Crop Management Methods on Red Soil Sloping Land

Abstract

The present invention relates to the technical field of crop management, and more specifically to a method for managing crops on red soil slopes. The method involves obtaining crop data, surrounding environment data, and soil data corresponding to crops planted in a target red soil slope area; identifying the crop data and soil data to determine the current drought level corresponding to the crops; identifying the crop data to determine the current pest and disease type and current pest and disease level corresponding to the crops; determining the cause of the drought level and the cause of the pest and disease level corresponding to the crops based on the surrounding environment data and soil data; outputting an irrigation strategy based on the current drought level and the cause of the drought level; and outputting pest and disease control measures corresponding to the crops based on the current pest and disease type, current pest and disease level, and the cause of the current pest and disease level. This method achieves comprehensive, scientific, and accurate management of crops on red soil slopes, improves efficiency, and saves time and labor costs.

Description

Technical Field

[0001] The present invention relates to the technical field of crop management, and in particular to a method for managing crops on red soil slopes. Background Art

[0002] Red soil slopes are widely distributed in southern my country, accounting for approximately 23% of the country's total land area. They are a vital land resource for agricultural production in the south. In terms of crop cultivation, these slopes primarily support rice, corn, sugarcane, and other food and cash crops. However, due to limited soil and topographical conditions, crop growth and development are significantly impacted, making it difficult to improve yield and quality. Currently, a series of technical measures are being implemented to manage crops on these slopes. For soil improvement, lime is applied to adjust soil pH, and organic fertilizers and biochar are added to enhance soil fertility and improve soil structure. For soil and water conservation, measures such as contour planting, terrace construction, and slope protection planting have reduced soil erosion to a certain extent.

[0003] However, existing technologies still have obvious shortcomings. Traditional lime application methods can easily cause large fluctuations in soil pH, affecting soil microbial activity and the absorption of other nutrients by crops. The application amount and time of organic fertilizers and biochar lack precise control, making it difficult to fully exert their soil improvement and fertilization effects. In terms of soil and water conservation, contour planting and terrace construction are expensive, and some areas are difficult to promote on a large scale due to terrain and funding constraints. The selection and configuration of slope protection plants are not scientific enough, and the protective effect is limited. In addition, most existing management technologies focus on solving single problems and lack comprehensive consideration of the crop growth environment on red soil slopes. It is impossible to achieve the coordinated development of soil improvement, soil and water conservation, and high-yield and high-quality crops.

[0004] Therefore, how to comprehensively manage crops on red soil slopes has become an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the present invention provides a method for managing crops on red soil slopes to solve the problem of how to comprehensively manage crops on red soil slopes.

[0006] In a first aspect, the present invention provides a method for managing crops on red soil slopes, the method comprising:

[0007] Acquire crop data, surrounding environment data, and soil data corresponding to crops planted in the target red soil slope area; the crop data includes crop image data, crop spectral data, and crop thermal infrared imaging data; the surrounding environment data includes at least one of light intensity, humidity, and wind speed; and the soil data includes at least one of soil physical property data, soil chemical property data, soil biological property data, and water property data;

[0008] Identify crop data and soil data to determine the current drought level corresponding to the crops;

[0009] Identify crop data to determine the current pest and disease type and level of the crop;

[0010] Determine the causes of drought levels and pest and disease levels corresponding to crops based on surrounding environmental data and soil data;

[0011] Output irrigation strategies based on the current drought level and the cause of the drought level; irrigation strategies include intermittent irrigation mode and continuous irrigation mode;

[0012] Based on the current pest and disease type, current pest and disease level, and the cause of the current pest and disease level, output the corresponding pest and disease control measures for crops.

[0013] The red soil slope crop management method provided in the embodiments of the present application obtains crop data, surrounding environment data, and soil data corresponding to crops planted in the target red soil slope area; identifies the crop and soil data to determine the current drought level corresponding to the crops. This ensures the accuracy of the current drought level determined, more accurately reflecting the actual degree of drought faced by crops than single-indicator assessments, and facilitates the timely implementation of appropriate drought relief measures. The crop data is identified to determine the current pest and disease type and level corresponding to the crops, ensuring the accuracy of the current pest and disease type and level determined, thereby providing a strong basis for formulating targeted pest and disease control strategies and avoiding inappropriate control measures due to misdiagnosis or misjudgment. Based on the surrounding environment and soil data, the causes of the drought level and pest and disease level corresponding to the crops are determined. This method can deeply explore the various environmental and soil factors that affect crop growth and clarify the mechanisms by which each factor affects the occurrence and development of drought and pests, thereby providing direction for fundamentally solving the problem and facilitating the implementation of more targeted and effective prevention and control measures. Based on the current drought level and the causes of the drought level, an irrigation strategy is output. The accuracy of the output irrigation strategies is guaranteed. Based on the current pest and disease type, current pest and disease level, and the causes of the current pest and disease level, corresponding pest and disease control measures for crops are output, ensuring the accuracy of the output pest and disease control measures. These management measures can precisely target the specific problems facing crops and their root causes, effectively improving the efficiency and quality of agricultural production, reducing the impact of drought and pests on crop yield and quality, and achieving sustainable agricultural development. This enables comprehensive, scientific, and accurate management of crops on red soil slopes, improving efficiency and saving time and labor costs.

[0014] In an optional embodiment, the soil data includes soil physical property data, soil chemical property data, and water property data. Identifying the crop data and soil data to determine the current drought level corresponding to the crop includes:

[0015] Inputting crop image data, crop spectral data, crop thermal infrared imaging data, soil physical property data, soil chemical property data, and water property data into a preset drought level determination model;

[0016] The preset drought level determination model extracts features from the input data and performs feature fusion to obtain a fused feature vector;

[0017] Perform feature extraction on the fused feature vector to generate the target feature vector;

[0018] Based on the target feature vector, the current drought level corresponding to the crops is output.

[0019] In an optional embodiment, performing feature extraction on the fused feature vector to generate a target feature vector includes:

[0020] Based on the variational autoencoder, the fused feature vector is mapped to the mean vector and variance vector in the latent space;

[0021] Sampling is performed based on the mean vector and the variance vector to obtain at least one sampling point; the sampling point contains mixed coding information of the multimodal data in the latent space; the mixed coding information includes multiple sub-features;

[0022] Based on the preset decoder, each sampling point is converted and combined to generate a target feature vector with the same dimension as the input data.

[0023] In an optional embodiment, based on a preset decoder, each sampling point is converted and combined to generate a target feature vector with the same dimension as the input data, including:

[0024] Each layer in the preset decoder calculates the similarity between each sub-feature in the sampling point and the pre-set drought key feature template;

[0025] According to the calculated similarity, the weight information of each sub-feature is determined;

[0026] Based on the weight information corresponding to each sub-feature, each sub-feature is weighted to obtain a weighted vector;

[0027] The weighted vector is convolved through the convolution layer in the preset decoder to generate the target feature vector.

[0028] In an optional embodiment, identifying crop data to determine the current pest type and pest level corresponding to the crop includes:

[0029] Extract features from crop image data, crop spectral data, and crop thermal infrared imaging data, and fuse the extracted features to generate initial fused features;

[0030] Extract the initial fusion features based on the spatiotemporal attention mechanism to generate the target fusion features;

[0031] The target fusion features are input into the preset pest and disease level determination model to output the current pest and disease type and current pest and disease level corresponding to the crop.

[0032] In an optional embodiment, feature extraction is performed on crop image data, crop spectral data, and crop thermal infrared imaging data, and the extracted features are fused to generate initial fused features, including:

[0033] Perform feature extraction on crop image data to obtain target image feature values;

[0034] The crop spectral data is encoded using a preset encoder to obtain a low-dimensional sparse representation; the low-dimensional sparse representation highlights the specific spectral band characteristics related to pests and diseases;

[0035] Arrange the crop thermal infrared imaging data in chronological order, and use the pixel points in each thermal infrared image in the crop thermal infrared imaging data as nodes;

[0036] Each node corresponds to a specific spatiotemporal location, and the node's attribute is the temperature value of the spatiotemporal location at the corresponding moment;

[0037] Determine the edge between two nodes based on the temperature change relationship between two nodes at adjacent time points and adjacent spatial positions;

[0038] According to the absolute value of the temperature difference between two nodes, the edge weight is determined to generate a space-time graph;

[0039] The target image feature values, low-dimensional sparse representation and spatiotemporal graph are fused to generate initial fused features.

[0040] In an optional embodiment, feature extraction is performed on crop image data to obtain target image feature values, including:

[0041] Extract features from crop image data to generate initial color features, initial texture features, and initial shape features;

[0042] Mapping the initial color features, initial texture features, and initial shape features to discrete value ranges to create an image feature search space;

[0043] Formulate encoding rules based on the discrete value range of the image feature search space;

[0044] Randomly assign an initial state to each initial qubit;

[0045] Initialize an initial quantum population; the initial quantum population includes multiple initial quantum states, each of which is represented by a group of initial quantum bits, representing a possible combination of image features;

[0046] According to the fitness function, the fitness evaluation results of the image feature combinations corresponding to each initial quantum state in the initial quantum population are calculated;

[0047] The rotation angle of each initial quantum bit is calculated based on the fitness evaluation results; the size and direction of the rotation angle depend on the fitness evaluation results of the current quantum state and the fitness evaluation results of the target state;

[0048] Performing a quantum rotating gate operation on each initial qubit according to the calculated rotation angle to obtain each updated qubit and an updated quantum state composed of the updated qubits;

[0049] Measure each updated quantum state, convert the updated quantum state into classical information, and obtain the classical state;

[0050] According to the encoding rules, the classical state is decoded into the initial image feature value;

[0051] The initial image feature values are screened to obtain the target image feature values.

