Corn disease and pest density prediction method and system based on big data
By generating standardized field-time multimodal feature data cubes and a self-supervised pre-trained network, pest and disease features are decoupled, solving the problems of weak generalization ability and insufficient reliability of existing models in multi-factor coupled environments, and achieving more stable and interpretable prediction of maize pest and disease density.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN ACAD OF AGRI SCI
- Filing Date
- 2026-02-24
- Publication Date
- 2026-05-12
AI Technical Summary
Existing maize disease and pest prediction models rely on high-cost labeled data and struggle to separate disease and pest features in crop growth environments with multiple coupled factors, resulting in weak generalization ability and insufficient prediction reliability.
By generating standardized field-time series multimodal feature data cubes, and utilizing a self-supervised pre-trained feature extraction network based on unlabeled data, we can decouple pest and disease, environmental contamination, and shared background features to generate predicted values for maize pest and disease density and provide visual early warning.
The model's stability and predictive reliability under different environments were improved, its sensitivity to subtle stresses was enhanced, interference from water and fertilizer stress and variety differences was eliminated, and the interpretability of the prediction results was improved.
Smart Images
Figure CN122020181A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural information technology, and in particular to a method and system for predicting the density of corn pests and diseases based on big data. Background Technology
[0002] In the field of agricultural information technology, crop disease and pest prediction based on multi-source data and artificial intelligence is an important research direction. Existing technologies usually integrate multi-source data such as satellite remote sensing, meteorology, and sensors. After preprocessing, feature extraction and prediction are performed using a deep learning model that combines convolutional and recurrent neural networks. A prediction model is built through supervised learning, realizing the mapping from data to disease and pest levels, and promoting the development of agricultural monitoring from qualitative to quantitative.
[0003] Existing technologies face two core bottlenecks: the models heavily rely on a large amount of precisely labeled pest and disease data, but such data is costly to acquire and has few samples, which limits the model's generalization ability in new environments. Crop growth abnormalities are the result of multiple factors working together, and existing models have difficulty effectively distinguishing pest and disease stress from other environmental stresses from mixed features. They are also prone to learning spurious correlations, which reduces the reliability and interpretability of predictions. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a big data-based method for predicting the density of maize diseases and pests to solve the problems of weak generalization ability and insufficient prediction reliability caused by the high cost of training data labeling and the difficulty in separating disease and pest-specific features from crop phenotypes coupled with multiple factors.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for predicting the density of maize pests and diseases based on big data, which includes accessing and preprocessing multi-source heterogeneous agricultural data to generate a standardized field-time series multimodal feature data cube. The standardized field-time series multimodal feature data cube is input into a general feature extraction network for feature encoding, and the primary fusion feature vector of each field is output. By utilizing unlabeled field-time multimodal feature data cubes, a general feature extraction network is optimized through a self-supervised pre-training task, outputting a high-quality primary fusion feature vector containing semantic information about crop physiological states. High-quality primary fusion feature vectors containing semantic information about crop physiological state are input into the causal decoupling module. The decoupling is mapped to the disease and pest feature subspace, the environmental mixing feature subspace, and the shared background feature subspace to obtain disease and pest related features. Based on the decoupled disease and pest related features, the predicted value of maize disease and pest density is generated. The predicted density of corn pests and diseases is processed to generate a visual early warning product.
[0007] As a preferred embodiment of the big data-based maize pest and disease density prediction method of the present invention, the method includes the following steps: accessing and preprocessing multi-source heterogeneous agricultural data to generate a standardized field-time series multimodal feature data cube: Receive multispectral and hyperspectral images, meteorological grid data, field spore capture data, and plant protection history records, and perform time alignment and spatial coordinate unification; Based on the field boundaries, the processed data is spatially cropped and attribute extracted to obtain the image sequence, meteorological sequence, spore data sequence and ledger record for each field. Cloud detection and filling are performed on the field image sequence to generate a cloudless image sequence, and vegetation index sequence is extracted using preset band combinations and mathematical transformation rules; Missing values were filled, outliers were removed and standardized from cloudless image sequences, meteorological sequences, spore data sequences, ledger records and vegetation index sequences. The sequences were then stacked and integrated according to a uniform time series length and feature dimension to generate a standardized field-time series multimodal feature data cube.
[0008] As a preferred embodiment of the big data-based maize pest and disease density prediction method of the present invention, the method includes the following steps: inputting a standardized field-time series multimodal feature data cube into a general feature extraction network for feature encoding, and outputting a primary fusion feature vector for each field. In the general feature extraction network, a spatial feature extraction module for extracting spatial features and a temporal feature extraction module for extracting temporal features are configured; The standardized field-time series multimodal feature data cube is input into the general feature extraction network; The spatial feature extraction module performs multi-scale convolution and pooling operations on the image data components in the standardized field-time series multimodal feature data cube to generate spatial features; The temporal feature extraction module sequentially inputs the sequence data components in the standardized field-temporal multimodal feature data cube into the recurrent neural network unit according to the time order. Through the gating mechanism and state transmission within the unit, the sequence information is gradually updated and fused to generate temporal features. Spatial features and temporal features are concatenated into vectors to form a primary fusion feature vector for each field.
[0009] As a preferred embodiment of the big data-based maize pest and disease density prediction method of the present invention, the method includes the following steps: utilizing unlabeled field-time series multimodal feature data cubes, optimizing a general feature extraction network through a self-supervised pre-training task, and outputting a high-quality primary fusion feature vector containing semantic information about crop physiological states. The general feature extraction network includes a spatial feature extraction module, a temporal feature extraction module, and a feature fusion module; The spatial feature extraction module extracts spatial features from the randomly masked unlabeled field-temporal multimodal feature data cube, the temporal feature extraction module extracts temporal features from the randomly masked unlabeled field-temporal multimodal feature data cube, and the feature fusion module fuses the spatial features and temporal features. The reconstruction head network then predicts the data content of the masked part based on the fused features. The spatial feature extraction module and the temporal feature extraction module extract spatial features and temporal features from the unlabeled field-temporal multimodal feature data cube, respectively, and align them to a shared feature space through the feature fusion module; The temporal feature extraction module extracts temporal features from the unlabeled field-temporal multimodal feature data cube injected with artificially synthesized anomaly patterns. The classification head network then judges the presence of anomalies and the location of anomalies in the sequence based on the temporal features. By jointly optimizing the total loss function of the multimodal mask reconstruction task, cross-modal contrast alignment task, and temporal anomaly detection task, the parameters of the spatial feature extraction module, temporal feature extraction module, feature fusion module, and related task head networks are updated, and a high-quality primary fusion feature vector containing semantic information about crop physiological state is output.
