Rice blast latent period diagnosis and prediction method and model based on double-flow data fusion

Abstract

This invention discloses a method and model for diagnosing and predicting the latent stage of rice blast based on dual-stream data fusion. The method includes: predicting time-series data based on the input of the dual-stream fused time-series feature vector into a gated loop module to obtain the hidden state output by the gated loop module. The gated loop module includes multiple preset DDC-PGRUs. The hidden state output by the gated loop module is input into two parallel fully connected layer branches, which respectively output the diagnostic results of the disease stage of rice at the current moment and the prediction results of the disease severity. The DDC-PGRU unit includes a trend gate and a latent gate. The calculation process of the trend gate introduces the rate of change of the input time-series feature vector at adjacent time points, and the calculation process of the latent gate introduces the energy accumulation value of the input time-series feature vector within a long time window. This invention realizes the synchronous mechanistic modeling of the "evolutionary direction" and "latent intensity" of the disease.

Description

Technical Field

[0001] This invention relates to the field of rice blast disease diagnosis and prediction technology, specifically to a method and model for diagnosing and predicting the incubation period of rice blast based on dual-stream data fusion. Background Technology

[0002] In the field of rice blast monitoring, especially for the incubation stage before symptoms appear, existing diagnostic techniques are evolving from single spectral index analysis to high-dimensional deep feature extraction. Because rice blast only manifests as microscopic physiological and biochemical changes in pigments, water content, and cell structure within the crop during the incubation period, it exhibits characteristics of "weak signal, strong time-varying nonlinearity, and high spatial heterogeneity" in canopy hyperspectral images.

[0003] To capture these subtle features, current methods primarily rely on two types of data processing techniques based on general models:

[0004] 1. Spectral feature classification techniques based on shallow machine learning: Early methods often extracted specific vegetation indices (such as NDVI, PRI) or screened sensitive bands, combining them with classifiers such as Support Vector Machine (SVM) and Random Forest (RF) for point-like identification. These methods mainly rely on manually designed features and ignore the temporal dimension of disease development, making it difficult to capture the dynamic evolution patterns during the incubation period.

[0005] 2. Spatiotemporal Feature Extraction Techniques Based on General-Purpose Deep Neural Networks: With the development of deep learning, researchers have begun to utilize "CNN+RNN" architectures to process hyperspectral time-series data. These methods use convolutional neural networks (CNNs) to extract spectral spatial features and recurrent neural networks (RNNs) to mine temporal dependencies, thereby improving the ability to capture temporal anomalous signals to some extent.

[0006] Although current methods based on general deep learning have introduced temporal modeling, directly transferring standard recurrent units used in natural language processing or general video analysis to the extremely complex crop pathology monitoring scenario presents fundamental mechanistic incompatibility and functional limitations. Specifically, existing technologies suffer from the following core shortcomings:

[0007] (1) The rigid framework of "single-dimensional time-series capture" makes it impossible to achieve synchronous decoupling of trend and potential: Existing time-series augmentation units (such as standard GRUs) are designed according to general statistical laws, and their gating mechanisms are mainly based on numerical fitting to control the retention and forgetting of information. This general logic leads to the model being able to capture only the dynamic change direction of the disease (i.e., the disease trend) or roughly extract the numerical intensity of features when processing high-dimensional disease data. It is completely unable to effectively separate and synchronously extract the information of the two orthogonal dimensions of "the speed of disease development" and "the accumulation of pathological energy" within a single computational unit.

[0008] (2) The lack of an effective measure of the "cumulative effect of disease" limits the model's ability to predict disease severity: The outbreak of rice blast depends on the energy accumulation of subtle pathological changes in the early stages. However, existing models lack explicit quantification of the intensity of this cumulative effect across time steps, which limits them to judging "whether the disease is present" based on the current state. They cannot establish a mapping relationship from "current cumulative amount" to "future outbreak level", thus failing to meet the core need of precision agriculture to quantitatively predict the severity of future disease. Summary of the Invention

[0009] To address the problems existing in the prior art, this invention provides a method and model for diagnosing and predicting the latent period of rice blast based on dual-stream data fusion. It utilizes a two-dimensional coupled enhanced gated cyclic unit (DDC-PGRU) to simultaneously quantify and capture the trend dynamics and potential accumulation during the crop disease process, thereby achieving an accurate early diagnosis of the latent period of rice blast and a method for predicting the future severity of disease. The technical solution is as follows:

[0010] Firstly, a method for diagnosing and predicting the incubation period of rice blast based on dual-stream data fusion is provided, comprising the following steps:

[0011] Based on the time series data of hyperspectral images of rice, spectral feature time series data and texture feature time series data were extracted respectively.

[0012] The time-series data of spectral features and the time-series data of texture features are represented by vectors respectively, and then concatenated and fused to obtain the dual-stream fused time-series feature vector;

[0013] Based on the dual-stream fusion time-series feature vector input gated loop module, time-series data prediction is performed to obtain the hidden state output by the gated loop module. The gated loop module includes multiple preset disease course time-series two-dimensional coupling enhanced gated loop units (DDC-PGRU).

[0014] The hidden state output by the gated loop module is input into two parallel fully connected layer branches, which output the diagnosis result of the disease stage of rice at the current moment and the prediction result of the disease severity.

