Winter sowing and spring wheat field dynamic diagnosis system based on deep learning

Abstract

The invention relates to a winter sowing and spring wheat field dynamic diagnosis system based on deep learning, in particular to the field of crop cultivation, and can realize early accurate identification and dynamic monitoring of winter sowing and spring wheat field diseases. Through multi-source data fusion and deep learning feature extraction, the detection sensitivity of tiny lesions is effectively improved; by means of an online learning mechanism, the system can adapt to crop growth and environment change, and the diagnosis performance is continuously optimized; the finally generated disease distribution thermodynamic diagram and the variable operation prescription diagram provide visual decision support for precise agricultural management, a closed loop from perception to decision is realized, and the disease prevention and control efficiency and the intelligent level of field management are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of crop cultivation, more particularly, the present application relates to a winter-sowing spring wheat field dynamic diagnosis system based on deep learning. BACKGROUND

[0002] As an important food crop in northern China, winter-sowing spring wheat has a long growth cycle and strong environmental adaptability, but it still faces many challenges in actual field management, especially in the critical mid-growth stage of wheat, such as the jointing stage to the heading stage, the plant density in the field increases significantly, the leaves overlap and shade each other, forming a complex canopy structure. During this period, the internal microenvironment of the crop population changes dramatically, and the ventilation and light transmission conditions are limited. Combined with the variable external climate conditions, such as the high temperature and humidity weather often seen in spring, it is easy to induce the initial occurrence and spread of typical diseases such as wheat rust and powdery mildew. At the early stage of disease occurrence, the disease spot area is small, and the color and texture characteristics are very similar to the natural changes of healthy leaves. In addition, random fluctuations in factors such as field light intensity and shooting angle further increase the difficulty of naked-eye observation and traditional equipment recognition. The above-mentioned complex internal structure and external environment jointly constitute the real scene of early disease identification, and put forward very high requirements for the accuracy and adaptability of the diagnosis technology.

[0003] At present, the technical methods applied to wheat disease identification mainly rely on traditional digital image processing and shallow machine learning models. Traditional image processing technology usually segments images based on pre-set thresholds, or extracts and classifies disease spot areas using manually designed features such as color and morphology. This kind of method relies heavily on expert prior knowledge, and its feature extraction rules are fixed, making it difficult to adapt to complex situations such as changes in actual field illumination, leaf shading, and disease spot morphology diversity. Although shallow machine learning models based on support vector machines have improved classification ability to some extent, their feature expression ability and model generalization performance are still limited, especially for the subtle and atypical visual features presented at the early stage of disease, their recognition sensitivity and robustness are obviously insufficient. In recent years, although deep learning technology has made significant progress in general image recognition, its direct application in agricultural field scenes still faces many bottlenecks. On the one hand, existing convolutional neural network models lack the ability to perceive subtle features in images, and the multi-scale feature fusion mechanism is imperfect, resulting in a high rate of missed detection of early disease spots. On the other hand, high-quality and large-scale field disease image datasets suitable for model training are scarce, and existing models are difficult to effectively handle image noise and intra-class differences introduced by environmental interference. Therefore, developing a system that can adapt to complex field environments and achieve accurate and real-time diagnosis of early wheat diseases has become an urgent need in the field of agricultural intelligent management. SUMMARY The present application aims at the technical problems existing in the prior art, and provides a winter-sowing spring wheat field dynamic diagnosis system based on deep learning, which solves the problems in the background art through a data synchronization and preprocessing module, a feature enhancement extraction module, an online learning and updating module, and a decision support and visualization module.

[0004] The technical scheme for solving the above technical problems is as follows: specifically comprising a data synchronization and preprocessing module, a feature enhancement extraction module, an online learning and updating module, and a decision support and visualization module, wherein; The data synchronization and preprocessing module is used for receiving visible light images and multispectral images collected from a UAV platform and environment data from a field sensor network, matching the environment data with the visible light images and the multispectral images through a space-time alignment operation based on geographic coordinates and time stamps, performing illumination normalization processing on the visible light images and the multispectral images, and outputting a standardized multi-modal data cube; The feature enhancement extraction module is connected to the data synchronization and preprocessing module, is used for receiving the multi-modal data cube, and performs space-time feature extraction by using a deep learning model based on a Transformer architecture, the deep learning model integrates a contrast learning strategy to optimize feature representation, and finally outputs a deep feature map containing subtle lesion features and a preliminary disease diagnosis result by the deep learning model; The online learning and updating module is connected to the feature enhancement extraction module, is used for receiving the deep feature map and the preliminary disease diagnosis result, receives visible light images, multispectral images, and environment data continuously collected by the UAV platform and the field sensor network through a streaming data processing platform, selects low-confidence samples for labeling by using a sample screening mechanism based on uncertainty estimation, and periodically performs incremental fine-tuning on the deep learning model in the feature enhancement extraction module; The decision support and visualization module is connected to the feature enhancement extraction module, is used for receiving the deep feature map, converts the deep feature map into a disease probability distribution map by using a classification head attached to the deep learning model in the feature enhancement extraction module, and generates a disease distribution heat map and a variable operation prescription map by combining a geographic information system for spatial mapping; In a preferred embodiment, the space-time alignment operation based on geographic coordinates and time stamps in the data synchronization and preprocessing module specifically operates as follows: Firstly, parallel receiving image data stream from unmanned aerial vehicle platform and environment data stream from field sensor network; the image data stream includes visible light image and multispectral image, and the environment data stream includes temperature data and humidity data; each of the visible light image, the multispectral image and the environment data is attached with geographical coordinates and acquisition time stamp provided by a high-precision positioning module; a unified reference time point and a predefined time window length are set, the acquisition time stamp carried by each of the visible light image, the multispectral image and the environment data is compared with the reference time point, and the visible light image, the multispectral image and the environment data whose acquisition time stamp falls within the closed interval with the reference time point as the center and the time window length as the span are screened out, and the screened visible light image, multispectral image and environment data are divided into the same batch processing unit to complete the initial time window alignment; Then, for the visible light image, the multispectral image and the environment data in the same batch processing unit, the records in the form of discrete geographical points in the environment data are processed by a spatial interpolation algorithm based on the geographical coordinates to generate a continuously distributed temperature field raster layer and a humidity field raster layer, and the temperature field raster layer and the humidity field raster layer together constitute an environment data raster layer, and the spatial resolution thereof is the same as that of the visible light image; At the same time, the unmanned aerial vehicle positioning and attitude system data are used to map each pixel point of the visible light image and the multispectral image in the same batch processing unit to the real geographical coordinates to complete orthographic correction and generate an image raster layer with geographical reference; Finally, the environment data raster layer and the image raster layer with geographical reference are superimposed to realize spatial matching at the pixel level.

[0005] In a preferred embodiment, the correction method based on atmospheric physical model is used to perform illumination normalization on each waveband of the visible light image and each waveband of the multispectral image, specifically: For each pixel of each waveband of the visible light image in the image raster layer with geographical reference, the original digital quantization value of the pixel in the waveband is taken as the original observation value, and the normalization process is as follows: A preset atmospheric path radiation value is deducted from the original observation value to obtain a first intermediate result; a product of the direct light intensity on the ground surface and the cosine value of the solar zenith angle determined by the image acquisition time and the geographical position of the pixel is calculated, and the product is added to the diffuse light intensity on the ground surface to obtain a second intermediate result; the first intermediate result is divided by the second intermediate result, and then multiplied by the cosine value of the ground surface incident angle determined by the ground surface slope and the solar azimuth to finally obtain the normalized value of the pixel in the waveband; The above normalization process is independently performed on each waveband of the multispectral image; After the above normalization processing, the output is the normalized visible light image data and the multispectral image data.

[0006] In a preferred embodiment, the construction process of the standardized multi-modal data cube is as follows: The normalized visible light image data and multispectral image data generated for the same geographical area, and the environment data raster layer obtained through the spatio-temporal alignment operation, are stacked and integrated according to a preset band order; wherein the normalized visible light image data includes red band, green band and blue band data, the multispectral image data includes multiple band data including near-infrared band and red edge band, and the environment data raster layer includes temperature field raster layer and humidity field raster layer. During stacking, the red band data, green band data and blue band data of the visible light image are first arranged as the first three dimensions of the data cube, then the above-mentioned multiple band data of the multispectral image are sequentially arranged, and finally the temperature field raster layer and the humidity field raster layer are arranged as the last dimensions of the multi-modal data cube. After integration, a three-dimensional multi-modal data cube is formed.

