level is calculated by multiplying the basic impact factor by the severity adjustment factor and combining it with the sensitivity parameter of the current fertility period to yield, and then quantified as the yield risk forecast. Acquire predicted environmental data for a predetermined future time period, and use the same process as for real-time environmental data to calculate the weighted stress index for future time points. Generate simulated values ​​of the second anomaly feature vector for the future. Combine this with the current first anomaly feature vector to extrapolate simulated values ​​of the first anomaly feature vector for future time points. The dynamic weight allocator is used to fuse the simulated feature vectors of the future, and the result is input into the decision logic network for forward reasoning to predict possible changes in the state of pests and diseases and generate a prediction of the development trend of pests and diseases. The preliminary disease and pest diagnosis results, the yield risk forecast, and the disease and pest development trend prediction are structured and organized, and the timestamp of the generated report, the target rice variety identifier, the current growth stage, and the variety benchmark data version information used are attached, and packaged into the integrated diagnosis and trend prediction report.

[0013] Furthermore, the method also includes report output and feedback steps: The generated integrated diagnostic and trend prediction report is converted into a file in a specified format or displayed in a visual interface; The complete context data of this diagnosis, including the real-time multimodal data, the first abnormal feature vector, the second abnormal feature vector, the comprehensive disease feature vector, the preliminary disease and pest diagnosis results, and the final generated report, are desensitized and formatted before being stored in the local diagnosis log database. Regularly perform statistical analysis on the cases in the local diagnostic log database. The results can serve as a basis for optimizing the variety benchmark data in the rice variety knowledge base, adjusting the rules in the decision logic network, or updating the influence coefficients in the historical yield database.

[0014] Furthermore, before the step of acquiring real-time multimodal data of the target rice plant at its current growth stage, the method also includes a data quality verification step: Check the online status and clock synchronization status of the image acquisition device and the environmental sensor; The received visible light image and near-infrared image are checked for sharpness and integrity. If the check fails, a re-acquisition or alarm is triggered. The received environmental data is checked for range rationality and temporal continuity, obvious outliers are removed, and attempts are made to complete the data by interpolation of adjacent data. Only after all data has passed quality verification will the subsequent feature extraction and comparison steps be performed.

[0015] This invention relates to the field of smart agriculture technology, and more particularly to a method for diagnosing and predicting trends of rice diseases and pests through multimodal data fusion. The method involves acquiring real-time multimodal data of a specific rice variety at its current growth stage and analyzing it in conjunction with baseline data of that variety retrieved from a pre-built knowledge base. Phenotypic features are extracted from the images and compared with standard features of the variety to generate a first abnormal feature vector. Environmental data is compared with the variety's environmental response threshold to generate a second abnormal feature vector. A dynamic weighted allocator fuses these two vectors based on growth stage and stress level to form a comprehensive disease feature vector. This vector is then input into a decision logic network built based on agronomical knowledge for reasoning to obtain preliminary diagnostic results. Finally, historical yield databases are used to assess yield risk, and future environmental trends are used to predict disease and pest development, generating an integrated diagnostic and trend prediction report. This invention achieves variety-specific accurate diagnosis and forward-looking early warning. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the diagnostic and trend prediction method for multimodal data fusion claimed in an embodiment of the present invention. Figure 2 This is a second flowchart of a diagnostic and trend prediction method for multimodal data fusion claimed in an embodiment of the present invention; Figure 3 This is a third flowchart of a diagnostic and trend prediction method for multimodal data fusion claimed in an embodiment of the present invention; Figure 4 The fourth flowchart is a diagnostic and trend prediction method for multimodal data fusion claimed in an embodiment of the present invention. Detailed Implementation

[0017] 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 a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] The terms "first," "second," and "third" in this application are 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," "second," or "third" may explicitly or implicitly include at least one of those features. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications in the embodiments of this application, such as up, down, left, right, front, back, etc., are only used to explain the relative positional relationships and movements between components in a specific orientation as shown in the accompanying drawings. If the specific orientation changes, the directional indications will change accordingly. Furthermore, the terms "including" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0019] References to embodiments herein mean that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] According to the first embodiment of the present invention, referring to Figure 1 A diagnostic and trend prediction method based on multimodal data fusion, comprising: S1. Obtain real-time multimodal data of the target rice plant during its current growth stage. Based on the pre-determined identifier of the target rice variety, retrieve and call up the corresponding variety baseline data from the pre-built rice variety knowledge base. S2, input the visible light image and near-infrared image from the real-time multimodal data to the pre-configured image feature extraction module to extract real-time image features; S3, based on the standard morphological spectral features in the variety benchmark data, compare the extracted real-time image features with the standard morphological spectral features of the corresponding standard growth period, calculate the deviation information of the image features, and generate the first abnormal feature vector based on the deviation information. S4. Compare the environmental data with the preset environmental response thresholds for the corresponding standard growth period in the variety benchmark data. Calculate the weighted stress index for different environmental stress factors based on the comparison results, and generate a second abnormal feature vector based on the weighted stress index. S5, invoke the dynamic weight allocator, determine the weight allocation strategy for fusing the first abnormal feature vector and the second abnormal feature vector based on the current reproductive period and the stress level indicated by the weighted stress index, and fuse the first abnormal feature vector and the second abnormal feature vector according to the weight allocation strategy to form a comprehensive disease feature vector. S6: Input the comprehensive disease feature vector into the decision logic network built based on agronomic knowledge, perform logical reasoning based on the input vector, and output preliminary disease and pest diagnosis results; S7. Access the historical yield database to obtain the yield impact coefficients of different rice varieties under historical pest and disease occurrence conditions. Based on the identification of the target rice variety and the preliminary pest and disease diagnosis results, match the corresponding historical impact coefficients from the historical yield database. S8 combines the historical impact coefficients and preliminary pest and disease diagnosis results to assess the potential impact of currently diagnosed pests and diseases on the final yield of the target rice plants, obtain yield risk forecasts, and generate integrated diagnosis and trend prediction reports by combining the extrapolation of environmental stress trends in the future preset time period.