[0052] In an optional embodiment, the target fusion feature is input into a preset pest and disease level determination model to output the current pest and disease type and current pest and disease level corresponding to the crop, including:

[0053] Obtain a target knowledge graph; the target knowledge graph is based on crop variety characteristics, pest and disease transmission patterns, and red soil slope environmental data;

[0054] Input the target fusion features into the preset pest and disease determination model, which outputs the initial determination result; the initial determination result includes the initial pest and disease type and initial pest and disease level corresponding to the crop;

[0055] Build query statements based on the initial pest and disease type, combined with crop varieties, growth stages, surrounding environment data, and soil data;

[0056] Based on the query statement, the target knowledge graph is queried to obtain the query results; the query results include the probability of occurrence of the initial pest and disease type and the severity level;

[0057] The initial judgment result and the query result are integrated to obtain the current insect pests and diseases and the current insect pests and diseases levels.

[0058] In an optional embodiment, determining the causes of drought levels and pest and disease levels corresponding to crops based on surrounding environment data and soil data includes:

[0059] Based on the current drought level, current pest and disease level, surrounding environment data and soil data as nodes, and according to the relationship between the current drought level, current pest and disease level, surrounding environment data and soil data, the corresponding nodes are connected with directed edges to generate a directed acyclic graph;

[0060] Input the directed acyclic graph into a feature extraction network in a preset causal relationship model;

[0061] The feature extraction network extracts features from the directed acyclic graph to obtain node features, edge features, and graph structure features corresponding to the directed acyclic graph, and inputs the node features, edge features, and graph structure features into the decision network in the preset causal relationship model;

[0062] The decision network determines the causes of drought levels and current pest and disease levels based on node features, edge features, and graph structure features.

[0063] In an optional embodiment, the decision network determines the cause of the drought level and the cause of the current pest and disease level based on node features, edge features, and graph structure features, including:

[0064] Construct the initial state corresponding to the state space in the decision network based on node features, edge features, and graph structure features;

[0065] Based on the initial state, the initial action is output from the action space corresponding to the decision network; the initial action represents the initial estimated value of the impact of a set of factors on the current drought level and the current pest and disease level;

[0066] The decision network starts from the nodes that affect the current drought level and the current pest and disease level, and derives the causal transmission path along the directed edges in the graph;

[0067] According to the strength of edge features and the position information of each node in the causal chain, key nodes are determined from each node;

[0068] According to each key node, the initial estimated value of the impact of each factor on the current drought level and insect pest level in the initial action is adjusted to obtain the target estimated value of the impact of each factor on the current drought level and insect pest level;

[0069] According to the target estimated values of each impact quantity, a target factor having a target estimated value of the impact quantity greater than a preset threshold value is determined among the factors;

[0070] Based on the target factors, determine the causes of drought levels and current pest and disease levels. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the specific embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0072] Figure 1 1 is a flow chart of a method for managing crops on red soil slopes according to an embodiment of the present invention;

[0073] Figure 2 2 is a flow chart of another method for managing crops on red soil slopes according to an embodiment of the present invention. DETAILED DESCRIPTION

[0074] To make the purpose, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of the present invention.

[0075] According to an embodiment of the present invention, an embodiment of a method for managing crops on red soil slopes is provided. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0076] In this embodiment, a method for managing crops on red soil slopes is provided, which can be used with the above-mentioned electronic device. Figure 1 FIG. 1 is a flow chart of a method for managing crops on red soil slopes according to an embodiment of the present invention. Figure 1 As shown, the process includes the following steps:

[0077] Step S101 , obtaining crop data, surrounding environment data, and soil data corresponding to crops planted in a target red soil slope area.

[0078] Crop data includes crop image data, crop spectral data, and crop thermal infrared imaging data. Ambient environment data includes at least one of light intensity, humidity, and wind speed. Soil data includes at least one of soil physical property data, soil chemical property data, soil biological property data, and moisture property data.

[0079] Specifically, electronic devices can collect crop data and surrounding environmental data based on smart wearable data collection devices equipped by farmers. These smart wearable data collection devices can integrate a micro-image collector, a spectral sensor, a thermal infrared detector, and an environmental monitoring module. The integrated micro-image collector can capture crop image data (such as images of leaf morphology and color changes at different growth stages); the spectral sensor can collect crop spectral data (spectral characteristics reflecting the crop's physiological state); and the infrared detector can collect thermal infrared imaging data (monitoring crop temperature and determining water stress). The environmental monitoring module can also collect ambient environmental data, including light intensity (light intensity at different time periods and slope directions), humidity (dynamic changes in air and soil moisture), wind speed (which affects crop transpiration and the spread of pests and diseases), and soil data (soil texture, nutrient content, pH, etc.).

[0080] Step S102: Identify the crop data and soil data to determine the current drought level corresponding to the crops.

[0081] Specifically, the electronic device may input crop data and soil data into a preset drought level determination model, and output the current drought level corresponding to the crops.

[0082] This step will be described in detail below.

[0083] Step S103: Identify the crop data to determine the current pest and disease type and current pest and disease level corresponding to the crop.

[0084] Specifically, the electronic device can identify crop image data, crop spectral data, and crop thermal infrared imaging data. Based on the identification results, it can then determine the current pest type and level of the crop. For example, yellow spots and curling on crop leaves may indicate viral infection, while holes and notches on leaves may indicate chewing pests such as cabbage loopers.

[0085] Pest and disease severity is assessed based on indicators such as the degree of damage to crops, the extent of the disease, and the number of pests. It is generally categorized into three levels: mild, moderate, and severe. For example, mild diseases may only cause symptoms on a few plants, with minimal impact on overall crop growth. Severe diseases, on the other hand, can affect a large area of plants, severely impacting crop yield and quality.

[0086] This step will be described in detail below.

[0087] Step S104: determining the causes of drought levels and insect pest levels corresponding to the crops based on the surrounding environment data and soil data.

[0088] Specifically, the electronic device can determine the causes of the drought level and the causes of the pest and disease level corresponding to the crops based on the relationship between the surrounding environment data and soil data and the current drought level, the current pest and disease type, and the current pest and disease level.

[0089] This step will be described in detail below.

[0090] Step S105: outputting an irrigation strategy based on the current drought level and the cause of the drought level.

[0091] Among them, irrigation strategies include intermittent irrigation mode and continuous irrigation mode.

[0092] Specifically, electronic devices can clearly identify the current drought level of crops. A higher drought level indicates more severe water stress for the crops and a more urgent need for irrigation. For example, a high drought level may mean a severe lack of soil moisture, significantly inhibiting crop growth. Furthermore, electronic devices can determine the cause of the drought level, such as rapid soil moisture infiltration, insufficient irrigation, or dry climate. This helps formulate targeted irrigation strategies. For example, if the drought is caused by rapid soil moisture infiltration, irrigation strategies should consider how to slow down moisture infiltration or increase irrigation frequency.

[0093] The electronic device then acquires the water requirements of the crop variety, its current growth stage, and real-time soil moisture data. Different crop varieties have different water requirements. For example, rice is a water-loving crop and requires more water during growth, while corn is more drought-tolerant and requires less water. Furthermore, crops have varying water requirements at different growth stages. For example, during the seedling stage, the root system is not yet fully developed, so water requirements are relatively low. However, during the flowering and fruiting stages, water requirements are typically higher. Real-time soil moisture data is acquired through devices such as soil moisture sensors. Based on this data, the theoretical water requirement (theoretical water requirement) is calculated to determine how much water is needed to meet the crop's growth needs at its current growth stage.

[0094] The electronic device calculates the initial irrigation volume based on the crop's water requirements, current growth stage, and real-time soil moisture data. The formula is as follows: Q = S × H × (Wtarget - Wcurrent). Q represents the initial irrigation volume (in cubic meters or liters); S represents the irrigated area (in square meters); H represents the depth of the soil layer where the crop's root system is primarily located (in meters); Wtarget represents the optimal soil moisture content for the crop at its current growth stage (expressed in volume percentage); and Wcurrent represents the real-time monitored soil moisture content (expressed in volume percentage).

[0095] The electronic device then corrects the theoretical water requirement based on the water infiltration rate of the red soil slope to determine the target irrigation rate. The rapid water infiltration rate of red soil slopes results in significant water loss, necessitating an appropriate increase in irrigation to compensate for this loss. Based on the specific water infiltration rate and the theoretical water requirement, the actual amount of water required after accounting for soil infiltration is calculated—the target irrigation rate.

[0096] When the current drought level is lower than the preset drought level, or the drought is primarily due to a short-term water shortage and the soil has relatively good water retention capacity, intermittent irrigation can be considered. For example, crops are experiencing a mild drought and the soil is heavy and has relatively slow water infiltration.

[0097] Implementation: Divide the target irrigation volume into several smaller portions and irrigate them multiple times at regular intervals. This prevents water loss from excessive irrigation at one time and allows crops ample time to absorb and utilize the water from each irrigation. For example, divide the target irrigation volume into 3-5 portions, irrigating every 1-2 days.

[0098] If the current drought level is greater than or equal to the preset drought level, crops are experiencing severe water stress, or the drought level is caused by long-term water shortages and extremely rapid soil water infiltration, continuous irrigation may be more appropriate. For example, crops are experiencing severe drought and the water infiltration rate on red soil slopes is very fast, resulting in poor soil water retention.

[0099] Implementation: Continuously irrigate over a period of time to maintain a high soil moisture content to quickly alleviate crop drought. However, care should be taken to control the irrigation rate to avoid problems such as water accumulation and soil erosion. Drip irrigation, sprinkler irrigation, and other methods can be used to provide a relatively even and continuous flow of water until the target recharge volume is reached or the soil moisture content reaches the appropriate range.

