Early disease prediction method and system driven by remote sensing change information of wheat stripe rust

Abstract

The invention provides a wheat stripe rust remote sensing change information-driven early disease prediction method and system, and the system comprises a multi-source data progressive fusion module which enables a natural image and a multispectral image to be spliced step by step into corresponding hierarchical features, and carries out the fusion and outputting of deep fusion features; the frequency domain decoupling-based change detection module is used for outputting a change probability graph of adjacent moments; constructing a plurality of groups of training samples and inputting the training samples into a conditional diffusion prediction model for training; to-be-predicted wheat remote sensing image data and corresponding meteorological data are collected, a change probability graph is obtained, the change probability graph, the corresponding meteorological data and diffusion mode prior data are fused to serve as a condition vector, the condition vector and random Gaussian noise are input into the trained condition diffusion prediction model together for denoising, and prediction denoising data are obtained; then a prediction change probability graph is obtained through a visual decoder; and on the basis, obtaining a prediction result of the severity and distribution range of the wheat stripe rust disease on the d-th day. The method can be used for predicting the early wheat stripe rust.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wheat stripe rust prediction, in particular to a wheat stripe rust early disease prediction method and system driven by remote sensing change information. BACKGROUND

[0002] Wheat production is long-term threatened by pests and diseases, among which wheat stripe rust is particularly severe. This disease is caused by the specific parasitic fungus Puccinia striiformis, which specifically infects the green tissues of the plant, leading to photosynthetic tissue yellowing and necrosis. In addition, wheat stripe rust has the characteristics of climate sensitivity and can cause regional outbreaks under certain environmental conditions, resulting in significant yield loss. Therefore, timely and accurate prediction of the occurrence and spread of stripe rust is of great significance for effectively preventing and controlling the disease and ensuring the yield and quality of wheat. Predicting the occurrence of stripe rust can provide a scientific basis for developing reasonable control measures, reducing pesticide use, reducing environmental pollution, and ensuring food security. However, the prediction of stripe rust faces great difficulties. Conventional detection methods rely on the identification of yellowing lesions, but by the time the lesions appear, the fungus has already spread through the vascular system, resulting in a loss of timeliness for control.

[0003] Traditional methods based on statistical analysis rely on the collection and analysis of a large amount of historical data, and establish the correlation between meteorological conditions and disease occurrence through manual feature extraction. However, due to the strong non-linear characteristics of the disease, these methods have limited data mining capabilities when faced with complex and variable environmental factors, resulting in less than ideal prediction results.

[0004] In contrast, deep learning prediction methods can automatically extract high-dimensional features from massive amounts of data, breaking through the bottleneck of manual feature selection in traditional methods and exhibiting stronger adaptability and higher prediction accuracy, thus gaining widespread application.

[0005] However, existing deep learning prediction methods have the following limitations:

[0006] 1. The existing multi-source data fusion method is single and has limited feature expression capability. Early disease mainly manifests as subtle changes in plant internal structure and physiological parameters, which are difficult to detect alone through natural images. Although existing methods attempt to fuse natural images and multispectral images to enhance detection capability, the fusion method is usually limited to image stacking or feature splicing, lacking in-depth mining of the correlation between different levels of features, resulting in limited fusion effect, feature redundancy or information loss, and affecting disease expression capability and model recognition accuracy.

[0007] 2. Lack of modeling capability for temporal change information, difficulty in identifying weak signals of disease latent period. The lesion characteristics of wheat stripe rust during the latent period are usually hidden, with small changes and low signal-to-noise ratio, which are easily overwhelmed by noise in the process of multi-source data fusion. However, existing deep learning methods are mostly based on direct detection framework, relying only on single temporal image to extract static features and identify diseases, failing to fully utilize the change information contained in multi-temporal remote sensing data, and difficult to achieve sensitive perception and accurate identification of early evolution of diseases.

[0008] 3. Insufficient ability to model images and environmental factors (such as meteorological data, diffusion mode prior knowledge, etc.) together, making it difficult to depict the dynamic propagation law of diseases. The occurrence and development of wheat stripe rust is influenced by time, space and environmental conditions, with obvious spatio-temporal dynamics and nonlinear characteristics. However, existing methods mostly use traditional time series modeling strategies, failing to effectively integrate remote sensing images and environmental factors, and lacking the ability to model disease transmission trends, making it difficult to meet the application needs of early warning and precise prevention and control. SUMMARY

[0009] To solve the technical problems existing in the prior art, the present application provides a wheat stripe rust remote sensing change information driven early disease prediction method and system, and the technical solution is as follows:

[0010] On the one hand, a wheat stripe rust remote sensing change information driven early disease prediction method is provided, which comprises:

[0011] S1, collecting and preprocessing historical remote sensing image data of wheat, corresponding historical meteorological data and diffusion mode prior data, the historical remote sensing image data comprising historical natural image data and historical multispectral image data;

[0012] S2, inputting the historical remote sensing image data into and training a multi-source data progressive fusion module, and fusing the multi-scale features of the preprocessed natural image and multispectral image by gradually splicing the corresponding level features, using a weighting strategy to enhance the response of the change area, and outputting deep fusion features of a unified scale ;

[0013] S3, inputting the into and training a change detection module based on frequency domain decoupling, and pairing each other in time sequence, respectively performing Fourier transform and passing through Gaussian high-pass filter and low-pass filter to obtain high-frequency decoupling features and low-frequency decoupling features, and fusing the difference results with the time domain features after inverse transform, to output change probability maps of adjacent time points , , ;

[0014] S4, constructing a plurality of groups of training samples in sequence according to a fixed length sliding window using a sliding window with a window size of d, encoding the last change probability map in each group of training samples through a visual encoder to obtain a change feature as noisy data, inputting the remaining change probability maps in each group of training samples, and corresponding historical meteorological data and diffusion mode prior data fusion as a conditional vector into a conditional diffusion prediction model to perform model training;

[0015] S5, collecting d-1 days of to-be-predicted wheat remote sensing image data and corresponding meteorological data, inputting the remote sensing image data into the trained multi-source data progressive fusion module and the change detection module based on frequency domain decoupling to obtain a change probability map , and corresponding meteorological data and diffusion mode prior data fusion as a conditional vector, and inputting the conditional diffusion prediction model trained into a denoising operation to obtain predicted denoising data , and then passing through a visual decoder to obtain a predicted change probability map on the dth day

[0016] S6, based on the multispectral image in the d-1 days of to-be-predicted wheat remote sensing image data, calculating an NDVI mask, retaining only the vegetation area, multiplying the change probability map and the NDVI mask pixel by pixel, and generating a lesion mask with a preset threshold, and calculating the proportion of lesion pixels to vegetation pixels as the prediction result of the severity and distribution range of wheat stripe rust disease on the dth day.