[0021] In this embodiment, specific data acquisition nodes deployed in the target rice field synchronously acquire real-time multimodal data sets of the target rice plants at their current specific growth stage. The data acquisition nodes integrate image acquisition units and environmental sensing units. The image acquisition units are hardware-synchronized to ensure the capture of visible light and near-infrared digital images from the same field of view at the same time. The visible light digital images contain red, green, and blue color channel information, while the near-infrared digital images characterize the crop's reflection intensity of near-infrared light. The environmental sensing units include temperature and relative humidity sensors, used to synchronously acquire air temperature and relative humidity data for the rice canopy microclimate area. All acquired data is appended with a unified timestamp and acquisition node location identifier.

[0022] Based on the unique identifier of the target rice variety, determined in advance through manual input or automatic identification, access a pre-built rice variety knowledge base stored on a remote server or local database. Within this knowledge base, the unique identifier is used as the query key to retrieve and fully access all variety benchmark data corresponding to the target rice variety. The variety benchmark data is a structured dataset whose core content includes: the definition of each standard growth stage of the variety within its complete growth cycle, and the corresponding standard morphological spectral characteristics and a preset set of environmental response thresholds for each standard growth stage. The standard morphological spectral characteristics quantitatively describe the typical numerical range of leaf color, texture, and near-infrared reflectance that a healthy plant should possess during this growth stage. The preset set of environmental response thresholds includes multiple temperature and humidity thresholds related to the occurrence of specific diseases; these thresholds are pre-set based on agronomic knowledge and historical data of the variety.

[0023] The acquired digital images in the visible light band and near-infrared band are input into an image feature extraction software module with pre-configured parameters. This module first performs spatial registration and alignment on the two images, and then automatically identifies the effective area of ​​the rice canopy in the image. Within this area, the module performs a series of feature calculation operations: extracting leaf color feature values ​​representing the leaf color state from the visible light image, extracting texture feature vectors representing the leaf surface structure, and extracting near-infrared reflectance feature values ​​representing the plant's physiological activity from the near-infrared image, and finally packaging them into a comprehensive real-time image feature vector.

[0024] From the loaded variety benchmark data, based on the current growth stage information, the corresponding standard morphological spectral features are matched and selected as the comparison benchmark. The generated real-time image feature vector is compared and the selected standard feature vector item by item. The comparison is not a simple equality judgment, but rather the relative deviation of each feature is calculated and the deviation is mapped to multiple preset abnormality levels, such as normal, slight deviation, and significant deviation. Finally, the abnormality levels corresponding to each feature are encoded into a discrete, fixed-dimensional vector in a predetermined order. This vector is defined as the first abnormal feature vector, which intuitively reflects the abnormal pattern of the real-time image features relative to the variety health benchmark.

[0025] The acquired real-time temperature and humidity data are compared with corresponding thresholds in a pre-set set of environmental response thresholds for the current growth stage. This comparison process involves determining whether the current environmental data falls within a threshold range that easily induces specific diseases. Based on the comparison results, a quantified stress index is calculated for both temperature and humidity. Then, according to the importance of different environmental factors to disease induction in agronomical knowledge, weights are assigned to these two stress indices, and they are weighted and summed to obtain a comprehensive weighted stress index. Based on the numerical range of this weighted stress index, the overall level of current environmental stress is determined and encoded into a vector reflecting the environmental induction risk of different disease types. This vector is defined as the second abnormal feature vector.

[0026] A software component called a dynamic weight allocator is invoked. The core of this component is a predefined strategy table that specifies the relative importance of the first abnormal feature vector (image feature anomaly information) and the second abnormal feature vector (environmental stress information) in the final diagnosis under different rice growth stages and environmental stress levels. Based on the current specific growth stage and the stress level obtained in step S105, the component queries this strategy table to obtain a set of specific fusion weight coefficients. Subsequently, these coefficients are used to weight and adjust the first and second abnormal feature vectors, and the two weighted vectors are concatenated and combined into a new, higher-dimensional vector. This vector is named the comprehensive disease feature vector, which integrates abnormal information from both the image and the environment.

[0027] The generated comprehensive disease feature vector is input into a decision logic network constructed based on agronomic pathology knowledge rules. This network contains a large number of diagnostic rules organized in a condition-conclusion format. After receiving the input vector, the network matches it with the condition parts of all rules, activating rule chains that meet the conditions. Through logical reasoning and conflict resolution mechanisms, the network outputs a structured preliminary diagnostic conclusion, which at least indicates the most likely type of pest or disease and its corresponding confidence level.