[0007] In a preferred embodiment, the deep learning model based on the Transformer architecture in the feature enhancement extraction module includes a convolutional neural network backbone for preliminary feature extraction, a fully connected network for environment data encoding, and an improved Transformer encoder. Secondly, the specific operation of the deep learning model based on the Transformer architecture for spatio-temporal feature extraction is as follows: Firstly, the received multi-modal data cube is processed by branching to separate the image data part composed of visible light image data and multispectral image data from the environment data part composed of environment data; Then, the image data part is preliminarily spatially feature-extracted by the convolutional neural network backbone to convert it into a spatial feature map, and the spatial feature map is flattened into a sequence of image feature tokens along the spatial dimension; At the same time, the environment data part is encoded by the fully connected network to generate an environment feature sequence. Then, the image feature token sequence and the environment feature sequence are jointly input into an improved Transformer encoder, the Transformer encoder comprising a cross-attention mechanism in each layer, wherein the image feature token sequence is taken as a query vector, the environment feature sequence is taken as a key vector and a value vector, the dot product of the query vector and the key vector is calculated, the dot product result is divided by the square root of the dimension of the key vector to perform scaling, then a Softmax normalization function is applied to the scaled result to obtain an attention weight, and the attention weight is weighted and summed with the value vector to generate an image feature representation modulated by environment information. Finally, the feature sequence output by the Transformer encoder is mapped to a two-dimensional spatial structure to form a deep feature map, and on this basis, the deep feature map is input into a lightweight classification head, the classification head being a fully connected layer or a one-by-one convolution layer, and the probability of each spatial position corresponding to different disease categories output by the classification head is taken as a preliminary disease diagnosis result.

[0008] In a preferred embodiment, the integrated contrast learning strategy is a specific process for optimizing the feature representation of a deep learning model based on a Transformer architecture, and the following operations are performed in the training phase of the deep learning model: A1, training sample pair construction: for a multi-modal data cube as an anchor sample, generate its positive sample pair through data augmentation techniques including slight rotation or color jittering, and select a multi-modal data cube confirming health from different fields or different growth periods as a negative sample; A2, environment-aware contrast loss function calculation and optimization: the optimization process is realized by minimizing an environment-aware contrast loss function, the value of the environment-aware contrast loss function being equal to the value of a standard contrast loss term plus the product of the value of a balance hyperparameter and the value of an environment consistency constraint term.

[0009] In a preferred embodiment, in the online learning and updating module, the specific operation of selecting low-confidence samples for annotation using a sample screening mechanism based on uncertainty estimation is as follows: First, for the new visible light images, multispectral images and environment data continuously received through the streaming data processing platform, the deep learning model in the feature enhancement extraction module is called for forward inference to obtain a preliminary disease diagnosis result for each sample, the preliminary disease diagnosis result representing the prediction confidence of the deep learning model for each disease category in the form of a prediction probability distribution; Then, the Monte Carlo Dropout method is used for uncertainty estimation, specifically: the Dropout layer activation is maintained during the inference process of the deep learning model, and multiple forward propagations are performed on the same sample to obtain multiple different prediction probability distributions. Then, two uncertainty indicators are calculated based on the plurality of different prediction probability distributions: The first indicator is prediction entropy, which is calculated as follows: First, the average probability distribution of the plurality of prediction probability distributions is calculated, that is, for each disease category, the arithmetic mean of the probability values of the category appearing in all prediction probability distributions is calculated to obtain the average probability of the category, thereby forming an average probability distribution; Then, the prediction entropy is calculated, which is equal to multiplying the average probability of each disease category by the inverse of the natural logarithm of the average probability of the category, and then summing the calculation results of all disease categories; The second indicator is cognitive uncertainty, which is calculated as follows: First, the average probability distribution of the plurality of prediction probability distributions is calculated, that is, for each disease category, the arithmetic mean of the probability values of the category appearing in all prediction probability distributions is calculated to obtain the average probability of the category, thereby forming an average probability distribution; Then, for each prediction probability distribution, the KL divergence between the prediction probability distribution and the average probability distribution is calculated, and the calculation process of the KL divergence is as follows: For each disease category, multiply the probability of the category in the prediction probability distribution by the natural logarithm of the ratio of the probability to the average probability of the category, and then sum the calculation results of all disease categories; Finally, the arithmetic mean of the KL divergences of all prediction probability distributions is calculated to obtain the cognitive uncertainty; Subsequently, the two uncertainty indicators, i.e., the prediction entropy and the cognitive uncertainty, are normalized respectively, and then weighted and summed according to the first weight and the second weight to obtain a comprehensive uncertainty score; Finally, a comprehensive uncertainty threshold is set, samples with a comprehensive uncertainty score higher than the comprehensive uncertainty threshold are marked as low-confidence samples, and are added to the queue of samples to be labeled, waiting for manual labeling by experts; after manual labeling, these low-confidence samples are marked as newly labeled samples.

[0010] In a preferred embodiment, the specific operation of periodically incrementally fine-tuning the deep learning model in the feature enhancement extraction module is as follows: First, a pre-set number of newly labeled samples accumulated from the queue of samples to be labeled are taken out to form an incremental learning training set for the current batch; Then, the incremental fine-tuning is performed by using an elastic weight consolidation algorithm, wherein the elastic weight consolidation algorithm is implemented by minimizing a total loss function, which is composed of two parts: a first part is a standard cross-entropy loss on the newly labeled samples, and a second part is a regularization term; the regularization term is calculated by: a preset regularization strength hyperparameter multiplied by a summation formula, which calculates, for each parameter of the deep learning model, a product of an importance weight of the parameter and a square of a difference between a current value of the parameter and an old value of the parameter before the incremental fine-tuning, and then sums all the product results of the parameters; During the training process, the parameters of the deep learning model are updated by using an optimization algorithm to minimize the total loss function; and after the incremental fine-tuning is completed, the parameters of the deep learning model in the feature enhancement extraction module are replaced by the updated parameters.

[0011] In a preferred embodiment, in the decision support and visualization module, the deep features are converted into a disease probability distribution map by using a classification head attached to the deep learning model in the feature enhancement extraction module, and the disease distribution heat map is generated by combining the spatial mapping of the geographic information system, and the specific operation is as follows: First, the received deep features are input into the classification head attached to the deep learning model in the feature enhancement extraction module, the classification head is a fully connected layer or a one-by-one convolution layer, and a Softmax activation function is connected after the classification head, and by linear transformation and Softmax activation function processing of the classification head, the high-dimensional feature vector at each spatial position in the deep features is converted into an initial disease probability distribution, which represents the initial probability of each spatial position belonging to each disease category; Then, the initial disease probability distribution is optimized for spatial consistency by using a conditional random field optimization algorithm, and the conditional random field optimization algorithm is implemented by constructing an energy function and minimizing the energy function; the energy function is composed of the sum of unary potential functions and the sum of binary potential functions; The value of the unary potential function is determined by the negative natural logarithm of the initial probability of the corresponding spatial position in the initial disease probability distribution; The binary potential function is composed of a label compatibility function multiplied by a weighted sum, which includes an appearance term multiplied by an appearance coefficient and a smoothing term multiplied by a smoothing coefficient; The appearance term is calculated by an exponential function of the reciprocal of the square of the difference between the color vectors of the spatial positions divided by the square of the color bandwidth parameter, multiplied by an exponential function of the reciprocal of the square of the difference between the position coordinates of the spatial positions divided by the square of the position bandwidth parameter; The smoothing term is calculated by an exponential function of the reciprocal of the square of the difference between the position coordinates of the spatial positions divided by the square of the smoothing bandwidth parameter; The label compatibility function takes a value of one when the disease class labels of adjacent spatial positions are different, and a value of zero when they are the same; After minimizing the energy function by an iterative optimization algorithm, an optimized disease probability distribution map with enhanced spatial consistency is obtained, and the disease occurrence probability value corresponding to each spatial position in the optimized disease probability distribution map is defined as the current disease probability value; Finally, the geographic position information of each spatial position in the optimized disease probability distribution map is registered with the geographic coordinates in the geographic information system, the probability value is mapped to the corresponding geographic spatial position, and a color rendering of the probability value is performed using a gradient color system from blue to red to generate the disease distribution heat map, where blue represents low probability and red represents high probability.

