Vegetable disease and pest prevention and control decision-making method and system based on multi-source data

Abstract

The invention relates to the technical field of agricultural information, in particular to a vegetable pest prevention and control decision-making method and system based on multi-source data. The method comprises the following steps: acquiring data of sound signals, electrochemical signals, spectral characteristics and microenvironment parameters related to vegetable diseases and insect pests; the data is preprocessed; extracting acoustic characteristic parameters and electrochemical characteristic parameters; obtaining a comprehensive feature vector; constructing a dynamic system model; acquiring time-space infection density of diseases and insect pests; constructing a topology perception prevention and control decision model; and generating a prevention and control decision in combination with an improved topological optimization algorithm. According to the method, acoustic and electrochemical data are considered, and an improved signal noise reduction and feature extraction algorithm is combined, so that a disease and pest dynamic propagation model is constructed; through introduction of topological persistence analysis and multi-scale optimization decision, accurate prediction and adaptive prevention and control of time-space infection density of diseases and insect pests are realized, the identification precision is significantly improved, the pesticide application amount is reduced by 28% or more, and perception-decision closed-loop intelligent prevention and control is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of agricultural information technology, in particular to a vegetable disease and pest control decision-making method and system based on multi-source data. BACKGROUND

[0002] With the rapid development of smart agriculture, accurate identification and efficient prevention and control of crop diseases and pests have become a key link to ensure agricultural production. Traditional disease and pest monitoring methods mainly rely on manual inspection and single image recognition technology, which has the problems of low recognition accuracy, great influence of environment and excessive use of pesticides. In the prior art, although there are fusion analysis methods based on unmanned aerial vehicle multi-spectral images, infrared thermal imaging and Internet of Things sensing data, there are still limitations such as single data modality, insufficient feature extraction and weak model generalization ability, which makes it difficult to achieve efficient and accurate disease and pest identification and prevention and control decision-making in complex farmland environment.

[0003] At present, the comparative document CN120339889A discloses a disease and pest identification and control method based on multi-modal edge computing, which fuses multi-spectral images, infrared thermal imaging and environmental sensing data, and uses a Transformer attention mechanism and a YOLOv7-Spectral model for target detection. However, this method still mainly uses optical and thermal imaging in data modality, lacks comprehensive utilization of new types of sensing data such as acoustic and electrochemical data; in terms of feature extraction, the spatio-temporal coupling characteristics of multi-dimensional signals are not fully considered; in terms of model construction, physical mechanisms and topological structure information are not effectively embedded, resulting in insufficient robustness and adaptability of the system in noisy interference and dynamic changing environment.

[0004] In view of the above-mentioned deficiencies, the present application introduces multi-source heterogeneous data of acoustic signals, electrochemical signals, spectral features and micro-environment parameters, and constructs an integrated analysis framework based on improved signal processing algorithms and multi-dimensional feature fusion; by introducing physical information neural networks and topological persistence analysis, high-precision modeling and spatial structure perception of the dynamic propagation process of diseases and pests are realized, and combined with multi-scale topological optimization and Kalman filter feedback mechanism, a closed-loop adaptive prevention and control path from perception to decision-making is formed. SUMMARY

[0005] In view of the defects in the prior art, the present application provides a vegetable disease and pest control decision-making method and system based on multi-source data.

[0006] In the first aspect, the present application provides a vegetable disease and pest control decision-making method based on multi-source data, which acquires data of sound signals, electrochemical signals, spectral features and micro-environment parameters related to vegetable diseases and pests; The data is preprocessed to obtain a preprocessing result; acoustic feature parameters and electrochemical feature parameters are extracted based on the preprocessing result; a comprehensive feature vector is obtained according to the acoustic feature parameters and the electrochemical feature parameters; a dynamic system model is constructed according to the comprehensive feature vector; the spatiotemporal infection density of the disease and insect pests is obtained by using the dynamic system model; a topological perception prevention and control decision model is constructed according to the spatiotemporal infection density of the disease and insect pests; and a vegetable disease and insect pest prevention and control decision is generated based on the topological perception prevention and control decision model and in combination with an improved topological optimization algorithm.

[0007] Optionally, the preprocessing of the data to obtain a preprocessing result includes: using an improved ensemble empirical mode decomposition-adaptive perfect de-noising algorithm to perform de-noising processing on the sound signal data to obtain an acoustic de-noising processing result; using an ensemble empirical mode decomposition and morphological filtering joint algorithm to perform baseline correction and de-noising processing on the electrochemical signal data to obtain a baseline correction and de-noising processing result; using an improved adaptive wavelet threshold de-noising algorithm to perform de-noising processing on the spectral feature data to obtain an optical de-noising processing result; and performing standardization processing on microenvironment parameter data by establishing a non-cyclic parameter standardization model and a cyclic parameter standardization model to obtain a standardization processing result, the microenvironment parameter data including data of temperature, humidity, light intensity, carbon dioxide concentration, volatile organic compound concentration, carbon monoxide concentration, environmental noise level, wind speed, wind direction, and electromagnetic radiation.

[0008] Optionally, based on the preprocessing result, the acoustic feature parameters and the electrochemical feature parameters are extracted. Based on the acoustic de-noising processing result, an acoustic feature extraction model is constructed; a multi-dimensional folding and fusion model is constructed through the acoustic feature extraction model; the acoustic feature parameters are extracted according to the multi-dimensional folding and fusion model; based on the optical de-noising processing result, a fractional radial basis encoder and a loss function are constructed; and the electrochemical feature parameters are extracted through the fractional radial basis encoder and the loss function.

[0009] Optionally, based on the acoustic de-noising processing result, the acoustic feature extraction model is constructed, including: reconstructing a single-channel signal into an acoustic matrix containing scale, time, and virtual sensor dimensions; based on the acoustic matrix, performing collaborative multi-dimensional decomposition using three orthogonal wavelet bases to form a decomposition tensor; and according to the decomposition tensor, performing coefficient screening and energy pooling by introducing a non-linear function of inter-subband energy correlation to generate an initial feature map that effectively captures signal time-space-frequency characteristics.

[0010] Optionally, the obtaining the comprehensive feature vector according to the acoustic feature parameter and the electrochemical feature parameter comprises: constructing an adversarial auto-encoder model; performing feature fusion and enhancement by using the adversarial auto-encoder model in combination with the acoustic feature parameter and the electrochemical feature parameter to obtain enhanced acoustic-electric fusion features; constructing a comprehensive feature vector model based on the acoustic-electric fusion features in combination with preprocessed spectral features and micro-environment parameter data; and obtaining the comprehensive feature vector by using the comprehensive feature vector model.

[0011] Optionally, a dynamic propagation model of the pest and disease is constructed according to the comprehensive feature vector, and the dynamic propagation model of the pest and disease satisfies the following expression:

[0012] wherein, is a pest and disease infection density, which is a spatiotemporal dynamic prediction target variable, is a two-dimensional spatial coordinate, is a time variable, is a characteristic dependence of a diffusion coefficient, is a characteristic modulation of an infection rate, is a crop characteristic dependence of a carrying capacity, is a nonlinear response term, is a mortality function, is the comprehensive feature vector.