[0015] The preset DDC-PGRU is an improvement based on the gated recurrent unit (GRU). The DDC-PGRU unit includes a trend gate and a potential gate. The calculation process of the trend gate introduces the rate of change of the input time-series feature vector at adjacent time points to capture high-frequency pathological mutations. The calculation process of the potential gate introduces the energy accumulation value of the input time-series feature vector within a long time window to perceive the historical accumulation state of latent pathological changes.

[0016] In some implementations, based on the time-series data of hyperspectral images of rice, spectral feature time-series data and texture feature time-series data are extracted, including:

[0017] Based on the time series data of hyperspectral images of rice canopy, black and white card correction and noise reduction were performed on the original images, and the average spectral reflectance data of the target region ROI was extracted.

[0018] Extract the spectral reflectance curves of the entire band or characteristic band, and calculate the first-order differential spectrum to form spectral characteristic time series data;

[0019] Based on the gray-level co-occurrence matrix algorithm, texture features of the target region ROI are extracted to form temporal data of texture features.

[0020] In some implementations, the method for obtaining the temporal feature vector of dual-stream fusion includes:

[0021] A one-dimensional convolutional neural network is used to encode the spectral feature time series data to obtain the spectral feature time series vector; a fully connected layer is used to perform nonlinear mapping on the texture feature time series data to obtain the texture feature time series vector.

[0022] Channel splicing is performed based on spectral feature time-series vectors and texture feature time-series vectors to obtain dual-stream fusion time-series feature vectors.

[0023] In some implementations, after channel concatenation based on spectral feature time-series vectors and texture feature time-series vectors, the method further includes: adjusting the concatenated features to a preset dimension through a linear mapping layer to generate the final dual-stream fusion time-series feature vector.

[0024] In some implementations, the preset DDC-PGRU includes:

[0025] Trend Gate : ;

[0026] in, ; , These are the two-stream fusion temporal feature vectors at time t-1 and time t, respectively. It is the hidden state at time t-1. It is a gradient-sensitive weight matrix. Use the Sigmoid activation function; , These are the input weight matrix and the cyclic weight matrix of the trend gate, respectively. It is the bias value of the trend gate;

[0027] Potential Gate : ,in, ; , These are the historical energy state vectors at time t-1 and time t, respectively. , These are the input weight matrix and the cyclic weight matrix of the potential gate, respectively. It is the bias value of the potential gate. For the cumulative sensitivity weight matrix, It is a latent amnesia factor that simulates the long-term retention characteristics of disease virulence in plants;

[0028] Candidate hidden state : ,in, , These are the input weight matrix and the cyclic weight matrix used to calculate the candidate hidden state, respectively. It calculates the bias value of the candidate hidden state;

[0029] Final hidden state : ,in, This is the calculation result of the standard update gate. ,in, , These are the learnable weight matrices of the standard update gate. It is the bias value of the standard update gate.

[0030] In some implementations, the hidden state output by the gated loop module is input to two parallel fully connected layer branches, including:

[0031] The first fully connected branch, which is based on the hidden state input of the gated loop module, outputs the diagnosis result of the disease stage of rice at the current moment. The first fully connected branch serves as a diagnostic classifier.

[0032] The hidden state input of the gated loop module is used as the output of the second fully connected branch to predict the severity of rice disease. The second fully connected branch serves as the level predictor.

[0033] In some implementations, the network layers, gated recurrent modules, and two parallel fully connected layer branches used for calculating the temporal feature vectors of dual-stream fusion are all obtained after training using pre-prepared training samples. The method for obtaining these training samples includes:

[0034] Starting from the tillering stage of rice, hyperspectral images of the rice canopy were continuously collected at preset time intervals, including the healthy stage before inoculation, the latent stage, and the symptom-manifesting stage, to obtain time series data of hyperspectral images of the rice canopy.

[0035] The original image was subjected to black and white card correction and noise reduction, and the average spectral reflectance data of the target region (ROI) was extracted.

[0036] Secondly, a diagnostic prediction model for the latent period of rice blast is provided based on dual-stream data fusion. This diagnostic prediction model is a spatiotemporal dual-stream fusion diagnostic network model, which sequentially includes:

[0037] The system consists of a parallel one-dimensional convolutional neural network and a fully connected layer. The one-dimensional convolutional neural network is used to represent the temporal data of spectral features into vectors, generating temporal vectors of spectral features. The fully connected layer is used to perform nonlinear mapping on the temporal data of texture features, generating temporal vectors of texture features.

[0038] The feature concatenation and dimension adaptation layer is used to concatenate the temporal vectors of spectral features and the temporal vectors of texture features through channels. Then, a linear mapping layer is used to adjust the concatenated features to a preset dimension to generate the final dual-stream fusion temporal feature vector.

[0039] The gated loop module includes multiple cascaded preset two-dimensional coupled enhanced gated loop units (DDC-PGRU) for disease progression time series processing. The multiple cascaded preset DDC-PGRU units are connected in series according to time steps to process time series data.

[0040] Two parallel fully connected layer branches are used. One fully connected layer branch is used to output the diagnosis result of the disease stage of rice at the current moment, and the other fully connected layer branch is used to output the prediction result of the disease severity of rice.