[0012] In a preferred embodiment, the specific operation of generating the variable operation prescription map is as follows: First, based on the optimized disease probability distribution map, the area with a disease occurrence probability higher than a preset high probability threshold is identified as a disease center area; Then, based on the historical disease distribution heat map data stored in time series, the maximum likelihood estimation algorithm is used to fit the key parameters of the spatial diffusion dynamics model, including the diffusion coefficient and the disease growth rate; Next, using the fitted spatial diffusion dynamics model, the range and intensity of disease spread from the disease center area to the outside in the future time window are predicted, a predicted disease risk distribution map is generated, and the disease occurrence risk probability value corresponding to each geographic position in the predicted disease risk distribution map is defined as the predicted disease risk value; Finally, a multi-objective optimization decision function is constructed to generate the variable operation prescription map, which specifically includes: F1, for each geographic position, the current disease probability value of the position is obtained from the optimized disease probability distribution map, and the predicted disease risk value of the position is obtained from the predicted disease risk distribution map; compare the current disease probability value and the predicted disease risk value, and select the larger one as the comprehensive risk indicator corresponding to the geographic position; F2, compare the comprehensive risk indicator with a preset economic threshold: when the comprehensive risk indicator is less than or equal to the economic threshold, it is determined that the geographic position does not need to be sprayed, and the recommended spraying amount is set to zero; when the comprehensive risk indicator is greater than the economic threshold, enter the spraying amount calculation process; In the drug application amount calculation process, first, the value of the sum of the economic threshold and a preset risk buffer threshold is calculated, the value is multiplied by a preset unit risk increment drug application coefficient to obtain a preliminary drug application amount calculation result, and the preliminary drug application amount calculation result cannot exceed a preset maximum drug application amount limit, if the comprehensive risk index does not exceed the sum of the economic threshold and the risk buffer threshold, it is still determined that the geographical position does not need to be applied, and the recommended drug application amount is set to zero; F3、After the calculation of all positions is completed, the recommended value is discretized and classified according to a preset drug application amount grade, and different colors or filling patterns are used for visual rendering in a geographic information system to generate a variable operation prescription map, which clearly identifies no need to apply, early warning monitoring area, light application area, moderate application area and heavy application area.

[0013] The beneficial effects of the present application are: early and accurate identification and dynamic monitoring of field diseases of winter-sown spring wheat can be achieved. Through multi-source data fusion and deep learning feature extraction, the detection sensitivity of subtle lesions is effectively improved; with the help of online learning mechanism, the system can adapt to crop growth and environmental changes, and continuously optimize the diagnosis performance; the finally generated disease distribution heat map and variable operation prescription map provide visual decision support for precision agriculture management, realize the closed loop from perception to decision, and significantly improve the disease prevention and control efficiency and the intelligent level of field management. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 The method flowchart of the present application is shown in the figure; Figure 2 The system structure block diagram of the present application is shown in the figure. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0016] In the description of the present application, the terms "first", "second" are used only for description purposes, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.

[0017] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0018] Example 1 This embodiment provides, for example Figures 1-2 The present invention discloses a field dynamic diagnostic system for winter-sown spring wheat based on deep learning, which specifically includes: a data synchronization and preprocessing module, a feature enhancement and extraction module, an online learning and updating module, and a decision support and visualization module, wherein; Data synchronization and preprocessing module: It is used to receive visible light images and multispectral images collected from UAV platform, as well as environmental data from field sensor network. Through spatiotemporal alignment operation based on geographic coordinates and timestamps, it matches environmental data with visible light images and multispectral images, performs illumination normalization processing on visible light images and multispectral images, and outputs standardized multimodal data cubes. Feature enhancement and extraction module: Connected to the data synchronization and preprocessing module, it receives multimodal data cubes and uses a deep learning model based on the Transformer architecture to extract spatiotemporal features. This deep learning model integrates a contrastive learning strategy to optimize feature representation. Finally, the deep learning model outputs a deep feature map containing subtle lesion features and preliminary disease diagnosis results. Online learning and update module: Connected to the feature enhancement and extraction module, it is used to receive deep feature maps and preliminary disease diagnosis results, and receive visible light images, multispectral images and environmental data continuously collected by UAV platform and field sensor network through streaming data processing platform. It adopts a sample screening mechanism based on uncertainty estimation to select low confidence samples for labeling, and performs incremental fine-tuning of the deep learning model in the feature enhancement and extraction module on a regular basis. Decision Support and Visualization Module: Connected to the Feature Enhancement and Extraction Module, it receives deep feature maps and uses the classification head attached to the deep learning model in the Feature Enhancement and Extraction Module to convert the deep feature maps into disease probability distribution maps. It then combines the geographic information system to perform spatial mapping and generate disease distribution heat maps and variable operation prescription maps.