[0013] Optionally, the constructing a topological perception prevention and control decision model according to the spatiotemporal infection density of the pest and disease comprises: calibrating parameters of the dynamic system model by using a physical information neural network to obtain a calibration result; constructing a multi-scale topological descriptor according to the calibration result and the spatiotemporal infection density of the pest and disease; constructing a topological persistence image according to the multi-scale topological descriptor; extracting a multi-scale topological feature by using the multi-scale topological descriptor and the topological persistence image; and constructing a topological perception prevention and control decision model based on the multi-scale topological feature.

[0014] Optionally, the topological perception prevention and control decision model is a random optimal control problem with a topological constraint, as follows:

[0015] satisfying the constraint:

[0016] wherein, denotes a prevention and control intensity at a position and a time , including a pesticide application amount, is a prevention and control efficiency function, a spatial white noise, , respectively are infection loss, prevention and control cost, prevention and control smoothness weight, is a terminal cost function, which represents an additional cost calculated according to the final infection state at the end of the decision cycle, is a spatial diffusion coefficient of pest individuals, is an intrinsic growth rate, is an environmental carrying capacity, is a natural mortality rate, is a nonlinear response term, is a noise intensity function.

[0017] Optionally, based on the topological awareness prevention and control decision model, a vegetable pest prevention and control decision is generated in combination with an improved topological optimization algorithm, including: based on the topological awareness prevention and control decision model, an improved multi-scale topological optimization algorithm is introduced to obtain an optimized prevention and control strategy; according to the prevention and control strategy, a Kalman filtering-topological fusion estimation algorithm is combined to update the state of the prevention and control strategy, and an adaptive vegetable pest prevention and control decision is obtained.

[0018] In a second aspect, the present application provides a vegetable pest prevention and control decision system based on multi-source data, comprising an input device, a processor, an output device and a memory, the input device, the processor, the output device and the memory are connected with each other, wherein the memory is used for storing a computer program, the computer program comprises program instructions, the processor is configured to call the program instructions, and the system uses the vegetable pest prevention and control decision method based on multi-source data.

[0019] The beneficial effects of the present application are as follows: (1) Multi-source data deep fusion improves recognition accuracy and robustness. By introducing multi-modal data of acoustics, electrochemistry, spectrum and microenvironment, combining improved signal denoising and feature extraction algorithm, the discrimination ability of pest characteristics and the anti-interference ability of the system are effectively improved.

[0020] (2) Physical mechanism and topological structure embedding realize high-precision dynamic modeling. A physical information neural network is used to construct a pest propagation dynamic model, and a topological persistence image and multi-scale topological features are combined to realize accurate prediction and structure perception of the spatio-temporal evolution of infection density, and the explanation and prediction ability of the model are significantly improved.

[0021] (3) Adaptive closed-loop decision-making optimizes prevention and control efficiency and resource utilization. Through the topological perception prevention and control decision-making model and the improved multi-scale topological optimization algorithm, combined with the Kalman filter-topological fusion estimator, dynamic adjustment and spatial precise control of the prevention and control strategy are realized, which significantly reduces the amount of pesticide use and improves the economic efficiency and environmental friendliness of the system while ensuring the prevention and control effect.

[0022] (4) End-to-end intelligent system promotes the upgrading of smart agriculture. By building a complete technology chain from data acquisition, signal processing, feature fusion, dynamic modeling to decision output, the whole process of intelligent and adaptive optimization of pest control is realized, providing a scalable and expandable system solution for smart agriculture. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of a vegetable pest control decision-making method based on multi-source data according to an embodiment of the present application; Figure 2 A structural schematic diagram of a vegetable pest control decision-making system based on multi-source data according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] The specific embodiments of the present application will be described in detail below. It should be noted that the embodiments described herein are only for illustration and do not limit the present application. In the following description, a large number of specific details are set forth in order to provide a thorough understanding of the present application. However, it is obvious to those skilled in the art that the present application does not have to be implemented with these specific details. In other instances, well-known circuits, software or methods are not specifically described in order not to obscure the present application.

[0025] Throughout the specification, references to "one embodiment", "an embodiment", "one example" or "an example" mean that a particular feature, structure, or characteristic described in connection with the embodiment or example is included in at least one embodiment of the application. Therefore, the appearance of the phrases "in one embodiment", "in an embodiment", "one example" or "an example" in various places throughout the specification are not necessarily all referring to the same embodiment or example. In addition, specific features, structures, or characteristics can be combined in any appropriate combination and / or subcombination in one or more embodiments or examples. In addition, those skilled in the art should understand that the diagrams provided herein are for illustrative purposes only and are not necessarily drawn to scale.

[0026] Please refer to Figure 1 Embodiments of the present application provide a vegetable pest control decision-making method based on multi-source data, which comprises the following steps: S1. Obtain data of sound signals, electrochemical signals, spectral features and microenvironment parameters related to vegetable pests.

[0027] wherein S1 further comprises the following steps: S11. Obtain data of sound signals and electrochemical signals related to vegetable diseases and insect pests.

[0028] In an embodiment, first, a high-sensitivity recording device is selected and deployed in the vegetable planting area to ensure coverage of different growth periods and high-incidence periods of diseases and insect pests, real-time collection of air-borne sound waves generated by insect pest activities as sound signals, and recording of environmental background noise as a control. Human interference should be avoided and the device should be kept stable during collection.

[0029] Further, an electrochemical sensor is inserted into the soil around the vegetable root system or the surface of the leaf to directly measure the electrochemical signals generated by the plant stress response, including changes in electrolyte concentration and pH fluctuations caused by diseases and insect pests. The insertion depth and contact area of the sensor should be controlled during collection to ensure data accuracy.

[0030] S12. Obtain data of spectral characteristics and microenvironment parameters related to vegetable diseases and insect pests.

[0031] In an embodiment, first, a hyperspectral imager is used to scan the vegetable leaves and stems under natural light conditions, covering the visible light to near-infrared band, ranging from 400 ~1000 nm, to obtain the changes in chlorophyll content, abnormal water distribution, or characteristic spectral reflection peaks of pathogenic bacteria caused by diseases and insect pests. The shooting angle and distance should be fixed during collection to avoid shadow interference.

[0032] Further, a multi-sensor integrated system is used to collect data of microenvironment parameters, including but not limited to temperature and humidity sensors, and the microenvironment parameters include temperature, humidity, light intensity, carbon dioxide concentration, volatile organic compound concentration, carbon monoxide concentration, environmental noise level, wind speed, wind direction, and electromagnetic radiation.