[0028] Access a historical yield database containing a large amount of historical case data. Using the target rice variety identifier and the identified pest and disease type as key query conditions, retrieve the average impact data of similar historical pest and disease cases on the final yield from this database, and extract a quantified historical impact coefficient. Combine this historical impact coefficient with the pest and disease severity information reflected in the current diagnosis, and use a pre-set assessment model to calculate the potential impact of the current pest and disease situation on the rice yield of the target field, and output it as a yield risk prediction in the form of percentage or risk level.

[0029] Based on preliminary diagnostic results and yield risk assessments, and incorporating weather forecasts of environmental conditions over a future period, this study simulates future environmental data to generate predictive environmental stress information. It also incorporates assumptions about the current state of pests and diseases to extrapolate potential future trends. Finally, all information—including current pest and disease status, future trend predictions, yield risk levels, key diagnostic criteria, and management recommendations—is integrated and formatted into a comprehensive, illustrated integrated diagnostic and trend prediction report for user review.

[0030] Furthermore, the image feature extraction module, dynamic weight allocator, decision logic network, historical yield database, and rice variety knowledge base together constitute the diagnosis and prediction system; The method also includes system initialization steps: Before performing the diagnosis on the target rice variety for the first time, the corresponding variety benchmark data is loaded from the rice variety knowledge base to the local cache based on the identification of the target rice variety, and the feature extraction parameters corresponding to the standard morphological spectral features in the image feature extraction module are configured.

[0031] Furthermore, S2 specifically includes the following steps: Spatial registration was performed on the synchronously acquired visible light and near-infrared images to ensure that pixels at the same spatial location correspond to the same plant parts. Based on the registered visible light images, a pre-trained plant segmentation model was used to identify the canopy region of the target rice plant, which was then used as the effective region for subsequent feature analysis. Within the effective area, the visible light image is converted from the original color space to the target color space, and the statistical values ​​of all pixels within the effective area in a specific channel of the target color space are calculated as leaf color features. Within the effective area, a multi-scale texture analysis operator is applied to the grayscale image of the visible light image to calculate the texture descriptor at each scale. The texture descriptors at multiple scales are then concatenated to form a multi-dimensional texture feature vector, which serves as the texture feature. Within the effective area, the pixel intensity values ​​of the corresponding area in the registered near-infrared image are read, and the statistical value of the near-infrared band reflectance within the effective area is calculated as the near-infrared reflectance feature. Leaf color features, texture feature vectors, and near-infrared reflectance features are spliced ​​and normalized in a preset order to form a unified real-time image feature vector for subsequent comparison.

[0032] Furthermore, referring to Figure 2 S3 specifically includes the following steps: Based on the current growth period information, find the standard morphological spectral characteristics corresponding to the standard growth period that is the same as or closest to the current growth period from the loaded variety benchmark data; The real-time image feature vector is subtracted element-by-element from the matched standard morphological spectral features to obtain the original deviation vector. For each element in the original deviation vector, the relative deviation ratio vector is obtained by dividing it by the reference value of the corresponding element in the standard morphological spectral features. For each type of feature deviation in the relative deviation ratio vector, a separate anomaly judgment interval is preset to distinguish between normal fluctuations and suspected anomalies. The relative deviation ratio vector is traversed, and each element is mapped to a discrete anomaly level based on the anomaly judgment interval into which its value falls. Using one-hot encoding or numerical encoding, the anomaly level corresponding to each feature is converted into a fixed-length sub-vector. All sub-vectors corresponding to all features are connected in a predetermined order to form the first anomaly feature vector, which reflects the number of features and the anomaly state of the monitored image.

[0033] In this embodiment, an internal clock is maintained or user-inputted information about the current growth stage, such as mid-tillering or heading stage, is received. This information is represented as a growth stage identifier or growth stage index. The standard growth stage list for the variety is retrieved from the loaded local cache. The current growth stage is compared with each standard growth stage in the list, using a matching strategy that prioritizes exact matching: if a standard growth stage with the exact same identifier exists in the list, it is directly selected; otherwise, if an exact match is not possible (e.g., the current stage is between two standard growth stages), the absolute difference between the current growth stage index and the indices of each standard growth stage in the list is calculated, and the standard growth stage with the smallest difference is selected as the matching growth stage.

[0034] In the storage structure of the varietal benchmark data, each standard growth period is associated with a standard morphological spectral feature. Based on the determined matching growth period, the corresponding standard feature vector is extracted from its data structure. The dimensions of this vector, the order of each dimension, and the physical meaning they represent—for example, dimensions 1-3 are leaf color features, dimensions 4-15 are texture features, and dimensions 16-18 are near-infrared reflectance features—must be completely consistent with the dimensional definitions of the generated real-time image feature vector to ensure comparability. This extracted vector serves as the gold standard or health benchmark for subsequent comparisons.