[0019] In this embodiment, it is particularly necessary to explain that the spatio-temporal alignment operation based on geographic coordinates and time stamps in the data synchronization and preprocessing module is specifically as follows: First, the image data stream from the unmanned aerial vehicle platform and the environment data stream from the field sensor network are received in parallel; the image data stream includes visible light images and multispectral images, and the environment data stream includes temperature data and humidity data; each of the visible light images, the multispectral images and the environment data is attached with geographic coordinates and a collection time stamp provided by the high-precision positioning module; a unified reference time point and a predefined time window length are set, the collection time stamp carried by each of the visible light images, the multispectral images and the environment data is compared with the reference time point, and the visible light images, the multispectral images and the environment data whose collection time stamps fall within the closed interval centered on the reference time point and having the time window length as the span are screened out, the screened-out visible light images, multispectral images and environment data are divided into the same batch processing unit, and the initial time window alignment is completed; the predefined time window length is set according to the data collection frequency and the field environment stability requirement, and a typical value is 60 seconds, which can effectively contain the small drift of the unmanned aerial vehicle GPS clock and the sensor data uploading delay, ensure the integrity of the spatio-temporal correlation data, and at the same time avoid too long time window to introduce too much environmental variation noise; Then, for the visible light images, the multispectral images and the environment data in the same batch processing unit, the records in the form of discrete geographic points in the environment data are processed by a spatial interpolation algorithm based on the geographic coordinates to generate a continuously distributed temperature field raster layer and a humidity field raster layer, the temperature field raster layer and the humidity field raster layer together constitute an environment data raster layer, and the spatial resolution thereof is the same as that of the visible light images; the spatial interpolation algorithm preferably adopts an inverse distance weighting method or a Kriging method, and the interpolation parameters are set according to the field sensor arrangement density and the spatial variation characteristics, for example, the power parameter of the inverse distance weighting method can be set to 2 to ensure that the generated environment data raster layer can truly reflect the spatial gradient distribution of temperature and humidity in the field and provide an accurate environment background field for subsequent fusion analysis; At the same time, the unmanned aerial vehicle positioning and attitude determination system data are used to map each pixel point of the visible light images and the multispectral images in the same batch processing unit to the real geographic coordinates, complete the orthographic correction, and generate an image raster layer with geographic reference; the orthographic correction process uses a digital elevation model to correct the geometric distortion of the images, eliminates the projection error caused by the terrain undulation and the sensor view angle, and makes each pixel point in the image raster layer correspond to the real ground position, thereby providing a geometric basis for pixel-level spatial matching; Finally, the environmental data raster layer is superimposed with the image raster layer with geographic reference to realize pixel-level spatial matching. The superimposing operation is realized by raster algebra operation in the geographic information system, ensuring that the image pixel value and the environmental variable value at the same geographic coordinate point are accurately corresponding, forming a spatio-temporally consistent data base. This step directly eliminates the analysis error caused by the spatio-temporal mismatch of data sources in the traditional method, and provides high-quality and aligned input for the subsequent deep learning model. The correction method based on the atmospheric physical model is used to perform illumination normalization processing on each band of the visible light image and each band of the multi-spectral image. Specifically, For each pixel of each band of the visible light image in the image raster layer with geographic reference, the original digital quantization value of the pixel in the band is taken as the original observation value, and the normalization process is as follows: The preset or real-time estimated atmospheric path radiation value is deducted from the original observation value to obtain a first intermediate result. The product of the ground surface horizontal direct light intensity and the cosine value of the solar zenith angle determined by the image acquisition time and the geographic position of the pixel is calculated, and the product is added to the ground surface horizontal diffuse light intensity to obtain a second intermediate result. The first intermediate result is divided by the second intermediate result, and then multiplied by the cosine value of the ground surface incident angle determined by the ground surface slope and the solar azimuth to finally obtain the normalized value of the pixel in the band. The estimation of the atmospheric path radiation value can be combined with the reflectivity characteristics of the dark pixels (such as dark green vegetation or shadow area) in the image or synchronous atmospheric optical thickness measurement data. The ground surface horizontal direct light intensity and diffuse light intensity can be calculated by using the atmospheric radiation transfer model such as MODTRAN or real-time solar radiation sensor data. The solar zenith angle and the ground surface incident angle are accurately calculated by the solar position model (such as SPA algorithm) according to the acquisition time, date and geographic coordinates (latitude and longitude) of the pixel. This physical correction process significantly reduces the influence of illumination changes on image color under different shooting times and weather conditions, makes the normalized pixel value closer to the true reflection characteristics of the ground object, and greatly improves the comparability and model generalization ability of disease features under different illumination scenes; The normalization process is independently performed for each waveband of the multispectral image, wherein the atmospheric path radiance value, the ground surface horizontal direct light intensity, and the ground surface horizontal diffuse light intensity parameters are independently calculated and assigned according to the spectral response characteristics of the waveband; the original digital quantization value of each waveband of the multispectral image is taken as the original observation value of the waveband, and the normalization process thereof is the same as that of the visible light image, but the parameter values of the wavebands are independently determined; the independence of the parameters of the wavebands is due to the differences in the spectral response functions of different wavebands and their interaction mechanisms with the atmosphere, for example, the atmospheric scattering effect of the near-infrared waveband is weaker than that of the blue light waveband, so the atmospheric path radiance value thereof is usually lower; this independent calibration of the wavebands preserves the accuracy of the reflectance information in each spectral interval from visible light to near-infrared, laying a foundation for subsequent fine disease diagnosis using multispectral features; After the above normalization processing, the light-normalized visible light image data and multispectral image data are output as input for subsequent processing, and the normalization data output in this step has light invariance, so that the subsequent feature extraction module can focus more on the physiological change characteristics of the crop itself rather than the environmental light differences, thereby effectively improving the detection sensitivity of early and subtle diseases; The construction process of the standardized multi-modal data cube is as follows: The light-normalized visible light image data and multispectral image data generated for the same geographical area, and the environmental data raster layers obtained through the spatio-temporal alignment operation, are stacked and integrated according to a predetermined waveband order; the light-normalized visible light image data includes red, green, and blue waveband data, the multispectral image data includes multiple waveband data including near-infrared and red edge wavebands, the environmental data raster layers include temperature field and humidity field raster layers, and the predetermined waveband order is usually arranged in ascending order of wavelength or importance, for example, red, green, blue, near-infrared, red edge, and finally temperature field and humidity field raster layers; this arrangement is in line with the data processing habits of most deep learning models, which helps the model to efficiently extract cross-modal features from optical to thermal infrared; During stacking, the red, green, and blue waveband data of the visible light image are first arranged as the first three dimensions of the data cube, then the above-mentioned multiple waveband data of the multispectral image are sequentially arranged, and finally the temperature field and humidity field raster layers are arranged as the last dimensions of the multi-modal data cube; this stacking logic ensures the structured organization of the data cube in the channel dimension, so that the feature vector corresponding to each spatial position (x, y) follows a consistent waveband sequence, and this structure is very suitable as input for models such as convolutional neural networks or Transformers, facilitating the model to learn the associated features between different data sources; After integration, a three-dimensional structure of the multi-modal data cube is formed, which is defined by geographical coordinates in the spatial dimension and contains all image bands and environmental variables represented by the temperature field and humidity field grid layers in the attribute dimension, so that each spatial position point in the multi-modal data cube corresponds to a feature vector composed of multi-source observation information, including visible light band values, multi-spectral band values, and values represented by temperature data and humidity data. The multi-modal data cube is the direct input of the subsequent feature enhancement extraction module, and its regular three-dimensional structure (height H x width W x channel number C) enables the deep learning model to simultaneously utilize spatial context and spectral environmental information for joint analysis. For example, the model can learn that "under certain temperature and humidity conditions, a specific spectral change pattern is more likely to indicate a certain type of disease." This deep fusion of multi-source information greatly enhances the accuracy and robustness of the diagnosis system, surpassing the capabilities of single data source analysis.

[0020] In this embodiment, it is specifically pointed out that the deep learning model based on the Transformer architecture in the feature enhancement extraction module includes a convolutional neural network backbone for preliminary feature extraction, a fully connected network for environmental data encoding, and an improved Transformer encoder. Secondly, the specific operation of the deep learning model based on the Transformer architecture for spatio-temporal feature extraction is as follows: Firstly, the received multi-modal data cube is processed by branch, separating the image data part composed of visible light image data and multi-spectral image data from the environmental data part composed of environmental data; Then, the image data part is preliminarily spatially feature-extracted by the convolutional neural network backbone, converted into a spatial feature map, and flattened into a sequence of image feature tokens along the spatial dimension. The convolutional neural network backbone can adopt the first four layers of the ResNet34 architecture, and the output spatial feature map can have a size of one-sixteenth of the input image size and a channel number of 512. After flattening the spatial feature map with a spatial size of H x W and a channel number of C, the length of the obtained image feature token sequence is H x W; At the same time, the environmental data part is encoded by the fully connected network to generate an environmental feature sequence. The fully connected network can include two linear layers with ReLU activation functions in between, encoding the environmental data (such as temperature and humidity) at each position into a 64-dimensional environmental feature vector, and all the environmental feature vectors at different positions form the environmental feature sequence. Then, the image feature token sequence and the environment feature sequence are jointly input into an improved Transformer encoder, the Transformer encoder comprises a cross-attention mechanism in each layer, wherein the image feature token sequence is taken as a query vector, the environment feature sequence is taken as a key vector and a value vector, the dot product of the query vector and the key vector is calculated, the dot product result is divided by the square root of the dimension of the key vector to perform scaling, then a Softmax normalization function is applied to the scaled result to obtain an attention weight, and the attention weight is weighted and summed with the value vector to generate an image feature representation modulated by the environment information, so as to realize the internal fusion of image information and environment conditions, the improved Transformer encoder can be stacked by 6 layers, the self-attention head and the cross-attention head of each layer can be set to 8 heads, the dimension of the key vector is usually 64, and the scaling factor is 8, the cross-attention mechanism enables the model to adaptively focus on the area more likely to be diseased in the image according to the environment context (such as high temperature and high humidity), for example, under the condition of high humidity, the model will give higher weight to the water stain-like features on the edges of the leaves; Finally, the feature sequence output by the Transformer encoder is mapped to a two-dimensional spatial structure to form a deep feature map, and on this basis, the deep feature map is input into a lightweight classification head, the classification head is a full connection layer or a one-by-one convolution layer, the probability of each spatial position corresponding to different disease categories output by the classification head is taken as a preliminary disease diagnosis result, the spatial size of the deep feature map is Figure 1 (HxW), and the number of channels is 512, the classification head can be specifically a one-by-one convolution layer, the number of convolution kernels of the one-by-one convolution layer is set to the number of disease categories (for example, 3 categories: healthy, rust, and powdery mildew), followed by a Softmax function, and the probability of each spatial position belonging to each disease category is output to form a preliminary disease diagnosis result; The specific process of integrating the contrast learning strategy to optimize the feature representation of the deep learning model based on the Transformer architecture is as follows: A1, training sample pair construction: for a multi-modal data cube as an anchor sample, a positive sample pair thereof is generated through a data enhancement technique including slight rotation or color jittering, and a multi-modal data cube confirmed to be healthy from different fields or different growth periods is selected as a negative sample, the angle range of slight rotation can be set to within 10 degrees, and color jittering can be a random disturbance of 5% of the maximum brightness value in the RGB color space channel, 1 positive sample can be generated for each anchor sample, and other healthy samples in the same training batch are randomly selected as negative samples, and the number of negative samples can be 15; A2, environment-aware contrastive loss function calculation and optimization: the optimization process is realized by minimizing an environment-aware contrastive loss function, the value of which is equal to the value of a standard contrastive loss term plus the product of the value of a balance hyperparameter and the value of an environmental consistency constraint term; wherein the standard contrastive loss term is used to reduce the distance between the anchor sample and the positive sample in the feature space, while increasing the distance between the anchor sample and the negative sample in the feature space; the environmental consistency constraint term is used to measure and punish the deep learning model for generating undesirable feature differences for samples under similar environmental conditions but with different health conditions, aiming to promote the deep learning model to learn a feature representation that is sensitive to disease rather than environmental changes; by adjusting the value of the balance hyperparameter, the relative weights of the standard contrastive loss term and the environmental consistency constraint term in the environment-aware contrastive loss function are balanced, and the standard contrastive loss term can take the form of an InfoNCE loss function. The environmental consistency constraint term can be calculated as follows: for a pair of samples with similar environmental conditions (such as temperature, humidity) (similarity higher than a preset threshold of 0.7), if their health conditions are different (one is diseased and the other is healthy), calculate their Euclidean distance in the feature space and encourage the distance to decrease; the empirical value range of the balance hyperparameter λ can be 0.1 to 1.0, and a typical effective value is 0.5; this joint loss function can effectively guide the model to ignore environmental fluctuations and focus on disease-related feature patterns; By iteratively minimizing the environment-aware contrastive loss function, the trained deep learning model outputs deep features with stronger discrimination ability for early subtle lesions, and provides a more reliable feature basis for preliminary disease diagnosis results. Iterative training can use the Adam optimizer, the initial learning rate can be set to 0.001, and learning rate decay can be performed when the validation set loss plateaus. The training goal is to minimize the total loss until the model converges, which usually requires 100 to 200 training cycles.