[0033] Specifically, by combining high-precision thermistors and capacitive humidity sensing elements, environmental temperature and relative humidity values are monitored and recorded in real time, with a sampling frequency of once per second to ensure data continuity and accuracy; at the same time, a photosensitive sensor is integrated to accurately measure the ambient light intensity using the photoelectric effect principle, with a sampling interval of every 30 seconds to capture subtle differences in light changes.

[0034] Further, a gas concentration sensor array is introduced, including but not limited to carbon dioxide sensors, volatile organic compound sensors, and carbon monoxide sensors, using non-dispersive infrared technology, metal oxide semiconductor technology, and electrochemical sensor technology respectively to monitor the concentration of various gas components in the air in real time, with a sampling frequency of once per minute.

[0035] Further, an integrated wind speed and direction sensor is added, which combines a three-cup anemometer with a wind vane to measure wind speed and direction in real-time, with a sampling frequency of once per minute, providing key data for understanding the impact of air flow on the microenvironment.

[0036] Further, an infrared thermal imager is added, which captures the temperature distribution on the surface of objects through non-contact means, forming thermal images that indirectly reflect the presence and activity of living organisms, with a sampling frequency set to once every 10 minutes.

[0037] Further, an electromagnetic radiation detector is introduced, which measures various types of electromagnetic radiation including radio frequency radiation, microwave radiation, and low-frequency electromagnetic fields, with a sampling frequency set to once per minute to ensure timely response to changes in the electromagnetic environment.

[0038] S2. Preprocess the data to obtain a preprocessing result.

[0039] Wherein, S2 further comprises the following steps: S21. Use the improved ensemble empirical mode decomposition-adaptive complete noise algorithm to denoise the sound signal data to obtain an acoustic denoising result.

[0040] Since the existing ensemble empirical mode decomposition algorithm relieves the mode aliasing problem by adding Gaussian white noise multiple times, in an embodiment, two key improvements are made, one is adaptive noise intensity, and the other is mode termination criterion, i.e. using dynamic noise amplitude and introducing an early termination condition based on spectral skewness.

[0041] Specifically, instead of using a fixed noise amplitude, the noise amplitude added for the first time is dynamically adjusted according to the instantaneous amplitude characteristics of the input signal :

[0042] Wherein, is a global adjustment coefficient, set to 0.2, is the signal length, is the Hilbert transform, represents the instantaneous envelope of the signal, is the signal sequence.

[0043] The advantage of this method is that the noise intensity is associated with the dynamic range of the signal itself, relatively strong noise is injected at weak signals to better extract details, and relatively weak noise is injected at strong signals to avoid excessive pollution.

[0044] ​​Further, an early termination condition based on spectral skewness is introduced. When the spectral skewness of a certain IMF component is lower than a pre-defined threshold , the following components are considered to contain no valid oscillatory mode and the decomposition is stopped. This effectively avoids meaningless low-frequency components caused by over-decomposition and improves the computational efficiency. The spectral skewness satisfies the following expression:

[0045] where is the power spectrum of the IMF component, and are the mean and standard deviation of the power spectrum, respectively, and the pre-defined threshold is an empirical value obtained from a large number of pure speech / sound samples.

[0046] Further, the improved EEMD algorithm decomposes the original signal into:

[0047] where is the original signal, is the th IMF component, is the residual component, is time, and is the set of IMF components.

[0048] Further, for the adaptive complete noise algorithm, a noise energy contribution rate is defined for each IMF component. The th noise energy contribution rate satisfies the following expression:

[0049] where is the expectation operator, is the intensity factor of the correlation penalty, which controls the weight, , are the shape and position parameters of the modal excitation, is the cross-correlation coefficient between the original signal and the IMF component , is the cross-correlation coefficient between the original signal and the IMF component , , is the index of the IMF component, is the total number of IMFs, indicates the contribution of the The result after soft threshold processing satisfies the following expression:

[0050] wherein, is a symbol function, is an initial threshold value calculated using an existing threshold function, which measures the information loss caused by direct threshold processing. The greater the loss, the more effective information the component contains, and the smaller the noise contribution.

[0051] Further, the global sensitivity coefficient is calculated as and the final processing threshold value for each eigenmodal component is calculated as , and its expression is as follows:

[0052] wherein, is a global sensitivity coefficient, which is adjusted according to the signal-to-noise ratio estimate.

[0053] Further, each eigenmodal component is processed using an improved continuous semi-soft threshold function to obtain the denoised component , which satisfies the following expression:

[0054] wherein, is a transition factor that controls the smoothness. The improvement of this function lies in the introduction of a smooth transition zone near the threshold, which realizes a natural transition from noise suppression to signal preservation, and greatly preserves weak but effective acoustic features such as friction and explosion sounds.

[0055] Further, all processed eigenmodal function components are added to the residual component to obtain the final sound denoised signal .

[0056] The method has the advantages of constructing a multi-dimensional noise evaluation model that fuses reconstruction error, signal correlation, and modal order priori, cooperating with an adaptive decomposition termination mechanism based on spectral skewness, and adopting a continuous semi-soft threshold function with an S-shaped transition zone for closed-loop feedback threshold adjustment, which realizes accurate separation of noise and signal features, efficiently denoises, and significantly improves the preservation of sound details.

[0057] S22. Utilize the set empirical mode decomposition and morphological filtering joint algorithm to perform baseline correction and denoising processing on the electrochemical signal, to obtain the baseline correction and denoising processing result of the electrochemical signal data.

[0058] In one embodiment, the original electrochemical signal is first subjected to ensemble empirical mode decomposition to obtain a series of intrinsic mode function components from high frequency to low frequency.

[0059] Further, a joint criterion based on information entropy and cross-correlation coefficient is introduced to automatically identify noise-dominant intrinsic mode function components, the identification function of which is as follows:

[0060] wherein, is the noise index of the th intrinsic mode function component, is the information entropy of the th intrinsic mode function component, is the cross-correlation coefficient of the original signal and the intrinsic mode function component, is time, is the standard deviation of the cross-correlation coefficient of all intrinsic mode function components and the original signal, is the index of the intrinsic mode function component. Further, morphological joint processing based on multi-scale structural elements is performed.

[0061] Specifically, a multi-scale adaptive morphological filtering model is constructed, the expression of which is as follows:

[0062]

[0063] wherein, is the processed intrinsic mode function component, represents closed operation followed by open operation with structural element SE , represents open operation followed by closed operation with structural element SE , is the signal threshold value, is the noise threshold value, is the noise index of the th intrinsic mode function component, is the intrinsic mode component, is time, is the structural element adaptively selected according to the oscillation characteristics of the th intrinsic mode function component.

[0064] Further, polynomial fitting is performed on the initial baseline estimate to obtain the final smoothed baseline .

[0065] ​​Furthermore, the final smoothed baseline is subtracted from the original signal, and the processed intrinsic mode function components are reconstructed with the baseline-corrected signal to obtain the final corrected and denoised signal. :

[0066] in, These are the eigenmode function components dominated by the signal.