[0010] As a preferred embodiment of the big data-based maize pest and disease density prediction method of the present invention, the method includes: inputting a high-quality primary fusion feature vector containing semantic information of crop physiological state into a causal decoupling module, and decoupling and mapping it to a pest and disease feature subspace, an environmental confounding feature subspace, and a shared background feature subspace to obtain pest and disease related features, including the following steps: The causal decoupling module includes a pest and disease feature mapping unit, an environmental confounding feature mapping unit, and a shared background feature mapping unit. High-quality primary fusion feature vectors containing semantic information about crop physiological states are input into the pest and disease feature mapping unit, the environmental confounding feature mapping unit, and the shared background feature mapping unit, respectively, to generate pest and disease feature subspace vectors, environmental confounding feature subspace vectors, and shared background feature subspace vectors, thus forming decoupled feature representations. Extract the pest and disease feature subspace vector as pest and disease related features from the pest and disease feature subspace vector, the environmental mixing feature subspace vector, and the shared background feature subspace vector.
[0011] As a preferred embodiment of the big data-based corn pest and disease density prediction method of the present invention, the method comprises the following steps: generating predicted corn pest and disease density values based on decoupled pest and disease related features. A multilayer perceptron regressor for density prediction is constructed based on the feature subspace vector of pests and diseases, and pest and disease related features are input into the multilayer perceptron regressor. The multilayer perceptron regressor performs nonlinear transformation and feature mapping on pest and disease related features to predict the density of corn pests and diseases.
[0012] As a preferred embodiment of the big data-based corn pest and disease density prediction method of the present invention, the method includes the following steps: processing the predicted corn pest and disease density values to generate a visual early warning product: The predicted density of maize diseases and pests is corrected for deviation based on historical error statistics to generate corrected predicted density values of maize diseases and pests. Uncertainty assessment based on Monte Carlo Dropout was performed on the corrected maize disease and pest density prediction values to generate confidence intervals for the maize disease and pest density prediction values. The corrected predicted density of corn pests and diseases is compared with a preset economic threshold to generate a preliminary warning level. Combining the confidence interval of the predicted density of corn pests and diseases with the preliminary warning level, a visualized early warning product is generated, showing the geographic spatial distribution, confidence range, and tiered prevention and control recommendations.
[0013] Secondly, the present invention provides a corn pest and disease density prediction system based on big data, including a processing module for accessing and preprocessing multi-source heterogeneous agricultural data to generate a standardized field-time series multimodal feature data cube. The feature encoding module inputs the standardized field-time series multimodal feature data cube into the general feature extraction network for feature encoding, and outputs the primary fusion feature vector for each field. The optimization module utilizes unlabeled field-time series multimodal feature data cubes to optimize the general feature extraction network through a self-supervised pre-training task, outputting a high-quality primary fusion feature vector containing semantic information about crop physiological states. The causal decoupling module inputs high-quality primary fusion feature vectors containing semantic information about crop physiological state into the causal decoupling module, decoupling and mapping them to the disease and pest feature subspace, the environmental mixing feature subspace, and the shared background feature subspace to obtain disease and pest related features. Based on the decoupled disease and pest related features, the predicted value of maize disease and pest density is generated. The visualization module processes the predicted density of corn pests and diseases to generate visualized early warning products.
[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the big data-based corn pest and disease density prediction method described in the first aspect of the present invention.
[0015] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the big data-based corn pest and disease density prediction method as described in the first aspect of the present invention.
[0016] The beneficial effects of this invention are as follows: By utilizing unlabeled big data and self-supervised pre-training tasks to optimize the feature extraction network, the network learns the intrinsic patterns between crop growth and multi-source data without the need for manual labeling, reducing dependence on scarce labeled data and enhancing the model's feature generalization ability and sensitivity to subtle stresses; through a causal decoupling mechanism, the fused features are separated into mutually independent subspaces of pests and diseases, environmental contamination, and shared background, actively identifying and extracting specific features triggered only by pest and disease biological stress, and removing the interference of water and fertilizer stress and varietal differences, thereby significantly improving the interpretability, reliability, and stability of the prediction results under different environments. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 This is a flowchart of a method for predicting the density of corn diseases and pests based on big data.
[0019] Figure 2 This is a schematic diagram of a corn pest and disease density prediction system based on big data. Detailed Implementation
[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0021] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0022] Secondly, the term "one embodiment" or "example" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the invention. The appearance of an embodiment in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that mutually excludes other embodiments.
[0023] Reference Figures 1-2 This is one embodiment of the present invention, which provides a method for predicting the density of corn pests and diseases based on big data, including the following steps: S1. Access and preprocess multi-source heterogeneous agricultural data to generate a standardized field-time series multimodal feature data cube.
[0024] S1.1 Receive multispectral and hyperspectral images, meteorological grid data, field spore capture data, and plant protection history records, and perform time alignment and spatial coordinate unification.
[0025] Furthermore, multispectral and hyperspectral imagery typically have specific imaging times and geographic projection coordinate systems, meteorological grid data comes from numerical forecasts or reanalysis products, and has regular spatial grids and time steps, field spore capture data is collected by sensors deployed in specific geographic locations, has a high temporal frequency but sparse spatial locations, while plant protection history logs are event-based data recorded by field plots, with discrete time points. Time alignment aims to unify all data into a common temporal reference framework.
[0026] Specifically, for example, using daily or ten-day periods as time nodes, for data with frequencies other than this time sequence, meteorological grid data can be matched through time interpolation, image data can select the product closest to the time point, spore capture data can be accumulated or averaged daily, and ledger records can be mapped to the corresponding time nodes, thereby ensuring that all data are comparable and fused in the time dimension. Spatial coordinate unification involves converting data with different spatial reference systems, such as multispectral and hyperspectral images and meteorological grid data, into a unified geographic coordinate system through resampling and projection transformation, and ensuring that it is consistent with the coordinate system of the field vector boundary. The geographical location of field spore capture data also needs to be converted to this coordinate system.