[0035] Element-wise difference calculation performs arithmetic subtraction on the vector, subtracting the corresponding element in the standard morphological spectral features at the same position from each element of the real-time image feature vector. This operation generates a new vector called the original deviation vector. Each element of the original deviation vector represents the absolute deviation of the real-time feature value from the standard feature value. The value can be positive or negative. Positive values ​​usually indicate that the feature value is higher than the standard, such as greener leaves or higher reflectance, while negative values ​​indicate that it is lower than the standard.

[0036] Because the baseline values ​​for different features vary in magnitude, absolute deviations are difficult to directly measure the severity of anomalies. Therefore, the original deviation vector is relativized. Specifically, each element of the original deviation vector is iterated over, and the absolute deviation of that element is divided by the baseline value of its corresponding element in the standard morphological spectral feature—that is, the typical value in a healthy state. This division yields a ratio representing the percentage of deviation relative to the baseline value, thus generating a relative deviation ratio vector. For example, if the baseline value for a leaf color feature is 0.5 and the measured value is 0.4, then the absolute deviation is -0.1, and the relative deviation ratio is -0.2, or -20%. This step makes feature deviations of different dimensions and orders of magnitude comparable.

[0037] Anomaly detection intervals are pre-loaded. During the initialization or knowledge base design phase, agronomic experts, based on varietal characteristics and historical experience, predefine multiple continuous numerical intervals for the relative deviation ratio of each characteristic, such as the H-mean of leaf color and the contrast of texture. These intervals are associated with different anomaly levels. For example, for the relative deviation ratio of the H-mean of leaf color, it might be set as follows: [-10%, +10%] as the normal interval, (10%, 25%] or [-25%, -10%) as the mild anomaly interval, (25%, 40%] or [-40%, -25%) as the moderate anomaly interval, and greater than 40% or less than -40% as the severe anomaly interval. These interval rules are loaded into the configuration file.

[0038] The anomaly level mapping is performed feature by feature, sequentially processing each element in the relative deviation ratio vector. For the currently processed i-th element, based on the feature type it represents, the anomaly judgment interval rule preset for that type of feature in step S403a is searched; then, it is determined which preset interval the relative deviation ratio value of the current element falls into. Once the interval is determined, the anomaly level label corresponding to that interval, such as normal or mildly abnormal, is assigned to the feature; this process is performed sequentially for all elements in the vector, and finally, each original feature is mapped to a discrete, semantically clear anomaly level.

[0039] The conversion from level to vector uses a fixed encoding scheme, converting the discrete anomaly level obtained from each feature into a numerical sub-vector for subsequent computer processing. For example, one-hot encoding can be used: assuming there are four anomaly levels—normal, mild, moderate, and severe—then normal is encoded as [1,0,0,0], mild as [0,1,0,0], and so on; or simple numerical encoding can be used, such as using numbers 1-4 to represent the four levels respectively. Regardless of the encoding method used, the length of the sub-vector corresponding to each feature is fixed: the length of the one-hot encoding is the number of levels, and the length of the numerical encoding is 1.

[0040] When constructing the first anomalous feature vector, the encoded sub-vectors of all features obtained in step S404a are concatenated in the same order as the original features. If one-hot encoding is used, the final first anomalous feature vector will be a high-dimensional sparse vector with a dimension equal to the number of features multiplied by the number of anomalous levels. If numerical encoding is used, a vector with a dimension equal to the number of features is formed, where each element is an integer representing a level. This final vector not only contains information about which features are anomalous, but more importantly, it records the severity of each feature's anomalousness in a structured way, providing clear and quantifiable input for subsequent logical decisions. This vector is a complete and standardized representation of image modal anomaly information.

[0041] Furthermore, S4 specifically includes: The temperature data is compared with the suitable temperature range in the environmental response threshold. If the real-time temperature exceeds the suitable temperature range, the temperature stress index is calculated based on the extent of the exceedance and the preset temperature stress coefficient. The humidity data is compared with the suitable humidity range in the environmental response threshold. If the real-time humidity exceeds the suitable humidity range, the humidity stress index is calculated based on the extent of the exceedance and the preset humidity stress coefficient. The weighted stress index is obtained by weighting the temperature stress index and the humidity stress index according to their relative importance in inducing specific diseases.

[0042] Furthermore, referring to Figure 3 S4 also includes: Multiple consecutive weighted stress index intervals are set, each interval corresponding to a preset environmental stress level; The calculated weighted stress index is mapped to the corresponding environmental stress level; The mapped environmental stress levels are encoded into a multi-dimensional vector, where each dimension represents a preset disease type or stress category. The value of this dimension represents the environmental stress level or the environmental induced risk level of the disease type derived from the environmental stress level. The resulting vector serves as the second abnormal feature vector.

[0043] In this embodiment, the stress level intervals are predefined during the design phase. Agronomic experts, based on historical data and pathological models, divide the numerical range of the Weighted Stress Index (WSI) into several continuous, non-overlapping intervals, assigning a semantic environmental stress level to each interval. For example, it can be divided as follows: [0, 1) corresponds to no stress, [1, 3) corresponds to mild stress, [3, 6) corresponds to moderate stress, and [6, +∞) corresponds to severe stress. The boundary values ​​and number of levels for these intervals can be adjusted according to different varieties or different growth stages and stored as configuration parameters.