[0017] On the other hand, a wheat stripe rust remote sensing change information driven early disease prediction system is provided, and the system comprises:

[0018] A collection and preprocessing module is configured to collect and preprocess historical wheat remote sensing image data, corresponding historical meteorological data, and diffusion mode prior data, wherein the historical remote sensing image data comprises historical natural image data and historical multispectral image data.

[0019] A first training module is configured to input the historical remote sensing image data into a multi-source data progressive fusion module and train the multi-source data progressive fusion module. The multiscale features of the preprocessed natural image and multispectral image are sequentially spliced and fused, a weighted strategy is used to enhance the response of the change area, and a unified scale deep fusion feature is output​​​​​​​​​​​ ;

[0020] The second training module is configured to train the change detection module based on frequency domain decoupling by inputting the change detection module based on frequency domain decoupling and the historical change probability map The change detection module based on frequency domain decoupling is configured to pair the historical change probability map and the current change probability map in time sequence, respectively perform Fourier transform and pass through Gaussian high-pass filtering and low-pass filtering to obtain high-frequency decoupling features and low-frequency decoupling features, fuse the high-frequency decoupling features and the low-frequency decoupling features after inverse transformation with a difference result of the time domain features, and output a change probability map of adjacent time points , , ;

[0021] The construction module is configured to sequentially slide in fixed lengths in the change probability map , , by using a sliding window with a window size d, to construct multiple groups of training samples, to take change features obtained by encoding the last change probability map in each group of training samples by a visual encoder as noisy data, to fuse the remaining change probability maps in each group of training samples, corresponding historical meteorological data and diffusion mode prior data as a conditional vector, and to input the conditional diffusion prediction model for model training.

[0022] The prediction module is configured to collect d-1 days of wheat remote sensing image data to be predicted and corresponding meteorological data, to input the remote sensing image data into the trained multi-source data progressive fusion module and the change detection module based on frequency domain decoupling to obtain a change probability map , , , to take the change probability map , , , corresponding meteorological data and diffusion mode prior data fusion as a conditional vector, and to input the trained conditional diffusion prediction model together with random Gaussian noise to perform a denoising operation to obtain predicted denoising data , and then pass through a visual decoder to obtain a predicted change probability map of the dth day ;

[0023] The calculation module is configured to calculate NDVI masks based on multispectral images in the d-1 days of wheat remote sensing image data to be predicted, to retain only vegetation regions, to multiply the change probability map and the NDVI masks pixel by pixel, to generate a lesion mask with a preset threshold, to count a proportion of lesion pixels to vegetation pixels as a prediction result of a severity and a distribution range of wheat stripe rust disease of the dth day.

[0024] ​In another aspect, an electronic device is provided, including a processor and a memory having at least one instruction stored therein, the at least one instruction being loaded and executed by the processor to implement the above-described early disease prediction method driven by remote sensing change information of wheat stripe rust.

[0025] In another aspect, a computer-readable storage medium is provided, having at least one instruction stored therein, the at least one instruction being loaded and executed by a processor to implement the above-described early disease prediction method driven by remote sensing change information of wheat stripe rust.

[0026] The technical solution provided by the present application brings at least the following beneficial effects:

[0027] 1. The multi-modal progressive fusion module based on adaptive strategy designed in the present application deeply fuses multi-source data such as natural images and multispectral images, realizes complementary utilization of different spectral information, and effectively improves the representation ability of disease characteristics.

[0028] 2. The change detection module based on frequency domain decoupling designed in the present application analyzes the difference information of double-time images in time domain, high frequency and low frequency, fully excavates the subtle change signals of wheat caused by diseases in the latent period, realizes sensitive identification and efficient data representation of early diseases.

[0029] 3. The conditional diffusion prediction model designed in the present application models the change trend of diseases over time, introduces meteorological data and diffusion mode prior knowledge as constraint vectors, improves the description ability of the model to time sequence characteristics, and realizes dynamic and accurate prediction of disease development. BRIEF DESCRIPTION OF DRAWINGS

[0030] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0031] Figure 1 is a flow chart of an early disease prediction method driven by remote sensing change information of wheat stripe rust provided by an embodiment of the present application;

[0032] Figure 2 is a pre-processing flow chart provided by an embodiment of the present application;

[0033] Figure 3 is a structural block diagram of a multi-source data progressive fusion module provided by an embodiment of the present application;

[0034] Figure 4is a feature extraction block structure block diagram provided by an embodiment of the application;

[0035] Figure 5 is a high-order feature enhancement schematic diagram provided by an embodiment of the application;

[0036] Figure 6 is a change detection module structure block diagram based on frequency domain decoupling provided by an embodiment of the application;

[0037] Figure 7 is a visual encoder structure block diagram provided by an embodiment of the application;

[0038] Figure 8 is a training flowchart of the conditional diffusion prediction model provided by an embodiment of the application;

[0039] Figure 9 is a denoising operation flowchart of the conditional diffusion prediction model provided by an embodiment of the application;

[0040] Figure 10 is a visual decoder structure block diagram provided by an embodiment of the application;

[0041] Figure 11 is a conditional diffusion prediction model structure block diagram provided by an embodiment of the application;

[0042] Figure 12 is a wheat stripe rust remote sensing change information driven early disease prediction system block diagram provided by an embodiment of the application;

[0043] Figure 13 is a structure schematic diagram of an electronic device provided by an embodiment of the application. DETAILED DESCRIPTION

[0044] In order to make the technical problems, technical solutions and advantages of the application clearer, specific embodiments will be described in detail below with reference to the drawings.

[0045] In order to overcome the limitations of the prior art in the prediction of wheat stripe rust, such as single data fusion mode, insufficient use of time series change information, and weak image and environmental factor collaborative modeling capability, the embodiment of the present application aims to provide a wheat stripe rust remote sensing change information driven early disease prediction method. The method innovatively predicts dynamic "change information" instead of traditional static "state information", which can more sensitively capture the weak signal of the latent period of the disease compared to directly detecting the state of the plant, thereby improving the accuracy and real-time performance of the prediction. First, the embodiment of the present application deeply integrates multi-source data such as natural images and multispectral images through a multi-source progressive fusion module based on an adaptive strategy, fully utilizes the feature complementarity between multi-source data to improve data quality. Second, based on the change detection framework, a time-frequency domain collaborative analysis method is used to analyze the difference information of double-time-phase images in time domain, high frequency and low frequency, fully excavate the subtle change signal of wheat caused by disease in the latent period, and realize sensitive identification of the change of early disease. Finally, a conditional diffusion prediction model is designed, which takes meteorological data (temperature, humidity, wind speed, etc.) and diffusion mode prior knowledge (such as disease transmission type, transmission stage, transmission direction, diffusion intensity level, etc.) as constraint vectors, combines with change probability map, and through the iterative process of forward noise addition-reverse noise removal, dynamically simulates the space-time diffusion law of stripe rust, thereby realizing high-quality prediction of the severity and range of wheat stripe rust.