[0100] Step S106: outputting pest control measures corresponding to the crops based on the current pest type, the current pest level, and the cause of the current pest level.

[0101] Specifically, electronic devices can output one of the following agricultural, physical, biological, and chemical control measures for crops based on the current pest type, pest level, and cause of the current pest level. Agricultural control measures include: rational crop rotation, clean fields, reasonable planting density, and scientific fertilization; physical control measures include: color plate trapping, light trapping, and high-temperature greenhouses; biological control measures include: utilizing natural enemies and using biological agents; and chemical control measures: when other control measures are ineffective or pests and diseases are severe, chemical pesticides can be used for control. However, care should be taken to select highly effective, low-toxic, and low-residue pesticides, and to strictly follow the instructions for use to ensure the quality and safety of agricultural products and the environment.

[0102] The red soil slope crop management method provided in the embodiments of the present application obtains crop data, surrounding environment data, and soil data corresponding to crops planted in the target red soil slope area; identifies the crop and soil data to determine the current drought level corresponding to the crops. This ensures the accuracy of the current drought level determined, more accurately reflecting the actual degree of drought faced by crops than single-indicator assessments, and facilitates the timely implementation of appropriate drought relief measures. The crop data is identified to determine the current pest and disease type and level corresponding to the crops, ensuring the accuracy of the current pest and disease type and level determined, thereby providing a strong basis for formulating targeted pest and disease control strategies and avoiding inappropriate control measures due to misdiagnosis or misjudgment. Based on the surrounding environment and soil data, the causes of the drought level and pest and disease level corresponding to the crops are determined. This method can deeply explore the various environmental and soil factors that affect crop growth and clarify the mechanisms by which each factor affects the occurrence and development of drought and pests, thereby providing direction for fundamentally solving the problem and facilitating the implementation of more targeted and effective prevention and control measures. Based on the current drought level and the causes of the drought level, an irrigation strategy is output. The accuracy of the output irrigation strategies is guaranteed. Based on the current pest and disease type, current pest and disease level, and the causes of the current pest and disease level, corresponding pest and disease control measures for crops are output, ensuring the accuracy of the output pest and disease control measures. These management measures can precisely target the specific problems facing crops and their root causes, effectively improving the efficiency and quality of agricultural production, reducing the impact of drought and pests on crop yield and quality, and achieving sustainable agricultural development. This enables comprehensive, scientific, and accurate management of crops on red soil slopes, improving efficiency and saving time and labor costs.

[0103] In this embodiment, a method for managing crops on red soil slopes is provided, which can be used with the above-mentioned electronic device. Figure 2FIG. 1 is a flow chart of a method for managing crops on red soil slopes according to an embodiment of the present invention. Figure 2 As shown, the process includes the following steps:

[0104] Step S201 , obtaining crop data, surrounding environment data, and soil data corresponding to crops planted in a target red soil slope area.

[0105] For details about this step, please refer to the above description of step S101 and will not be repeated here.

[0106] Step S202 : Identify the crop data and soil data to determine the current drought level corresponding to the crops.

[0107] Specifically, the soil data includes soil physical property data, soil chemical property data, and water property data. The above step S202 may include the following steps:

[0108] Step S2021 , inputting crop image data, crop spectral data, crop thermal infrared imaging data, soil physical property data, soil chemical property data, and water property data into a preset drought level determination model.

[0109] Specifically, the electronic device may input crop image data, crop spectrum data, crop thermal infrared imaging data, soil physical property data, soil chemical property data, and water property data into a preset drought level determination model.

[0110] In step S2022 , the preset drought level determination model extracts features from the input data and performs feature fusion to obtain a fused feature vector.

[0111] Specifically, the pre-set drought level determination model can use a pre-trained ResNet model to extract the texture, shape, and color features of crop leaves in crop image data. These features can reflect the health and growth status of the crops. For crop spectral data, the pre-set drought level determination model uses principal component analysis (PCA) to extract the main features of the spectral curve, reducing the data dimension while retaining important spectral information. For soil data, features such as soil porosity, pH, and organic matter content are extracted based on its physical, chemical, and moisture properties.

[0112] Then, based on the multimodal fusion network in the pre-set drought level determination model, different types of features are fused to obtain an initial feature vector. This multimodal fusion network can adopt a deep neural network structure, such as a multilayer perceptron (MLP) or a long short-term memory (LSTM). During the fusion process, different weights are assigned to different types of features, and adaptive adjustments are made based on their importance to the drought level determination.

[0113] Then, feature compression and dimensionality reduction techniques such as autoencoder or linear discriminant analysis (LDA) are used to compress the high-dimensional initial feature vector into a low-dimensional fused feature vector.

[0114] Step S2023: extract features from the fused feature vector to generate a target feature vector.

[0115] Specifically, the above step S2023 may include the following steps:

[0116] In step a1, the fused feature vector is mapped to the mean vector and variance vector in the latent space based on the variational autoencoder.

[0117] Specifically, the fused feature vector is input into the variational encoder network. The variational encoder network processes the input fused feature vector through a series of linear transformations and nonlinear activation functions, and ultimately outputs two vectors: the mean vector and the logarithmic variance vector.

[0118] Step a2: performing sampling processing based on the mean vector and the variance vector to obtain at least one sampling point.

[0119] Among them, the sampling points contain mixed coding information of multimodal data in the latent space; the mixed coding information includes multiple sub-features.

[0120] Specifically, the mean vector and variance vector obtained by the variational autoencoder define the Gaussian distribution in the latent space, namely zN(μ,σ 2 I), where z represents the sample in the latent space and I is the identity matrix. This distribution property is the basis for subsequent sampling, indicating that the sample points will be generated from a Gaussian distribution with the specified mean and variance.

[0121] Since the variance vector σ 2 It represents the variance in each dimension. In order to perform sampling, the standard deviation vector σ needs to be calculated. By taking the square root of the variance in each dimension, that is, , we get the standard deviation vector. Each element of this vector corresponds to the standard deviation of a dimension in the latent space, which determines the fluctuation range of the sampling in that dimension.

[0122] Then, use the random number generator to generate standard Gaussian distribution random numbers with the same dimension as the latent space. ,Right now -N(0,I). The standard Gaussian distribution is a normal distribution with a mean of 0 and a variance of 1. These random numbers will be used to introduce randomness into the sampling process.

[0123] According to the reparameterization technique, the sampling point z is sampled from the defined Gaussian distribution, and the calculation formula is z=μ+σ⊙ , where ⊙ represents element-wise multiplication. By using this formula, the mean vector μ, standard deviation vector σ and the standard Gaussian distribution random number Combined with the mean and variance, we get the sampling point z. This calculation method not only uses the mean and variance to determine the center position and fluctuation range of the sampling point, but also introduces randomness through random numbers, so that the results obtained from each sampling are different.

[0124] The resulting sampling point z resides in a latent space, containing mixed encodings of multimodal data within the latent space. This mixed encoding represents an abstract representation of the original multimodal data (such as crop image data, spectral data, thermal infrared imaging data, soil data, etc.) encoded using a variational autoencoder. Each dimension in the latent space corresponds to one or more factors influencing the crop's state. Therefore, the different element values of sampling point z constitute multiple sub-features. These sub-features reflect information such as crop growth status, environmental factors, and pest and disease risks from different perspectives, providing a highly summarized and refined representation of the original multimodal data.

[0125] Finally, the sampling steps above can be repeated multiple times to generate multiple sampling points, depending on actual needs. Different sampling points represent different states in the latent space. These sampling points allow for a more comprehensive exploration of the latent space, mining the information contained in multimodal data and providing rich data support for subsequent data analysis and decision-making.

[0126] Step a3: Based on a preset decoder, each sampling point is converted and combined to generate a target feature vector with the same dimension as the input data.

[0127] Specifically, the above step a3 may include the following steps:

[0128] In step a31 , each layer in the preset decoder is configured to calculate the similarity between each sub-feature in the sampling point and a preset drought key feature template.

[0129] Specifically, starting from the input layer of the preset decoder, the sampling points are used as input. After processing through the first layer of the network, a set of sub-features is obtained. The similarity between these sub-features and the drought key feature template is then calculated using a preset similarity algorithm. The preset similarity algorithm can be any similarity algorithm such as the Euclidean distance algorithm, the Manhattan distance algorithm, and the cosine similarity algorithm.

[0130] Step a32: Determine the weight information of each sub-feature based on the calculated similarity.

[0131] Specifically, the electronic device assigns weight information to each sub-feature according to the calculated similarity.

[0132] Step a33: performing weighted processing on each sub-feature based on the weight information corresponding to each sub-feature to obtain a weighted vector.

[0133] Specifically, the electronic device performs weighted processing on each sub-feature based on weight information corresponding to each sub-feature to obtain a weighted vector.

[0134] In step a34, the weighted vector is convolved through a convolution layer in a preset decoder to generate a target feature vector.

[0135] Specifically, a convolutional layer performs a convolution operation on a weighted vector (represented as a feature map in data such as images). The convolution kernel slides over the weighted vector, extracting local features through the convolution operation and combining these local features to generate a target feature vector. Convolutional layers effectively capture the local structure and spatial relationships of features, making them particularly effective for processing spatially structured data, such as those associated with crop images. Furthermore, convolutional layers can further reduce feature dimensionality through operations such as pooling, thereby reducing computational effort.

[0136] Step S2024: output the current drought level corresponding to the crop based on the target feature vector.

[0137] Specifically, the target feature vector is mapped into a probability distribution space, outputting the probability distribution of crops experiencing different drought levels (such as mild drought, moderate drought, and severe drought). Then, based on this probability distribution, the drought level with the highest probability value is determined as the current drought level. This probabilistic output method not only provides an estimate of the drought level but also reflects the uncertainty of the estimate, providing richer information for agricultural decision-making.