[0021] In this embodiment, it is specifically pointed out that in the online learning and updating module, the specific operation of selecting low-confidence samples for labeling using the sample screening mechanism based on uncertainty estimation is as follows: First, for the new visible light images, multispectral images and environmental data continuously received by the streaming data processing platform, the deep learning model in the feature enhancement extraction module is called for forward inference to obtain the preliminary disease diagnosis result of each sample. The preliminary disease diagnosis result represents the prediction confidence of the deep learning model for each disease class in the form of a prediction probability distribution. The streaming data processing platform can use the Apache Kafka framework, and the data reception frequency can be configured to receive a batch of new data every 5-10 minutes. During forward inference, the deep learning model works in inference mode and does not update parameters. Then, the Monte Carlo Dropout method is used for uncertainty estimation, specifically: keeping the Dropout layer active during the inference process of the deep learning model, performing multiple forward propagations on the same sample to obtain multiple different prediction probability distributions, the number of times of forward propagation is denoted as a preset integer T greater than 1, the dropout rate of the Dropout layer can be set to 0.1 to 0.3, the value range of the number of times of forward propagation T is usually 10-50 times, preferably 20 times, to balance between calculation cost and estimation accuracy; Then, two uncertainty indicators are calculated based on the multiple different prediction probability distributions: The first indicator is prediction entropy, which is used to measure the total uncertainty of the deep learning model prediction, and the calculation process is: First, the average probability distribution of the multiple prediction probability distributions is calculated, that is, for each disease category, the probability value of the category appearing in all prediction probability distributions is calculated to obtain the average probability of the category, thereby forming the average probability distribution; Then, the prediction entropy is calculated, which is equal to multiplying the average probability of each disease category by the inverse of the natural logarithm of the average probability of the category, and then summing the calculation results of all disease categories; the higher the value of the prediction entropy, the greater the total uncertainty of the deep learning model prediction for the sample, and the value range of the prediction entropy is between 0 and log(FG), where FG is the number of disease categories, and when the probabilities of all categories are equal, the prediction entropy is maximum, indicating that the model is completely uncertain; The second indicator is cognitive uncertainty, which is used to measure the uncertainty of the deep learning model due to lack of knowledge, and the calculation process is: First, the average probability distribution of the multiple prediction probability distributions is calculated, that is, for each disease category, the probability value of the category appearing in all prediction probability distributions is calculated to obtain the average probability of the category, thereby forming the average probability distribution; Then, for each prediction probability distribution, the KL divergence between the prediction probability distribution and the average probability distribution is calculated, and the calculation process of the KL divergence is: For each disease category, multiply the probability of the category in the prediction probability distribution by the natural logarithm of the ratio of the probability to the average probability of the category, and then sum the calculation results of all disease categories; Finally, the arithmetic mean of the KL divergences of all prediction probability distributions is calculated to obtain the cognitive uncertainty; the higher the value of the cognitive uncertainty, the greater the uncertainty of the deep learning model due to lack of relevant knowledge, and the cognitive uncertainty mainly reflects the uncertainty of the parameters of the deep learning model, which will significantly increase when the deep learning model lacks training for a certain type of sample; Subsequently, the two uncertainty indicators, i.e., the prediction entropy and the cognitive uncertainty, are normalized to adjust their value ranges to between zero and one, and then are weighted and summed according to preset first and second weights to obtain a comprehensive uncertainty score, wherein the first and second weights are both constants greater than zero, and the normalization can adopt a min-max normalization method. The first weight (corresponding to the prediction entropy) is usually set to 0.6-0.8, and the second weight (corresponding to the cognitive uncertainty) is set to 0.2-0.4, and the sum of the two weights is 1. This weight distribution emphasizes that the overall uncertainty of the model is more important; Finally, a comprehensive uncertainty threshold is set, samples with a comprehensive uncertainty score higher than the comprehensive uncertainty threshold are marked as low-confidence samples and are added to the to-be-labeled sample queue to wait for manual labeling by experts; after manual labeling, these low-confidence samples are marked as newly-labeled samples for subsequent incremental fine-tuning, and the value range of the comprehensive uncertainty threshold is 0.5-0.8, preferably 0.65. When the comprehensive uncertainty score exceeds the threshold, it indicates that the reliability of the deep learning model for judging the sample is low, and expert intervention is needed for labeling. This mechanism can significantly improve the labeling efficiency and reduce about 40% of the manual labeling workload. The specific operation of periodically performing incremental fine-tuning on the deep learning model in the feature enhancement extraction module is as follows: First, a pre-set number of newly-labeled samples accumulated from the to-be-labeled sample queue that has been manually labeled are taken out to form an incremental learning training set for the current batch, and the "pre-set number" is set to 50-200 samples, or the incremental learning is triggered when the newly-labeled samples accumulate to a certain time (such as 24 hours); Then, the elastic weight consolidation algorithm is used for incremental fine-tuning, wherein the elastic weight consolidation algorithm is realized by minimizing a total loss function composed of two parts: the first part is the standard cross-entropy loss on the newly-labeled samples, and the second part is a regularization term; the calculation process of the regularization term is as follows: A preset regularization strength hyperparameter is multiplied by a summation formula, which calculates the product of the importance weight of each parameter of the deep learning model and the square of the difference between the current value of the parameter and the old value of the parameter before incremental fine-tuning, and then sums all the product results of the parameters, and the value range of the regularization strength hyperparameter is 10-1000, preferably 100. The regularization strength hyperparameter controls the balance between new and old knowledge, and the larger the value, the stronger the protection of old knowledge; When calculating the importance weight of each parameter in the deep learning model, the importance weight is obtained by calculating the diagonal elements of the Fisher information matrix of the parameter on the old task training data. The larger the diagonal element of the Fisher information matrix, the more important the parameter is to the old task and should be kept as stable as possible during incremental learning. In the training process, the above total loss function is minimized using an optimization algorithm to update the parameters of the deep learning model, so that the deep learning model adapts to new samples while avoiding catastrophic forgetting of old knowledge; after the incremental fine-tuning is completed, the updated parameters are used to replace the parameters of the deep learning model in the feature enhancement extraction module, and then the performance of the updated deep learning model is evaluated on a reserved validation set to ensure that its performance is improved or at least not degraded; and the metadata of this update is recorded, including the update time, the amount of data used, the performance change, for monitoring and iterative optimization, the optimization algorithm can use the Adam optimizer, the learning rate is set to 0.0001-0.001, the validation set should contain samples of new and old categories, and the evaluation indicators include accuracy, F1 score, etc. Metadata recording helps to analyze the model evolution trend and provides a basis for subsequent parameter optimization. The incremental learning cycle is usually once a day, which can be adjusted according to the actual data volume.