[0041] Thirdly, an electronic device is provided, the electronic device comprising:

[0042] processor;

[0043] Memory used to store processor-executable instructions;

[0044] The processor executes the executable instructions to implement the rice blast incubation period diagnosis and prediction method based on dual-stream data fusion as described in the first aspect.

[0045] Fourthly, a computer-readable storage medium is provided having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method for diagnosing and predicting the latent period of rice blast based on dual-stream data fusion as described in the first aspect.

[0046] The present invention provides a diagnostic prediction method and model for rice blast latent period based on dual-stream data fusion, which has the following beneficial effects: By constructing a diagnostic prediction network based on dual-stream data fusion and embedding a "disease course temporal two-dimensional coupled enhanced gated recurrent unit (DDC-PGRU)," the orthogonal decoupling of the symptom evolution direction and latent intensity is achieved at the mathematical mechanism level, thereby achieving the dual goals of accurate identification of latent period and reliable prediction of future disease severity; by using the "latency gate" to quantify the energy accumulation of latent period, the model is endowed with the ability to predict "future disease severity," realizing the leap from qualitative identification to quantitative early warning; the dual-stream architecture combined with the trend perception capability of DDC-PGRU can effectively filter the interference of normal crop growth and environmental noise, significantly improving the diagnostic stability under complex environments. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating the method for diagnosing and predicting the latent period of rice blast based on dual-stream data fusion in the embodiments of this application.

[0048] Figure 2 This is a schematic diagram of the structure of the rice blast incubation period diagnosis and prediction model based on dual-stream data fusion in the embodiments of this application;

[0049] Figure 3 This is a schematic diagram of the structure of the disease progression time-coupling enhanced gated recurrent unit (DDC-PGRU) in the embodiments of this application;

[0050] Figure 4 This is a comparison chart of the original hyperspectral reflectance curves of rice canopy at different disease stages;

[0051] Figure 5 This is a comparison chart of first-order differential spectral curves;

[0052] Figure 6 This is a comparison chart of the response mechanisms of trend gates and potential gates throughout the incubation period. Detailed Implementation

[0053] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0054] See Figure 1 This application provides a method for diagnosing and predicting the incubation period of rice blast based on dual-stream data fusion, comprising the following steps:

[0055] Step 1: Based on the time series data of hyperspectral images of rice, extract the time series data of spectral features and the time series data of texture features respectively;

[0056] Step 2: Based on the temporal data of spectral features and the temporal data of texture features, vector representations are generated and then concatenated and fused to obtain the dual-stream fused temporal feature vector;

[0057] Step 3: Based on the dual-stream fusion time-series feature vector input, the gated loop module is used to predict time-series data and obtain the hidden state output by the gated loop module. The gated loop module includes multiple preset disease course time-series two-dimensional coupling enhanced gated loop units (DDC-PGRU).

[0058] Step 4: Based on the hidden state output by the gated loop module, input it into two parallel fully connected layer branches, and output the diagnosis result of the disease stage of rice at the current moment and the prediction result of the disease severity.

[0059] In step 3, the preset DDC-PGRU is obtained based on the gated recurrent unit GRU. The DDC-PGRU unit includes a trend gate and a potential gate. The calculation process of the trend gate introduces the rate of change of the input time-series feature vector at adjacent time points to capture high-frequency pathological mutations. The calculation process of the potential gate introduces the energy accumulation value of the input time-series feature vector within a long time window to perceive the historical accumulation state of latent pathology.

[0060] In this embodiment, the model layers involved in steps 2 to 4 above (the network layer used for calculating the temporal feature vector of dual-stream fusion, the gated recurrent module, and the two parallel fully connected layer branches) are all obtained by joint training using pre-prepared training samples. The model structure used in steps 2 to 4 above is denoted as the spatiotemporal dual-stream fusion diagnostic network model. The spatiotemporal dual-stream fusion diagnostic network model (see model structure) Figure 2 The process from input to output includes:

[0061] The system employs parallel one-dimensional convolutional neural networks (1D-CNN) and fully connected layers. The 1D-CNN is used to represent spectral feature time-series data into vectors, extracting local morphological features (slope changes of absorption peaks) between bands using convolutional kernels and compressing them from the high-dimensional spectral space to the feature space to generate spectral feature time-series vectors. The fully connected layers are used to perform nonlinear mapping on texture feature time-series data, mapping discrete GLCM statistical indices into semantic vectors that match the spectral feature dimensions, thus generating texture feature time-series vectors.

[0062] Next is the feature concatenation and dimension adaptation layer, which concatenates the spectral feature time-series vector and the texture feature time-series vector through channels. Then, a linear mapping layer adjusts the concatenated features to a preset dimension to generate the final dual-stream fusion time-series feature vector.

[0063] Following that is the gated loop module, which includes multiple cascaded, pre-defined, two-dimensional coupled enhanced gated loop units (DDC-PGRU) for disease progression time series processing.

[0064] Following that are two parallel fully connected layer branches. The first fully connected layer branch is used to output the diagnosis result of the disease stage of rice at the current moment, and the second fully connected layer branch is used to output the prediction result of the severity of disease in rice.