[0138] Step S203: Identify the crop data to determine the current pest and disease type and current pest and disease level corresponding to the crop.

[0139] Specifically, the above step S203 may include the following steps:

[0140] Step S2031 : extracting features from the crop image data, crop spectral data, and crop thermal infrared imaging data, and fusing the extracted features to generate initial fused features.

[0141] Specifically, the above step S2031 may include the following steps:

[0142] Step b1: extracting features from crop image data to obtain target image feature values.

[0143] Specifically, the above step b1 may include the following steps:

[0144] Step b11: extracting features from the crop image data to generate initial color features, initial texture features, and initial shape features.

[0145] Specifically, electronic devices can convert crop image data from the common RGB color space to a preset color space to obtain initial color features. Preset color spaces can include HSV (hue, saturation, brightness) or Lab (brightness, green-red axis, blue-yellow axis), among others. Electronic devices can also calculate GLCM feature values (such as contrast, correlation, energy, and entropy) on the crop image data to obtain initial texture features. Contrast reflects the intensity of grayscale variations within the image, correlation indicates the directionality of the texture, energy indicates texture uniformity, and entropy indicates texture complexity. Electronic devices use edge detection algorithms to detect edges in the crop image and then use contour tracking algorithms to obtain the crop's contour curve to obtain initial shape features. The contour curve intuitively represents the crop's shape boundary, providing a foundation for subsequent shape analysis.

[0146] Step b12: Mapping the initial color features, initial texture features, and initial shape features to discrete value ranges to create an image feature search space.

[0147] Specifically, the electronic device maps the initial color, texture, and shape features to discrete value ranges to construct an image feature search space. This converts continuous feature values into discrete states, facilitating subsequent encoding and processing using quantum bits. Taking color features as an example, the value range of each channel is quantized into several levels. This allows for the classification of color information, enabling clearer differentiation and processing of different color features during subsequent searches.

[0148] Step b13: Formulate encoding rules based on the discrete value range of the image feature search space.

[0149] Specifically, electronic devices can use quantum bits (qubits) to encode each feature dimension in the image feature search space. A qubit can be in a state of 0, 1, or a superposition of both. By combining multiple qubits, a high-dimensional feature space can be represented.

[0150] The electronic device then assigns a corresponding encoding rule to each qubit based on the discrete value range of the image feature search space. For example, a binary feature can be represented by a single qubit, with a 0 state representing a feature value of 0 and a 1 state representing a feature value of 1. Multi-valued features can be encoded using multiple qubits.

[0151] Step b14, randomly assigning an initial state to each initial quantum bit.

[0152] Specifically, each initial qubit is randomly assigned an initial state, typically a superposition of 0 and 1. This is done because before the search for feature combinations begins, there is no prior information about which combination is optimal. Random initialization allows the algorithm to conduct extensive exploration across the entire feature space. Using quantum gate operations such as Hadamard gates to initialize the qubits to a superposition state leverages the properties of quantum gate operations to precisely control the qubit's state.

[0153] Step b15: initialize an initial quantum population.

[0154] The initial quantum population includes multiple initial quantum states, each of which is represented by a set of initial quantum bits, representing a possible combination of image features.

[0155] Specifically, the electronic device can initialize an initial quantum population based on the initial state randomly assigned to each initial quantum bit.

[0156] Step b16: Calculate the fitness evaluation result of the image feature combination corresponding to each initial quantum state in the initial quantum population according to the fitness function.

[0157] Specifically, the electronic device can generate a fitness function based on the spot recognition accuracy. Then, based on this fitness function, it calculates the fitness evaluation results of the image feature combinations corresponding to each initial quantum state in the initial quantum population. By evaluating the spot recognition accuracy of feature combinations corresponding to different quantum states, it can be determined which feature combinations are most conducive to spot recognition.

[0158] Step b17: Calculate the rotation angle of each initial quantum bit based on the fitness evaluation result.

[0159] The size and direction of the rotation angle depend on the fitness evaluation results of the current quantum state and the fitness evaluation results of the target state.

[0160] Specifically, the electronic device can calculate the difference between the fitness evaluation result and the target fitness. Specifically, the difference ΔF=Ftarget-Fcurrent. Among them, Ftarget is the target fitness and Fcurrent is the fitness evaluation result. Then, the direction and size of the rotation angle θ are determined according to the positive and negative and size of the difference ΔF. For example: if ΔF>0, it means that the current quantum state still has room for improvement. In order to move closer to the target state, the rotation angle is determined according to a certain proportional relationship. A proportional coefficient k (0 <k<1), , and the direction of rotation is determined based on the characteristic differences between the current quantum state and the target state (for example, by comparing the values of each feature in the feature combination corresponding to the quantum state). If ΔF=0, it means that the current quantum state has reached the target state or is in a local optimum. At this time, a smaller random rotation angle can be selected to avoid falling into the local optimum and continue to explore other possible feature combinations. If ΔF<0, it means that the current quantum state may have deviated from the target state and needs to be adjusted in the opposite direction. Similarly, the size of the rotation angle is determined based on |ΔF| and the proportional coefficient k, and the appropriate rotation direction is determined.

[0161] Step b18: performing a quantum rotating gate operation on each initial quantum bit according to the calculated rotation angle to obtain each updated quantum bit and an updated quantum state composed of the updated quantum bits.

[0162] Specifically, for each initial qubit , multiply it by the quantum rotation gate matrix R(θ), where θ is the rotation angle, to get the updated quantum bit The specific calculation process is as follows:

[0163]

[0164] Through the above calculations, we get the updated quantum bit , its coefficient and It reflects the change of the quantum bit state caused by the rotation operation.

[0165] Step b19: measure each updated quantum state, convert the updated quantum state into classical information, and obtain a classical state.

[0166] Specifically, electronic devices can measure the updated quantum state using specialized quantum measurement devices (such as the measurement module in a quantum computer). These devices interact with qubits to measure the quantum state. The measurement device then outputs a measurement result based on the qubit's state.

[0167] The measurement result is the qubit's eigenstate (|0〉 or |1〉), which can be directly mapped to a binary bit (0 or 1) in classical information. For multiple qubit measurements, combining the results of each qubit in sequence yields a classical binary string, the classical state. For example, measuring an updated quantum state consisting of n qubits yields the result |b1b2…bn〉 (bi∈{0,1}, i=1,2,…,n). Converting this to classical information yields the binary string b1b2…bn. This binary string can then be mapped to a specific combination of image features according to pre-defined encoding rules.

[0168] Through the above measurement and transformation operations on the updated quantum state, the quantum state information obtained by quantum calculation is converted into classical information, realizing the conversion from quantum state to classical state, laying the foundation for further using this information to solve practical problems.

[0169] Step b110: decoding the classical state into initial image feature values according to the encoding rules.

[0170] Specifically, encoding rules are developed to map the various values of image features into a form that can be processed by quantum computing (such as the state of a quantum bit). Encoding rules include binary encoding, numerical encoding, one-hot encoding, and others. Electronic devices need to determine the image features corresponding to each part in the classical state based on the encoding rules. For example, if the encoding rule is to use n quantum bits to encode color features, texture features, and shape features respectively, then the binary string of the classical state can be divided according to a certain number of bits, corresponding to different features. Assuming that the color feature is encoded using the first m1 bits, the texture feature is encoded using the next m2 bits, and the shape feature is encoded using the last m3 bits (m1+m2+m3=n), we can split the classical state string into the corresponding parts.

[0171] If binary encoding is used, for the color feature encoding portion, the binary number is converted to a decimal number. Then, based on the previously divided discrete intervals, the corresponding color feature value, such as hue, saturation, or brightness, is determined. For example, if the hue value is encoded using an 8-bit binary number in the range [0, 255], the actual hue value range is [0, 360]. After converting the binary number to a decimal number x, the actual hue value y = 255360x.

[0172] For features encoded numerically (such as shape features), the eigenvalue is determined directly based on the correspondence between the numerical value and the feature category in the encoding rules. For one-hot encoding, the position with a value of 1 in the encoding vector is found, and the eigenvalue is determined based on the correspondence between that position and the feature value. For example, for a color feature, the one-hot encoding vector [0,1,0] represents the color value corresponding to the second position, and the specific color feature value (e.g., green) is determined according to the encoding rules.

[0173] Finally, the eigenvalues decoded from each part of the classical state are integrated to form a complete set of initial image eigenvalues. This set includes a variety of image feature information such as color eigenvalues, texture eigenvalues, and shape eigenvalues, which can reflect the basic characteristics of crop images.

[0174] Step b111: Screen the initial image feature values to obtain target image feature values.

[0175] Specifically, the electronic device can check for redundancy between the initial image feature values. If two or more features have a strong linear correlation (correlation coefficient close to 1 or -1), they may contain similar information and may be redundant. For example, in color features, two components in certain color spaces may be highly correlated, so only one component can be retained.

[0176] The electronic device then uses dimensionality reduction techniques such as principal component analysis to remove redundant features and obtain the target image's eigenvalues. PCA can transform multiple correlated features into a small number of uncorrelated principal components that retain most of the information from the original features. By selecting an appropriate number of principal components, redundancy can be removed and the feature space simplified.

[0177] Step b2: Encode the crop spectral data using a preset encoder to obtain a low-dimensional sparse representation.

[0178] Among them, low-dimensional sparse representation highlights the specific spectral band characteristics related to pests and diseases.

[0179] Specifically, the electronic device inputs the crop spectral data into a preset encoder. During the forward propagation process of the preset encoder, the crop spectral data is processed through a series of neuron layers. Each layer transforms the crop spectral data, gradually extracting higher-level features. For crop spectral data, the preset encoder learns the correlation and underlying structure between different spectral bands. Through operations such as convolutional layers, the information of spatially adjacent spectral bands can be integrated to extract more representative features. At the same time, in order to achieve low-dimensional sparse representation, some regularization methods may be used, such as L1 regularization. This can make most elements in the feature representation learned by the preset encoder zero, thereby achieving sparsity and highlighting key features related to pests and diseases.