[0027] S1.2. Using the field boundaries as a reference, spatial cropping and attribute extraction are performed on the processed data to obtain the image sequence, meteorological sequence, spore data sequence and ledger record for each field.
[0028] Furthermore, data extraction and organization are performed using the vector boundary of each independent physical field as a spatial unit. Spatial clipping operations cut multispectral and hyperspectral images, meteorological grid data, and other planar or raster data under a unified coordinate system according to the boundary range of each field, thereby obtaining the image patch sequence corresponding to each field. Attribute extraction is performed on point data such as field spore capture data and tabular plant protection history records. The former assigns the captured data to the specific field formation sequence through the inclusion relationship between spatial location and field boundary, while the latter is directly associated and extracted based on field identifiers.
[0029] Specifically, each field has a set of time-aligned and spatially matched data sequences, including image sequences, meteorological sequences, spore data sequences, and ledger records. This transforms macro-level regional data into micro-level datasets with specific fields as the analysis objects, establishing the field as the core analysis unit. The organization method closely aligns with the actual scenarios of agricultural production management.
[0030] S1.3 Perform cloud detection and filling on the field image sequence to generate a cloudless image sequence, and extract the vegetation index sequence using preset band combinations and mathematical transformation rules.
[0031] Furthermore, cloud detection algorithms identify pixels covered by clouds in the image and use temporal interpolation methods, such as linear or nonlinear filling based on cloudless images from previous and later periods, to generate a sequence of cloudless images, thereby ensuring the continuity and availability of image data in time. Vegetation index sequences are extracted using preset band combinations and mathematical transformation rules, such as calculating the normalized vegetation index using the reflectance of near-infrared and red light bands.
[0032] Specifically, it solves the data quality problem. More importantly, by extracting vegetation index sequences, it transforms the original spectral reflectance information into biophysical parameters that can directly and sensitively reflect the vegetation growth status, biomass, and canopy structure. As one of the core inputs for subsequent analysis, the vegetation index sequences can be more directly correlated with crop physiological status and pest and disease stress, improving the interpretability of feature information and its relevance to agricultural knowledge, and providing key time series signals for dynamic changes in crop health.
[0033] S1.4. Missing values are filled, outliers are removed and standardized from cloudless image sequences, meteorological sequences, spore data sequences, ledger records and vegetation index sequences. The sequences are stacked and integrated according to a uniform time series length and feature dimension to generate a standardized field-time series multimodal feature data cube.
[0034] Furthermore, data quality cleaning is performed on cloudless image sequences, meteorological sequences, spore data sequences, ledger records, and vegetation index sequences. This includes completing missing values using interpolation methods, identifying and removing or reasonably correcting outliers that significantly deviate from the normal range, standardizing all features to eliminate dimensional differences, truncating or completing all these cleaned and standardized sequences according to a uniform time series length, and stacking and integrating features from different sources along the feature dimension to finally generate a standardized field-time series multimodal feature data cube.
[0035] Specifically, a highly structured and standardized multi-dimensional data object was constructed, using fields as samples, time as the order axis, and multiple features as dimensions. This perfectly encapsulates the multi-angle state information of each field during its temporal evolution. This format greatly facilitates the input processing of subsequent deep learning models, enabling convolutional neural networks to handle its spatial-spectral dimension and recurrent neural networks to handle its temporal dimension. This provides an ideal data foundation for efficient and unified spatiotemporal feature learning and is a key guarantee for the smooth progress of the entire methodology.
[0036] S2. Input the standardized field-time series multimodal feature data cube into the general feature extraction network for feature encoding, and output the primary fusion feature vector of each field.
[0037] S2.1 In the general feature extraction network, configure a spatial feature extraction module for extracting spatial features and a temporal feature extraction module for extracting temporal features.
[0038] S2.2 Input the standardized field-time series multimodal feature data cube into the general feature extraction network.
[0039] Furthermore, the standardized field-temporal multimodal feature data cube is input into a general feature extraction network as a whole, starting the feature extraction process. This maintains the integrity of the data object, enabling the subsequent spatial feature extraction module and temporal feature extraction module to read their respective required components in parallel or on demand from the same standardized and regular data source. This ensures that spatial features and temporal features originate from the same spatiotemporal entity, laying an alignment foundation for subsequent feature fusion and avoiding spatiotemporal misalignment or semantic deviation problems that may arise from processing data from different sources separately.
[0040] S2.3 The spatial feature extraction module performs multi-scale convolution and pooling operations on the image data components in the standardized field-time series multimodal feature data cube to generate spatial features.
[0041] Furthermore, the spatial feature extraction module performs multi-scale convolution and pooling operations on the image-type data components in the data cube, such as each band of a multispectral image or a vegetation index layer. Multi-scale convolution captures features at different spatial scales, from subtle leaf textures to large-scale canopy structures, by using convolution kernels of different sizes. For example, small-sized convolution kernels can identify tiny discoloration spots on leaf edges, while large-sized convolution kernels can perceive the overall uniformity of crop growth within a field.
[0042] Specifically, pooling operations gradually reduce the spatial resolution of the feature map, achieving spatial downsampling. This expands the receptive field while extracting more spatially robust features and reducing computational complexity. Through this series of operations, the complex spatial patterns and local correlations contained in image-like data are effectively encoded into a condensed, high-dimensional spatial feature vector. This vector represents the comprehensive state of the field in the visual spectral space. By simulating the hierarchical processing mechanism of the visual system, spatial hierarchical features beneficial to the task are automatically learned and abstracted from the pixel-level raw data.
[0043] S2.4 The temporal feature extraction module takes the sequence data components in the standardized field-temporal multimodal feature data cube, inputs them sequentially into the recurrent neural network unit according to the time order, and gradually updates and merges the sequence information through the gating mechanism and state transmission inside the unit to generate temporal features.