[0065] In this embodiment, the trend gate is used to sense the rate of change of spectral texture features in adjacent time steps to perceive the "speed of disease development," and the potential gate is used to sense the degree of accumulation of historical energy to achieve an effective measurement of the "cumulative effect of disease" and establish a mapping relationship from "current accumulation" to "future outbreak level." However, this approach is insufficient to meet the core requirement of quantitative prediction of future disease severity in precision agriculture. This application's specially designed disease progression time-series dual-dimensional coupled enhanced gated recurrent unit (DDC-PGRU) breaks through the single-dimensional limitations of traditional models. By innovatively constructing a dual-dimensional coupling mechanism of "trend gate + potential gate," it achieves for the first time a synchronous mechanistic modeling of the "evolutionary direction" and "latency intensity" of the disease, thus balancing accurate identification of the incubation period with reliable prediction of disease severity.

[0066] In one embodiment, step 1 above, based on the hyperspectral image time series data of rice, extracts spectral feature time series data and texture feature time series data respectively, including:

[0067] Step 11: Based on the time series data of hyperspectral images of rice canopy, perform black and white board correction and noise reduction on the original images, extract the average spectral reflectance data of the target region ROI, and construct a time series hyperspectral dataset of the entire disease course.

[0068] Step 12: Extract the spectral reflectance curves of the entire band or characteristic band, and calculate the first-order differential spectrum to form spectral characteristic time series data;

[0069] Step 13: Based on the gray-level co-occurrence matrix algorithm, extract the texture features of the target region ROI to form texture feature time series data. The texture features include: contrast, energy, entropy and correlation.

[0070] In this embodiment of the application, in order to comprehensively capture disease characteristics, based on the data obtained in step 11, two parallel time-series data streams (i.e., "spectral-texture" dual-stream data) are constructed as inputs to the spatiotemporal dual-stream fusion diagnostic network model:

[0071] (1) Spectral Stream ): Extract the spectral reflectance curves of the entire band or characteristic bands, and calculate the first-order differential spectrum to form time-series data that can characterize the physiological and biochemical characteristics of crops, such as pigment degradation and water stress.

[0072] (2) Texture Stream Based on the Gray Co-occurrence Matrix (GLCM) algorithm, texture features of the corresponding ROI are extracted to form time-series data that can characterize the microscopic morphological evolution of crop leaf surfaces.

[0073] In one implementation, step 2 above, the method for obtaining the temporal feature vector of dual-stream fusion, includes:

[0074] Step 21: Encode the spectral feature time series data using a one-dimensional convolutional neural network to obtain the spectral feature time series vector;

[0075] Step 22: Apply a fully connected layer to perform nonlinear mapping on the temporal data of texture features to obtain the temporal vector of texture features;

[0076] Step 23: Channel splicing is performed based on spectral feature time-series vector and texture feature time-series vector to obtain dual-stream fusion time-series feature vector.

[0077] In one implementation, after performing channel concatenation based on spectral feature time-series vectors and texture feature time-series vectors in step 23, the method further includes: adjusting the concatenated features to a preset dimension through a linear mapping layer to generate the final dual-stream fusion time-series feature vector.

[0078] In this embodiment of the application, in order to provide the subsequent core inference module with a time-series input that is dimensionally unified and semantically rich, the dual-stream data is subjected to basic encoding and concatenation:

[0079] (1) Spectral stream coding: Spectral data is processed using a one-dimensional convolutional neural network (1D-CNN). Local morphological features (slope changes of absorption peaks) between bands are extracted using convolutional kernels and compressed from the high-dimensional spectral space to the feature space to generate a spectral feature time-series vector.

[0080] (2) Texture Stream Coding: The texture data is nonlinearly mapped using a fully connected layer. Discrete GLCM statistical indices are mapped into semantic vectors that match the spectral feature dimensions, generating temporal vectors of texture features.

[0081] (3) Feature fusion and dimension adaptation: The encoded spectral feature time-series vector and texture feature time-series vector are concatenated by channels, and the fused features are adjusted to the preset dimension through a linear mapping layer to generate the final two-stream fused time-series feature vector. This serves as the input for the subsequent DDC-PGRU.

[0082] In one implementation, see Figure 3 Step 3 above, the disease progression time-series two-dimensional coupled enhanced gated recurrent unit (DDC-PGRU), includes:

[0083] Trend Gate : ;in, ; , These are the two-stream fusion temporal feature vectors at time t-1 and time t, respectively. It is the hidden state at time t-1. It is a gradient-sensitive weight matrix. Use the Sigmoid activation function; , These are the input weight matrix and the cyclic weight matrix of the trend gate, respectively. It is the bias value of the trend gate;

[0084] Potential Gate : ,in, ; , These are the historical energy state vectors at time t-1 and time t, respectively. , These are the input weight matrix and the cyclic weight matrix of the potential gate, respectively. It is the bias value of the potential gate. For the cumulative sensitivity weight matrix, It is a latent amnesia factor that simulates the long-term retention characteristics of disease virulence in plants;

[0085] Candidate hidden state : ,in, , These are the input weight matrix and the cyclic weight matrix used to calculate the candidate hidden state, respectively. It calculates the bias value of the candidate hidden state;

[0086] Final hidden state : ,in, This is the calculation result of the standard update gate. ,in, , These are the learnable weight matrices of the standard update gate. It is the bias value of the standard update gate.