[0180] After multiple layers of processing by the preset encoder, a low-dimensional representation of the crop spectral data is ultimately obtained. This low-dimensional representation maps the original high-dimensional spectral data to a low-dimensional space. It retains the most important information in the original data while, through sparsity constraints, highlighting the specific spectral band characteristics associated with pests and diseases. For example, in the low-dimensional representation, the eigenvalues corresponding to spectral bands sensitive to pests and diseases may be relatively large, while the eigenvalues corresponding to other bands unrelated to pests and diseases may be suppressed to near-zero values.

[0181] Step b3: Arrange the crop thermal infrared imaging data in chronological order, and use each pixel point in the thermal infrared image of the crop thermal infrared imaging data as a node.

[0182] Specifically, the electronic device can arrange crop thermal infrared imaging data in chronological order. Each thermal infrared image is then parsed into a two-dimensional array, with each array element representing a pixel. Each pixel in the thermal infrared image is considered a node.

[0183] In step b4, each node corresponds to a specific spatiotemporal position, and the attribute of the node is the temperature value of the spatiotemporal position at the corresponding moment.

[0184] Specifically, each node corresponds to a specific spatiotemporal location, determined by the pixel's spatial coordinates (row and column indices) in the image and the time the image was captured. Electronic equipment then converts the pixel values into actual temperature values based on the thermal infrared camera's calibration parameters. Different thermal infrared cameras may use different conversion formulas. The converted temperature values are used as node attributes. Node information can be stored using a dictionary or a custom data structure, where the key is the node's identifier (e.g., a combination of spatiotemporal location) and the value is the corresponding temperature value.

[0185] Step b5: determining an edge between two nodes based on the temperature change relationship between the two nodes at adjacent time points and adjacent spatial positions.

[0186] Specifically, for each node, we need to find the nodes at adjacent time points and spatial locations. Spatially, adjacent locations typically refer to the four adjacent pixels above, below, to the left, and to the right of a pixel; temporally, adjacent time points refer to images taken before and after. If two nodes meet the adjacency criteria, an edge is established between them, indicating a temperature change relationship between them.

[0187] Step b6: Determine the edge weight based on the absolute value of the temperature difference between the two nodes and generate a spatiotemporal graph.

[0188] Specifically, for two nodes connected by a determined edge, the absolute value of their temperature difference is calculated, and the calculated absolute value of the temperature difference is used as the weight of the edge to generate a space-time graph.

[0189] Step b7: Fusing the target image feature values, low-dimensional sparse representation, and spatiotemporal graph to generate initial fused features.

[0190] Specifically, electronic devices can encode the target image feature values, low-dimensional sparse representations, and spatiotemporal graph features separately, converting them into a "genetic" form that is easy to manipulate. For numerical target image feature values and low-dimensional sparse representations, binary encoding can be used. For spatiotemporal graph features, since they contain node and edge information, information such as node attributes and edge weights can be serialized and encoded.

[0191] The encoded characteristic genes are randomly combined to generate a certain number of individuals, forming the initial population. Each individual represents a possible feature fusion method. Individuals for the crossover operation are selected from the current population according to a specific selection strategy. For these selected individuals, a crossover point is randomly determined, and partial gene segments are exchanged to generate two new individuals. Then, individuals are randomly selected from the population for mutation with a certain mutation probability (usually a small value, such as 0.01-0.1). For these selected individuals, the mutation position in the gene sequence is randomly determined. At this determined mutation position, the gene is randomly mutated. Mutation introduces random variation to the population, preventing the algorithm from falling into local optima, uncovering unexpected feature patterns, and increasing the diversity of feature combinations.

[0192] A fitness function is designed based on task objectives such as the accuracy of crop disease and pest diagnosis and the reliability of growth status assessment. For example, the fused features are input into a trained disease and pest diagnosis model, with the model's diagnostic accuracy for test samples used as the fitness value. The fitness value of each individual in the population is calculated, and individuals are ranked from high to low in terms of fitness. Based on a set screening ratio, individuals with lower fitness are eliminated, while those with higher fitness are retained for the next generation of the population. This screening mechanism retains and strengthens feature combinations that contribute positively to the task, while gradually eliminating those that are detrimental to the task objective, thereby promoting the continuous evolution of the fused features.

[0193] Repeated crossover, mutation, and screening operations continuously generate new populations, allowing the feature combination to continuously evolve during the iterative process. Each generation inherits the excellent features of the previous generation and introduces new feature patterns through crossover and mutation, gradually developing towards a more optimal feature fusion method. Iteration termination conditions are set, such as reaching a preset maximum number of iterations or the fitness value not significantly improving over multiple generations. When the termination conditions are met, the individual with the highest fitness is selected from the last generation of the population, and its corresponding feature combination is decoded and output as the final initial fusion feature.

[0194] Step S2032: extract the initial fusion features based on the spatiotemporal attention mechanism to generate target fusion features.

[0195] Specifically, the electronic device can arrange the initial fusion features in chronological order to construct a feature sequence. Then, the electronic device can use recurrent neural networks (RNN) and its variants to model the feature sequence. After being processed by networks such as RNN, the feature representation of each time step is obtained. Then, a time attention module is used to calculate the attention weight of the time dimension. The calculation process is as follows: suppose the feature sequence after RNN processing is H=[h1,h2,…,hT], where ht represents the feature vector of the tth time step, and T is the total number of time steps. Input ht into the MLP to obtain the attention score et=wTtanh(Wht+b), where W and w are weight matrices and vectors, and b is the bias vector. Normalize the attention score to obtain the attention weight of the time dimension .

[0196] Then, the electronic device performs weighted summation on the features of each time step according to the calculated time attention weight to obtain the aggregated features in the time dimension. This step enables the model to focus on features at moments in the time series that are more important for the task.

[0197] In addition, the electronic device can divide the initial fusion features containing spatial information (such as the spatiotemporal features corresponding to thermal infrared imaging data) according to the spatial position. For the features of each spatial unit, a spatial attention module is used to calculate its importance. Suppose the feature of a spatial unit is x i , the feature map f is obtained through the convolution layer and activation function i =σ(W conv* x i +b conv ), where W conv is the convolution kernel, b conv is the bias, σ is the activation function. i Input to a global average pooling layer to obtain a scalar value g i , which represents the comprehensive feature strength of the spatial unit. Then the spatial attention weight is calculated through a fully connected layer and activation function , where N is the total number of spatial units. According to the spatial attention weight, the features of each spatial unit are weighted and summed to obtain the aggregated features in the spatial dimension This allows for highlighting features in spatially critical regions that are critical to the task.

[0198] Finally, the electronic device aggregates the time dimension into features h time and the feature h after spatial dimension aggregation space Fusion is performed. The fused feature h fusionFurther processing is performed, such as feature transformation through a fully connected layer, to obtain the final target fusion features.

[0199] Step S2033: Input the target fusion features into a preset pest and disease level determination model, and output the current pest and disease type and current pest and disease level corresponding to the crop.

[0200] Specifically, the above step S2033 may include the following steps:

[0201] Step c1, obtain the target knowledge graph.

[0202] Among them, the target knowledge graph is based on the characteristics of crop varieties, the spread of pests and diseases, and red soil slope environmental data.

[0203] Specifically, electronic devices can collect a wide range of crop health-related data. Natural language processing techniques and data mining algorithms are then used to extract knowledge from this collected data. Entities are extracted from text data, such as crop varieties, pest and disease names, and environmental factors. Relationships between entities are identified, such as "a certain crop variety is susceptible to a certain pest and disease" or "a certain pest and disease spreads faster at a certain temperature." Attribute information is extracted, such as the symptoms of the pest and disease, and the correlation between the probability of occurrence and environmental factors. Finally, the electronic devices integrate the extracted knowledge to eliminate duplication and inconsistencies. This knowledge is represented using standard formats such as the Resource Description Framework (RDF) and stored in a graph database (such as Neo4j) to construct a crop health knowledge graph. This graph structure intuitively displays the relationships between various knowledge types, facilitating subsequent query and analysis.

[0204] Step c2: input the target fusion features into a preset pest and disease determination model, and the preset pest and disease determination model outputs an initial determination result.

[0205] Among them, the initial judgment results include the initial pest and disease types and initial pest and disease levels corresponding to the crops.

[0206] Specifically, the electronic device inputs the target fusion features into the preset pest and disease determination model, first passing through the input layer of the preset pest and disease determination model, and then performing feature extraction and abstraction in the hidden layer. Taking CNN as an example, the convolution kernel of the convolution layer slides on the feature map to extract local features related to pests and diseases, such as the edge texture of the lesions, color patches, etc.; the pooling layer reduces the dimension of the extracted features, retaining important features while reducing the amount of calculation. For RNN and its variants, through the recurrent connection between neurons, the time series features are analyzed to capture the dynamic trends of pests and diseases over time, such as the spread of the disease, changes in the breeding cycle of pests, etc. As the number of network layers increases, the model continuously abstracts and combines features to form a more advanced and discriminative feature representation.

[0207] After multiple layers of processing, features of different types and levels are fused within a pre-defined pest and disease identification model. These fused features encompass multi-dimensional information about crop pests and diseases across space, time, and spectrum. The pre-defined pest and disease identification model makes comprehensive decisions based on these features. A fully connected layer maps these fused features to the output layer, where the number of neurons is set based on the type and level of pest and disease classification. The output layer's results are processed using an activation function (such as the Softmax function, which converts neuron outputs into probability distributions) to determine the probability of crops belonging to different pest and disease categories. The pre-defined pest and disease identification model selects the category with the highest probability as the initial pest and disease category.