[0046] The early disease prediction method for wheat stripe rust remote sensing change information driven provided by the embodiment of the present application can be realized by an electronic device, which can be a terminal or a server. Figure 1 The method flowchart is shown, and the processing flow can include the following steps:

[0047] S1, collect and pre-process wheat historical remote sensing image data, corresponding historical meteorological data and diffusion mode prior data, the historical remote sensing image data including historical natural image data and historical multispectral image data;

[0048] The embodiment of the application utilizes the unmanned aerial vehicle carrying MicaSense Altum-PT multi-spectral camera to carry out low-altitude aerial photography on farmland area, simultaneously obtains high-resolution data of blue light, green light, red light, red edge, near-infrared waveband and natural image, accurately controls flight position and aerial photography path through real-time dynamic difference technology, ensures time and space consistency of data collection, covers the entire farmland area, executes aerial photography task at the same time every day during the collection process, so as to reduce the influence of light change and environmental factors on data quality, and generates a complete area coverage map by using image stitching technology. In order to further improve the geometric matching accuracy of data, high-reflectivity ground calibration points are arranged in the farmland area as the geometric correction reference points of image data, and the registration consistency of multi-temporal data is ensured. In the data collection process, the daily collection time point is taken as the time sequence index , the initial reference day is taken as , and the subsequent days are taken in turn , the obtained blue light, green light, red light, red edge, near-infrared waveband image sequence and natural image sequence are respectively represented as , , , , , in addition, the environmental information of the farmland area on the same day is recorded, including precipitation , temperature and wind speed , to supplement the sensing timing data of disease growth.

[0049] As shown in Figure 2 , the preprocessing of the embodiment of the application includes:

[0050] Firstly, the natural image and the multi-spectral image are radiometrically corrected, and the high-reflectivity ground calibration points are used for geometric correction. Then, in view of the resolution difference between the natural image and the multi-spectral image, a bicubic interpolation algorithm is used to resample the natural image, realize pixel-level registration of the heterogeneous images, stack the blue light image , the green light image , the red light image , the red edge image and the near-infrared image , combine the images of different wavebands along the channel dimension to form a five-channel multi-spectral image . This operation provides standardized input data for subsequent multi-source data fusion and feature extraction, and divides the data into training set and verification set in the ratio of 9:1.

[0051] S2, input the historical remote sensing image data into the multi-source data progressive fusion module and train the module, input the preprocessed natural image​ and multispectral images Multi-scale features are sequentially spliced ​​together and fused with corresponding hierarchical features. A weighted strategy is used to enhance the response in changing regions, outputting a deep fused feature at a unified scale. ;

[0052] Optionally, such as Figure 3 As shown, the multi-source data progressive fusion module gradually fuses perceptual information from natural images and multispectral images (at the same acquisition time point) at different levels through multi-level feature extraction and high-order feature enhancement techniques. The processing includes:

[0053] natural images In the multi-level feature extraction submodule, the input first passes through a 3D convolutional layer with a kernel of 3×3×3, and then passes through three feature extraction blocks in sequence, gradually extracting the first-level features, second-level features, and third-level features of the natural image from the shallow layer to the deep layer.

[0054] Multispectral images at the same acquisition time point The same operation was performed to obtain the first-level features, second-level features, and third-level features of the multispectral image;

[0055] Next, a progressive fusion of multi-source features is performed according to an adaptive strategy, including:

[0056] First, the first-level, second-level, and third-level features of the natural image and the multispectral image are concatenated to form first-level shallow fusion features, second-level shallow fusion features, and third-level shallow fusion features;

[0057] Next, the first-level shallow fusion feature is used as a low-order feature of the shallow layer, and the second-level shallow fusion feature is used as a high-order feature of the deep layer. They are then combined to perform high-order feature enhancement to obtain hidden shallow fusion features. The second-level shallow fusion feature is used as a low-order feature of the shallow layer, and the third-level shallow fusion feature is used as a high-order feature of the deep layer. They are then combined to perform high-order feature enhancement to obtain hidden deep fusion features.

[0058] Finally, the hidden shallow-layer fusion features are used as low-order features of the shallow layer, and the hidden deep-layer fusion features are used as high-order features of the deep layer. These are then combined to perform high-order feature enhancement, ultimately yielding the deep fusion features. (t ranges from 0 to n).

[0059] Optionally, such as Figure 4As shown, the feature extraction block, based on multi-scale convolutional fusion, introduces channel attention and spatial attention mechanisms. Channel attention focuses on the dependencies between channels, enhancing the responsiveness to key channel features by weighting the importance of each channel. Spatial attention focuses on the spatial distribution of the feature map, enhancing the perception of lesion areas and abnormal physiological locations. The two work synergistically to enhance information through channel selection and spatial focusing, thereby improving the expression quality of feature extraction. The processing includes:

[0060] The input shallow layer of the previous level features (such as...) Figure 3 The training process (using natural image primary features) first involves three branches: each branch uses convolutional layers of different sizes, each convolutional layer is followed by a ReLU activation function, and then undergoes layer normalization to stabilize the training process. These three branches generate three feature maps, which are then used... , , express;

[0061] Then, the , , Adaptive fusion is performed through a channel attention mechanism to obtain fused features. By performing adaptive fusion through a spatial attention mechanism, fused features are obtained. The fusion feature and fusion features Pixel-wise multiplication is performed, and features are integrated through a 1×1 convolution operation to output the next level of deep features (such as...). Figure 3 (Natural image secondary features)

[0062] The channel attention mechanism includes:

[0063] The , , Global information of the feature map is obtained through global average pooling and global max pooling. The results of global average pooling and global max pooling are then fed into a multilayer perceptron and summed. Finally, the weight of each channel is calculated using the sigmoid activation function. , The calculated weights Corresponding Multiply to obtain the weighted features Finally, all weighted features , , Add them together to obtain the final fusion feature. The formula is as follows:

[0064]

[0065]

[0066]

[0067] in represents the Sigmoid activation function, MLP represents the multilayer perceptron, GAP represents global average pooling, and GMP represents global max pooling.

[0068] The spatial attention mechanism includes:

[0069] right Perform average pooling and max pooling along the channel dimension respectively to obtain two spatial feature maps. and This reflects the average and maximum responses at each location. These two feature maps are then concatenated along the channel dimension, and a 7×7 convolutional layer is used to extract fused features. The spatial attention weights are then obtained after passing through a sigmoid activation function. Then spatial attention weights With features Multiply to obtain the weighted features Finally, all weighted features , , Adding them together yields the final fusion feature. The formula is as follows:

[0070]

[0071]

[0072]

[0073] in This represents the Sigmoid function, where C represents the number of channels. This represents averaging the feature map along the channel dimension. This represents performing the maximum operation on the feature map along the channel dimension. This represents a two-dimensional convolution operation with a kernel size of 7×7.