[0044] In each diagnostic cycle, real-time mapping compares the calculated real-time WSI value with predefined intervals. Using either sequential comparison or interval lookup algorithms, it determines which predefined interval the WSI value falls into, and then assigns the corresponding environmental stress level label, such as moderate stress, to the current environmental state.

[0045] The risk dimension predefines a list of potentially environmentally induced diseases or stresses that require attention. For example, the list might include: rice blast risk, sheath blight risk, bacterial leaf blight risk, high-temperature heat damage risk, and low-temperature chilling damage risk. Each item in the list corresponds to a dimension of the final vector.

[0046] The conversion rules from environmental severity level to risk value store one rule for each disease or stress type in the list, describing how to derive the specific environmental induced risk level for that disease from a general environmental stress level. This takes into account the differences in sensitivity of different diseases to environmental factors. For example: Rule 1: If the primary concern is rice blast and the environmental stress level is moderate, then the value of the rice blast risk dimension is set to high or represented by the number 3.

[0047] Rule 2: For the same moderate stress, the risk of sheath blight may be set to a medium number 2, because the current temperature conditions may not be optimal.

[0048] Rule 3: The risk dimension of high temperature heat damage may be directly determined and mapped to a level by the temperature stress index I_temp, instead of directly using the comprehensive WSI.

[0049] Vector assignment and generation involves iterating through all predefined disease / stress dimensions. For each dimension, a specific transformation rule is applied. The input is the obtained environmental stress level and the original temperature / humidity stress index, which may be used. Through rule calculation or table lookup, the risk level value to be assigned to that dimension can be a discrete label or a numerical level. The risk level values ​​obtained from all dimensions are arranged into a vector according to the fixed order defined by the dimensions. For example, a vector containing 5 risks may be represented as: [High, Medium, Low, None, Medium] or its numerical encoding form [3, 2, 1, 0, 2]. This vector is the second anomaly feature vector.

[0050] Furthermore, the process by which the dynamic weight allocator determines the weight allocation strategy includes: The dynamic weight allocator has a built-in weight configuration table that defines the fusion weight coefficients of the first and second abnormal feature vectors under different reproductive periods and different stress levels. The dynamic weight allocator queries the weight configuration table based on the input current reproductive period information and the degree of stress reflected by the second abnormal feature vector or directly indicated by the weighted stress index to obtain a set of corresponding fusion weight coefficients. The first and second abnormal feature vectors are fused according to a weighting strategy, specifically as follows: Each element of the first abnormal feature vector is multiplied by the corresponding first weight coefficient, and each element of the second abnormal feature vector is multiplied by the corresponding second weight coefficient. The two weighted vectors are then concatenated to form a dimensionally expanded comprehensive disease feature vector.

[0051] Furthermore, the decision logic network is a production rule-based system that performs logical reasoning and outputs preliminary pest and disease diagnosis results, as well as subsequent evaluation and report generation, specifically including the following steps: The decision logic network contains agronomic diagnostic rules, and the condition part of each rule is associated with a specific pattern or value range of the comprehensive disease feature vector; The comprehensive disease feature vector is matched with the condition parts of all rules to filter out all rules whose conditions are met. Each triggered rule outputs one or more candidate pest / disease types and their confidence factors. All candidate pest and disease types output by triggered rules are integrated, and the confidence factors of the same pest and disease type are aggregated and calculated. Based on the sorting of the aggregated confidence factors, the pest and disease type with the highest confidence is selected as the preliminary pest and disease diagnosis result. Using the identifier of the target rice variety and the types of pests and diseases identified in the preliminary pest and disease diagnosis results as the query keys, a search is performed in the historical yield database; The historical yield database is stored in the form of a data structure, in which each record is associated with a variety identifier, a type of pest or disease, and a yield impact coefficient derived from statistical analysis of historical data; If a perfect match is found, extract its stored production impact coefficient; If no exact match is found, then according to the preset approximate matching rules, find records of similar diseases and pests of the same variety or other varieties of the same diseases and pests, and estimate the historical influence coefficient by interpolation or averaging. Using the historical impact coefficient obtained through matching or estimation as the basic impact factor, we analyze the characteristic values ​​representing the severity of pests and diseases in the comprehensive disease feature vector, and calculate the ratio of the current severity to the historical typical severity as the severity adjustment factor. The potential impact level is calculated by multiplying the basic impact factor and the severity adjustment factor, and combining them with the sensitivity parameter of the current fertility period to yield, and then quantified as a yield risk forecast. Acquire predicted environmental data for a predetermined future time period, and use the same process as for real-time environmental data to calculate the weighted stress index for future time points. Generate simulated values ​​of the second anomaly feature vector for the future. Combine this with the current first anomaly feature vector to extrapolate simulated values ​​of the first anomaly feature vector for future time points. By using a dynamic weight allocator to fuse future simulated feature vectors, and inputting them into a decision logic network for forward inference, the possible changes in the state of pests and diseases are predicted, and a forecast of the development trend of pests and diseases is generated. The preliminary disease and pest diagnosis results, yield risk estimates, and disease and pest development trend predictions are structured and organized, and the report generation timestamp, target rice variety identifier, current growth stage, and variety benchmark data version information are attached, and packaged into an integrated diagnosis and trend prediction report.