[0067] The advantages of this method lie in its ability to achieve accurate and automatic identification of noise components through a combination of empirical mode decomposition and information entropy-cross-correlation coefficient joint criteria. By constructing a multi-scale adaptive morphological filtering model, the noise index is dynamically embedded into the morphological operation selection logic. When the noise index is within the range of the signal and noise threshold, the weighted fusion of closing-open and opening-closing operations is automatically triggered, effectively avoiding signal distortion or noise residue caused by a single structural element, and achieving conformal denoising of weak electrochemical features in a strong noise background. By combining baseline correction technology based on polynomial fitting, the interference of baseline drift on signal reconstruction is eliminated. Through the collaborative reconstruction of the dominant signal component and the baseline correction signal, the signal-to-noise ratio and waveform fidelity are significantly improved, providing reliable technical support for high-precision electrochemical detection.

[0068] S23. An improved adaptive wavelet threshold denoising algorithm is used to denoise the spectral feature data to obtain the optical denoising result.

[0069] In one embodiment, the spectral feature data is first decomposed using wavelet decomposition to obtain multi-scale wavelet coefficients, and the statistical characteristics of the wavelet coefficients at each scale are calculated, including local variance. Noise variance and local signal-to-noise ratio ,in, Indicates scale index.

[0070] Furthermore, wavelet coefficient processing is performed based on an improved adaptive threshold function, which is as follows:

[0071] in, For the processed first Scale wavelet coefficients, These are the original wavelet coefficients. Based on the threshold, This is the local signal-to-noise ratio adjustment factor. For adaptive weights, Adjusting parameters for local entropy As a noise suppression factor, Adjust parameters for kurtosis skewness. This is a dynamic threshold.

[0072] The improvement of the improved adaptive wavelet threshold denoising algorithm lies in the construction of a new nonlinear fractional threshold function and a multi-dimensional adaptive mechanism. First, the fractional structure with dynamic saturation characteristics is adopted to replace the traditional soft and hard threshold function, realizing continuous fine adjustment of the wavelet coefficients. Second, the threshold parameters are deeply coupled with the local statistical characteristics of each scale, such as local signal-to-noise ratio, information entropy, kurtosis and skewness, so that the algorithm has accurate situational awareness capability. Third, by introducing a periodic structure perception term , the sinusoidal function is used to detect the signal hiding structure in the wavelet domain, realizing the intelligent filtering breakthrough from amplitude dependence to signal form combination.

[0073] Further, the wavelet coefficients after processing are reconstructed to obtain the denoised spectral data, and the denoising effect is evaluated to ensure that the optical denoising processing result effectively removes noise while preserving spectral characteristics.

[0074] The method has the advantages that the dynamic saturation characteristic fundamentally solves the inherent contradiction between denoising and signal fidelity, and can nearly distortionlessly preserve strong signals representing real spectral characteristics; the multi-dimensional adaptive mechanism ensures that the algorithm can achieve optimal signal-to-noise separation in different spectral regions, such as smooth base and sharp peak, and improves the overall denoising effect. It should be noted that the periodic perception term can effectively protect the fine structure and weak features of the spectrum, avoid being filtered out as noise, and thus exhibit higher peak signal-to-noise ratio and better visual fidelity in the output result.

[0075] S24. Standardizing the microenvironment parameter data to obtain a standardized result of the microenvironment parameter data.

[0076] In one embodiment, each parameter is standardized using a formula with a cycle adjustment. The microenvironment parameters are divided into non-cyclic parameters and cyclic parameters, wherein the wind direction parameter is a cyclic parameter and the remaining parameters are non-cyclic parameters.

[0077] Specifically, a non-cyclic parameter standardization model formula and a cyclic parameter standardization model formula are established, the non-cyclic parameter standardization model formula is as follows:

[0078] wherein, is an original parameter value, is a non-cyclic parameter standardized value, is a parameter standard deviation, is a parameter mean value; and the cyclic parameter standardization model formula is as follows:

[0079] wherein, a circular parameter standardized value, a wind direction arc value, a mean wind direction arc, a circular standard deviation.

[0080] Further, the standardized processing results of all micro-environment parameters are obtained to form a standardized data set.

[0081] The method has the advantages that by introducing the circular statistical processing of wind direction data, the inapplicability of the existing Z-score to the circular parameter is solved; by embedding an exponential weight term in the formula, the standardization process is dynamically adjusted, the sensitivity to extreme values is reduced, and the robustness is improved; by uniformly processing linear and circular parameters, more accurate comprehensive standardized results are generated, which are suitable for complex micro-environment analysis.

[0082] S3. Extracting acoustic feature parameters and electrochemical feature parameters based on the preprocessing results.

[0083] S3 further includes the following steps: S31. Constructing an acoustic feature extraction model based on multi-dimensional wavelet packet transform based on the acoustic denoising processing results.

[0084] In one embodiment, an acoustic feature extraction model based on multi-dimensional wavelet packet decomposition is established, which first reconstructs a single-channel signal into an acoustic matrix containing scale, time and virtual sensor dimensions, then performs collaborative multi-dimensional decomposition using three orthogonal wavelet bases to form a decomposition tensor. Finally, by introducing a nonlinear function of inter-subband energy correlation for coefficient screening and energy pooling, an initial feature map is generated which can effectively capture the time-space-frequency characteristics of the signal.

[0085] Specifically, first, the single-channel signal after acoustic denoising is reconstructed into a multi-dimensional acoustic matrix satisfying the following expression:

[0086] wherein, is a scale modulation factor, is a time frame index, is a virtual sensor index, is a mother wavelet function, , , , is a space-time-frequency coupling parameter, respectively, a basic time interval, a scale-dependent delay, a delay between virtual sensors, and a frequency modulation coefficient, used to control the stretching and twisting of the multi-dimensional structure.

[0087] Further, the multi-dimensional matrix a multi-dimensional decomposition tensor is formed by using different wavelet basis functions simultaneously for wavelet packet decomposition in each dimension , whose expression is as follows:

[0088] wherein, , , is a multi-dimensional decomposition level, respectively corresponding to the scale, time and virtual space dimensions; , 、 is a multi-dimensional position index, 、 、 is a collaborative wavelet packet basis, which is three different orthogonal wavelet basis functions for capturing the characteristics of the signal in different dimensions.

[0089] Further, a nonlinear function based on the energy correlation between subbands is used for coefficient screening and energy pooling for each subband of the decomposition tensor to form an initial feature map .

[0090]

[0091] wherein, is an adaptive gain factor, which controls the steepness of the nonlinear function, is the standard deviation of the subband coefficients, is a smoothing constant, is the correlation coefficient between subbands, which is used to measure the energy correlation between different subbands, is a multi-dimensional decomposition level.

[0092] S32. Extracting acoustic feature parameters through the acoustic feature extraction model.