[0074] Optionally, such as Figure 5 As shown, the higher-order feature enhancement achieves progressive fusion of multi-level features through spatial scale alignment and dual-path convolution extraction. Simultaneously, it introduces average pooling and covariance pooling to jointly model global and local statistical information, enhancing the sensitivity of the fused features to spatial variations and structural differences, and improving the stability and discriminative ability of feature representation. The processing includes:

[0075] Deep high-order features (such as) Figure 3 Secondary shallow features), through upsampling operations, are recovered to the same level as the shallow low-order features (such as... Figure 3 The first-level shallow features in the model have the same spatial resolution so that they can be fused at the same scale;

[0076] Next, the high-order and low-order features are respectively entered into a dual-parallel feature extraction operation: one path directly extracts spatial features through 3×3 convolution, while the other path generates weights through 3×3 convolution and the Sigmoid activation function for spatial weighted modulation.

[0077] Next, the convolutional output of each path is multiplied pixel by pixel with the corresponding weight to achieve local weighted fusion. The results of the two paths are then added together to generate fused features.

[0078] Then, the fused features are fed into a feature recalibration module. This module first adjusts the channel dimensions using a 1×1 convolution to extract a compact global feature map. Next, the global feature map is fed into average pooling and max pooling branches respectively: the average pooling branch captures global contextual information, while the max pooling branch highlights salient region or edge features. After the two types of information are added, a 3-convolutional layer is used to fuse the local spatial structure, ultimately outputting the higher-order enhanced features (such as...). Figure 3 (Hidden shallow fusion features in the middle).

[0079] S3, the above Input and train a frequency-domain decoupling-based change detection module, Pairing the data in chronological order, Fourier transforms are performed on each pair, followed by Gaussian high-pass and low-pass filtering to obtain high-frequency and low-frequency decoupling features. These features are then inversely transformed and fused with the time-domain feature difference results to output a probability map of changes between adjacent time points. , , ;

[0080] Optionally, such as Figure 6 As shown, the frequency-domain decoupling-based change detection module separates high-frequency and low-frequency change components through Fourier transform and filters. By analyzing the differences in time, high-frequency, and low-frequency information of dual-phase images, it fully explores the subtle change signals caused by diseases in wheat during the incubation period. The processing includes:

[0081] Deeply fused features paired in chronological order (e.g.) and , and

[0082] (etc.), sequentially passing through the same shared frequency domain decoupled encoder, and then through transposed convolutional decoding with a kernel of 3, to output the probability map of changes between adjacent time steps. , , The change probability map is an image of the same size as the input image, where each pixel takes a value in the range [0, 1], representing the probability that the position of each pixel changes between adjacent time points in the image;

[0083] Among them, for adjacent deep fusion features and The frequency domain decoupling encoder first maps the components to the frequency domain using Fourier transform, and then decomposes them into high-frequency decoupling features using Gaussian high-pass filtering and Gaussian low-pass filtering. , and low-frequency decoupling characteristics , ;

[0084] Then, the high-frequency characteristic changes are calculated. and low-frequency characteristic changes ;

[0085] Then, the high-frequency feature changes were analyzed. and low-frequency characteristic changes After being transformed back to the time domain by inverse Fourier transform, the high-frequency and low-frequency variation features are obtained by sequentially passing the two-dimensional convolution with a kernel of 3, the ReLU activation function, and normalization.

[0086] Deeply fused features that have not undergone frequency domain decoupling and The temporal variation features are obtained by sequentially passing the data through a 2D convolution with a kernel of 3, the ReLU activation function, and normalization, followed by element-wise subtraction.

[0087] The high-frequency variation features, low-frequency variation features, and time-domain variation features are added and fused element by element, and then decoded using a transposed convolution with a kernel of 3 to output a variation probability map. .

[0088] S4. Using a sliding window of size d in the probability map of change. , , Multiple training samples are constructed sequentially by sliding the window with a fixed length (for example, setting the sliding window size to 6, and the first training sample being a probability map). , , , , , Then, the sliding window is moved forward one step, and the second set of training samples is the probability map of change. , , , , , Repeat the above steps; the last set of samples is a probability plot. , , , , , The last change probability map in each training sample is encoded by a visual encoder to obtain the change features as noise-added data. The remaining change probability maps in each training sample, along with the corresponding historical meteorological data and prior data of diffusion mode, are fused as a condition vector and input into the conditional diffusion prediction model for model training.

[0089] Optionally, such as Figure 7 As shown, the visual encoder extracts features from the probability map of change through local convolution and residual connections, and introduces a query-key-value mechanism to achieve dynamic weighted modeling, effectively enhancing the representation ability of key regions and patterns in the probability map of change, and improving the modeling accuracy of disease temporal evolution features. The processing includes:

[0090] The probability diagram of the input change , , , First, flatten them into one-dimensional vectors. , , , Next, feature extraction is performed four times, and then the feature changes are obtained through a multilayer perceptron. , , , ;

[0091] The feature extraction operation includes:

[0092] The one-dimensional vector , , , First, it is normalized, then subjected to a one-dimensional convolution with a kernel of 3, and then connected via residuals. , , , The sum is used as the input to the second residual network;

[0093] The second residual network generates queries, keys, and values ​​from the input through a convolution operation with a kernel of 1, weights them through an attention mechanism, and then adds them to the input through a residual connection.

[0094] Optionally, such as Figure 8 As shown, the training of the conditional diffusion prediction model is divided into a forward noise addition process and a noise prediction learning process. It simulates the disease change process, makes predictions based on multi-temporal change probability map sequences, and encodes meteorological data and prior knowledge of diffusion modes into constraint vectors to improve the model prediction accuracy and quality.

[0095] The forward noise addition process includes:

[0096] First, perform the first set of training samples, and then generate a probability plot. , , , , , Forward noise addition process:

[0097] Probability of change diagram Change features are obtained by encoding using a visual encoder. , Considered as an initial feature during forward noise addition. ;

[0098] Through a Markov process with T time steps, the initial features are... Add Gaussian noise until it is converted into a pure noise distribution. This yields a series of noisy data. ,in, , , The labels used for training the noise prediction are used to calculate the loss with the noise prediction values, and the weights of the conditional diffusion prediction model are adjusted accordingly, as expressed by the formula:

[0099]

[0100] in The cumulative noise attenuation factor is defined as follows: ; The single-step noise attenuation factor is defined as follows: ; Indicates noise scheduling parameters; Indicates Gaussian noise; Represents a normal distribution; Represents the identity matrix;

[0101] For each time step t, The noise prediction learning process, through the conditional diffusion prediction model, extracts and fuses feature information from noisy data and conditional vectors, learns the mapping relationship from feature information to noise distribution, and outputs the predicted value of the noise added at the corresponding time step, including:

[0102] The corresponding meteorological data and prior knowledge of diffusion mode are processed separately with time step t. The meteorological data is processed by conditional encoding consisting of one-dimensional 3×3 convolution and fully connected layers. The prior knowledge of diffusion mode is processed by one-hot encoding and fully connected layers. The time step t is processed by sinusoidal embedding and fully connected layers. Then the three are concatenated to obtain the constraint vector.