[0087] Specifically, this application features a specially designed two-dimensional coupled enhanced gated recurrent unit (DDC-PGRU) for disease progression timing, using the aforementioned dual-stream fused timing feature vector. As input, a DDC-PGRU unit is used to extract two-dimensional temporal features of the disease course. This unit calculates the hidden state at each time step. At this time, perform the following sub-steps to decouple the disease course information:

[0088] (1) "Trend Gate" "Calculation" – Capturing High-Frequency Pathological Mutations

[0089] The core function of trend gates is to sense the rate of change of spectral or texture features in adjacent time steps. First, the first-order difference vector of the features is calculated. :

[0090]

[0091] Subsequently, using a unique gradient-sensitive weight matrix Perform a linear transformation on the difference vector and calculate the activation by combining the current input with the historical state:

[0092]

[0093] in, , These are the two-stream fusion temporal feature vectors at time t-1 and time t, respectively. It is the hidden state at time t-1. It is a gradient-sensitive weight matrix. Use the Sigmoid activation function; , These are the input weight matrix and the cyclic weight matrix of the trend gate, respectively. It is the bias value of the trend gate; It is for differential features The learnable weight matrix is ​​set to adjust the dimensions of the difference features to fit the computation of the gating unit, enabling the model to perceive the rate of change of features. When physiological indicators change drastically, reaching the latency threshold, As the number increases, the trend gate is activated, indicating that the model focuses on the "onset trend".

[0094] (2) "Potential Gate" ) Calculation - Quantification of Low-Frequency Energy Accumulation

[0095] The core function of the potential gate is to sense the energy accumulation of pathogens within a long time window. First, the historical energy state vector is calculated using the sliding cumulative operator. :

[0096]

[0097] Subsequently, using a unique cumulative sensitive weight matrix Transform the accumulated state and perform activation calculation:

[0098]

[0099] in It is a latent amnesia factor that simulates the long-term retention characteristics of disease virulence in plants; It is aimed at historical accumulation characteristics The learnable weight matrix is ​​set to adjust the dimensionality of the cumulative features, enabling the model to perceive the degree of historical energy accumulation.

[0100] Even at the current moment The value did not exceed the symptomatic threshold. If there has been a persistent, weak abnormality over a period of time, It will continue to accumulate. Under the influence of It will be activated. This is equivalent to an "integrator" to prevent the model from ignoring weak but persistent latent pathological signals.

[0101] (3) Performing two-dimensional coupling and hidden state update: Unlike the single feature integration method of traditional models, DDC-PGRU uses the above two gating mechanisms to modulate the input information stream and the historical memory stream respectively, generating a hidden state containing two-dimensional disease course information:

[0102] Two-dimensional coupled computation:

[0103]

[0104] Trend Gate Determines the current input Which "dynamic mutations" are worth incorporating into new candidate states? Potential gates The hidden state of the previous moment was determined. What "cumulative risks" are worth retaining?

[0105] Final hidden state output: Combined with standard update gate Generate the final hidden state .

[0106]

[0107] The final hidden state It simultaneously encodes the transient outbreak trend and steady-state accumulation potential of the disease, providing a complete feature set for subsequent dual-task prediction.

[0108] In one implementation, step 4 above, based on the hidden state output by the gated loop module, is input into two parallel fully connected layer branches, including:

[0109] Step 41: Based on the hidden state input of the gated loop module, the first fully connected branch outputs the diagnosis result of the disease stage of rice at the current moment, and the first fully connected branch serves as the diagnostic classifier.

[0110] Step 42: Based on the hidden state input of the gated loop module, the second fully connected branch outputs the prediction result of the disease severity of rice, and the second fully connected branch serves as the level predictor.

[0111] In this embodiment, the final hidden state output by the disease progression time-coupled enhanced gated recurrent unit (DDC-PGRU) contains two-dimensional information about the disease progression (evolutionary trend and latent potential). The final result is output through two lightweight parallel fully connected layer branches:

[0112] (1) Diagnostic branch: This branch is designed to decode The dynamic characteristics are dominated by "trend gates". Through classification mapping, the disease stage (healthy / latent / manifest stage) of rice at the current moment is output, realizing accurate identification of early disease signals.

[0113] (2) Prediction branch: This branch is designed to decode The cumulative characteristics dominated by the "latency gate" are used to establish a mapping relationship between the characteristics and the future severity of the disease by utilizing the integral effect of the potential gate on historical latent stress. The predicted disease level (0-9 grade standard classification) after the onset of symptoms is output, so as to achieve a prospective early warning of the severity of the disease.

[0114] It should be noted that the network layers used in step 2 (dual-stream fusion temporal feature vector calculation), the gated recurrent module in step 3, and the two parallel fully connected layer branches in step 4 are all obtained after training using pre-prepared training samples. The methods for obtaining these training samples include:

[0115] Step A1: Starting from the tillering stage of rice, collect hyperspectral images of the rice canopy at preset time intervals, including the healthy stage before inoculation, the latent stage, and the symptom-showing stage, to obtain time series data of hyperspectral images of the rice canopy.