[0044] Furthermore, the temporal feature extraction module processes the sequence-type data components in the data cube, such as the average temperature time series, precipitation time series, and spore capture quantity time series for a certain pixel or field. Based on the chronological order, these sequence data are sequentially input into the recurrent neural network unit. The gating mechanisms within the recurrent neural network unit, such as the input gate, forget gate, and output gate, dynamically control the flow of information and the updating of memory.
[0045] Specifically, when processing data at each time step, the unit combines the current input with the hidden state from the previous time step, and uses gating to determine which historical information needs to be retained and which needs to be forgotten, as well as how to generate the current output and pass the hidden state to the next time step. This allows the time-series feature extraction module to gradually and cumulatively fuse the information of the entire sequence, capturing not only short-term time dependencies, such as the impact of temperature changes over several consecutive days on pests and diseases, but also long-term evolution patterns, such as the crop's response trajectory to the environment throughout the growing season. At the end of the sequence, the time-series feature vector generated by the module condenses the dynamic changes and evolution patterns of the sequence throughout the entire observation period.
[0046] S2.5. Concatenate the spatial features and temporal features into vectors to form the primary fusion feature vector for each field.
[0047] Furthermore, the spatial feature vector generated by the spatial feature extraction module and the temporal feature vector generated by the temporal feature extraction module are concatenated. The concatenation operation preserves the complete feature information of both the spatial and temporal feature vectors, without prematurely compressing information or performing nonlinear transformations during the fusion stage. This ensures that all discriminative features from the spatial and temporal dimensions can flow equally into the subsequent processing stages, forming a primary fusion feature vector that simultaneously encodes a composite representation of the field's spatial structural characteristics and temporal dynamic characteristics.
[0048] Specifically, for example, it includes the spatial distribution pattern of suspected lesions in the field, as well as the temporal trend of regional temperature and humidity conditions that are conducive to disease development. The early fusion strategy provides a comprehensive and multi-dimensional starting point for subsequent self-supervised pre-training or causal decoupling, enabling a more comprehensive understanding of the field status from a combined spatial and temporal perspective, thereby supporting more accurate predictions.
[0049] S3. Using unlabeled field-time series multimodal feature data cubes, the general feature extraction network is optimized through a self-supervised pre-training task to output a high-quality primary fusion feature vector containing semantic information about crop physiological state.
[0050] S3.1 The general feature extraction network includes a spatial feature extraction module, a temporal feature extraction module, and a feature fusion module.
[0051] S3.2 The spatial feature extraction module extracts spatial features from the randomly masked unlabeled field-temporal multimodal feature data cube, and the temporal feature extraction module extracts temporal features from the randomly masked unlabeled field-temporal multimodal feature data cube. The feature fusion module fuses the spatial features and temporal features, and the reconstruction head network predicts the data content of the masked part based on the fused features.
[0052] Furthermore, a random masking operation is applied to the input unlabeled field-time series multimodal feature data cube. The masking strategy can be to randomly mask a certain proportion of spatial blocks or time segments in the data cube. For example, randomly masking all satellite image data of a field for several consecutive days, or randomly masking part of the meteorological variable sequence within a specific time period. The masked data cube is then fed into the spatial feature extraction module and the temporal feature extraction module, respectively. The spatial feature extraction module tries its best to extract the remaining effective spatial features from the partially masked image data, while the temporal feature extraction module extracts the usable temporal features from the partially masked sequence data. The feature fusion module fuses the two features extracted from the incomplete data. An independent reconstruction head network, usually composed of several fully connected layers, receives this fused feature vector and attempts to accurately predict the original data values of the masked part, such as the pixel values of the masked image blocks or the values of the masked meteorological sequence.
[0053] Specifically, in order to accurately reconstruct the masked portion, the spatial feature extraction module must learn to utilize the surrounding spatial information that is not masked, the temporal feature extraction module must learn to utilize the preceding and following temporal information that is not masked, and the feature fusion module and the reconstruction head network must learn to effectively combine spatiotemporal cues. By minimizing the difference between the reconstructed value and the true value, the network is guided to learn robust spatiotemporal representations and cross-modal associations in the data. This representation can capture the overall state of the crop growth environment, thereby improving the quality of the output features.
[0054] S3.3, the spatial feature extraction module and the temporal feature extraction module extract spatial features and temporal features from the unlabeled field-temporal multimodal feature data cube, respectively, and align them to the shared feature space through the feature fusion module.
[0055] Furthermore, a cross-modal contrastive alignment task is constructed and executed. In this task, the same unlabeled field-temporal multimodal feature data cube is simultaneously input into both the spatial feature extraction module and the temporal feature extraction module. The spatial feature extraction module extracts spatial feature vectors from the image-like data components, while the temporal feature extraction module extracts temporal feature vectors from the sequence-like data components. The feature fusion module is used as a shared mapper in this task, mapping the spatial and temporal feature vectors to a common, low-dimensional embedding space. The goal of contrastive learning is that, within the embedding space, the distance between spatial and temporal feature embeddings from the same field at the same time should be as small as possible, forming positive sample pairs; while the distance between feature embeddings from different fields or different times should be as large as possible, forming negative sample pairs.
[0056] Specifically, by using the prior knowledge that multimodal observations of the same entity at the same time should be semantically aligned as a supervisory signal, and employing a loss function such as noise contrast estimation to optimize network parameters, the network is forced to discover and encode essential semantic information that coexists in different modal data and is related to the state of a specific field at a specific time. For example, the network needs to learn that the feature embedding of a specific canopy spectral pattern should be close to the feature embedding of the specific temperature and humidity condition sequence that caused the pattern to appear in the embedding space. Through this cross-modal alignment constraint, the feature representation learned by the network can better bridge different data sources, capture more consistent and discriminative crop physiological state semantics, and thus generate high-quality primary fusion feature vectors.
[0057] S3.4 The temporal feature extraction module extracts temporal features from the unlabeled field-temporal multimodal feature data cube injected with artificially synthesized anomaly patterns. The classification head network then judges the presence of anomalies and their locations in the sequence based on the temporal features.