[0093] In one embodiment, the initial feature map is folded and fused in different dimensions to generate a compact and information-rich joint feature vector . This process constructs a multi-dimensional folding and fusion model, which satisfies the following expression:

[0094] wherein, is the index of the feature vector, is an entropy weight coefficient, is a two-dimensional entropy function, which calculates the information entropy of all time-space subbands at this scale, is a gradient operator along the scale dimension for capturing the multi-scale variation trend of the feature, is a stability constant, are maximum levels of the time and spatial dimensions, respectively, is a coordinate mapping symbol, is the time dimension, is the spatial dimension, is an output index maps back to an effective index of the time dimension .

[0095] Further, according to the multi-dimensional folding and fusion model, the calculated vector is the final extracted acoustic feature parameter.

[0096] The method has the advantages that by constructing a virtual sound field dimension, a one-dimensional signal is upgraded to multi-dimensional analysis, the spatial scattering characteristics hidden in the signal can be captured; by adopting a plurality of wavelet bases for collaborative decomposition and combining sub-band correlation for adaptive pooling, joint extraction of cross-dimensional features is realized, and the representation robustness is significantly improved; by introducing a nonlinear fusion mechanism with information entropy as the weight, the signal complexity and structural changes are encoded in the feature vector, deep acoustic fingerprints that are difficult to obtain by existing methods are generated, and the feature discrimination is stronger.

[0097] S33. Based on the optical noise reduction processing result, an electrochemical feature extraction model based on nonlinear principal component analysis is constructed.

[0098] In one embodiment, first, the electrochemical data matrix after optical noise reduction is input, and noise variance estimation is obtained.

[0099] Further, a fractional radial basis encoder is constructed, which outputs feature The calculation formula of the feature is as follows:

[0100] wherein, , are weights, is input feature data, , are basis function centers, , are width parameters, is a denominator regularization constant, is the number of radial basis functions, is the index of the basis function.

[0101] Further, optimization training is performed, wherein a weighted loss function is minimized as follows:

[0102] wherein, is the total number of samples, is the sample index, is the variance of the th sample, is the th raw electrochemical sample data, is the encoder function that maps the raw data to the feature space, is the decoder function that reconstructs the raw data from the features, is the regularization coefficient, is the feature dimension, is the feature index, is the th feature vector.

[0103] It is noted that the optimization training process is as follows: the inverse variance weighted reconstruction error is used to ensure that the high-variance sample weight is low; at the same time, the weighted regularization term is applied to punish the absolute value of the feature vector , promoting feature sparsity; the gradient descent method is used to jointly optimize the encoder and decoder parameters, balancing data reconstruction accuracy and feature simplicity.

[0104] It is noted that the fractional radial basis encoder and its corresponding optimization training process together constitute the electrochemical feature extraction model.

[0105] S34. Extracting electrochemical feature parameters through the electrochemical feature extraction model.

[0106] In one embodiment, the preprocessed electrochemical data is input into the trained electrochemical feature extraction model, and the feature vector is directly extracted through the encoder.

[0107] The method has the advantages that by using the fractional radial basis function structure, the ability to capture electrochemical complex nonlinear features is significantly enhanced, end-to-end optimization of noise reduction and feature extraction is realized, and the accuracy and robustness of feature extraction are improved.

[0108] S4. Obtaining a comprehensive feature vector according to the acoustic feature parameters and the electrochemical feature parameters.

[0109] ​The comprehensive feature vector is a discriminative representation that integrates multi-source information. Robust encoding is achieved through a dynamically mean-adjusted and periodically activated adversarial autoencoder. The fusion of acoustic and electrochemical features is enhanced by a gated cross-attention mechanism. Spectral and microenvironment parameters are effectively integrated using multi-head attention. Finally, layer normalization and nonlinear transformation are used to improve the discriminability of feature representation and the generalization ability of the system.

[0110] In one embodiment, an adversarial autoencoder is first constructed, which realizes robust feature encoding and enhanced reconstruction through dynamic mean adjustment and periodic activation functions.

[0111] Further, the acoustic features and the electrochemical features are encoded by the adversarial autoencoder into and , respectively.

[0112] Further, the and are fused using a gated cross-attention mechanism:

[0113] wherein is the temperature parameter controlling the smoothness of the attention distribution, is the gating weight matrix, is the activation function, is the concatenation operation, is the fused feature, is the transpose symbol of the matrix, is the scaling factor, is the feature dimension index.

[0114] Further, the fused feature is enhanced by a nonlinear transformation:

[0115] wherein is the enhanced feature, is the enhancement weight matrix, is the bias vector, is the enhancement intensity coefficient, is a constant.

[0116] Further, a multi-head attention mechanism is introduced to calculate the query vector , the key vector , and the value vector , as follows:

[0117] wherein , 、 is a projection weight matrix, is a pre-processed spectral feature, is a micro-environment parameter.

[0118] Further, the output of multi-head attention is performed :

[0119] wherein, is the number of heads, is the head index, is the dimension of the key, is a mask matrix, is the value vector of the th attention head, is the query vector of the th attention head, is the key vector of the th attention head.

[0120] Further, a comprehensive feature vector model is constructed:

[0121] wherein, is a comprehensive feature vector, is an output weight matrix, is an output projection bias vector, denotes layer normalization, denotes the output symbol of the multi-head attention mechanism.

[0122] The method has the advantages that robust feature coding is realized through the dynamic mean adjustment and the periodically activated adversarial autoencoder; the acoustic and electrochemical feature fusion quality is enhanced by combining the gated cross-attention and the nonlinear transformation; the multi-head attention mechanism is further introduced to effectively integrate the spectrum and the micro-environment parameter, thereby significantly improving the discriminability of feature expression and the system generalization capability.

[0123] S5. A dynamic system model is constructed according to the comprehensive feature vector.

[0124] In one embodiment, first, a physical information neural network is adopted to construct a dynamic system model according to the comprehensive feature vector which is also called a disease and pest dynamic transmission model, and the dynamic system model satisfies the following expression:

[0125] wherein, is the disease and pest infection density, which is the target variable of the spatiotemporal dynamic prediction, ​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​is a two-dimensional spatial coordinate, is a time variable, is a characteristic dependence of the diffusion coefficient, is a characteristic modulation of the infection rate, is a characteristic dependence of the carrying capacity of the crop, is a nonlinear response term, is a mortality function, is a comprehensive characteristic vector.

[0126] It should be noted that the dynamic system model characterizes the distribution and evolution of the disease and pests in space-time through the infection density, and its composition includes a diffusion term describing the spatial propagation of the disease and pests, a growth term simulating the process of resource-limited growth of the disease, a mortality term reflecting the natural decay of the disease, and a nonlinear response term depicting complex interactive effects. All model parameters are dynamically modulated by a comprehensive characteristic vector, thereby realizing data-driven accurate prediction.

[0127] S6. Obtain the spatio-temporal infection density of the disease and pests by the dynamic system model.