[0103] Transformation probability diagram , , , , After concatenation along the channel dimension, the historical vector is generated sequentially through two-dimensional convolution, flattening into a one-dimensional vector, and fully connected operations;

[0104] The constraint vector and the historical vector are concatenated to form the condition vector. ;

[0105] Add noise to the data and condition vector Simultaneously, the conditional diffusion prediction model is input, and the output is the prediction noise. ;

[0106] Then, the predicted denoised data is calculated using the following denoising formula. :

[0107]

[0108] Calculate and predict denoised data With tags The mean squared error loss between the two is optimized by backpropagation to improve the model weights;

[0109] After repeating the above steps to train all groups of training samples, the training is complete.

[0110] S5. Collect remote sensing image data of the wheat to be predicted for d-1 days (e.g., collect data for 5 days if d=6) and the corresponding meteorological data. Input the remote sensing image data into the trained multi-source data progressive fusion module and the frequency domain decoupling-based change detection module to obtain the change probability map. , , , will the , , The corresponding meteorological data and prior data on diffusion patterns are fused together as a conditional vector, and then input along with random Gaussian noise into the trained conditional diffusion prediction model to perform a denoising operation, resulting in denoised prediction data. Then, through a visual decoder, the probability map of change for the predicted day d (e.g., if d=6, then the predicted day 6) is obtained. ;

[0111] Optionally, such as Figure 9 As shown, the denoising operation in S5 progressively predicts the probability map of change on day d through a reverse denoising mechanism, performs predictions based on historical multi-temporal probability maps, and encodes meteorological data and prior knowledge of diffusion patterns into constraint vectors to improve the model's prediction accuracy and quality. Specifically, this includes:

[0112] The change probability diagram , , After being stitched along the channels, historical vectors are generated sequentially through two-dimensional convolution, flattening, and fully connected operations;

[0113] The corresponding meteorological data, diffusion mode prior data, and time steps are processed separately and then concatenated to obtain the constraint vector;

[0114] The historical vector and the constraint vector are concatenated to form the condition vector.

[0115] The conditional vector and random Gaussian noise are input together into the trained conditional diffusion prediction model, using random Gaussian noise as the initial... Perform the inverse denoising process to obtain the predicted noise. And calculated using the denoising formula ,Will The predicted denoised data is obtained by repeating the process T times as input for the next iteration. ;

[0116] The result is then processed by a visual decoder to obtain the predicted probability map of change on day d. ,like Figure 10 As shown, it specifically includes:

[0117] The input prediction denoising data After being reshaped into a two-dimensional vector, the data is decoded through six stacked feature recovery operations. By introducing a query-key-value structure and residual connections, the predicted denoised data is gradually recovered. The spatial structure information effectively enhances the ability to reconstruct fine-grained disease changes, enabling more accurate prediction of change probabilities. Then, after a two-dimensional convolution with a kernel of 3, the predicted change probability map for day d is obtained. ;

[0118] The feature recovery operation includes:

[0119] The two-dimensional vector is first subjected to a transpose convolution with a kernel of 4, and then the query, key, and value are generated by convolution operations with a kernel of 1×1. The values ​​are then weighted by an attention mechanism and finally added to the two-dimensional vector through a residual connection.

[0120] Optionally, such as Figure 11 As shown, the conditional diffusion prediction model, by integrating multiple time-series statistical features, comprehensively captures the trends, critical points, and abnormal fluctuations in the disease evolution process, and introduces conditional vectors to participate in attention calculation, enabling climate and prior knowledge factors to dynamically adjust the time-series feature recovery process. The processing includes:

[0121] The input noisy data and the constraint vector Encoding operations are performed using a stacked encoding structure, including:

[0122] First, a one-dimensional convolution with a kernel of 3 is applied, followed by average pooling, minimum pooling, and maximum pooling, and then concatenated. Next, key and value are generated through convolution operations, and the constraint vector is then... The query is generated by a one-dimensional convolution operation with a kernel of 1, weighted by an attention mechanism, and finally downsampled.

[0123] Then, a decoding operation is performed using a four-stacked decoding structure to output the prediction noise. Then, the predicted denoised data is calculated using the denoising formula. ;

[0124] The decoding operation includes:

[0125] The output of the encoded structure is first subjected to a one-dimensional convolution with a kernel of 3, then concatenated after average pooling, minimum pooling, and maximum pooling, and finally generated as keys and values ​​through convolution operations with a kernel of 1. The constraint vector is then... The query is generated by a one-dimensional convolution operation with a kernel of 1, weighted by an attention mechanism, and then upsampled.

[0126] S6. Based on the multispectral image in the wheat remote sensing image data to be predicted on day d-1 (e.g., day 5 if d=6), calculate the NDVI mask, retaining only the vegetation area, and then... The NDVI mask is multiplied pixel by pixel, and a lesion mask is generated with a preset threshold. The proportion of lesion pixels to vegetation pixels is counted as the prediction result of the severity and distribution range of wheat stripe rust on day d.

[0127] The NDVI mask calculation formula of this invention is as follows:

[0128]

[0129] NIR represents near-infrared reflectance, and RED represents red reflectance; both can be obtained from multispectral images. Extract the corresponding values. The NDVI range is between -1 and 1. An NDVI > 0.2 indicates that the pixel is vegetation. The NDVI mask value is 1 in vegetation areas and 0 in non-vegetation areas.

[0130] Next, the NDVI mask and the probability of change map are compared. Pixel-by-pixel multiplication is performed, retaining only changes in vegetation areas, followed by statistical analysis to predict wheat stripe rust disease, with a preset threshold. Set to 0.3, and compare all pixel values ​​with the threshold. Comparison, when above the threshold If a pixel is considered a lesion, the value of a suspected lesion pixel is 1, and the value of other areas is 0. A lesion mask map is generated. At the same time, the proportion of lesions is calculated, that is, the total number of lesion pixels is divided by the total number of pixels in the vegetation area to predict the severity and extent of the disease.

[0131] like Figure 12 As shown, this embodiment of the invention also provides an early disease prediction system for wheat stripe rust driven by remote sensing change information, the system comprising:

[0132] The acquisition and preprocessing module 1210 is used to acquire and preprocess historical remote sensing image data of wheat, corresponding historical meteorological data, and prior data of diffusion modes. The historical remote sensing image data includes historical natural image data and historical multispectral image data.