[0052] Furthermore, referring to Figure 4 The method also includes report output and feedback steps: Convert the generated integrated diagnostic and trend prediction reports into files of a specified format or display them in a visual interface; The complete contextual data of this diagnosis, including real-time multimodal data, first abnormal feature vector, second abnormal feature vector, comprehensive disease feature vector, preliminary disease and pest diagnosis results, and the final generated report, are anonymized and formatted before being stored in the local diagnosis log database. Regularly perform statistical analysis on cases in the local diagnostic log database. The results can serve as a basis for optimizing the variety baseline data in the rice variety knowledge base, adjusting the rules in the decision logic network, or updating the influence coefficients in the historical yield database.

[0053] In this embodiment, the internal structured report data generated in step S806 is rendered into the final user deliverable based on a preset output configuration or user request. For web or mobile application interfaces, a template engine is used to generate dynamic HTML pages from the data, which are then displayed in the user dashboard. For scenarios requiring archiving or distribution, a document generation library, such as one used for PDF generation, is invoked to convert the report into a PDF document. Simultaneously, a machine-readable format, such as JSON, is always retained for API calls or data exchange.

[0054] After format conversion, the report is delivered to users through multiple channels, primarily including: displaying the report in a list and providing detailed information in the report center module of the diagnostic graphical user interface; sending PDF report attachments via email to pre-subscribed farmers or agronomists; and pushing the report's core conclusions and risk warnings as concise text to users' mobile applications through a messaging service interface. This ensures users can access diagnostic information promptly and conveniently.

[0055] After a complete diagnostic cycle, all key inputs, intermediate results, and final outputs involved in the diagnosis are automatically collected, forming a diagnostic case data package. The specific contents include: raw visible light and near-infrared images (thumbnails or paths can be stored), raw environmental data; processed real-time image feature vectors, first anomaly feature vectors, second anomaly feature vectors, and comprehensive disease feature vectors; key rule trigger records in the decision logic network reasoning process; preliminary diagnostic results and confidence levels; parameters for yield risk prediction calculation process; and the final integrated report (full text or index).

[0056] Before storage, the data packets are anonymized to remove or obfuscate specific field owner information that may involve personal privacy or trade secrets, while retaining crop and environmental data for analysis. At the same time, all elements in the data packets are formatted according to a predefined log data model to ensure that their fields and types conform to the table structure design of the local diagnostic log database.

[0057] The formatted diagnostic case data package is written as a complete record to a dedicated diagnostic history table in a local relational database or time-series database. Each record contains index fields such as a unique diagnostic ID, timestamp, and variety information, as well as a BLOB binary large object field or associated sub-table storing detailed context data. This establishes a traceable and auditable diagnostic case library.

[0058] Regularly analyze task scheduling and set up a background scheduled task, such as once a month or quarter, to perform batch statistical analysis on cases in the local diagnostic log database.

[0059] Variety benchmark validation analysis involves statistically analyzing the actual distribution range of various image features and environmental data for a specific variety across different growth stages and in cases diagnosed as healthy. This actual distribution is then compared to preset standard morphological spectral features and environmental response thresholds in a knowledge base. If persistent deviations are found, such as the actual value of a certain leaf color feature in healthy plants generally being 10% higher than the standard value, adjustment suggestions are generated to optimize the variety's benchmark data in the knowledge base, making it more consistent with reality.

[0060] The rule effectiveness evaluation analyzes the historical trigger frequency and diagnostic accuracy of each decision rule. Accuracy can be indirectly assessed by the recording of subsequent manually verified field survey results or by the consistency of multiple rounds of diagnosis. Rules that are frequently triggered but whose final conclusions are often overridden by higher confidence conclusions, or rules that are almost never triggered, are marked to prompt experts to review the rationality or necessity of their condition settings, thereby optimizing the decision logic network.

[0061] Impact coefficient calibration involves comparing the predicted historical yield impact coefficients with the actual yield reductions for field cases where final actual yield records are available. Prediction errors are calculated, and error patterns under different variety-disease combinations are statistically analyzed. Based on this analysis, the impact coefficients stored in the historical yield database are calibrated or confidence intervals are provided to make them more realistic.

[0062] An optimization suggestion report is generated, summarizing the results of the above statistical analysis into a knowledge base optimization suggestion report for administrators or agronomic experts. The report clearly identifies specific data items, rules, or coefficients that need review and may be updated, and includes supporting statistical analysis data. Experts review and make decisions based on this report, completing iterative updates to the rice variety knowledge base, decision logic network, and historical yield database, thereby achieving continuous self-improvement of diagnostic capabilities.

[0063] Furthermore, before acquiring real-time multimodal data of the target rice plants at the current growth stage, the method also includes a data quality verification step: Check the online status and clock synchronization status of the image acquisition equipment and environmental sensors; The received visible light and near-infrared images are checked for sharpness and integrity. If the check fails, a re-acquisition or alarm is triggered. The received environmental data is checked for range rationality and temporal continuity, obvious outliers are removed and attempts are made to complete the data by interpolation of adjacent data. Only after all data has passed quality verification will the subsequent feature extraction and comparison steps be performed.