[0116] Step A2: Perform black and white card correction and noise reduction on the original image, and extract the average spectral reflectance data of the target region ROI.

[0117] This application provides an electronic device, comprising: a processor; and a memory for storing processor-executable instructions; wherein the processor executes the executable instructions to implement the aforementioned method for diagnosing and predicting the latent period of rice blast based on dual-stream data fusion. The processor and the memory are connected, for example, via a bus. Optionally, the electronic device may further include a transceiver, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver is not limited to one, and the structure of the electronic device does not constitute a limitation on the embodiments of this application. The processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). In some embodiments, the processor may further include an Artificial Intelligence (AI) processor for processing computational operations related to machine learning. The memory may include one or more computer-readable storage media, which may be non-transitory. The memory may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash storage devices.

[0118] This application provides a computer-readable storage medium storing computer instructions that, when executed by a processor, implement the steps of the above-described method for diagnosing and predicting the latent period of rice blast based on dual-stream data fusion. Those skilled in the art will recognize that the functions described in the above examples can be implemented using hardware, software, firmware, or any combination thereof. When implemented in software, these functions can be stored in a computer-readable storage medium or transmitted as one or more instructions or code on the computer-readable storage medium. Computer-readable storage media include computer storage media and communication media, wherein communication media include any medium that facilitates the transfer of computer programs from one location to another. Storage media can be any available medium accessible to general-purpose or special-purpose computers.

[0119] The following provides a specific embodiment of the method for diagnosing and predicting the latent period of rice blast based on dual-stream data fusion, along with corresponding experimental verification results.

[0120] 1. Experimental Data Acquisition and Preprocessing

[0121] This embodiment first establishes a standard observation field under controlled greenhouse conditions, selects the CO39 variety, which is susceptible to rice blast, for pot cultivation, and uses the spore suspension spraying method for artificial inoculation to establish a precise "incubation timeline". Using hyperspectral imaging equipment (900-1700nm) and high-resolution image acquisition equipment, rice canopy data were collected daily from the day of inoculation until the end of the symptom-manifesting period, thus forming a time-series dataset that includes the entire stages of health, latent, and symptom-manifesting stages.

[0122] After acquiring the raw data, this embodiment constructs multidimensional feature engineering and dual-stream input:

[0123] Spectral flow ( The differential enhancement construction involves first smoothing the original spectral data using SG (Savitzky-Golay) filtering to remove instrument noise. To meet the gradient sensing requirements of the trend gate in the subsequent DDC-PGRU, this embodiment further calculates the first-order differential sequence of the spectrum. Specifically, the derivative of each band is calculated using the central difference method. This operation not only effectively eliminates background baseline drift but also significantly amplifies the subtle slope changes in the chlorophyll red-edge position. Finally, the smoothed reflectance spectrum and the first-order differential spectrum are concatenated along the channel dimension to form an input stream rich in "intensity + gradient" information. See also... Figure 4 This is a comparison of the original hyperspectral reflectance curves of rice canopy at different disease stages; see also Figure 5 The image shows a comparison of first-order differential spectral curves.

[0124] Texture Flow ( The statistical construction involved extracting regions of interest (ROIs) from the acquired images and converting them to grayscale. Subsequently, based on the Gray-Level Co-occurrence Matrix (GLCM) algorithm, four core statistics were extracted: Contrast, Energy, Entropy, and Correlation. The design logic of this feature combination is that contrast reflects the clarity of lesion edges, while entropy reflects the degree of disorder in leaf texture, i.e., the entropy increase caused by disease infection. This set of features can effectively quantify the evolutionary process of micromorphology.

[0125] 2. Construction of a spatiotemporal dual-stream fusion diagnostic network

[0126] like Figure 2The network model structure shown in this invention, which constructs a spatiotemporal dual-stream fusion diagnostic network architecture, is implemented based on deep learning frameworks such as PyTorch and mainly includes three functional modules: parallel encoding, feature fusion, and core inference.

[0127] 2.1 Parallel Encoding

[0128] For the parallel feature encoder, the network employs a TimeDistributed mechanism to ensure independent processing at each time step t. The spectral encoding branch uses a one-dimensional convolutional neural network (1D-CNN), preferably configured with smaller kernels to capture fine spectral fingerprints between adjacent bands, and connected to max-pooling layers to compress data dimensionality while preserving significant pathological features. The texture encoding branch uses a fully connected network (FC Network), containing several fully connected layers and a non-linear activation function (ReLU), aiming to map low-dimensional discrete statistical indicators (GLCM features) to a high-dimensional semantic space, aligning its feature representation capabilities with those of the spectral branch.

[0129] 2.2 Feature Fusion

[0130] After encoding is complete, the process proceeds to the feature fusion and adaptation layer. For example... Figure 2 As shown, the system concatenates the two feature vectors and, to address the issue of inconsistent heterogeneous feature distributions, connects a linear adapter layer after concatenation. This layer projects the fused features onto a unified latent feature space using a learnable weight matrix, generating a standardized temporal input. This provides high-quality input for subsequent core units.