[0133] The first training module 1220 is used to input the historical remote sensing image data and train the multi-source data progressive fusion module to process the preprocessed natural images. and multispectral images Multi-scale features are sequentially spliced ​​together and fused with corresponding hierarchical features. A weighted strategy is used to enhance the response in changing regions, outputting a deep fused feature at a unified scale. ;

[0134] The second training module 1230 is used to train the... Input and train a frequency-domain decoupling-based change detection module, Pairing the data in chronological order, Fourier transforms are performed on each pair, followed by Gaussian high-pass and low-pass filtering to obtain high-frequency and low-frequency decoupling features. These features are then inversely transformed and fused with the time-domain feature difference results to output a probability map of changes between adjacent time points. , , ;

[0135] Module 1240 is used to construct the probability map of change using a sliding window of size d. , , Multiple training samples are constructed sequentially by sliding along a fixed length. The change feature obtained by encoding the last change probability map in each training sample through a visual encoder is used as noise-added data. The remaining change probability maps in each training sample, along with the corresponding historical meteorological data and prior data of diffusion mode, are fused as a conditional vector and input into the conditional diffusion prediction model for model training.

[0136] Prediction module 1250 is used to collect remote sensing image data of wheat to be predicted for day d-1 and corresponding meteorological data. The remote sensing image data is input into the trained multi-source data progressive fusion module and the frequency domain decoupling-based change detection module to obtain the change probability map. , , , will the , , The corresponding meteorological data and prior data on diffusion patterns are fused together as a conditional vector, and then input along with random Gaussian noise into the trained conditional diffusion prediction model to perform a denoising operation, resulting in denoised prediction data. Then, through a visual decoder, the predicted probability map of change on day d is obtained. ;

[0137] Calculation module 1260 is used to calculate an NDVI mask based on the multispectral image in the wheat remote sensing image data to be predicted on day d-1, retaining only the vegetation area. The NDVI mask is multiplied pixel by pixel, and a lesion mask is generated with a preset threshold. The proportion of lesion pixels to vegetation pixels is counted as the prediction result of the severity and distribution range of wheat stripe rust on day d.

[0138] The wheat stripe rust early disease prediction system driven by remote sensing change information provided in this embodiment of the invention has a functional structure that corresponds to the wheat stripe rust early disease prediction method driven by remote sensing change information provided in this embodiment of the invention, and will not be described again here.

[0139] Figure 13This is a schematic diagram of the structure of an electronic device 1300 provided in an embodiment of the present invention. The electronic device 1300 may vary considerably due to different configurations or performance. It may include one or more central processing units (CPUs) 1301 and one or more memories 1302. The memory 1302 stores at least one instruction, which is loaded and executed by the processor 1301 to implement the steps of the above-mentioned early disease prediction method driven by remote sensing change information of wheat stripe rust.

[0140] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to complete the aforementioned wheat stripe rust remote sensing change information-driven early disease prediction method. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device.

[0141] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0142] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for early disease prediction of wheat stripe rust driven by remote sensing change information, characterized in that, The method includes: S1. Collect and preprocess historical remote sensing image data of wheat, corresponding historical meteorological data, and prior data on diffusion modes. The historical remote sensing image data includes historical natural image data and historical multispectral image data. S2. Input the historical remote sensing image data and train the multi-source data progressive fusion module to process the preprocessed natural images. and multispectral images Multi-scale features are sequentially spliced ​​together and fused with corresponding hierarchical features. A weighted strategy is used to enhance the response in changing regions, outputting a deep fused feature at a unified scale. ; S3, the above Input and train a frequency-domain decoupling-based change detection module, Pairing the data in chronological order, Fourier transforms are performed on each pair, followed by Gaussian high-pass and low-pass filtering to obtain high-frequency and low-frequency decoupling features. These features are then inversely transformed and fused with the time-domain feature difference results to output a probability map of changes between adjacent time points. , , ; S4. Using a sliding window of size d in the probability map of change. , , Multiple training samples are constructed sequentially by sliding along a fixed length. The change feature obtained by encoding the last change probability map in each training sample through a visual encoder is used as noise-added data. The remaining change probability maps in each training sample, along with the corresponding historical meteorological data and prior data of diffusion mode, are fused as a conditional vector and input into the conditional diffusion prediction model for model training. S5. Collect remote sensing image data of wheat to be predicted for day d-1 and corresponding meteorological data. Input the remote sensing image data into the trained multi-source data progressive fusion module and the frequency domain decoupling-based change detection module to obtain the change probability map. , , , will the , , The corresponding meteorological data and prior data on diffusion patterns are fused together as a conditional vector, and then input along with random Gaussian noise into the trained conditional diffusion prediction model to perform a denoising operation, resulting in denoised prediction data. Then, through a visual decoder, the predicted probability map of change on day d is obtained. ; S6. Based on the multispectral image in the wheat remote sensing image data to be predicted on day d-1, calculate the NDVI mask, retaining only the vegetation area, and then... The NDVI mask is multiplied pixel by pixel, and a lesion mask is generated with a preset threshold. The proportion of lesion pixels to vegetation pixels is counted as the prediction result of the severity and distribution range of wheat stripe rust on day d.

2. The method according to claim 1, characterized in that, The multi-source data progressive fusion module gradually fuses perceptual information from natural images and multispectral images at different levels through multi-level feature extraction and high-order feature enhancement techniques. The processing includes: natural images In the multi-level feature extraction submodule, the input first passes through a 3D convolutional layer with a kernel of 3×3×3, and then passes through three feature extraction blocks in sequence, gradually extracting the first-level features, second-level features, and third-level features of the natural image from the shallow layer to the deep layer. Multispectral images at the same acquisition time point The same operation was performed to obtain the first-level features, second-level features, and third-level features of the multispectral image; Next, a progressive fusion of multi-source features is performed according to an adaptive strategy, including: First, the first-level, second-level, and third-level features of the natural image and the multispectral image are concatenated to form first-level shallow fusion features, second-level shallow fusion features, and third-level shallow fusion features; Next, the first-level shallow fusion feature is used as a low-order feature of the shallow layer, and the second-level shallow fusion feature is used as a high-order feature of the deep layer. They are then combined to perform high-order feature enhancement to obtain hidden shallow fusion features. The second-level shallow fusion feature is used as a low-order feature of the shallow layer, and the third-level shallow fusion feature is used as a high-order feature of the deep layer. They are then combined to perform high-order feature enhancement to obtain hidden deep fusion features. Finally, the hidden shallow-layer fusion features are used as low-order features of the shallow layer, and the hidden deep-layer fusion features are used as high-order features of the deep layer. These are then combined to perform high-order feature enhancement, ultimately yielding the deep fusion features. .