[0064] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0065] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0066] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A diagnostic and trend prediction method based on multimodal data fusion, characterized in that, The method includes: S1. Obtain real-time multimodal data of the target rice plant during its current growth stage. Based on the pre-determined identifier of the target rice variety, retrieve and call up the corresponding variety baseline data from the pre-built rice variety knowledge base. S2, input the visible light image and near-infrared image from the real-time multimodal data into the pre-configured image feature extraction module to extract real-time image features; S3, based on the standard morphological spectral features in the variety benchmark data, the extracted real-time image features are compared with the standard morphological spectral features of the corresponding standard growth period, the deviation information of the image features is calculated, and a first abnormal feature vector is generated based on the deviation information. S4, compare the environmental data with the preset environmental response threshold corresponding to the standard growth period in the variety benchmark data, calculate the weighted stress index for different environmental stress factors based on the comparison results, and generate a second abnormal feature vector based on the weighted stress index. S5, invoke the dynamic weight allocator, determine the weight allocation strategy for fusing the first abnormal feature vector and the second abnormal feature vector according to the current reproductive period and the stress level indicated by the weighted stress index, and fuse the first abnormal feature vector and the second abnormal feature vector according to the weight allocation strategy to form a comprehensive disease feature vector; S6, input the comprehensive disease feature vector into the decision logic network constructed based on agronomic knowledge, perform logical reasoning based on the input vector, and output preliminary disease and pest diagnosis results; S7. Access the historical yield database to obtain the yield impact coefficients of different rice varieties under historical pest and disease occurrence conditions. Based on the identification of the target rice variety and the preliminary pest and disease diagnosis results, match the corresponding historical impact coefficients from the historical yield database. S8. Combining the historical impact coefficients and the preliminary disease and pest diagnosis results, assess the potential impact of the currently diagnosed diseases and pests on the final yield of the target rice plants, obtain a yield risk forecast, and generate an integrated diagnosis and trend prediction report by combining the extrapolation of environmental stress trends in the future preset time period.

2. The method according to claim 1, characterized in that, The image feature extraction module, the dynamic weight allocator, the decision logic network, the historical yield database, and the rice variety knowledge base together constitute the diagnosis and prediction system. The method also includes a system initialization step: Before performing the diagnosis on the target rice variety for the first time, the corresponding benchmark data of the variety is loaded from the rice variety knowledge base to the local cache according to the identifier of the target rice variety, and the feature extraction parameters corresponding to the standard morphological spectral features in the image feature extraction module are configured.

3. The method according to claim 2, characterized in that, S2 specifically includes the following steps: Spatial registration is performed on the synchronously acquired visible light image and near-infrared image to ensure that pixels at the same spatial location correspond to the same plant part. Based on the registered visible light image, a pre-trained plant segmentation model is used to identify the canopy region of the target rice plant, which is then used as the effective region for subsequent feature analysis. Within the effective area, the visible light image is converted from the original color space to the target color space, and the statistical values ​​of all pixels within the effective area in a specific channel of the target color space are calculated as leaf color features. Within the effective area, a multi-scale texture analysis operator is applied to the grayscale image of the visible light image to calculate the texture descriptor at each scale. The texture descriptors at multiple scales are then concatenated to form a multi-dimensional texture feature vector, which serves as the texture feature. Within the effective area, the pixel intensity value of the corresponding area in the registered near-infrared image is read, and the statistical value of the near-infrared band reflectance within the effective area is calculated as the near-infrared reflectance feature. The leaf color feature, the texture feature vector, and the near-infrared reflectance feature are spliced ​​and normalized in a preset order to form a unified real-time image feature vector for subsequent comparison.

4. The method according to claim 3, characterized in that, S3 specifically includes the following steps: Based on the current growth period information, find the standard morphological spectral characteristics corresponding to the standard growth period that is the same as or closest to the current growth period from the loaded variety benchmark data; The real-time image feature vector is subtracted element-wise from the matched standard morphological spectral features to obtain the original deviation vector. Each element in the original deviation vector is divided by the reference value of the corresponding element in the standard morphological spectral features to obtain the relative deviation ratio vector. For each type of feature deviation in the relative deviation ratio vector, a separate anomaly judgment interval is preset to distinguish between normal fluctuations and suspected anomalies. The relative deviation ratio vector is traversed, and each element is mapped to a discrete anomaly level according to the anomaly judgment interval into which its value falls. Using one-hot encoding or numerical encoding, the anomaly level corresponding to each feature is converted into a sub-vector of fixed length. All sub-vectors corresponding to all features are connected in a predetermined order to form the first anomaly feature vector, which reflects the number of features and the anomaly state of the monitored image.

5. The method according to claim 1, characterized in that, S4 specifically includes: The temperature data is compared with the suitable temperature range in the environmental response threshold. If the real-time temperature exceeds the suitable temperature range, the temperature stress index is calculated based on the extent of the exceedance and the preset temperature stress coefficient. The humidity data is compared with the suitable humidity range in the environmental response threshold. If the real-time humidity exceeds the suitable humidity range, the humidity stress index is calculated based on the extent of the exceedance and the preset humidity stress coefficient. The weighted stress index is obtained by weighting and summing the temperature stress index and the humidity stress index according to their relative importance in inducing specific diseases.