[0058] Furthermore, the sequence data components in a portion of the unlabeled field-time-series multimodal feature data cube are modified by injecting artificially synthesized anomalous patterns that conform to agricultural pathology. For example, a rapid decline inflection point caused by disease infection can be simulated on a normal vegetation index growth curve, or a brief pulse simulating high-temperature heat damage can be inserted into a stable temperature sequence. The modified sequences are mixed with the original normal sequences, and the presence and location of anomalies are marked. The time-series feature extraction module processes these sequences that may contain anomalies and extracts time-series features. A classification head network, such as a network containing fully connected layers and a softmax output layer, uses the time-series features as input to perform a binary classification task to determine whether the entire sequence is abnormal, or to perform a sequence labeling task to determine whether each time step belongs to an anomaly point.
[0059] Specifically, the temporal feature extraction module is made highly sensitive to deviations from normal patterns during crop growth. For accurate classification, the module must deeply understand the temporal patterns and rhythms of normal crop growth, enabling it to keenly identify abnormal fluctuations that do not conform to these patterns. The ability to identify abnormal patterns is crucial for the early detection of pest and disease stress. Through task optimization, the temporal feature extraction module can encode more stress-related discriminative information in its extracted temporal features, thus contributing to the final high-quality primary fusion feature vector.
[0060] S3.5 By jointly optimizing the total loss function of the multimodal mask reconstruction task, cross-modal contrast alignment task, and temporal anomaly detection task, the parameters of the spatial feature extraction module, temporal feature extraction module, feature fusion module, and related task head network are updated, and a high-quality primary fusion feature vector containing semantic information of crop physiological state is output.
[0061] Furthermore, a reconstruction loss, such as mean squared error loss, is defined for the multimodal mask reconstruction task to measure the difference between the predicted values of the reconstruction head network and the true mask values. A contrastive loss, such as InfoNCE loss, is defined for the cross-modal contrast alignment task to measure the similarity relationship between positive and negative sample pairs in the embedding space. A classification loss, such as cross-entropy loss, is defined for the temporal anomaly detection task to measure the difference between the anomaly labels predicted by the classification head network and the true labels. These three loss functions are weighted and summed to obtain the total loss function. During training, in each iteration, based on the current batch of unlabeled field-temporal multimodal feature data cubes, the losses of the three tasks are obtained simultaneously. The gradient of the total loss relative to the spatial feature extraction module, temporal feature extraction module, feature fusion module, and the parameters of each task head network is obtained through backpropagation. Finally, an optimization algorithm is used to update all parameters.
[0062] Specifically, the general feature extraction network provides multi-faceted and complementary learning objectives. The network parameters need to simultaneously adapt to the requirements of data reconstruction, modality alignment, and anomaly detection. This is equivalent to constraining and guiding feature learning from three dimensions: data integrity, semantic consistency, and pattern health. Through this intensive and comprehensive self-supervised pre-training, the parameters of the general feature extraction network are optimized to extract extremely rich and robust feature representations from unlabeled data. When processing input data, the final output vector of the feature fusion module, namely the high-quality primary fusion feature vector, not only integrates spatiotemporal information but also encodes deep semantic information that has been validated through multiple tasks and is highly correlated with crop physiological states and potential stresses. This provides a superior feature foundation for subsequent causal decoupling and accurate prediction.
[0063] S4. Input the high-quality primary fusion feature vector containing semantic information of crop physiological state into the causal decoupling module, and decouple and map it to the disease and pest feature subspace, the environmental mixed feature subspace and the shared background feature subspace to obtain disease and pest related features.
[0064] S4.1 The causal decoupling module includes a pest and disease feature mapping unit, an environmental confounding feature mapping unit, and a shared background feature mapping unit.
[0065] S4.2 Input the high-quality primary fusion feature vector containing semantic information of crop physiological state into the pest and disease feature mapping unit, the environmental confounding feature mapping unit and the shared background feature mapping unit respectively, and generate pest and disease feature subspace vector, environmental confounding feature subspace vector and shared background feature subspace vector respectively, forming the decoupled feature representation.
[0066] Furthermore, the same high-quality primary fusion feature vector containing semantic information about crop physiological state is copied three times and input into the pest and disease feature mapping unit, the environmental confounding feature mapping unit, and the shared background feature mapping unit, respectively. Each mapping unit processes this same input independently. The pest and disease feature mapping unit performs a series of weighted sums and nonlinear activations through its internal fully connected layers, outputting a fixed-dimensional vector called the pest and disease feature subspace vector. This vector aims to encode only the feature information related to the occurrence and development of pests and diseases. The environmental confounding feature mapping unit outputs the environmental confounding feature subspace vector through the same process, aiming to encode changes caused by abiotic stress. The shared background feature mapping unit outputs the shared background feature subspace vector, encoding common information unrelated to stress.
[0067] Specifically, these three subspace vectors together constitute the decoupled feature representation of the original high-quality primary fusion feature vector. The processing logic of source splitting uses the original feature vector as an information source, which is analyzed simultaneously by three parallel, functionally specialized processing paths. Each path is trained to focus on capturing and amplifying a specific type of information while suppressing other types of information. For example, when the input features simultaneously contain information about chlorophyll reduction caused by disease and leaf wilting caused by water shortage, the disease and pest feature mapping unit will strengthen the pattern related to the former, while the environmental contamination feature mapping unit will strengthen the pattern related to the latter. The mixed features are actively and guidedly decomposed into different semantic subspaces, realizing the decoupling of features at the representation level.
[0068] S4.3 Extract the pest and disease feature subspace vector as pest and disease related features from the pest and disease feature subspace vector, the environmental mixing feature subspace vector, and the shared background feature subspace vector.
[0069] Furthermore, after further decoupling the three feature subspace vectors—namely, the pest and disease feature subspace vector, the environmental confounding feature subspace vector, and the shared background feature subspace vector—an explicit selection operation is performed. The pest and disease feature subspace vector is directly extracted and designated as the pest and disease related feature. The environmental confounding feature subspace vector and the shared background feature subspace vector are retained or discarded after this step, but do not participate in the final regression prediction.