3. The method according to claim 2, characterized in that, The feature extraction block, based on multi-scale convolutional fusion, introduces channel attention and spatial attention mechanisms. Channel attention focuses on the dependencies between channels, enhancing the responsiveness to key channel features by weighting the importance of each channel. Spatial attention focuses on the spatial distribution of the feature map, enhancing the perception of lesion areas and abnormal physiological locations. The synergistic effect of these two mechanisms enhances the information of channel selection and spatial focusing, improving the expression quality of feature extraction. The processing includes: The input shallow layer's previous-level features are first processed through three branches: each branch uses convolutional layers of different sizes, each convolutional layer is followed by a ReLU activation function, and then layer normalization is applied to stabilize the training process. These three branches generate three feature maps, which are then processed using... , , express; Then, the , , Adaptive fusion is performed through a channel attention mechanism to obtain fused features. By performing adaptive fusion through a spatial attention mechanism, fused features are obtained. The fusion feature and fusion features Pixel-wise multiplication is performed, and features are integrated through a 1×1 convolution operation to output the next level of deep features; The channel attention mechanism includes: The , , Global information of the feature map is obtained through global average pooling and global max pooling. The results of global average pooling and global max pooling are then fed into a multilayer perceptron and summed. Finally, the weight of each channel is calculated using the sigmoid activation function. , The calculated weights Corresponding Multiply to obtain the weighted features Finally, all weighted features , , Add them together to obtain the final fusion feature. The formula is as follows: ; ； ； in represents the Sigmoid activation function, MLP represents the multilayer perceptron, GAP represents global average pooling, and GMP represents global max pooling. The spatial attention mechanism includes: right Perform average pooling and max pooling along the channel dimension respectively to obtain two spatial feature maps. and This reflects the average and maximum responses at each location. These two feature maps are then concatenated along the channel dimension, and a 7×7 convolutional layer is used to extract fused features. The spatial attention weights are then obtained after passing through a sigmoid activation function. Then spatial attention weights With features Multiply to obtain the weighted features Finally, all weighted features , , Adding them together yields the final fusion feature. The formula is as follows: ； ； ； in This represents the Sigmoid function, where C represents the number of channels. This represents averaging the feature map along the channel dimension. This represents performing the maximum operation on the feature map along the channel dimension. This represents a two-dimensional convolution operation with a kernel size of 7×7.

4. The method according to claim 2, characterized in that, The higher-order feature enhancement achieves progressive fusion of multi-level features through spatial scale alignment and dual-path convolution extraction. Simultaneously, it introduces average pooling and covariance pooling to jointly model global and local statistical information, enhancing the sensitivity of the fused features to spatial variations and structural differences, and improving the stability and discriminative ability of feature representation. The processing includes: The high-order features of the deep layers are restored to the same spatial resolution as the low-order features of the shallow layers through upsampling operations, so that they can be fused at the same scale; Next, the high-order and low-order features are respectively entered into a dual-parallel feature extraction operation: one path directly extracts spatial features through 3×3 convolution, while the other path generates weights through 3×3 convolution and the Sigmoid activation function for spatial weighted modulation. Next, the convolutional output of each path is multiplied pixel by pixel with the corresponding weight to achieve local weighted fusion. The results of the two paths are then added together to generate fused features. Then, the fused features are fed into the feature recalibration module. The feature recalibration module first adjusts the channel dimension through 1×1 convolution to extract a compact global feature map. Then, the global feature map is fed into the average pooling and max pooling branches respectively: the average pooling branch captures global context information, while the max pooling branch highlights salient regional feature information or edge feature information. After the two types of information are added, the local spatial structure is fused through a 3-convolutional layer, and finally, the higher-order enhanced features are output.

5. The method according to claim 1, characterized in that, The frequency-domain decoupling-based change detection module separates high-frequency and low-frequency change components through Fourier transform and filters. By analyzing the differences in time, high-frequency, and low-frequency information of the dual-phase images, it fully explores the subtle change signals caused by diseases in wheat during the latent period. The processing includes: The deep fusion features, paired in time sequence, are sequentially passed through the same shared frequency domain decoupled encoder, and then through a transposed convolutional decoder with a kernel of 3, outputting a probability map of changes between adjacent time steps. , , The change probability map is an image of the same size as the input image, where each pixel takes a value in the range [0, 1], representing the probability that the position of each pixel changes between adjacent time points in the image; Among them, for adjacent deep fusion features and The frequency domain decoupling encoder first maps the components to the frequency domain using Fourier transform, and then decomposes them into high-frequency decoupling features using Gaussian high-pass filtering and Gaussian low-pass filtering. , and low-frequency decoupling characteristics , ; Then, the high-frequency characteristic changes are calculated. and low-frequency characteristic changes ; Then, the high-frequency feature changes were analyzed. and low-frequency characteristic changes After being transformed back to the time domain by inverse Fourier transform, the high-frequency and low-frequency variation features are obtained by sequentially passing the two-dimensional convolution with a kernel of 3, the ReLU activation function, and normalization. Deeply fused features that have not undergone frequency domain decoupling and The temporal variation features are obtained by sequentially passing the data through a 2D convolution with a kernel of 3, the ReLU activation function, and normalization, followed by element-wise subtraction. The high-frequency variation features, low-frequency variation features, and time-domain variation features are added and fused element by element, and then decoded using a transposed convolution with a kernel of 3 to output a variation probability map. .

6. The method according to claim 1, characterized in that, The visual encoder extracts features from the probability map of change through local convolution and residual connections, and introduces a query-key-value mechanism to achieve dynamic weighted modeling. This effectively enhances the representation ability of key regions and patterns in the probability map of change, and improves the modeling accuracy of disease temporal evolution features. The processing includes: The probability diagram of the input change , , , First, flatten them into one-dimensional vectors. , , , Next, feature extraction is performed four times, and then the feature changes are obtained through a multilayer perceptron. , , , ; The feature extraction operation includes: The one-dimensional vector , , , First, it is normalized, then subjected to a one-dimensional convolution with a kernel of 3, and then connected via residuals. , , , The sum is used as the input to the second residual network; The second residual network generates queries, keys, and values ​​from the input through a convolution operation with a kernel of 1, weights them through an attention mechanism, and then adds them to the input through a residual connection.