[0128] In one embodiment, the spatio-temporal infection density of the disease and pests is obtained by the dynamic system model according to the evolution of According to the spatio-temporal infection density, the spatio-temporal prediction result of the spread of the disease and pests can be further obtained.

[0129] In a specific case of cucumber downy mildew, a cucumber field of m is set as the prediction area, the spatial coordinates , the simulation time span is 30 days ( days), and the prediction target variable is the lesion density on the cucumber leaves (unit: lesion number / cm2). Based on the comprehensive characteristic vector which includes static environmental characteristics such as soil type and planting density of the area, and dynamic characteristics such as hourly temperature and humidity spatio-temporal data, the model parameters are specifically set as follows: the diffusion coefficient (m2 / day), which is used to simulate the diffusion capacity change of the pathogen under day and night and weather cycle; the infection rate (per day), wherein is the relative humidity, indicating that the infection rate increases significantly when the humidity remains above 80%; the environmental carrying capacity (lesion number / cm2), which is determined by the crop variety and health status; the mortality rate (per day), wherein is the temperature (℃), indicating that the mortality rate of the pathogen increases when the temperature exceeds 30℃; the nonlinear response term (disease spot number per square centimeter per day) to describe the growth inhibition due to leaf resource competition or plant defense response under high infection density. An initial infection source with a radius of 5 meters is set at the center of the region at the initial moment, and the initial infection density is set as 1 disease spot per square centimeter. (disease spot number per square centimeter), is the coordinate of the center. By solving the dynamic system model, the spatio-temporal evolution map of the disease infection density in the next 30 days can be obtained, and the spatio-temporal prediction result of the spread can be obtained. Specifically, it is predicted that the disease will spread to the east side boundary in the southeast direction on the 10th day, and the infection density of about 60% of the whole field will exceed 5 disease spots per square centimeter on the 20th day, which provides a decision basis for precise pesticide application and prevention and control.

[0130] S7. Construct a topological perception prevention and control decision model according to the spatio-temporal infection density of the disease and pest.

[0131] In one embodiment, the parameters of the dynamic system model are calibrated by a physical information neural network to obtain a calibration result; the optimization function of the calibration is as follows:

[0132] wherein, is the parameter to be learned of the physical information system, including the parameters obtained in step S1, used to fit functions such as , , , , , is a diffusion coefficient, is an infection rate, is an environmental carrying capacity, which depends on the characteristics of the vegetable, is a natural mortality function of the disease and pest individuals, is a nonlinear response term, including human and natural factors, is a comprehensive feature vector, , is a regularization coefficient.

[0133] Further, according to the optimization function, a real-time calibration result is obtained.

[0134] Further, according to the calibration result, combined with the spatio-temporal infection density of the disease and pest, multi-scale topological features are extracted.

[0135] It should be noted that the multi-scale is reflected by systematically observing the topological features such as the complete life cycle from birth to persistence to death of holes and connected regions through continuously changing infection density thresholds, so as to capture the spatial pattern from local hotspots to global distribution.

[0136] ​Specifically, firstly, a multi-scale topology descriptor is constructed, wherein the multi-scale topology descriptor... As shown below:

[0137] in, Indicates the first Birth time of dimensional topological features Indicates the first The death time of 3D topological features Indicates the first The number of dimensional topological features, The index of the topological feature.

[0138] Further, a topologically persistent image is constructed, wherein the topologically persistent image The expression is as follows:

[0139] in, The dimension of the topological feature. The lifetime range for topological features. For the first Persistent graphs with 3D topological features Maintaining the durability weight function, This represents the projection of the midpoint of the characteristic life interval onto the birth axis. For the persistence of features, , For coordinates in a persistent image, denoted as the standard deviation of the Gaussian kernel.

[0140] Furthermore, multi-scale topological features are extracted based on the multi-scale topological descriptor and the topological persistence image. The multi-scale topological descriptor is a feature representation constructed based on persistent homology theory. It characterizes the multi-scale topological structure of data by recording the "birth" and "death" times of topological features, such as connected components and holes, at different scales. This descriptor systematically summarizes the lifetime information of topological features across various dimensions. The topological persistence image maps the lifetimes of the aforementioned topological features to a two-dimensional image space, using the birth time and persistence of the features (i.e., the difference between the birth and death times) as coordinates. A Gaussian kernel function and a weighting function are used to transform discrete topological features into a continuous grayscale image. This visual representation more intuitively encodes the distribution and importance of topological features.

[0141] Taking cucumber images as an example, multi-scale topological features are extracted: First, a 0-2 dimensional topological descriptor is constructed, and the birth time of each feature is recorded. and time of death Then, a topologically persistent image is generated for each dimension. feature points Finally, the multi-scale topological features of cucumber are obtained by fusing the persistence images of three dimensions into a Gaussian-weighted image.

[0142] Further, based on the multi-scale topological features, a topological-aware prevention and control decision model is constructed.

[0143] Specifically, the prevention and control decision is modeled as a stochastic optimal control problem with topological constraints:

[0144] The constraints are satisfied:

[0145] wherein, represents the prevention and control intensity at the position and the time , including the amount of pesticide applied, is a prevention and control efficiency function, is a spatial white noise, , are the infection loss, prevention and control cost, and prevention and control smoothness weight, respectively, is a terminal cost function, representing an additional cost calculated according to the final infection state at the end of the decision period, is a spatial diffusion coefficient of pest individuals, is an intrinsic growth rate, is an environmental carrying capacity, is a natural mortality rate, is a nonlinear response term, is a noise intensity function.

[0146] S8. Based on the topological-aware prevention and control decision model, a vegetable pest prevention and control decision is generated in combination with an improved topological optimization algorithm.

[0147] S8 further includes the following steps: S81. Based on the topological-aware prevention and control decision model, an improved multi-scale topological optimization algorithm is introduced to obtain an optimized prevention and control strategy.

[0148] In one embodiment, based on the topological-aware prevention and control decision model, an improved multi-scale topological optimization algorithm, also known as a novel topological gradient descent algorithm, is introduced to embed topological features into the gradient update of the prevention and control decision, specifically as follows:

[0149] wherein, is the prevention and control strategy function at the th iteration, which is a spatially distributed intensity field; is the prevention and control strategy function at the The prevention and control strategy function at the next iteration Step size; Let be the objective function, and let be the total cost; These are the topological gradient weight coefficients; Let the topological gradient represent the prevention and control strategy. How will tiny changes affect the infection field? The topological features are weighted by their importance. The topological gradient term is included. Defined as:

[0150] in, Dimensions representing topological features Represents a topologically persistent image. For the first Maintaining the permanent weighted sum, It is the integral variable.

[0151] Furthermore, obtain the optimized prevention and control strategies. .

[0152] It is important to note that the optimized prevention and control strategies It is a dynamic field of prevention and control intensity that is distributed in space and time. It specifically indicates the optimal control intensity to be implemented at different locations and times, such as testing density and isolation intensity, with the aim of achieving the most effective epidemic topology control with the lowest total cost.