[0208] In step c3, a query statement is constructed based on the initial pest and disease type and combined with the crop variety, growth stage, surrounding environment data, and soil data.

[0209] Specifically, the electronic device can specify the specific query target based on the initial pest and disease type. For example, if it is wheat rust, the query target may be to obtain the cause of the pest, symptoms, prevention and control methods, and prevention and control precautions for the current crop variety, growth stage, and environmental conditions.

[0210] The electronic device then rationally integrates the crop variety, growth stage, surrounding environment data, and soil data into the query statement, thereby constructing the query statement. These data elements can be arranged in a certain logical order to make the query statement clearer and more organized. For example: "For [specific crop variety], in [growth stage], the ambient temperature is [X]°C, the humidity is [X]%, the sunlight duration is [X] hours, the wind speed is [X] level, the farmland is surrounded by [surrounding vegetation conditions], the soil pH is [X], the fertility is [X], and the air permeability is [X]. For [initial pest and disease type], please inquire about its cause, symptoms, and effective prevention and control methods."

[0211] Step c4: perform a query in the target knowledge graph based on the query statement to obtain the query results.

[0212] The query results include the probability of occurrence and severity of the initial pest and disease type.

[0213] Specifically, the electronic device converts natural language query statements into a graph query language supported by the target knowledge graph. During this conversion process, a corresponding query expression is constructed based on the structure of the knowledge graph and the relationships between nodes and edges. The converted graph query language is input into the graph database to execute the query operation. The graph database traverses the knowledge graph based on the query statement, searches for nodes and edges that meet the conditions, and extracts relevant information. The query process may involve integrating information from multiple nodes and edges. Through the association relationships of the target knowledge graph, information from different sources is analyzed for correlation to obtain more comprehensive knowledge.

[0214] Extract information related to the initial pest and disease type, probability of occurrence, and severity level from the query results. This information may be stored in the knowledge graph in various formats, such as node attribute values and edge weights. Finally, the extracted results are further processed, including data formatting, probability adjustment (calibration based on actual conditions), and clarification of severity levels. Finally, the processed results are presented in a suitable format, such as a table showing the initial pest and disease type, probability of occurrence, and severity level, for easy user understanding and use.

[0215] Step c5: The initial determination result and the query result are integrated to obtain the current pests and diseases and the current pest and disease levels.

[0216] Specifically, the electronic device may assign different weights to the initial determination result and the query result based on their evaluations. Then, based on the weights assigned to the initial determination result and the query result, the electronic device may fuse the initial determination result and the query result to obtain the current pest and disease level.

[0217] For example, if the preset pest and disease identification model has been verified by a large number of experiments and has high accuracy, a higher weight can be assigned to the initial judgment result; and if the data in the target knowledge graph is rich and authoritative, the query results can also be given appropriate weights accordingly.

[0218] Step S204: determining the causes of drought levels and insect pest levels corresponding to the crops based on the surrounding environment data and soil data.

[0219] Specifically, the above step S204 may include the following steps:

[0220] Step S2041, based on the current drought level, the current pest and disease level, the surrounding environment data and the soil data as nodes, according to the relationship between the current drought level, the current pest and disease level, the surrounding environment data and the soil data, the corresponding nodes are connected with directed edges to generate a directed acyclic graph.

[0221] Specifically, the electronic device can use drought severity, current pest and disease level, surrounding environment data, and soil data as nodes. Then, based on the relationship between the current drought level and other nodes, the electronic device connects the current drought level node with corresponding nodes using directed edges. For example, drought can affect soil moisture, causing it to dry out. Therefore, a directed edge exists from the current drought level node to the soil moisture node in the soil data, representing the impact of drought on soil moisture.

[0222] The relationship between the current pest and disease level and other nodes is shown using directed edges connecting the current pest and disease level node to the corresponding node. For example, the occurrence and development of pests and diseases may be affected by environmental factors. Pests and diseases may also affect crop growth, which in turn affects soil fertility. Diseased or infested crops may not fully absorb soil nutrients, leading to changes in soil fertility. Therefore, there is a directed edge from the current pest and disease level node to the soil fertility node in the soil data.

[0223] Furthermore, directed edges are used to connect the corresponding nodes based on the relationship between the surrounding environment data and the soil data. For example, the precipitation in the surrounding environment directly affects the soil moisture content, so there is a directed edge from the precipitation node in the surrounding environment data to the soil moisture node in the soil data.

[0224] Finally, the electronic device represents each node with a graph based on the determined nodes and directed edges to generate a directed acyclic graph.

[0225] Step S2042: input the directed acyclic graph into a feature extraction network in a preset causal relationship model.

[0226] Specifically, the electronic device may input the directed acyclic graph into a feature extraction network in a preset causal relationship model.

[0227] In step S2043, the feature extraction network extracts features from the directed acyclic graph to obtain node features, edge features, and graph structure features corresponding to the directed acyclic graph, and inputs the node features, edge features, and graph structure features into the decision network in the preset causal relationship model.

[0228] Specifically, the feature extraction network extracts features from each node in the directed acyclic graph, determining the node's corresponding numerical features, categorical features, and association information between adjacent nodes. The numerical features, categorical features, and association information between adjacent nodes are then fused to generate node features. The feature extraction network then analyzes the attributes of the edges in the directed acyclic graph, such as the strength of the causal relationship represented by the edge and the direction of influence, to obtain the initial features of the edge. The feature extraction network considers the edge's contextual information within the entire directed acyclic graph and, through interaction with adjacent edges and nodes, learns more comprehensive features of the edge, thereby dynamically adjusting the weights corresponding to the initial features of the edge to obtain edge features.

[0229] The feature extraction network analyzes the topology of the DAG and extracts its overall structural features. The feature extraction network learns structural features at different levels of the DAG's hierarchy in a layered manner. Finally, after analyzing the topology and hierarchy, the feature extraction network generates a global feature vector representing the entire DAG, known as the graph structural feature. This vector contains information about the graph's structure and the relationships between nodes, providing an abstract representation of the graph as a whole.

[0230] The extracted node features, edge features, and graph structure features are then integrated to form a comprehensive feature vector. This integrated feature vector is then fed into a decision network within a pre-defined causal model. The decision network is typically a neural network-based structure, such as a multilayer perceptron or recurrent neural network. The decision network performs calculations and reasoning based on the input feature vector, predicting the causal relationships between different factors, crop growth trends, and potential decision-making measures.

[0231] Step S2044: The decision network determines the cause of the drought level and the cause of the current insect and disease level based on the node features, edge features, and graph structure features.

[0232] Specifically, the above step S2044 may include the following steps:

[0233] Step d1: Construct the initial state corresponding to the state space in the decision network based on the node features, edge features and graph structure features.

[0234] Specifically, the decision network can fuse node features, edge features, and graph structure features. These three vectors can be concatenated to form a comprehensive feature vector. This fused comprehensive feature vector is then mapped into the decision network's state space as the initial state. The state space is an abstract space used to describe the state of the system in the decision network. The initial state determines the starting point of the decision-making process. By mapping the features of the directed acyclic graph into the state space, the decision network can use this information for subsequent reasoning and decision-making, analyzing the causal relationships between different factors and their impact on crop growth.

[0235] In step d2, based on the initial state, the initial action is output from the action space corresponding to the decision network.

[0236] The initial action represents the initial estimated value of the impact of a group of factors on the current drought level and the current pest and disease level.

[0237] Specifically, after the decision network receives the initial state information, it is passed from the input layer to the hidden layer. In the hidden layer, neurons perform a weighted summation of the initial state inputs and perform nonlinear transformations using activation functions (such as ReLU and Sigmoid), thereby extracting higher-level, more abstract feature representations. These feature representations contain a deeper understanding of the current state of the crop and the relationships between various factors.

[0238] After processing by the hidden layer, the information is passed to the output layer. Based on the hidden layer's output, neurons in the output layer calculate the score or probability of each action in the action space. Finally, based on the action scores or probabilities, the decision network selects the action with the highest score or greatest probability from the action space as the initial action output.

[0239] In step d3, the decision network starts from the nodes that affect the current drought level and the current pest and disease level, and derives the causal transmission path along the directed edges in the graph.

[0240] Specifically, starting from the starting node, the decision network gradually traces the causal relationships between factors based on the direction of the edges in the directed acyclic graph. Directed edges represent the direction of influence between factors. For example, a directed edge from the temperature node to the current pest and disease level node indicates that changes in temperature will affect the current pest and disease level. The decision network analyzes the effect of each node on subsequent nodes along these directed edges. During the inference process, the decision network considers the interactions of multiple factors. A node may be affected by multiple predecessor nodes and may also affect multiple subsequent nodes. For example, the soil fertility node may be affected by multiple factors such as fertilizer application and rainfall, which in turn affects crop growth and, in turn, the current pest and disease level. The decision network comprehensively analyzes these complex interactions to accurately deduce the causal transmission path. The decision network conducts multi-level inference, gradually expanding from directly affected nodes to indirectly affected nodes. For example, changes in temperature may first affect the physiological state of crops (such as photosynthesis rate), which in turn affects the crop's resistance and ultimately the current pest and disease level. The decision network follows this multi-layered path, deeply exploring the potential causal relationships between various factors. During the derivation process, the decision network records the nodes and edges it passes through, forming a complete causal transmission path. These paths can be represented as graphs or stored in data structures (such as linked lists and trees) for subsequent analysis and processing.

[0241] Step d4: Determine the key nodes from each node based on the strength of the edge features and the position information of each node in the causal chain.