7. The method according to claim 1, characterized in that, The training of the conditional diffusion prediction model is divided into a forward noise addition process and a noise prediction learning process. It simulates the disease change process, makes predictions based on multi-temporal change probability map sequences, and encodes meteorological data and prior knowledge of diffusion modes into constraint vectors to improve the model's prediction accuracy and quality. The forward noise addition process includes: First, perform the first set of training samples, and then generate a probability plot. , , , , , Forward noise addition process: Probability of change diagram Change features are obtained by encoding using a visual encoder. , Considered as an initial feature during forward noise addition. ; Through a Markov process with T time steps, the initial features are... Add Gaussian noise until it is converted into a pure noise distribution. This yields a series of noisy data. ,in, , , The labels used for training the noise prediction are used to calculate the loss with the noise prediction values, and the weights of the conditional diffusion prediction model are adjusted accordingly, as expressed by the formula: ； in The cumulative noise attenuation factor is defined as follows: ; The single-step noise attenuation factor is defined as follows: ; Indicates noise scheduling parameters; Indicates Gaussian noise; Represents a normal distribution; Represents the identity matrix; For each time step t, The noise prediction learning process, through the conditional diffusion prediction model, extracts and fuses feature information from noisy data and conditional vectors, learns the mapping relationship from feature information to noise distribution, and outputs the predicted value of the noise added at the corresponding time step, including: The corresponding meteorological data and prior knowledge of diffusion mode are processed separately with time step t. The meteorological data is processed by conditional encoding consisting of one-dimensional 3×3 convolution and fully connected layers. The prior knowledge of diffusion mode is processed by one-hot encoding and fully connected layers. The time step t is processed by sinusoidal embedding and fully connected layers. Then the three are concatenated to obtain the constraint vector. Transformation probability diagram , , , , After concatenation along the channel dimension, the historical vector is generated sequentially through two-dimensional convolution, flattening into a one-dimensional vector, and fully connected operations; The constraint vector and the historical vector are concatenated to form the condition vector. ; Add noise to the data and condition vector Simultaneously, the conditional diffusion prediction model is input, and the output is the prediction noise. ; Then, the predicted denoised data is calculated using the following denoising formula. : ； Calculate and predict denoised data With tags The mean squared error loss between the two is optimized by backpropagation to improve the model weights; After repeating the above steps to train all groups of training samples, the training is complete.

8. The method according to claim 7, characterized in that, The denoising operation in S5 involves progressively predicting the probability map of change on day d using a reverse denoising mechanism. Prediction is based on historical multi-temporal probability maps, and prior knowledge of meteorological data and diffusion patterns is encoded into constraint vectors to improve model prediction accuracy and quality. Specifically, this includes: The change probability diagram , , After being stitched along the channels, historical vectors are generated sequentially through two-dimensional convolution, flattening, and fully connected operations; The corresponding meteorological data, diffusion mode prior data, and time steps are processed separately and then concatenated to obtain the constraint vector; The historical vector and the constraint vector are concatenated to form the condition vector. The conditional vector and random Gaussian noise are input together into the trained conditional diffusion prediction model, using random Gaussian noise as the initial... Perform the inverse denoising process to obtain the predicted noise. And calculated using the denoising formula ,Will The predicted denoised data is obtained by repeating the process T times as input for the next iteration. ; The result is then processed by a visual decoder to obtain the predicted probability map of change on day d. Specifically, it includes: The input prediction denoising data After being reshaped into a two-dimensional vector, the data is decoded through six stacked feature recovery operations. By introducing a query-key-value structure and residual connections, the predicted denoised data is gradually recovered. The spatial structure information effectively enhances the ability to reconstruct fine-grained disease changes, enabling more accurate prediction of change probabilities. Then, after a two-dimensional convolution with a kernel of 3, the predicted change probability map for day d is obtained. ; The feature recovery operation includes: The two-dimensional vector is first subjected to a transpose convolution with a kernel of 4, and then the query, key, and value are generated by convolution operations with a kernel of 1×1. The values ​​are then weighted by an attention mechanism and finally added to the two-dimensional vector through a residual connection.

9. The method according to claim 7, characterized in that, The conditional diffusion prediction model, by integrating multiple time-series statistical features, comprehensively captures trends, critical points, and abnormal fluctuations in the disease evolution process. It also introduces conditional vectors to participate in attention calculations, enabling climate and prior knowledge factors to dynamically adjust the time-series feature recovery process. The processing includes: The input noisy data and the constraint vector Encoding operations are performed using a stacked encoding structure, including: First, a one-dimensional convolution with a kernel of 3 is applied, followed by average pooling, minimum pooling, and maximum pooling, and then concatenated. Next, key and value are generated through convolution operations, and the constraint vector is then... The query is generated by a one-dimensional convolution operation with a kernel of 1, weighted by an attention mechanism, and finally downsampled. Then, a decoding operation is performed using a four-stacked decoding structure to output the prediction noise. Then, the predicted denoised data is calculated using the denoising formula. ; The decoding operation includes: The output of the encoded structure is first subjected to a one-dimensional convolution with a kernel of 3, then concatenated after average pooling, minimum pooling, and maximum pooling, and finally generated as keys and values ​​through convolution operations with a kernel of 1. The constraint vector is then... The query is generated by a one-dimensional convolution operation with a kernel of 1, weighted by an attention mechanism, and then upsampled.

10. A wheat stripe rust early disease prediction system driven by remote sensing change information, characterized in that, The system includes: The acquisition and preprocessing module is used to acquire and preprocess historical remote sensing image data of wheat, corresponding historical meteorological data, and prior data on diffusion modes. The historical remote sensing image data includes historical natural image data and historical multispectral image data. The first training module is used to input the historical remote sensing image data and train the multi-source data progressive fusion module, which then processes the pre-processed natural image data. and multispectral images Multi-scale features are sequentially spliced ​​together and fused with corresponding hierarchical features. A weighted strategy is used to enhance the response in changing regions, outputting a deep fused feature at a unified scale. ; The second training module is used to train the... Input and train a frequency-domain decoupling-based change detection module, Pairing the data in chronological order, Fourier transforms are performed on each pair, followed by Gaussian high-pass and low-pass filtering to obtain high-frequency and low-frequency decoupling features. These features are then inversely transformed and fused with the time-domain feature difference results to output a probability map of changes between adjacent time points. , , ; Build a module for using a sliding window of size d in the probability map of change. , , Multiple training samples are constructed sequentially by sliding along a fixed length. The change feature obtained by encoding the last change probability map in each training sample through a visual encoder is used as noise-added data. The remaining change probability maps in each training sample, along with the corresponding historical meteorological data and prior data of diffusion mode, are fused as a conditional vector and input into the conditional diffusion prediction model for model training. The prediction module collects remote sensing image data of wheat to be predicted for d-1 days and corresponding meteorological data. It then inputs the remote sensing image data into a trained multi-source data progressive fusion module and a frequency-domain decoupling-based change detection module to obtain a change probability map. , , , will the , , The corresponding meteorological data and prior data on diffusion patterns are fused together as a conditional vector, and then input along with random Gaussian noise into the trained conditional diffusion prediction model to perform a denoising operation, resulting in denoised prediction data. Then, through a visual decoder, the predicted probability map of change on day d is obtained. ; The calculation module is used to calculate an NDVI mask based on the multispectral image in the wheat remote sensing image data to be predicted on day d-1, retaining only the vegetation area. The NDVI mask is multiplied pixel by pixel, and a lesion mask is generated with a preset threshold. The proportion of lesion pixels to vegetation pixels is counted as the prediction result of the severity and distribution range of wheat stripe rust on day d.

Images