[0153] S82. Based on the control strategy, and combined with the Kalman filter-topology fusion estimation algorithm, update the state of the control strategy to obtain an adaptive vegetable pest and disease control decision.

[0154] In one embodiment, the information obtained in step S8 The system inputs data into the execution system, including intelligent spraying machines and drones, and monitors the control effects in real time.

[0155] Furthermore, the system state is updated using the Kalman filter-topology fusion estimation algorithm, and the update formula is as follows:

[0156] in, for Estimated pest and disease infection density at any given time. for Estimated pest and disease infection density at any given time. This is a prediction function for the physical model, which predicts the current state based on the state at the previous time step and the implemented control strategies. For Kalman gain, is a real-time measurement value, and involves the parameter data acquired in step S1, is an observation function, is a topological feedback intensity coefficient, represents a topological change amount.

[0157] Further, according to the result of the updating, an adaptive vegetable pest and disease prevention and control decision is obtained.

[0158] It should be noted that the pest and disease infection density state estimation value updated in real time based on the Kalman filter-topological fusion estimator can generate a dynamically adjusted adaptive vegetable pest and disease prevention and control decision, which is essentially a continuously optimized space-time-varying prevention and control intensity distribution map. In combination with the multi-scale topological perception optimization algorithm, the random optimal control problem is re-solved, thereby outputting the optimal prevention and control strategy to be executed in the next period. This strategy not only depends on the real-time infection density, but also more sensitively responds to changes in the spatial topological structure of the infection area, such as the appearance of new isolated lesions or the merging of major infection areas. Accordingly, the type, dosage, and timing of pesticide application in different regions are precisely adjusted, ultimately forming a dynamic precision operation scheme that can prospectively respond to the spatial evolution of the epidemic and achieve the optimal balance between suppressing pests and diseases and reducing pesticide use.

[0159] Compared with the prior art, the improvements of this method are as follows: by adopting the method of combining physical information neural networks with topological feature embedding, a more accurate characterization of pest and disease dynamics is achieved; in terms of decision basis, the limitations of threshold-based or rule-based methods are broken through, and multi-scale topological perception stochastic control is introduced, enabling the decision-making process to consider spatial structure and random disturbances; in terms of parameter dynamics, the lag of static or empirically adjusted parameters is overcome, and online adaptive updating of model parameters is achieved through real-time physical information neural network calibration and topological feedback mechanism; in terms of spatial adaptability, the extensive mode of uniform control or simple zoning is completely changed, and non-uniform prevention and control is driven by topological gradients, achieving precise matching of pesticide application strategies and complex spatial infection patterns; in terms of noise resistance, the noise sensitivity of existing technologies is addressed, and the stability of prevention and control in uncertain environments is significantly improved through the fusion of stochastic control theory and topological white noise filtering. These improvements have four positive effects: first, the prevention and control precision is significantly improved by identifying the spatial topological structure, avoiding excessive or insufficient prevention and control; second, the non-uniform precision control can reduce pesticide use by 20% to 30%; third, multi-source information fusion greatly enhances the robustness to meteorological mutations and sensor noise; and fourth, real-time topological evolution and calibration achieve complete closed-loop optimization from perception to decision-making, achieving truly adaptive prevention and control.

[0160] In an actual case of cucumber pest and disease prevention and control, a high-sensitivity microphone array is first deployed in the field to collect aphid feeding sound waves (frequency range ) and leaf electrochemical signals ( fluctuation Simultaneous use of a hyperspectral imager ( ) Capture the chlorophyll absorption valley (680nm) and water stress characteristics caused by downy mildew ( ), combined with temperature and humidity sensors ( Twelve types of microenvironment parameters, including [list of parameters]. An improved ensemble empirical mode decomposition algorithm was used to reduce acoustic signal noise (signal-to-noise ratio improved by 12 dB), and nonlinear principal component analysis was employed to extract a 6-dimensional feature vector from electrochemical data. The fused comprehensive features were input into a physical information neural network dynamic model, outputting a heatmap of powdery mildew infection density for the next 72 hours (spatial resolution of 0.5 m). Based on topological persistence images, three high-risk diffusion hubs were identified, and differentiated application schemes were generated through multi-scale topology optimization: UAV spraying was used in topologically connected regions. ), targeted drip irrigation for isolated lesions ( / plant). After real-time correction using Kalman filtering, pesticide use was reduced by 28% and the control accuracy rate reached 94.2% in three growing season trials.

[0161] Please see Figure 2 , Figure 2 This is a schematic diagram of a vegetable pest and disease control decision-making system based on multi-source data according to an embodiment of the present invention. The system includes an input device, a processor, an output device, and a memory. The input device, processor, output device, and memory are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to call the program instructions. The system uses the aforementioned vegetable pest and disease control decision-making method based on multi-source data.

[0162] In this embodiment, the input device includes an acoustic acquisition unit, an electrochemical sensing unit, an optical acquisition unit, and a microenvironment monitoring unit. The functions of the input device are as follows: synchronously acquiring multimodal data with timestamps accurate to the millisecond level, having a self-calibration function to ensure data accuracy, supporting remote configuration and firmware upgrades, and having a built-in data cache to prevent data loss due to network interruption.

[0163] The output device includes a decision display terminal, an execution control interface, and a data output interface; the functions of the output device are as follows: real-time display of pest and disease risk assessment results, generation of visual operation instructions and prevention and control maps, support for multi-device collaborative control, and provision of audible and visual early warning prompts.

[0164] The processor includes a main processing unit, a coprocessing unit and an interface control unit; the functions of the processor are as follows: parallel processing of multiple sensor data streams, real-time running of complex machine learning algorithms, support for edge computing, reduction of cloud dependence, heat management mechanism to ensure long-term stable operation.

[0165] The memory includes running memory, program storage, data storage and cache memory; the functions of the memory are as follows: hierarchical storage architecture, optimized data access efficiency, support for data compression and encrypted storage, power failure protection mechanism, remote data synchronization and backup.

[0166] In summary, the application effectively improves the discrimination ability of disease and pest characteristics and the anti-interference ability of the system by introducing multi-modal data of acoustics, electrochemistry, spectrum and microenvironment, and combining improved signal noise reduction and feature extraction algorithms; by using a physical information neural network to construct a disease and pest propagation dynamic model, combining a topological persistent image and multi-scale topological features, the precise prediction and structural perception of the spatio-temporal evolution of infection density are realized, and the interpretability and prediction ability of the model are significantly improved; by using a topological perception prevention and control decision model and an improved multi-scale topological optimization algorithm, combining a Kalman filter-topological fusion estimator, dynamic adjustment and spatial precise control of the prevention and control strategy are realized, while ensuring the prevention and control effect, the amount of pesticide used is significantly reduced, and the economy and environmental friendliness of the system are improved; by constructing a complete technical chain from data acquisition, signal processing, feature fusion, dynamic modeling to decision output, the whole process of intelligent and adaptive optimization of disease and pest prevention and control is realized, and a popular and scalable system solution for smart agriculture is provided.