[0022] In this embodiment, it is specifically necessary to explain that in the decision support and visualization module, the classification head attached to the deep learning model in the feature enhancement extraction module is used to convert the deep feature map into a disease probability distribution map, and combined with the geographic information system for spatial mapping, the specific operation of generating the disease distribution heat map is as follows: First, input the received deep feature map into the classification head attached to the deep learning model in the feature enhancement extraction module. The classification head is a fully connected layer or a one-by-one convolution layer, followed by a Softmax activation function. Through the linear transformation of the classification head and the Softmax activation function processing, the high-dimensional feature vector of each spatial position (also called pixel) in the deep feature map is converted into an initial disease probability distribution. The initial disease probability distribution represents the initial probability of each spatial position belonging to each disease category. The output dimension of the fully connected layer or the one-by-one convolution layer is set to the number of disease categories, usually 3-5 categories, including healthy, rust, powdery mildew and other common winter-sown spring wheat diseases. The Softmax function ensures that the output value is between 0 and 1 and the sum is 1, forming an effective probability distribution; Then, the conditional random field optimization algorithm is used to optimize the spatial consistency of the initial disease probability distribution, and the conditional random field optimization algorithm optimizes by constructing an energy function and minimizing the energy function; the energy function is composed of the sum of unary potential functions and the sum of binary potential functions, and the conditional random field optimization can use the mean field approximation algorithm for efficient inference, and the iteration number is usually set to 5-10 times; the weight of the unary potential function can be set to 1.0, and the weight of the binary potential function can be set to 3.0-5.0 to emphasize the importance of spatial consistency constraints; The unary potential is used to constrain the optimized probability distribution not to deviate from the initial disease probability distribution, and its value is determined by the negative natural logarithm of the initial probability of the corresponding spatial position in the initial disease probability distribution. The unary potential directly depends on the initial confidence output by the deep learning model, ensuring that the optimization process does not sacrifice classification accuracy; The binary potential is composed of a label compatibility function multiplied by a weighted sum, which includes an appearance coefficient multiplied by an appearance term and a smoothing coefficient multiplied by a smoothing term. The ratio of the appearance coefficient and the smoothing coefficient is usually set between 1:2 and 1:3, for example, the appearance coefficient is 1.0 and the smoothing coefficient is 2.5, to balance the influence of appearance similarity and spatial proximity; The appearance term is calculated by the exponential function of the inverse of the square of the difference between the color vectors of the spatial positions divided by the square of the color bandwidth parameter, multiplied by the exponential function of the inverse of the square of the difference between the position coordinates of the spatial positions divided by the square of the position bandwidth parameter. The color bandwidth parameter ranges from 0.1 to 0.3, and the position bandwidth parameter ranges from 3 to 10 pixel distances. These parameters control the sensitivity of the similarity measure, with smaller values being more sensitive to differences; The smoothing term is calculated by the exponential function of the inverse of the square of the difference between the position coordinates of the spatial positions divided by the square of the smoothing bandwidth parameter. The smoothing bandwidth parameter is usually the same as or slightly larger than the position bandwidth parameter, with a value of 5-15 pixel distances, to define the effective spatial neighborhood range; The label compatibility function takes a value of one when the disease class labels of adjacent spatial positions are different, and a value of zero when they are the same. This function is a form of the Potts model, which only imposes a penalty when different labels are adjacent, encouraging the formation of continuous regions of the same class; The appearance term encourages spatial positions with similar color features and close spatial positions to have the same disease class label. The smoothing term encourages spatial positions adjacent to each other to have the same disease class label. The color bandwidth parameter, position bandwidth parameter, and smoothing bandwidth parameter control the sensitivity of color similarity and spatial proximity. The appearance coefficient and smoothing coefficient are weight parameters used to balance the importance of the appearance term and the smoothing term. This design allows the optimized probability distribution to maintain clear image boundaries (appearance term effect) and smooth the internal regions (smoothing term effect), effectively eliminating isolated noise points; After minimizing the energy function through the iterative optimization algorithm, an optimized disease probability distribution map with enhanced spatial consistency is obtained. The disease occurrence probability value corresponding to each spatial position in the optimized disease probability distribution map is defined as the current disease probability value. After optimization, the visual effect of the optimized disease probability distribution map is significantly improved, the connectivity of the same class region is enhanced, and the spatial distribution of the probability value is more continuous, which helps to improve the reliability of subsequent decision-making; Finally, the geographical position information of each spatial position in the optimized disease probability distribution map is registered with the geographical coordinates in the geographic information system, the probability values are mapped to the corresponding geographical spatial positions, and the probability values are color rendered using a gradient color system from blue to red, generating a disease distribution heat map, where blue represents low probability and red represents high probability. The geographical coordinate registration adopts bilinear interpolation method to ensure that the image pixels correspond to the geographical coordinates accurately. In the color rendering, the probability value 0 corresponds to pure blue, the probability value 1 corresponds to pure red, and the intermediate values use linear interpolation color; At the same time, the disease distribution heat map has dynamic display function, and can store heat map data of different collection time points in time sequence; for a specific geographical area in the geographic information system, the time sequence change curve of disease severity in the area can be calculated, and the spread trend and propagation speed of the disease in the area can be intuitively displayed in the form of animation by continuously playing the heat maps of different time points. The time sequence data storage interval can be once a day, and the dynamic display frame rate can be adjusted to 1-5 frames per second. The time sequence change curve can display the area average probability value, disease area ratio and other indicators, providing data support for disease warning; The specific operation of generating the variable operation prescription map is as follows: Firstly, based on the optimized disease probability distribution map, the area with disease occurrence probability higher than the preset high probability threshold is identified as the disease center area, the preset high probability threshold is 0.7-0.9, preferably 0.8, and the pixel points with probability value higher than the threshold are clustered into areas, and the area with area greater than a certain threshold (such as 10 square meters) is identified as the meaningful disease center; Then, based on the historical disease distribution heat map data stored in time sequence, the maximum likelihood estimation algorithm is used to fit the key parameters of the spatial diffusion dynamics model, including diffusion coefficient and disease growth rate, and the spatial diffusion dynamics model can adopt reaction-diffusion equation. The time span of historical data is suggested to be 7-14 days, and the maximum likelihood estimation is solved by gradient descent method, and the iteration is converged; Then, using the fitted spatial diffusion dynamics model, the range and intensity of disease spread from the disease center area to the outside in the future time window are predicted, a predicted disease risk distribution map is generated, and the disease occurrence risk probability value corresponding to each geographical position in the predicted disease risk distribution map is defined as the predicted disease risk value. The future time window is usually set to 24-72 hours, i.e. the next expected operation period. The prediction result includes the geographical range where the disease may spread and the expected occurrence probability at that place; Finally, a multi-objective optimization decision function is constructed to generate the variable operation prescription map, which specifically includes: F1, for each geographical location, obtaining the current disease probability value of the location from the optimized disease probability distribution map and obtaining the predicted disease risk value of the location from the predicted disease risk distribution map; comparing the current disease probability value and the predicted disease risk value, and selecting the larger value as the comprehensive risk index corresponding to the geographical location, which ensures that the decision considers both the current actual situation and the potential future risk, and follows the principle of prudence; F2, comparing the comprehensive risk index with a preset economic threshold: when the comprehensive risk index is less than or equal to the economic threshold, it is determined that the geographical location does not need to be sprayed, and the recommended spraying amount is set to zero; when the comprehensive risk index is greater than the economic threshold, the spraying amount calculation process is entered, the economic threshold is determined according to the disease type, the wheat growth period and the expected loss, and the value range is usually 0.2-0.4, for example, for the rust disease at the heading stage, the economic threshold can be set to 0.3; In the spraying amount calculation process, first calculate the value of the sum of the economic threshold and a preset risk buffer threshold that exceeds the comprehensive risk index, multiply the value by a preset unit risk increment spraying coefficient to obtain a preliminary spraying amount calculation result, and the preliminary spraying amount calculation result cannot exceed a preset maximum spraying amount limit, if the comprehensive risk index does not exceed the sum of the economic threshold and the risk buffer threshold, it is still determined that the geographical location does not need to be sprayed, and the recommended spraying amount is set to zero, the risk buffer threshold is used to prevent excessive reaction when the risk is slightly higher than the economic threshold, the value range is 0.05-0.1, the unit risk increment spraying coefficient converts the probability difference into the actual spraying amount (such as milliliter / hectare), which is calibrated according to the pesticide efficacy and crop tolerance, and the maximum spraying amount limit is set by the pesticide label recommended dose and environmental protection requirements to prevent pesticide damage and environmental residues; F3, after the calculation of all locations is completed, the recommended values are discretely classified according to a preset spraying amount level, and different colors or filling patterns are used for visual rendering in the geographic information system to generate a variable operation prescription map, which clearly identifies the no-spraying area, the early warning monitoring area, the light spraying area, the moderate spraying area and the heavy spraying area, the spraying amount level can be set to 4-5 levels, in the visual rendering, green color can be used to represent the no-spraying area, yellow color can be used to represent the early warning monitoring area, orange, red and dark red can be used to represent the light, moderate and heavy spraying areas respectively, the prescription map can be directly imported into the intelligent spraying equipment to guide the variable operation.