[0242] Specifically, the strength of edge features reflects the closeness or influence of the causal relationship between two nodes. The position information of a node in the causal chain describes the role and status of the node in the entire causal network. The strength of all edge features is evaluated and ranked. Edges with higher strength indicate that there is a stronger causal relationship between the two nodes they connect, and the nodes connected by these edges are more likely to be key nodes. The directionality of the edge is also important because it determines the direction of transmission of the causal relationship. The strength of the edge pointing from the starting node to other nodes reflects the driving effect of the starting node on the subsequent nodes; while the strength of the edge pointing to the target node reflects the degree of influence of various factors on the target node.

[0243] The starting node is typically the source of a causal relationship and plays a crucial role in driving the entire causal chain. If a starting node is connected to multiple subsequent nodes via high-strength edges, it is likely a key node. Intermediate nodes serve as a link between upstream and downstream nodes in a causal chain. If an intermediate node is connected to multiple high-strength edges with different causal directions—that is, it receives influence from multiple upstream nodes and transmits influence to multiple downstream nodes—it is often a key node. End nodes are the final destination of causal relationships. While they do not directly influence other nodes themselves, they are the destination of the entire causal chain. The feature strength of the edges connecting the end node reflects the contribution of each factor to the final outcome. If the edge feature strength between a node and the end node is particularly high, it indicates that the node has a significant impact on the final outcome and is a key node.

[0244] Key nodes are determined based on a pre-set evaluation criterion, taking into account both edge feature strength and node location information. This criterion can be a quantitative indicator, such as a threshold for edge feature strength. A node is identified as a key node when the number of high-strength edges connecting it to other nodes exceeds a certain percentage, and the node's position in the causal chain meets the characteristics of a key node (e.g., a starting node, an important intermediate node, or a node with a strong connection to the terminal node).

[0245] Step d5: According to each key node, the initial estimated value of the impact of each factor on the current drought level and pest and disease level in the initial action is adjusted to obtain the target estimated value of the impact of each factor on the current drought level and pest and disease level.

[0246] Specifically, for key nodes related to the current drought level, initial estimated values of the impact of various factors corresponding to the key nodes in the initial action on the current drought level are adjusted according to the impact of the key nodes on the current drought level.

[0247] For key nodes related to the current pest and disease level, the initial estimated value of the impact of each factor corresponding to the key node on the current pest and disease level in the initial action is adjusted according to the impact of the key node on the current pest and disease level. In addition, multiple key nodes will affect the current drought level and the current pest and disease level at the same time, and there will also be interactions between them. Therefore, when adjusting the estimated value of the impact, these synergistic effects should be comprehensively considered. When evaluating the impact of irrigation volume on the current drought level, it is necessary not only to consider the direct effect of irrigation on soil moisture and crop drought conditions, but also to consider the indirect impact of changes in irrigation volume on the current pest and disease level by affecting crop growth conditions.

[0248] Step d6: According to the target estimated values of the influence quantities, a target factor having a target estimated value of the influence quantity greater than a preset threshold is determined from among the factors.

[0249] Specifically, the target estimated value of the impact amount corresponding to each factor is compared with the preset threshold. When the target estimated value of the impact amount of a factor on the current drought level and / or pest and disease level is greater than the preset threshold, the factor is identified as a target factor.

[0250] Step d7: Determine the cause of the drought level and the cause of the current insect and disease level based on the target factors.

[0251] Specifically, all target factors related to drought severity and their influencing mechanisms are comprehensively considered to comprehensively summarize the reasons for increases or decreases in drought severity. For example, an increase in drought severity may be the result of a combination of factors, such as low precipitation, insufficient irrigation, loose soil texture, and high temperatures leading to rapid evaporation. Regarding pest and disease severity, a comprehensive analysis of all relevant target factors identifies the causes of changes in pest and disease severity. These factors, such as high pest density, irrational pesticide use, poor crop disease resistance, and high field humidity conducive to pathogen growth, all contribute to the occurrence and progression of pests and diseases.

[0252] Step S205: outputting an irrigation strategy based on the current drought level and the cause of the drought level; the irrigation strategy includes an intermittent supplementary irrigation mode and a continuous irrigation mode.

[0253] For details about this step, please refer to the above description of step S105 and will not be repeated here.

[0254] Step S206: Outputting pest control measures corresponding to the crops based on the current pest type, the current pest level, and the cause of the current pest level.

[0255] For details about this step, please refer to the above description of step S106 and will not be repeated here.

[0256] The red soil slope crop management method provided in the embodiment of the present application inputs crop image data, crop spectral data, crop thermal infrared imaging data, soil physical property data, soil chemical property data, and water property data into a preset drought level determination model. This avoids the limitation of relying solely on a single data type for judgment, thereby more accurately assessing the drought level of crops. The preset drought level determination model extracts features from the input data and performs feature fusion to obtain a fused feature vector. This makes the data easier to process and analyze, while fully exploring the potential relationships between the data, further improving the representativeness and effectiveness of the data. Based on the variational autoencoder, the fused feature vector is mapped to the mean vector and variance vector in the latent space. Dimensionality reduction and feature compression of the high-dimensional fused feature vector are achieved. At the same time, the variational autoencoder can also learn the probability distribution of the data, providing a probabilistic basis for subsequent sampling operations. Sampling is performed based on the mean vector and variance vector to obtain at least one sampling point. Different feature combinations can be obtained from the latent space, avoiding falling into local optimal solutions and improving the generalization ability of the model. Each layer in the pre-set decoder calculates the similarity between each sub-feature at a sampling point and a pre-set drought key feature template, accurately measuring the degree of correlation between each sub-feature and the drought key feature. Based on the calculated similarity, weight information is determined for each sub-feature. This allows for the rational allocation of importance to each sub-feature in the feature representation based on its correlation with the drought key feature. Based on the corresponding weight information for each sub-feature, a weighted vector is generated. This weighting process enhances the weighted vector in the weighted vector for sub-features that are most influential in determining drought severity. This weighted vector is then convolved with the convolutional layer in the pre-set decoder to generate a target feature vector. The convolutional layer effectively extracts local information and spatial structure of features. Through these operations, the generated target feature vector comprehensively considers the weight information of each sub-feature and their interrelationships, providing a more comprehensive and accurate representation of the crop drought condition. Ultimately, the current drought severity output based on this target feature vector is more accurate and reliable, providing strong support for decision-making in agricultural production. Based on the target feature vector, the current drought severity corresponding to the crop is output. It can directly provide a clear basis for decision-making in agricultural production. Accurate drought level assessment can help farmers timely understand the moisture status of crops and take appropriate irrigation or other drought relief measures, thereby contributing to the rational use of water resources, improving crop yields and quality, and reducing losses caused by drought.

[0257] In addition, feature extraction is performed on crop image data to generate initial color, texture, and shape features, providing a rich information foundation for subsequent accurate pest and disease identification. The initial color, texture, and shape features are mapped to discrete value ranges to create an image feature search space. This makes feature representation more standardized and orderly, while also limiting the feature value range and reducing feature uncertainty. Encoding rules are formulated based on the discrete value range of the image feature search space, providing a clear basis for qubit encoding and decoding. Each initial qubit is randomly assigned an initial state, allowing the algorithm to search from multiple different starting points in the search space, increasing the randomness and comprehensiveness of the search. An initial quantum population is initialized. Based on a fitness function, the fitness evaluation results of the image feature combinations corresponding to each initial quantum state in the initial quantum population are calculated. The fitness function quantitatively evaluates the effectiveness of each image feature combination for identifying crop pests and diseases. The rotation angle of each initial qubit is calculated based on the fitness evaluation results. The algorithm adaptively adjusts the search direction based on the current search situation, evolving toward a direction with higher fitness, thereby more quickly finding the optimal image feature combination. A quantum rotating gate operation is performed on each initial qubit based on the calculated rotation angle to obtain the updated qubits and the updated quantum state composed of them, thereby achieving efficient exploration of the search space. Each updated quantum state is measured and converted into classical information to obtain a classical state. According to the encoding rules, the classical state is decoded into the initial image feature value. This achieves the conversion from quantum information to image features. The initial image feature values are screened to obtain the target image feature values. This screening can remove some features that have a smaller contribution to pest and disease identification, further optimize the image features, and improve the quality and representativeness of the features. Then, a preset encoder is used to encode the crop spectral data to obtain a low-dimensional sparse representation, allowing the preset pest and disease grade determination model to focus more on spectral regions sensitive to pests and diseases. Crop thermal infrared imaging data is arranged in chronological order, with each pixel in the thermal infrared image as a node. Each node corresponds to a specific spatiotemporal location, and its attribute is the temperature value at that time. Edges between two nodes are determined based on the temperature change relationship between two nodes at adjacent time points and spatial locations. The edge weight is determined based on the absolute value of the temperature difference between the two nodes, generating a spatiotemporal graph. This method fully considers the temporal and spatial continuity and dynamics of the data, clearly depicting the temporal and spatial trends of crop surface temperature. Initial fused features are generated by fusion of target image feature values, low-dimensional sparse representations, and the spatiotemporal graph. This reduces the bias and uncertainty that can arise from a single data source. Feature extraction from the initial fused features is then performed using a spatiotemporal attention mechanism to generate target fused features.This improves the quality and effectiveness of the features and ensures the accuracy of the target fusion features obtained. Obtain the target knowledge graph. Input the target fusion features into the preset pest and disease determination model, which outputs the initial judgment result, ensuring the accuracy of the output initial judgment result and providing a basic result for subsequent in-depth analysis. This saves time and effort in manual analysis, and the judgment using the model has a certain degree of objectivity and accuracy. With the initial pest and disease type as the core, a query statement is constructed in combination with the crop variety, growth stage, surrounding environment data, and soil data. Based on the query statement, a query is performed in the target knowledge graph to obtain the query results. The query results further enrich the understanding of the pest and disease situation, taking into account the possibility and severity of pests and diseases under specific environmental and crop conditions, and provide more basis for accurately assessing the pest and disease situation. The initial judgment result and the query result are fused to obtain the current pest and disease and the current pest and disease level. It can fully combine the preliminary judgment of the model and the rich domain knowledge in the knowledge graph, complement and verify each other, so as to obtain more accurate and more actual current pests and diseases and current pest and disease levels, improve the accuracy and reliability of pest and disease diagnosis, and provide a more accurate basis for subsequent effective management measures.