[0131] 2.3 Internal Operation Mechanism of the DDC-PGRU Core Unit

[0132] In this embodiment, the DDC-PGRU is a key component for achieving the purpose of this invention, and its internal structure is as follows: Figure 3 As shown. Let the input feature dimension be... The hidden state dimension is This unit achieves two-dimensional decoupling through its unique parameter matrix and computation path:

[0133] First, the key parameter matrix is ​​initialized. This embodiment constructs a matrix with dimension [missing information]. The unique gradient sensitivity matrix and unique cumulative sensitivity matrix Both employ orthogonal initialization to maintain gradient propagation stability. Its function is to linearly project the difference features onto the trend gate activation space, while Its function is to linearly project historically accumulated features onto the potential gate activation space. Furthermore, the latent forgetting factor... It is preferable to set a large initial value (0.9) and set it as a trainable parameter, which allows the model to adaptively learn the decay rate of disease energy.

[0134] In the step-by-step derivation process at each time step t, DDC-PGRU executes the following logic:

[0135] (1) Trend feature extraction and activation: Calculate the difference between the current input and the input at the previous time step. The model utilizes Identification High-frequency abrupt changes in the trend gate occur when the slope of the red edge of the spectrum changes drastically. Activated by highlighting ( ).

[0136] (2) Potential feature extraction and activation: update historical cumulative items The model utilizes measure The integral strength, even with a small change at a single point, will increase if there is a persistent latent stress. It will also be activated ( ).

[0137] (3) Two-dimensional coupled update: Finally, the current input is weighted using a trend gate. Using the potential gate weighted historical memory This achieves a deep integration of "the instantaneous evolution trend of the disease" and "the historical accumulated potential of the disease," generating candidate hidden states. This gives it both a keen sensitivity to mutation signals and a long-term memory of latent stress. ).

[0138] 2.4 Dual-task collaborative training strategy

[0139] During the optimization process, a joint loss function was constructed. ,in Classification loss used for diagnosing fully connected branches of a classifier , This is the regression loss used in the fully connected branch of the level predictor. Among them, the cross-entropy loss... Main driving trend gate ( The parameter updates make it sensitive to sudden changes; while the mean squared error loss Main driving potential gate ( The parameters of the system are updated to make it sensitive to the cumulative amount. This is achieved by adjusting the balance coefficient. , This allows the model to achieve an optimal balance between the two tasks.

[0140] 3. Implementation Results and Mechanism Verification

[0141] The model verification results based on this embodiment are as follows: Figure 6 As shown, the effectiveness of the mechanism of this invention is fully demonstrated:

[0142] Regarding the trend gate response mechanism: On day 3 after artificial inoculation (i.e., the critical point of transition from the latent to the symptomatic stage), the trend gate activation value in DDC-PGRU showed a significant peak. This demonstrates the effectiveness of the introduced difference term. sum matrix The model was successfully endowed with the ability to capture subtle physiological mutations, enabling early warning up to 48 hours before symptoms appear.

[0143] Regarding the response mechanism of the latent potential gate: As the infestation time progresses, even when the external texture has not changed, the activation value of the latent potential gate shows a continuous upward trend. This proves the effectiveness of the introduced cumulative term. sum matrix The energy accumulation of the disease was successfully quantified, enabling the model to accurately predict the severity of disease after symptoms appear based on the accumulated information during the incubation period.

[0144] This invention is not limited to the specific embodiments described above. Any modifications made by those skilled in the art based on the above concept without creative effort are within the scope of protection of this invention.

Claims

1. A method for diagnosing and predicting latent period of rice blast based on dual-stream data fusion, characterized in that, Includes the following steps: Based on the time series data of hyperspectral images of rice, spectral feature time series data and texture feature time series data were extracted respectively. The time-series data of spectral features and the time-series data of texture features are represented by vectors respectively, and then concatenated and fused to obtain the dual-stream fused time-series feature vector; Based on the dual-stream fusion time-series feature vector input gated loop module, time-series data prediction is performed to obtain the hidden state output by the gated loop module. The gated loop module includes multiple preset disease course time-series two-dimensional coupling enhanced gated loop units (DDC-PGRU). The hidden state output by the gated loop module is input into two parallel fully connected layer branches, which output the diagnosis result of the disease stage of rice at the current moment and the prediction result of the disease severity. The preset DDC-PGRU is an improvement based on the gated recurrent unit GRU. The DDC-PGRU unit includes a trend gate and a potential gate. The calculation process of the trend gate introduces the rate of change of the input time-series feature vector at adjacent time points to capture high-frequency pathological mutations. The calculation process of the potential gate introduces the energy accumulation value of the input time-series feature vector within a long time window to perceive the historical accumulation state of latent pathology. The default DDC-PGRU includes: Trend Gate : ; in, ; , These are the two-stream fusion temporal feature vectors at time t-1 and time t, respectively. It is the hidden state at time t-1. It is a gradient-sensitive weight matrix. Use the Sigmoid activation function; , These are the input weight matrix and the cyclic weight matrix of the trend gate, respectively. It is the bias value of the trend gate; It is for differential features The learnable weight matrix is ​​set to adjust the dimension of the difference features to fit the calculation of the gating unit, so that the model can perceive the rate of change of the features. Potential Gate : ,in, ; , These are the historical energy state vectors at time t-1 and time t, respectively. , These are the input weight matrix and the cyclic weight matrix of the potential gate, respectively. It is the bias value of the potential gate. For the cumulative sensitivity weight matrix, It is a latent amnesia factor that simulates the long-term retention characteristics of disease virulence in plants; It is aimed at historical accumulation characteristics The learnable weight matrix is ​​set to adjust the dimension of the cumulative features, enabling the model to perceive the degree of accumulation of historical energy. Candidate hidden state : ,in, , These are the input weight matrix and the cyclic weight matrix used to calculate the candidate hidden state, respectively. It calculates the bias value of the candidate hidden state; Final hidden state : ,in, This is the calculation result of the standard update gate. ,in, , These are the learnable weight matrices of the standard update gate. It is the bias value of the standard update gate.