[0023] It should be noted that in the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0024] Those skilled in the art will appreciate that embodiments of the present application can be devised for a variety of applications. It is intended that the present application be limited only by the scope of the appended claims, and it is intended that various modifications and alterations made by those skilled in the art be considered as within the scope of the present application. The embodiments of the present application will be described with reference to the attached drawings identified below.

[0025] The present application is described in reference to the drawings using a flowchart and / or a block diagram of the method, apparatus (system) and computer program product according to embodiments of the application. It will be understood that each block of the flowchart and / or block diagram, and combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0026] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0027] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks. Figure 1 one or more functions specified in the flowchart and / or block diagram block or blocks.

[0028] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such modifications and variations as fall within the scope of the present application.

[0029] Obviously, many modifications and variations of the present application are possible in light of the above teachings. It is, therefore, to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A field dynamic diagnostic system for winter-sown spring wheat based on deep learning, characterized in that, Specifically, it includes: The system includes a data synchronization and preprocessing module, a feature enhancement and extraction module, an online learning and updating module, and a decision support and visualization module. Data synchronization and preprocessing module: It is used to receive visible light images and multispectral images collected from UAV platform, as well as environmental data from field sensor network. Through spatiotemporal alignment operation based on geographic coordinates and timestamps, it matches environmental data with visible light images and multispectral images, performs illumination normalization processing on visible light images and multispectral images, and outputs standardized multimodal data cubes. Feature enhancement and extraction module: Connected to the data synchronization and preprocessing module, it receives multimodal data cubes and uses a deep learning model based on the Transformer architecture to extract spatiotemporal features. This deep learning model integrates a contrastive learning strategy to optimize feature representation. Finally, the deep learning model outputs a deep feature map containing subtle lesion features and preliminary disease diagnosis results. Online learning and update module: Connected to the feature enhancement and extraction module, it is used to receive deep feature maps and preliminary disease diagnosis results, and receive visible light images, multispectral images and environmental data continuously collected by UAV platform and field sensor network through streaming data processing platform. It adopts a sample screening mechanism based on uncertainty estimation to select low confidence samples for labeling, and performs incremental fine-tuning of the deep learning model in the feature enhancement and extraction module on a regular basis. Decision Support and Visualization Module: Connected to the Feature Enhancement and Extraction Module, it receives deep feature maps and uses the classification head attached to the deep learning model in the Feature Enhancement and Extraction Module to convert the deep feature maps into disease probability distribution maps. It then combines the geographic information system to perform spatial mapping and generate disease distribution heat maps and variable operation prescription maps.

2. The field dynamic diagnostic system for winter-sown spring wheat based on deep learning according to claim 1, characterized in that: In the data synchronization and preprocessing module, the spatiotemporal alignment operation based on geographic coordinates and timestamps specifically involves the following steps: First, image data streams from the UAV platform and environmental data streams from the field sensor network are received in parallel. The image data streams include visible light images and multispectral images, while the environmental data streams include temperature and humidity data. Each data item in the visible light images, multispectral images, and environmental data is accompanied by geographic coordinates and acquisition timestamps provided by a high-precision positioning module. A unified reference time point and a predefined time window length are set. The acquisition timestamps carried by each visible light image, multispectral image, and environmental data item are compared with the reference time point. Visible light images, multispectral images, and environmental data items whose acquisition timestamps fall within a closed interval centered on the reference time point and spanning the time window length are selected. These selected visible light images, multispectral images, and environmental data items are divided into the same batch processing unit to complete the initial time window alignment. Next, for the visible light images, multispectral images and environmental data in the same batch of processing units, the records in the environmental data that exist in the form of discrete geographic points are processed by spatial interpolation algorithms based on their geographic coordinates to generate continuously distributed temperature field raster layers and humidity field raster layers. The temperature field raster layer and humidity field raster layer together constitute the environmental data raster layer, and its spatial resolution is the same as that of the visible light images. At the same time, using the data from the UAV positioning and attitude determination system, each pixel of the visible light image and multispectral image in the same batch processing unit is mapped to the real geographic coordinates to complete orthorectification and generate an image raster layer with geographic reference. Finally, the environmental data raster layer is overlaid with a georeferenced image raster layer to achieve pixel-level spatial matching.

3. The field dynamic diagnostic system for winter-sown spring wheat based on deep learning according to claim 2, characterized in that: An illumination normalization process is performed on each band of the visible light image and each band of the multispectral image using a correction method based on an atmospheric physics model. Specifically: For each pixel in each band of a visible light image in a georeferenced image raster layer, the raw digital quantization value of that pixel in that band is used as the raw observation value, and its normalization process is as follows: The first intermediate result is obtained by subtracting the preset atmospheric path radiation value from the original observation value; the product of the direct light intensity on the horizontal surface of the ground and the cosine of the solar zenith angle determined by the image acquisition time and the pixel's geographical location is calculated, and this product is added to the diffuse light intensity on the horizontal surface of the ground to obtain the second intermediate result; the first intermediate result is divided by the second intermediate result, and then multiplied by the cosine of the ground incidence angle determined by the ground slope and the solar azimuth to finally obtain the normalized value of the pixel in this band; The above normalization process is performed independently for each band of the multispectral image; After the above normalization process, the output is the normalized visible light image data and multispectral image data.

4. The field dynamic diagnostic system for winter-sown spring wheat based on deep learning according to claim 3, characterized in that: The construction process of the standardized multimodal data cube is as follows: The light-normalized visible light image data and multispectral image data generated for the same geographic area, as well as the environmental data raster layer obtained through spatiotemporal alignment operations, are stacked and integrated according to a preset band order. The light-normalized visible light image data includes red, green, and blue band data, the multispectral image data includes multiple band data including near-infrared and red-edge bands, and the environmental data raster layer includes a temperature field raster layer and a humidity field raster layer. When stacking, the red, green, and blue band data of the visible light image are first used as the first three dimensions of the data cube. Then, the above multiple band data of the multispectral image are arranged sequentially. Finally, the temperature field raster layer and the humidity field raster layer are used as the last dimension of the multimodal data cube. After integration, a three-dimensional multimodal data cube is formed.

5. The field dynamic diagnostic system for winter-sown spring wheat based on deep learning according to claim 4, characterized in that: In the feature enhancement and extraction module, the deep learning model based on the Transformer architecture includes a convolutional neural network backbone for preliminary feature extraction, a fully connected network for environmental data encoding, and an improved Transformer encoder. Secondly, the specific operations for spatiotemporal feature extraction using a deep learning model based on the Transformer architecture are as follows: First, the received multimodal data cube is branched to separate the image data portion, which consists of visible light image data and multispectral image data, from the environmental data portion, which consists of environmental data. Next, the image data is processed by a convolutional neural network backbone to perform preliminary spatial feature extraction, which converts it into a spatial feature map. This spatial feature map is then flattened along the spatial dimension into a sequence of image feature tokens. Meanwhile, the environmental data is encoded through a fully connected network to generate environmental feature sequences; Then, the image feature token sequence and the environment feature sequence are input into the improved Transformer encoder, which contains a cross-attention mechanism in each layer. The image feature token sequence is used as the query vector, and the environment feature sequence is used as the key vector and value vector. The dot product of the query vector and the key vector is calculated, and the dot product result is divided by the square root of the key vector dimension for scaling. Then, the scaling result is normalized by the Softmax function to obtain the attention weight. Finally, the attention weight is weighted and summed with the value vector to generate the image feature representation modulated by the environment information. Finally, the feature sequence output by the Transformer encoder is remapped to a two-dimensional spatial structure to form a depth feature map. Based on this, the depth feature map is input into a lightweight classification head, which is either a fully connected layer or a 1:1 convolutional layer. The classification head outputs the probability of each spatial location corresponding to different disease categories, which serves as the preliminary disease diagnosis result.

6. The field dynamic diagnostic system for winter-sown spring wheat based on deep learning according to claim 5, characterized in that: The specific process of the ensemble contrastive learning strategy to optimize the feature representation of the Transformer-based deep learning model involves performing the following operations during the training phase of the deep learning model: A1. Training Sample Pair Construction: For a multimodal data cube that serves as the anchor sample, positive sample pairs are generated using data augmentation techniques, including slight rotation or color jitter, and confirmed healthy multimodal data cubes from different fields or different growth stages are selected as negative samples. A2. Calculation and optimization of contrastive loss function for environmental perception: The optimization process is achieved by minimizing a contrastive loss function for environmental perception. The value of the contrastive loss function for environmental perception is equal to the value of a standard contrastive loss term, plus the product of the value of a balancing hyperparameter and the value of an environmental consistency constraint term.