[0070] Specifically, by limiting the decision-making basis of the prediction model to the identified causal features of pests and diseases, the path of prediction through irrelevant or confusing factors such as environmental confounding features is logically cut off. For example, during the training process, even if the environmental confounding feature subspace vector contains features that are highly correlated with drought and may co-occur with a certain disease, the final prediction only uses the pest and disease feature subspace vector. This false statistical correlation cannot be utilized, and the mapping relationship between the real pest and disease signals and density must be learned. This makes the prediction results insensitive to changes in confounding factors and more accurately reflects the degree of occurrence of pests and diseases themselves, thus achieving the fundamental purpose of feature decoupling.
[0071] S5. Based on the decoupled pest and disease related features, generate predicted values for maize pest and disease density.
[0072] S5.1 Construct a multilayer perceptron regressor for density prediction based on the feature subspace vector of pests and diseases, and input pest and disease related features into the multilayer perceptron regressor.
[0073] Furthermore, the dimension of the input layer of the multilayer perceptron regressor is strictly matched with the dimension of the disease and pest feature subspace vector. The multilayer perceptron regressor is usually composed of an input layer, at least one hidden layer and an output layer connected sequentially. Each layer contains several neurons, and the layers are fully connected through a learnable weight matrix and bias vector. The construction process involves determining the number of hidden layers, the number of neurons in each layer and selecting an appropriate nonlinear activation function, such as a rectified linear unit or a hyperbolic tangent function. Subsequently, the disease and pest related features and the extracted disease and pest feature subspace vector are fed as input data into the input layer of the multilayer perceptron regressor.
[0074] Specifically, a prediction function is created to map purified pest and disease features to continuous density values. After the aforementioned complex feature extraction, optimization, and decoupling process, the pest and disease related features are already a representation that has removed environmental clutter and focuses on the causal relationship of the target. A relatively simple multilayer perceptron regressor is built on the basis of pure features, which can be understood as building an accurate mapper in a clean environment. This avoids coupling the complex feature decoupling task and the nonlinear regression task in the same complex network, making the model structure clearer and the task division clear. This reduces the learning difficulty of the regressor itself and ensures that its prediction depends entirely on causal features, thereby improving the stability and interpretability of the prediction. The high-level feature semantics and the specific regression task are connected through a dedicated, structured network, completing the key transformation from abstract features to specific numerical prediction.
[0075] S5.2, The multilayer perceptron regressor performs nonlinear transformation and feature mapping on pest and disease related features to predict the density of corn pests and diseases.
[0076] Furthermore, the multilayer perceptron regressor performs forward propagation on the input pest and disease-related features to generate predicted values for maize pest and disease density. The input feature vector is first multiplied by the weight matrix from the input layer to the first hidden layer, and a bias vector is added. It is then transformed by a nonlinear activation function, linearly combining the input features and mapping them to a new nonlinear feature space. The new feature vector is then used as input, and the process of multiplying by the weight matrix of the next hidden layer, adding a bias, and passing through the activation function is repeated layer by layer until it reaches the output layer. The output layer typically uses a linear activation function or no activation function, and its output is a single continuous scalar value, which is the predicted value for maize pest and disease density. The entire forward propagation process realizes a series of hierarchical nonlinear transformations and feature mappings of the input pest and disease-related features.
[0077] Specifically, the multilayer perceptron, through its deep structure, can learn complex, non-linear functional relationships between pure causal features and density values. For example, the feature subspace vector of pests and diseases may encode abstract information such as the intensity and range of disease infection and the combination patterns of different symptoms. The hidden layers of the multilayer perceptron regressor can learn to combine and weight these abstract patterns, mapping them to a density estimate that reflects the overall degree of damage. It utilizes the powerful function approximation ability of deep neural networks, but limits this ability to decoupled, semantically clear input features. This allows the network to focus on learning how causal features affect the core mapping of density, without having to distinguish which features are truly related to pests and diseases. This ensures that the effective capacity of the model is concentrated on improving regression accuracy, thereby generating more accurate and reliable corn pest and disease density predictions based on pure pest and disease features.
[0078] S6. Process the predicted density of corn pests and diseases to generate a visual early warning product.
[0079] S6.1 Perform deviation correction on the predicted density of corn diseases and pests based on historical error statistics to generate the corrected predicted density of corn diseases and pests.
[0080] Furthermore, a historical prediction-actual value pairing dataset is constructed by collecting the predicted maize pest and disease density outputs from the multilayer perceptron regressor over a historical period and the actual pest and disease density values of the corresponding fields obtained through field surveys. The error distribution pattern between the predicted and actual values in this dataset is analyzed, such as the average deviation of the predicted value relative to the actual value in different density intervals, or by fitting a correction function that describes the relationship between the prediction deviation and the prediction value itself or certain auxiliary variables. Then, for a new predicted maize pest and disease density, a deviation correction amount is obtained based on the density interval or the fitted correction function. This correction amount is then added to or subtracted from the original predicted maize pest and disease density to generate a corrected predicted maize pest and disease density.
[0081] Specifically, systematic biases arise due to the imperfect representativeness of the training data distribution or the approximation errors of the model itself. These biases may manifest as overall overestimation or underestimation, or as different shifts at different prediction levels. Correcting these biases based on historical error statistics is equivalent to adding a post-processing calibration step to the model's predictions. S6.2. Perform an uncertainty assessment based on Monte Carlo Dropout on the corrected maize disease and pest density prediction values to generate confidence intervals for the maize disease and pest density prediction values.
[0082] Furthermore, Monte Carlo Dropout technology is used to quantify the uncertainty of the corrected maize pest and disease density predictions. In the trained multilayer perceptron regressor, this is achieved by keeping the Dropout layer active during the inference phase and repeating the forward propagation process multiple times. For the same input, namely pest and disease related features, the network weights remain unchanged, but each time the forward propagation occurs, the Dropout layer randomly discards a portion of the neuron outputs according to the probability set during training. After repeated inference, a set of slightly different corrected maize pest and disease density predictions is obtained. Based on this set of predictions, the statistical distribution is obtained, such as the range of the mean plus or minus two standard deviations, which constitutes the confidence interval of the maize pest and disease density predictions. This is the Dropout technique commonly used in the training phase to prevent overfitting.