[0167] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the application, and they should be covered in the scope of the claims and the specification of the application.

Claims

1. A decision-making method for vegetable pest and disease control based on multi-source data, characterized in that, The method includes the following steps: To acquire data on sound signals, electrochemical signals, spectral characteristics, and microenvironment parameters related to vegetable diseases and pests; The data is preprocessed to obtain the preprocessing result; Based on the preprocessing results, acoustic and electrochemical characteristic parameters are extracted. Based on the acoustic and electrochemical characteristic parameters, a comprehensive feature vector is obtained; Based on the comprehensive feature vector, a dynamic system model is constructed; Using the aforementioned dynamic system model, the spatiotemporal infection density of pests and diseases can be obtained; Based on the spatiotemporal infection density of the pests and diseases, a topology-sensing prevention and control decision model is constructed. Based on the aforementioned topology-aware prevention and control decision model, and combined with an improved topology optimization algorithm, vegetable pest and disease control decisions are generated.

2. The vegetable pest and disease control decision-making method based on multi-source data according to claim 1, characterized in that, The data is preprocessed to obtain the following preprocessing results: An improved ensemble empirical mode decomposition-adaptive complete noise algorithm is used to denoise audio signal data, resulting in acoustic noise reduction. A combined algorithm of ensemble empirical mode decomposition and morphological filtering is used to perform baseline correction and denoising on electrochemical signal data, and the baseline correction and denoising results are obtained. An improved adaptive wavelet threshold denoising algorithm is used to denoise spectral feature data to obtain optical denoising results. By establishing non-cyclic parameter standardization models and cyclic parameter standardization models, the microenvironment parameter data are standardized to obtain the standardized results. The microenvironment parameter data includes data on temperature, humidity, light intensity, carbon dioxide concentration, volatile organic compound concentration, carbon monoxide concentration, environmental noise level, wind speed, wind direction, and electromagnetic radiation.

3. The vegetable pest and disease control decision-making method based on multi-source data according to claim 1, characterized in that, Based on the preprocessing results, the extraction of acoustic and electrochemical characteristic parameters includes: Based on the acoustic noise reduction processing results, an acoustic feature extraction model is constructed. A multi-dimensional folding and fusion model is constructed using the acoustic feature extraction model. Based on the multi-dimensional folding and fusion model, acoustic feature parameters are extracted; Based on the optical noise reduction processing results, a fractional radial basis encoder and a loss function are constructed. Electrochemical characteristic parameters are extracted using the fractional radial basis encoder and the loss function.

4. The vegetable pest and disease control decision-making method based on multi-source data according to claim 3, characterized in that, Based on the acoustic noise reduction processing results, the acoustic feature extraction model is constructed as follows: The single-channel signal is reconstructed into an acoustic matrix that includes scale, time, and virtual sensor dimensions; Based on the acoustic matrix, three orthogonal wavelet bases are used for cooperative multidimensional decomposition to form a decomposed tensor. Based on the decomposed tensor, coefficients are selected and energy pooling is performed by introducing a nonlinear function of energy correlation between subbands to generate an initial feature map that effectively captures the spatiotemporal-frequency characteristics of the signal.

5. A vegetable pest and disease control decision-making method based on multi-source data according to claim 1, characterized in that, Based on the acoustic and electrochemical characteristic parameters, the comprehensive feature vector is obtained by: Construct an adversarial autoencoder model; Using the adversarial autoencoder model, combined with the acoustic feature parameters and the electrochemical feature parameters, feature fusion and enhancement are performed to obtain enhanced acoustic-electrochemical fusion features; Based on the aforementioned acoustic-electric fusion characteristics, and combined with the preprocessed spectral characteristics and microenvironmental parameter data, a comprehensive feature vector model is constructed. The comprehensive feature vector is obtained through the comprehensive feature vector model; the comprehensive feature vector model adopts a multi-head attention mechanism.

6. The vegetable pest and disease control decision-making method based on multi-source data according to claim 1, characterized in that, Based on the comprehensive feature vector, the dynamic system model is constructed as follows: Based on the comprehensive feature vector, a dynamic transmission model of pests and diseases is constructed, which satisfies the following expression: ； in, For pest and disease infection density, i.e., spatiotemporal dynamic prediction target variable. Two-dimensional spatial coordinates, For time variables, The characteristic dependence of the diffusion coefficient, Characteristic modulation for infection rate, Crop characteristics that determine carrying capacity For nonlinear response terms, Let the mortality rate function be... This is a comprehensive feature vector.

7. The vegetable pest and disease control decision-making method based on multi-source data according to claim 1, characterized in that, Based on the spatiotemporal infection density of the pests and diseases, a topology-sensing prevention and control decision-making model is constructed, including: The parameters of the dynamic system model are calibrated using a physical information neural network to obtain calibration results. Based on the calibration results and the spatiotemporal infection density of pests and diseases, a multi-scale topological descriptor is constructed. Based on the multi-scale topology descriptor, a topologically persistent image is constructed; Multi-scale topological features are extracted using the multi-scale topological descriptor and the topological persistence image; Based on the aforementioned multi-scale topological features, a topology-aware prevention and control decision model is constructed.

8. A vegetable pest and disease control decision-making method based on multi-source data according to claim 7, characterized in that, The topology-aware control decision model is a stochastic optimal control problem with topological constraints, as follows: ； Satisfy constraints: ； in, Indicates the location and time The intensity of prevention and control For the control efficiency function, It is spatial white noise. , These are the weights for infection loss, prevention and control costs, and the smoothness of prevention and control. Let be the terminal cost function, representing the additional cost calculated based on the final infection status at the end of the decision-making cycle. This represents the spatial diffusion coefficient of individual pests and diseases. The intrinsic growth rate For environmental carrying capacity, The natural mortality rate, For nonlinear response terms, This is a noise intensity function.

9. A vegetable pest and disease control decision-making method based on multi-source data according to claim 1, characterized in that, Based on the aforementioned topology-aware pest and disease control decision model, and combined with an improved topology optimization algorithm, vegetable pest and disease control decisions are generated, including: Based on the aforementioned topology-aware prevention and control decision model, an improved multi-scale topology optimization algorithm is introduced to obtain an optimized prevention and control strategy. Based on the aforementioned prevention and control strategy, and combined with the Kalman filter-topology fusion estimation algorithm, the state of the prevention and control strategy is updated to obtain an adaptive vegetable pest and disease control decision.

10. A vegetable pest and disease control decision-making system based on multi-source data, wherein the system uses the vegetable pest and disease control decision-making method based on multi-source data as described in any one of claims 1 to 9, characterized in that, The system includes an input device, a processor, an output device, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions, and the processor is configured to invoke the program instructions.

Images