6. The method according to claim 5, characterized in that, S4 also includes: Multiple consecutive weighted stress index intervals are set, each interval corresponding to a preset environmental stress level; The calculated weighted stress index is mapped to the corresponding environmental stress level; The mapped environmental stress level is encoded into a multi-dimensional vector, where each dimension represents a preset disease type or stress category. The value of this dimension represents the environmental stress level or the environmental induced risk level of the disease type derived from the environmental stress level. The resulting vector serves as the second abnormal feature vector.

7. The method according to claim 1, characterized in that, The process by which a dynamic weight allocator determines its weight allocation strategy includes: The dynamic weight allocator has a built-in weight configuration table that defines the fusion weight coefficients of the first abnormal feature vector and the second abnormal feature vector under different reproductive periods and different stress levels. The dynamic weight allocator queries the weight configuration table based on the input current reproductive period information and the degree of stress reflected by the second abnormal feature vector or directly indicated by the weighted stress index to obtain a corresponding set of fusion weight coefficients. The process of fusing the first abnormal feature vector and the second abnormal feature vector according to the weight allocation strategy is as follows: Each element of the first abnormal feature vector is multiplied by the corresponding first weight coefficient, and each element of the second abnormal feature vector is multiplied by the corresponding second weight coefficient. The two weighted vectors are then concatenated to form the dimensionally expanded comprehensive disease feature vector.

8. The method according to claim 1, characterized in that, The decision logic network is a production rule-based system that performs logical reasoning and outputs preliminary pest and disease diagnosis results, as well as subsequent evaluation and report generation. Specifically, it includes the following steps: The decision logic network includes agronomic diagnostic rules, and the condition part of each rule is associated with a specific pattern or value range of the comprehensive disease feature vector; The comprehensive disease feature vector is matched with the condition parts of all rules to filter out all rules whose conditions are met. Each triggered rule outputs one or more candidate pest / disease types and their confidence factors. All candidate pest and disease types output by the triggered rules are integrated, the confidence factors of the same pest and disease type are aggregated and calculated, and the pest and disease type with the highest confidence is selected as the preliminary pest and disease diagnosis result according to the sorting of the aggregated confidence factors. Using the identifier of the target rice variety and the pest and disease types identified in the preliminary pest and disease diagnosis results as the query keys, a search is performed in the historical yield database; The historical yield database is stored in the form of a data structure, in which each record is associated with a variety identifier, a type of pest or disease, and a yield impact coefficient obtained through statistical analysis of historical data; If a perfect match is found, extract its stored production impact coefficient; If no exact match is found, then according to the preset approximate matching rules, find records of similar diseases and pests of the same variety or other varieties of the same diseases and pests, and estimate the historical influence coefficient by interpolation or averaging. Using the historical impact coefficient obtained through matching or estimation as the basic impact factor, the characteristic values ​​representing the severity of pests and diseases in the comprehensive disease feature vector are analyzed, and the ratio of the current severity to the historical typical severity is calculated as the severity adjustment factor. The potential impact level is calculated by multiplying the basic impact factor by the severity adjustment factor and combining it with the sensitivity parameter of the current fertility period to yield, and then quantified as the yield risk forecast. Acquire predicted environmental data for a predetermined future time period, and use the same process as for real-time environmental data to calculate the weighted stress index for future time points. Generate simulated values ​​of the second anomaly feature vector for the future. Combine this with the current first anomaly feature vector to extrapolate simulated values ​​of the first anomaly feature vector for future time points. The dynamic weight allocator is used to fuse the simulated feature vectors of the future, and the result is input into the decision logic network for forward reasoning to predict possible changes in the state of pests and diseases and generate a prediction of the development trend of pests and diseases. The preliminary disease and pest diagnosis results, the yield risk forecast, and the disease and pest development trend prediction are structured and organized, and the timestamp of the generated report, the target rice variety identifier, the current growth stage, and the variety benchmark data version information used are attached, and packaged into the integrated diagnosis and trend prediction report.

9. The method according to claim 8, characterized in that, The method also includes report output and feedback steps: The generated integrated diagnostic and trend prediction report is converted into a file in a specified format or displayed in a visual interface; The complete context data of this diagnosis, including the real-time multimodal data, the first abnormal feature vector, the second abnormal feature vector, the comprehensive disease feature vector, the preliminary disease and pest diagnosis results, and the final generated report, are desensitized and formatted before being stored in the local diagnosis log database. Regularly perform statistical analysis on the cases in the local diagnostic log database. The results can serve as a basis for optimizing the variety benchmark data in the rice variety knowledge base, adjusting the rules in the decision logic network, or updating the influence coefficients in the historical yield database.

10. The method according to claim 1, characterized in that, The method further includes a data quality verification step before the step of acquiring real-time multimodal data of the target rice plant at its current growth stage: Check the online status and clock synchronization status of the image acquisition device and the environmental sensor; The received visible light image and near-infrared image are checked for sharpness and integrity. If the check fails, a re-acquisition or alarm is triggered. The received environmental data is checked for range rationality and temporal continuity, obvious outliers are removed, and attempts are made to complete the data by interpolation of adjacent data. Only after all data has passed quality verification will the subsequent feature extraction and comparison steps be performed.

Images