2. The method for predicting latent period of rice blast based on dual-stream data fusion according to claim 1, characterized in that, Based on the time-series data of hyperspectral images of rice, spectral feature time-series data and texture feature time-series data were extracted, including: Based on the time series data of hyperspectral images of rice canopy, black and white card correction and noise reduction were performed on the original images, and the average spectral reflectance data of the target region ROI was extracted. Extract the spectral reflectance curves of the entire band or characteristic band, and calculate the first-order differential spectrum to form spectral characteristic time series data; Based on the gray-level co-occurrence matrix algorithm, texture features of the target region ROI are extracted to form temporal data of texture features. 3.The rice blast latent period diagnosis and prediction method based on dual-stream data fusion according to claim 1, characterized in that, Methods for obtaining temporal feature vectors from two-stream fusion include: A one-dimensional convolutional neural network is used to encode the spectral feature time series data to obtain the spectral feature time series vector; a fully connected layer is used to perform nonlinear mapping on the texture feature time series data to obtain the texture feature time series vector. Channel splicing is performed based on spectral feature time-series vectors and texture feature time-series vectors to obtain dual-stream fusion time-series feature vectors.

4. The method according to claim 3, wherein the method is based on dual-stream data fusion for latent period diagnosis and prediction of rice blast. After channel concatenation based on spectral feature time-series vectors and texture feature time-series vectors, the process also includes: adjusting the concatenated features to a preset dimension through a linear mapping layer to generate the final dual-stream fusion time-series feature vector.

5. The method for diagnosing and predicting the latent period of rice blast based on dual-stream data fusion according to claim 1, characterized in that, The hidden state output from the gated recurrent module is input into two parallel fully connected layer branches, including: The first fully connected branch, which is based on the hidden state input of the gated loop module, outputs the diagnosis result of the disease stage of rice at the current moment. The first fully connected branch serves as a diagnostic classifier. The hidden state input of the gated loop module is used as the output of the second fully connected branch to predict the severity of rice disease. The second fully connected branch serves as the level predictor.

6. The method for diagnosing and predicting the latent period of rice blast based on dual-stream data fusion according to claim 1, characterized in that, The network layers, gated recurrent modules, and two parallel fully connected layer branches used in the calculation of the temporal feature vectors for dual-stream fusion are all obtained after training using pre-prepared training samples. The methods for obtaining these training samples include: Starting from the tillering stage of rice, hyperspectral images of the rice canopy were continuously collected at preset time intervals, including the healthy stage before inoculation, the latent stage, and the symptom-manifesting stage, to obtain time series data of hyperspectral images of the rice canopy. The original image was subjected to black and white card correction and noise reduction, and the average spectral reflectance data of the target region (ROI) was extracted.

7. A diagnostic prediction system for the latent period of rice blast based on dual-stream data fusion according to any one of the diagnostic prediction methods described in claims 1-6, characterized in that, The diagnostic prediction system includes a spatiotemporal dual-stream fusion diagnostic network model, which comprises, in sequence: The system consists of a parallel one-dimensional convolutional neural network and a fully connected layer. The one-dimensional convolutional neural network is used to represent the temporal data of spectral features into vectors, generating temporal vectors of spectral features. The fully connected layer is used to perform nonlinear mapping on the temporal data of texture features, generating temporal vectors of texture features. The feature concatenation and dimension adaptation layer is used to concatenate the temporal vectors of spectral features and the temporal vectors of texture features through channels. Then, a linear mapping layer is used to adjust the concatenated features to a preset dimension to generate the final dual-stream fusion temporal feature vector. The gated loop module includes multiple cascaded preset two-dimensional coupled enhanced gated loop units (DDC-PGRU) for disease progression time series processing. The multiple cascaded preset DDC-PGRU units are connected in series according to time steps to process time series data. Two parallel fully connected layer branches are used. One fully connected layer branch is used to output the diagnosis result of the disease stage of rice at the current moment, and the other fully connected layer branch is used to output the prediction result of the disease severity of rice.

8. An electronic device, comprising: The electronic device includes: processor; Memory used to store processor-executable instructions; The processor executes the executable instructions to implement the rice blast incubation period diagnosis and prediction method based on dual-stream data fusion as described in any one of claims 1-6.

9. A computer readable storage medium having stored thereon computer instructions, wherein, When the instruction is executed by the processor, it implements the steps of the method for diagnosing and predicting the latent period of rice blast based on dual-stream data fusion as described in any one of claims 1-6.

Images