[0258] Finally, a directed acyclic graph (DAG) is generated, with the current drought level, current pest and disease level, surrounding environmental data, and soil data as nodes. Directed edges connect the corresponding nodes based on the relationships between the current drought level, current pest and disease level, surrounding environmental data, and soil data. This allows complex multi-source data relationships to be represented in an intuitive graphical structure. The DAG is then fed into a feature extraction network within a pre-defined causal model, automatically extracting key features from the graph data. The feature extraction network extracts features from the DAG, obtaining node features, edge features, and graph structural features corresponding to the DAG. These features are then fed into a decision network within the pre-defined causal model. This allows the decision network to comprehensively consider all aspects of the data and more accurately identify causal relationships within the data. Based on the node features, edge features, and graph structural features, an initial state corresponding to the state space of the decision network is constructed, providing a foundation for subsequent decision-making. Based on this initial state, an initial action is output from the action space corresponding to the decision network, providing a starting point for more precise analysis and adjustments. This helps to preliminarily determine which factors may have a significant impact on drought and pest and disease levels. Starting from nodes that influence the current drought and pest levels, the decision network infers causal transmission paths along the directed edges in the graph. This provides a deep understanding of the causal mechanisms of the entire system and identifies potential causal paths, providing strong support for accurately determining the causes of drought and pest levels. Key nodes are identified from each node based on the strength of edge features and the position of each node in the causal chain. This allows accurate identification of factors that play a key role in the overall causal relationship. Based on each key node, the initial estimate of each factor's impact on the current drought and pest levels in the initial action is adjusted to obtain a target estimate of each factor's impact. This corrects for biases in the initial estimate and improves the accuracy of quantifying the impact of each factor, providing more reliable data support for subsequent identification of target factors and accurate analysis of the causes of drought and pest levels. Based on the target estimate of each impact, factors with a target estimate exceeding a preset threshold are identified. This allows screening of factors with significant impact on the current drought and pest levels. Based on the target factors, the causes of the drought level and the current insect and disease level are determined, ensuring the accuracy of the determined causes of the drought level and the current insect and disease level.

[0259] Although the embodiments of the present invention have been described with reference to the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention. Such modifications and variations are all within the scope defined by the appended claims.

Claims

1. A method for managing crops on red soil slopes, characterized in that: Methods include: Acquire crop data, surrounding environment data, and soil data corresponding to crops planted in the target red soil slope area; the crop data includes crop image data, crop spectral data, and crop thermal infrared imaging data; the surrounding environment data includes at least one of light intensity, humidity, and wind speed; and the soil data includes at least one of soil physical property data, soil chemical property data, soil biological property data, and water property data; Identify crop data and soil data to determine the current drought level corresponding to the crops; Extract features from crop image data to generate initial color features, initial texture features, and initial shape features; Mapping the initial color features, initial texture features, and initial shape features to discrete value ranges to construct an image feature search space; Formulate encoding rules based on the discrete value range of the image feature search space; Randomly assign an initial state to each initial qubit; Initialize an initial quantum population; Based on the fitness function, the fitness evaluation results of the image feature combinations corresponding to each initial quantum state in the initial quantum population are calculated; Calculate the difference between the fitness evaluation result and the target fitness, and calculate the rotation angle of each initial quantum bit based on the difference; Perform a quantum rotating gate operation on each initial quantum bit according to the rotation angle to obtain each updated quantum bit and an updated quantum state composed of the updated quantum bits; Measure each updated quantum state, convert the updated quantum state into classical information, and obtain the classical state; According to the encoding rules, the classical state is decoded into the initial image feature value; Screen the initial image feature values to obtain the target image feature values; The crop spectral data is encoded using a preset encoder to obtain a low-dimensional sparse representation; Arrange the crop thermal infrared imaging data in chronological order, and use the pixel points in each thermal infrared image in the crop thermal infrared imaging data as nodes; Each node corresponds to a specific spatiotemporal location, and the node's attribute is the temperature value of the spatiotemporal location at the corresponding moment; Determine the edge between two nodes based on the temperature change relationship between two nodes at adjacent time points and adjacent spatial positions; According to the absolute value of the temperature difference between two nodes, the edge weight is determined to generate a space-time graph; Perform feature fusion on the target image feature value, low-dimensional sparse representation and spatiotemporal graph to generate initial fusion features; Extract the initial fusion features based on the spatiotemporal attention mechanism to generate the target fusion features; Input the target fusion features into the preset pest and disease level determination model to output the current pest and disease type and current pest and disease level corresponding to the crop; Determine the causes of drought levels and pest and disease levels corresponding to crops based on surrounding environmental data and soil data; Output irrigation strategies based on the current drought level and the cause of the drought level; irrigation strategies include intermittent irrigation mode and continuous irrigation mode; Based on the current pest and disease type, current pest and disease level, and the cause of the current pest and disease level, output the corresponding pest and disease control measures for crops.

2. The method according to claim 1, characterized in that The soil data includes soil physical property data, soil chemical property data, and water property data. Identifying the crop data and the soil data to determine the current drought level corresponding to the crop includes: Inputting the crop image data, the crop spectral data, the crop thermal infrared imaging data, the soil physical property data, the soil chemical property data, and the water property data into a preset drought level determination model; The preset drought level determination model extracts features from input data and performs feature fusion to obtain a fused feature vector; Performing feature extraction on the fused feature vector to generate a target feature vector; Based on the target feature vector, the current drought level corresponding to the crop is output.

3. The method according to claim 2, characterized in that The step of extracting features from the fused feature vector to generate a target feature vector includes: Mapping the fused feature vector to a mean vector and a variance vector in a latent space based on a variational autoencoder; Sampling is performed based on the mean vector and the variance vector to obtain at least one sampling point; the sampling point contains mixed coding information of the multimodal data in the latent space; the mixed coding information includes multiple sub-features; Based on a preset decoder, each of the sampling points is converted and combined to generate the target feature vector with the same dimension as the input data.

4. The method according to claim 3, characterized in that The converting and combining the sampling points based on the preset decoder to generate the target feature vector having the same dimension as the input data includes: Each layer of the preset decoder calculates the similarity between each sub-feature in the sampling point and a preset drought key feature template; Determining weight information of each sub-feature based on the calculated similarity; Based on the weight information corresponding to each sub-feature, weighting the sub-features to obtain a weighted vector; The weighted vector is convolved through a convolution layer in the preset decoder to generate the target feature vector.

5. The method according to claim 1, wherein The initial quantum population includes multiple initial quantum states, each of which is represented by a group of initial quantum bits, representing a possible combination of image features.

6. The method according to claim 1, characterized in that The step of inputting the target fusion feature into a preset pest and disease level determination model and outputting the current pest and disease type and the current pest and disease level corresponding to the crop includes: Obtaining a target knowledge graph; the target knowledge graph is based on crop variety characteristics, pest and disease transmission patterns, and red soil slope environmental data; Inputting the target fusion feature into a preset pest and disease determination model, the preset pest and disease determination model outputs an initial determination result; the initial determination result includes the initial pest and disease type and initial pest and disease level corresponding to the crop; Constructing a query statement based on the initial pest and disease type and combining the crop variety, growth stage, surrounding environment data, and soil data; Querying the target knowledge graph based on the query statement to obtain query results; the query results include the probability of occurrence and severity level of the initial pest and disease type; The initial determination result and the query result are integrated to obtain the current pests and diseases and the current pest and disease level.

7. The method according to claim 1, characterized in that The determining of the drought level cause and the pest and disease level cause corresponding to the crop based on the surrounding environment data and the soil data includes: Based on the current drought level, the current pest and disease level, the surrounding environment data, and the soil data as nodes, and according to the relationship between the current drought level, the current pest and disease level, the surrounding environment data, and the soil data, corresponding nodes are connected with directed edges to generate a directed acyclic graph; Inputting the directed acyclic graph into a feature extraction network in a preset causal relationship model; The feature extraction network extracts features from the directed acyclic graph to obtain node features, edge features, and graph structure features corresponding to the directed acyclic graph, and inputs the node features, edge features, and graph structure features into a decision network in the preset causal relationship model; The decision network determines the cause of the drought level and the cause of the current insect and disease level based on the node features, the edge features, and the graph structure features.

8. The method according to claim 7, characterized in that The decision network determines the cause of the drought level and the cause of the current insect and disease level based on the node features, the edge features, and the graph structure features, including: Constructing an initial state corresponding to a state space in the decision network according to the node features, the edge features, and the graph structure features; Based on the initial state, outputting an initial action from the action space corresponding to the decision network; the initial action represents an initial estimated value of the impact of a group of factors on the current drought level and the current pest and disease level; The decision network starts from the nodes that affect the current drought level and the current pest and disease level, and derives a causal transmission path along the directed edges in the graph; Determining key nodes from the nodes based on the strength of the edge features and the position information of each node in the causal chain; Adjusting the initial estimated value of the impact of each factor on the current drought level and insect pest level in the initial action according to each of the key nodes to obtain a target estimated value of the impact of each factor on the current drought level and insect pest level; According to the target estimated values of the impact quantities, a target factor having a target estimated value of the impact quantity greater than a preset threshold value is determined from among the factors; The cause of the drought level and the cause of the current insect and disease level are determined based on the target factors.

Images