[0083] Specifically, the randomness introduced by Dropout is equivalent to sampling the network ensemble. The dispersion of multiple inference results reflects the model's hesitation or uncertainty regarding the prediction outcome under that input feature. For example, when the input feature is ambiguous or near the model's decision boundary, the results of multiple Dropout inferences may differ greatly, resulting in a wider confidence interval, and vice versa. By generating confidence intervals, not only are point predictions provided, but also a quantitative indicator of the possible range of variation of the prediction value is provided, greatly enhancing the information content and practicality of the prediction results. This allows users to distinguish between high-confidence predictions and low-confidence predictions, thereby making more prudent decisions.
[0084] S6.3. Compare the corrected maize pest and disease density prediction with the preset economic threshold to generate a preliminary warning level. Combine the confidence interval of the maize pest and disease density prediction and the preliminary warning level to generate a visualized early warning product with geographic spatial distribution, confidence range and graded prevention and control recommendations.
[0085] Furthermore, the corrected predicted density of maize pests and diseases is compared with multiple economic thresholds pre-set based on crop variety, growth stage, and cost-effectiveness analysis of prevention and control. For example, density thresholds are set for several levels of occurrence, such as light, moderate, and severe occurrence. By comparison, a preliminary warning level is determined for each field, and the confidence interval of the preliminary warning level and the predicted density of maize pests and diseases is comprehensively considered.
[0086] Specifically, for example, although the predicted density of a field may reach the threshold for moderate occurrence, its confidence interval is wide, and the lower limit may fall within the range of mild occurrence. This indicates that the prediction has significant uncertainty. Finally, the geographical location of all fields, the corrected predicted values of maize pest and disease density, the confidence interval, and the comprehensive assessment of the warning level are spatially visualized and rendered using Geographic Information System (GIS) technology. This can generate a spatial distribution map of pest and disease density, using color gradients to represent density levels; confidence interval bands or error bars can be overlaid on the map; and different symbols or colors can be used to mark different zones according to the final warning level, and associated with the generation of tiered prevention and control recommendation texts, such as monitoring and observation, preparation for prevention and control, and immediate prevention and control.
[0087] It should be noted that transforming complex numerical prediction results into intuitive, comprehensive, and actionable decision support information, and dynamically combining point prediction, uncertainty measurement, and economic decision thresholds, does not simply involve issuing warnings based on a single predicted value. Instead, it integrates the reliability information of the prediction, thereby generating more refined and robust warning strategies. For example, for fields with high predicted values but wide confidence intervals, the warning suggestion may be high risk; for fields with moderate predicted values but narrow confidence intervals, the warning suggestion may be more specific. The generated visual warning products not only show what is happening but also explain how certain they are and suggest what to do, thus providing a complete information chain support for precision plant protection from perception to decision-making.
[0088] This embodiment also provides a big data-based maize pest and disease density prediction system, including: a processing module that accesses and preprocesses multi-source heterogeneous agricultural data to generate a standardized field-time series multimodal feature data cube; The feature encoding module inputs the standardized field-time series multimodal feature data cube into the general feature extraction network for feature encoding, and outputs the primary fusion feature vector for each field. The optimization module utilizes unlabeled field-time series multimodal feature data cubes to optimize the general feature extraction network through a self-supervised pre-training task, outputting a high-quality primary fusion feature vector containing semantic information about crop physiological states. The causal decoupling module inputs high-quality primary fusion feature vectors containing semantic information about crop physiological state into the causal decoupling module, decoupling and mapping them to the disease and pest feature subspace, the environmental mixing feature subspace, and the shared background feature subspace to obtain disease and pest related features. Based on the decoupled disease and pest related features, the predicted value of maize disease and pest density is generated. The visualization module processes the predicted density of corn pests and diseases to generate visualized early warning products.
[0089] This embodiment also provides a computer device applicable to the big data-based corn pest and disease density prediction method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the big data-based corn pest and disease density prediction method proposed in the above embodiment.
[0090] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0091] This embodiment also provides a storage medium storing a computer program. When executed by a processor, the program implements the method for predicting the density of corn pests and diseases based on big data, as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0092] In summary, this invention utilizes unlabeled big data and self-supervised pre-training tasks to optimize the feature extraction network. Without requiring manual annotation, the network learns the intrinsic patterns between crop growth and multi-source data, reducing dependence on scarce labeled data and enhancing the model's feature generalization ability and sensitivity to subtle stresses. Through a causal decoupling mechanism, fused features are separated into independent subspaces of pests and diseases, environmental contamination, and shared background. This actively identifies and extracts specific features triggered solely by pest and disease biological stress, removing interference from water and fertilizer stress and varietal differences. Consequently, the interpretability, reliability, and stability of prediction results under different environments are significantly improved.
[0093] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for predicting the density of corn pests and diseases based on big data, characterized in that: This includes accessing and preprocessing multi-source heterogeneous agricultural data to generate standardized field-time series multimodal feature data cubes; The standardized field-time series multimodal feature data cube is input into a general feature extraction network for feature encoding, and the primary fusion feature vector of each field is output. By utilizing unlabeled field-time multimodal feature data cubes, a general feature extraction network is optimized through a self-supervised pre-training task, outputting a high-quality primary fusion feature vector containing semantic information about crop physiological states. High-quality primary fusion feature vectors containing semantic information about crop physiological state are input into the causal decoupling module. The decoupling is mapped to the disease and pest feature subspace, the environmental mixing feature subspace, and the shared background feature subspace to obtain disease and pest related features. Based on the decoupled disease and pest related features, the predicted value of maize disease and pest density is generated. The predicted density of corn pests and diseases is processed to generate a visual early warning product.
2. The method for predicting the density of corn diseases and pests based on big data as described in claim 1, characterized in that: The process of accessing and preprocessing multi-source heterogeneous agricultural data to generate standardized field-time series multimodal feature data cubes includes the following steps: Receive multispectral and hyperspectral images, meteorological grid data, field spore capture data, and plant protection history records, and perform time alignment and spatial coordinate unification; Based on the field boundaries, the processed data is spatially cropped and attribute extracted to obtain the image sequence, meteorological sequence, spore data sequence and ledger record for each field. Cloud detection and filling are performed on the field image sequence to generate a cloudless image sequence, and vegetation index sequence is extracted using preset band combinations and mathematical transformation rules; Missing values were filled, outliers were removed and standardized from cloudless image sequences, meteorological sequences, spore data sequences, ledger records and vegetation index sequences. The sequences were stacked and integrated according to a uniform time series length and feature dimension to generate a standardized field-time series multimodal feature data cube.