7. The field dynamic diagnostic system for winter-sown spring wheat based on deep learning according to claim 6, characterized in that: In the online learning and update module, the specific operation of selecting low-confidence samples for annotation using a sample screening mechanism based on uncertainty estimation is as follows: First, for new visible light images, multispectral images and environmental data continuously received through the streaming data processing platform, the deep learning model in the feature enhancement and extraction module is invoked to perform forward inference to obtain the preliminary disease diagnosis results for each sample. The preliminary disease diagnosis results are represented by the prediction confidence of the deep learning model for each disease category in the form of a prediction probability distribution. Then, the Monte Carlo Dropout method is used for uncertainty estimation. Specifically, the Dropout layer is kept active during the inference process of the deep learning model, and the same sample is forwarded multiple times to obtain multiple different predicted probability distributions. Next, based on these multiple different prediction probability distributions, two uncertainty indices are calculated: The first indicator is the prediction entropy, and its calculation process is as follows: First, calculate the average probability distribution of multiple predicted probability distributions. That is, for each disease category, calculate the arithmetic mean of the probability values ​​of that category appearing in all predicted probability distributions to obtain the average probability of that category, thus forming the average probability distribution. Then, the prediction entropy is calculated. The prediction entropy is equal to the average probability of each disease category multiplied by the negative of the natural logarithm of the average probability of that category, and then the results of all disease categories are summed. The second indicator is cognitive uncertainty, which is calculated as follows: First, calculate the average probability distribution of multiple predicted probability distributions. That is, for each disease category, calculate the arithmetic mean of the probability values ​​of that category appearing in all predicted probability distributions to obtain the average probability of that category, thus forming the average probability distribution. Then, for each predicted probability distribution, the KL divergence between the predicted probability distribution and the average probability distribution is calculated. The calculation process for the KL divergence is as follows: For each disease category, multiply the probability of that category in the predicted probability distribution by the natural logarithm of the ratio of that probability to the average probability of that category, and then sum the results for all disease categories. Finally, the arithmetic mean of the KL divergences of all predicted probability distributions is calculated to obtain the cognitive uncertainty; Subsequently, the two uncertainty indicators, namely prediction entropy and cognitive uncertainty, are normalized respectively, and then weighted and summed according to the preset first and second weights to obtain a comprehensive uncertainty score. Finally, a comprehensive uncertainty threshold is set, and samples with comprehensive uncertainty scores higher than the threshold are marked as low-confidence samples and added to the queue of samples to be labeled, waiting for experts to manually label them; after manual labeling, these low-confidence samples are marked as newly labeled samples.

8. The field dynamic diagnostic system for winter-sown spring wheat based on deep learning according to claim 7, characterized in that: The specific operation of periodically performing incremental fine-tuning of the deep learning model in the feature enhancement extraction module is as follows: First, new labeled samples accumulated to a predetermined number are taken from the manually labeled sample queue to form the incremental learning training set for the current batch. Then, the elastic weight consolidation algorithm is used for incremental fine-tuning. This algorithm minimizes a total loss function, which consists of two parts: the first part is the standard cross-entropy loss on the newly labeled samples, and the second part is a regularization term. The calculation process for the regularization term is as follows: A pre-defined regularization strength hyperparameter is multiplied by a summation formula that calculates the product of the importance weight of each parameter in the deep learning model and the square of the difference between the current value of the parameter and the old value of the parameter before incremental fine-tuning. Then, the results of such products of all parameters are summed. During training, an optimization algorithm is used to minimize the total loss function mentioned above, thereby updating the parameters of the deep learning model; after incremental fine-tuning, the updated parameters are used to replace the parameters of the deep learning model in the feature enhancement extraction module.

9. A field dynamic diagnostic system for winter-sown spring wheat based on deep learning according to claim 8, characterized in that: In the decision support and visualization module, the specific operation of converting the deep feature map into a disease probability distribution map using the classification head attached to the deep learning model in the feature enhancement extraction module, and then combining it with the geographic information system for spatial mapping to generate a disease distribution heat map is as follows: First, the received deep feature map is input into the classification head attached to the deep learning model in the feature enhancement and extraction module. This classification head is a fully connected layer or a one-to-one convolutional layer, followed by a Softmax activation function. Through the linear transformation of the classification head and the processing of the Softmax activation function, the high-dimensional feature vector of each spatial location in the deep feature map is converted into an initial disease probability distribution. This initial disease probability distribution represents the initial probability that each spatial location belongs to each disease category. Then, the initial disease probability distribution is spatially consistent with the conditional random field optimization algorithm. This algorithm achieves optimization by constructing an energy function and minimizing it. The energy function is composed of the sum of univariate potential functions and the sum of binary potential functions. The value of the univariate potential function is determined by the negative natural logarithm of the initial probability of the corresponding spatial location in the initial disease probability distribution; The binary potential function is composed of a tag compatibility function multiplied by a weighted sum, which includes appearance coefficient multiplied by appearance term and smoothing coefficient multiplied by smoothing term; The appearance item is calculated by multiplying the square of the difference between the spatial position color vectors by the exponential function of the negative of the square of the color bandwidth parameter, and then by the exponential function of the square of the difference between the spatial position coordinates by the negative of the square of the position bandwidth parameter. The smoothing term is calculated by dividing the square of the difference between the spatial location coordinates by an exponential function of the negative of the square of the smoothing bandwidth parameter. The label compatibility function takes a value of one when the disease category labels in adjacent spatial locations are different, and a value of zero when they are the same. After minimizing the energy function through an iterative optimization algorithm, an optimized disease probability distribution map with enhanced spatial consistency is obtained, and the disease occurrence probability value corresponding to each spatial location in the optimized disease probability distribution map is defined as the current disease probability value. Finally, the geographic location information of each spatial location in the optimized disease probability distribution map is registered with the geographic coordinates in the geographic information system, the probability value is mapped to the corresponding geographic spatial location, and the probability value is rendered with a gradient color system from blue to red to generate the disease distribution heat map, where blue represents low probability and red represents high probability.

10. A field dynamic diagnostic system for winter-sown spring wheat based on deep learning according to claim 9, characterized in that: The specific steps for generating the variable job prescription diagram are as follows: First, based on the optimized disease probability distribution map, areas where the disease occurrence probability is higher than a preset high probability threshold are identified as disease center areas; Then, based on the historical disease distribution heatmap data stored in time series, the key parameters of the spatial diffusion dynamics model, including the diffusion coefficient and disease growth rate, are fitted using the maximum likelihood estimation algorithm. Next, using the fitted spatial diffusion dynamics model, the range and intensity of the disease spreading outward from the disease center area within the future time window are predicted, generating a predicted disease risk distribution map, and the probability value of disease occurrence corresponding to each geographical location in the predicted disease risk distribution map is defined as the predicted disease risk value. Finally, a multi-objective optimization decision function is constructed to generate a variable job prescription map, specifically including: F1. For each geographical location, obtain the current disease probability value from the optimized disease probability distribution map and the predicted disease risk value from the predicted disease risk distribution map; compare the current disease probability value and the predicted disease risk value, and select the larger value as the comprehensive risk index corresponding to the geographical location. F2. Compare the comprehensive risk index with the preset economic threshold: when the comprehensive risk index is less than or equal to the economic threshold, it is determined that no pesticide application is required for this geographical location, and the recommended application amount is set to zero; when the comprehensive risk index is greater than the economic threshold, the pesticide application amount calculation process begins. During the calculation of the dosage, the value of the comprehensive risk index exceeding the sum of the economic threshold and a preset risk buffer threshold is first calculated. This value is then multiplied by a preset unit risk increment dosage coefficient to obtain the preliminary dosage calculation result. The preliminary dosage calculation result must not exceed a preset maximum dosage limit. If the comprehensive risk index does not exceed the sum of the economic threshold and the risk buffer threshold, it is still determined that no pesticide application is required for this geographical location, and the recommended dosage is set to zero. F3. After completing the calculations for all locations, the recommended values ​​are discretized and classified according to the preset application rate levels, and visualized and rendered in the geographic information system using different colors or fill patterns to generate the variable operation prescription map. This prescription map clearly identifies areas that do not require application, early warning monitoring areas, light application areas, moderate application areas, and heavy application areas.

Images