3. The method for predicting the density of maize diseases and pests based on big data as described in claim 2, characterized in that: The normalized field-time series multimodal feature data cube is input into a general feature extraction network for feature encoding, and the primary fused feature vector for each field is output, including the following steps: In the general feature extraction network, a spatial feature extraction module for extracting spatial features and a temporal feature extraction module for extracting temporal features are configured; The standardized field-time series multimodal feature data cube is input into the general feature extraction network; The spatial feature extraction module performs multi-scale convolution and pooling operations on the image data components in the standardized field-time series multimodal feature data cube to generate spatial features; The temporal feature extraction module sequentially inputs the sequence data components in the standardized field-temporal multimodal feature data cube into the recurrent neural network unit according to the time order. Through the gating mechanism and state transmission within the unit, the sequence information is gradually updated and fused to generate temporal features. Spatial features and temporal features are concatenated into vectors to form a primary fusion feature vector for each field.
4. The method for predicting the density of corn diseases and pests based on big data as described in claim 3, characterized in that: Using unlabeled field-time multimodal feature data cubes, a general feature extraction network is optimized through a self-supervised pre-training task to output a high-quality primary fusion feature vector containing semantic information about crop physiological states. The process includes the following steps: The general feature extraction network includes a spatial feature extraction module, a temporal feature extraction module, and a feature fusion module; The spatial feature extraction module extracts spatial features from the randomly masked unlabeled field-temporal multimodal feature data cube, the temporal feature extraction module extracts temporal features from the randomly masked unlabeled field-temporal multimodal feature data cube, and the feature fusion module fuses the spatial features and temporal features. The reconstruction head network then predicts the data content of the masked part based on the fused features. The spatial feature extraction module and the temporal feature extraction module extract spatial features and temporal features from the unlabeled field-temporal multimodal feature data cube, respectively, and align them to a shared feature space through the feature fusion module; The temporal feature extraction module extracts temporal features from the unlabeled field-temporal multimodal feature data cube injected with artificially synthesized anomaly patterns. The classification head network then judges the presence of anomalies and the location of anomalies in the sequence based on the temporal features. By jointly optimizing the total loss function of the multimodal mask reconstruction task, cross-modal contrast alignment task, and temporal anomaly detection task, the parameters of the spatial feature extraction module, temporal feature extraction module, feature fusion module, and related task head networks are updated, and a high-quality primary fusion feature vector containing semantic information about crop physiological state is output.
5. The method for predicting the density of corn diseases and pests based on big data as described in claim 4, characterized in that: High-quality primary fusion feature vectors containing semantic information about crop physiological states are input into the causal decoupling module. The decoupling mapping is then performed to the pest and disease feature subspace, the environmental confounding feature subspace, and the shared background feature subspace to obtain pest and disease-related features, including the following steps: The causal decoupling module includes a pest and disease feature mapping unit, an environmental confounding feature mapping unit, and a shared background feature mapping unit. High-quality primary fusion feature vectors containing semantic information about crop physiological states are input into the pest and disease feature mapping unit, the environmental confounding feature mapping unit, and the shared background feature mapping unit, respectively, to generate pest and disease feature subspace vectors, environmental confounding feature subspace vectors, and shared background feature subspace vectors, thus forming decoupled feature representations. Extract the pest and disease feature subspace vector as pest and disease related features from the pest and disease feature subspace vector, the environmental mixing feature subspace vector, and the shared background feature subspace vector.
6. The method for predicting the density of corn diseases and pests based on big data as described in claim 5, characterized in that: Based on the decoupled pest and disease-related features, a predicted value for maize pest and disease density is generated, including the following steps: A multilayer perceptron regressor for density prediction is constructed based on the feature subspace vector of pests and diseases, and pest and disease related features are input into the multilayer perceptron regressor. The multilayer perceptron regressor performs nonlinear transformation and feature mapping on pest and disease related features to predict the density of corn pests and diseases.
7. The method for predicting the density of maize diseases and pests based on big data as described in claim 6, characterized in that: The process of processing predicted corn pest and disease density values to generate a visual early warning product includes the following steps: The predicted density of maize diseases and pests is corrected for deviation based on historical error statistics to generate corrected predicted density values of maize diseases and pests. Uncertainty assessment based on Monte Carlo Dropout was performed on the corrected maize disease and pest density prediction values to generate confidence intervals for the maize disease and pest density prediction values. The corrected predicted density of corn pests and diseases is compared with a preset economic threshold to generate a preliminary warning level. Combining the confidence interval of the predicted density of corn pests and diseases with the preliminary warning level, a visualized early warning product is generated, showing the geographic spatial distribution, confidence range, and tiered prevention and control recommendations.
8. A big data-based corn pest and disease density prediction system, based on the big data-based corn pest and disease density prediction method according to any one of claims 1 to 7, characterized in that: This includes a processing module that accesses and preprocesses multi-source heterogeneous agricultural data to generate a standardized field-time series multimodal feature data cube. The feature encoding module inputs the standardized field-time series multimodal feature data cube into the general feature extraction network for feature encoding, and outputs the primary fusion feature vector for each field. The optimization module utilizes unlabeled field-time series multimodal feature data cubes to optimize the general feature extraction network through a self-supervised pre-training task, outputting a high-quality primary fusion feature vector containing semantic information about crop physiological states. The causal decoupling module inputs high-quality primary fusion feature vectors containing semantic information about crop physiological state into the causal decoupling module, decoupling and mapping them to the disease and pest feature subspace, the environmental mixing feature subspace, and the shared background feature subspace to obtain disease and pest related features. Based on the decoupled disease and pest related features, the predicted value of maize disease and pest density is generated. The visualization module processes the predicted density of corn pests and diseases to generate visualized early warning products.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the big data-based corn pest and disease density prediction method as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the big data-based corn pest and disease density prediction method as described in any one of claims 1 to 7.