A method and system for monitoring rice growth status based on multispectral remote sensing

Abstract

This invention relates to the field of remote sensing monitoring technology, specifically to a method and system for monitoring the growth status of rice based on multispectral remote sensing. The method includes: acquiring multispectral remote sensing sequence data of a target paddy field and performing time-series synchronization processing, unifying it to a standard time scale to generate a time-aligned set; extracting rice canopy growth characteristics through spectral feature evolution analysis and matching them with a rice growth knowledge graph to generate a set of semantic vectors representing growth status; spatially matching and fusing the semantic vectors with a soil moisture dataset to construct a fused feature field encompassing crop growth and the soil environment; decoding and outputting quantitative values ​​of growth status to drive a growth simulation model to obtain predicted values; and feeding back the predicted values ​​to correct subsequent time-series synchronization processing. This method achieves time-series data standardization, semanticization of growth information, and multi-source feature spatial fusion, forming a closed-loop iterative monitoring system, improving the accuracy of growth status characterization and the reliability of prediction.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring technology, and in particular to a method and system for monitoring the growth status of rice based on multispectral remote sensing. Background Technology

[0002] Traditional rice growth monitoring often uses discrete, multi-temporal, multispectral remote sensing data. Data from different acquisition times are not time-series synchronized, resulting in inconsistent time bases, temporal misalignments, and difficulty in forming continuous and comparable time-series observation sequences. Spectral features are only analyzed through numerical statistics and index calculations, without semantic interpretation using specialized growth knowledge maps. Descriptions of growth states lack mechanistic explanations, and the correlation between features and actual growth stages is weak.

[0003] Rice growth characteristics and soil moisture data were processed independently, without precise matching and deep fusion within a unified spatial grid. Crop status information and soil environmental information were fragmented and could not form an integrated representation. After decoding the growth status, it was only used for one-way output, and the prediction results had no feedback correction mechanism, so they could not be used to influence the previous time-series data processing. The monitoring and simulation process was a one-way execution flow, with insufficient overall coherence and adaptability.

[0004] This invention performs time-series synchronization processing on multispectral remote sensing sequences and unifies them to a standard time scale. It generates semantic vectors by combining growth knowledge graphs, integrates growth semantics with soil moisture space to construct a fusion feature field, and corrects the time-series synchronization process through predicted value feedback to achieve closed-loop iterative dynamic monitoring. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a method and system for monitoring rice growth status based on multispectral remote sensing.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for monitoring the growth status of rice based on multispectral remote sensing, comprising: Acquire multispectral remote sensing sequence data of the target paddy field within a specific monitoring period, perform time-series synchronization processing on the multispectral remote sensing sequence data, unify the original data from different acquisition times to a standard time scale, and generate a time-aligned remote sensing sequence set. Spectral feature evolution analysis was performed on the time-aligned remote sensing sequence set to extract the rice canopy growth evolution features corresponding to each standard time scale; Retrieve semantic descriptions of growth states that match the growth evolution characteristics from the rice growth knowledge graph database, and generate a set of semantic vectors of growth states; The set of semantic vectors representing the growth state is spatially matched with the externally input soil moisture dataset, and feature fusion is performed within the spatial grid to generate a fused feature field containing the rice growth state and the soil environment. The fused feature field is subjected to growth state decoding processing, and the quantitative value of rice growth state corresponding to the standard time scale is output. The quantitative value of rice growth state is input into the growth process simulation model, and the growth process simulation model is driven to calculate the predicted value of growth state at the next monitoring time. The predicted growth status is returned to the time-series synchronization processing step to correct the processing flow of the subsequently acquired multispectral remote sensing sequence data.

[0007] As a further aspect of the present invention, the multispectral remote sensing sequence data is subjected to time-series synchronization processing to unify the original data from different acquisition times to a standard time scale, generating a time-aligned remote sensing sequence set, including the following steps: The multispectral remote sensing sequence data includes spatial location coding, time-series labels, and spectral reflectance values; Receive raw multispectral remote sensing sequence data, which carries timestamp information for multiple acquisition times; A standard time scale axis is constructed, which consists of multiple pre-set equally spaced monitoring points; For each data acquisition moment, the time difference between the acquisition timestamp of the raw data and the two nearest monitoring points on the standard time scale axis is calculated. Based on the time difference, the original data is resampled in the time dimension, and the spectral values ​​of the original data are interpolated to the corresponding two standard monitoring points. For each standard monitoring point, collect all spectral values ​​that have been resampled and allocated to the standard monitoring point, calculate the average spectral value of the standard monitoring point, and generate the standardized spectral data unit corresponding to the standard time scale. The standardized spectral data units of all standard monitoring points are arranged in chronological order and assembled into the time-aligned remote sensing sequence set. The resampling process specifically includes: using a cubic spline interpolation algorithm, with the time difference as the interpolation weight, to smoothly interpolate the original spectral curve between two standard monitoring points to generate a transitional spectral numerical curve; and extracting the spectral values ​​located at the standard monitoring point from the transitional spectral numerical curve.

[0008] As a further aspect of the present invention, spectral feature evolution analysis is performed on the time-aligned remote sensing sequence set to extract the rice canopy growth evolution characteristics corresponding to each standard time scale, which is achieved through the following steps: Spectral data corresponding to consecutive standard time scales in the time-aligned remote sensing sequence set are selected to form a spectral evolution data segment; Principal component transformation is performed on the spectral evolution data segment to eliminate the correlation between spectral bands and obtain multiple independent evolutionary principal components; Calculate the rate of change of each principal component of evolution along the time dimension of the spectral evolution data segment, and generate a set of rate of change vectors; Identify the principal components in the set of rate of change vectors whose change magnitude exceeds a preset threshold, and mark them as sensitive principal components; Temporal pattern mining is performed on the sensitive evolutionary principal components to extract the trend pattern features of their spectral values ​​changing over time. The trend pattern features include linear growth and periodic fluctuations. The trend pattern features are encoded into a fixed-length feature vector, which serves as the rice canopy growth and evolution features corresponding to the standard time scale.

[0009] As a further aspect of the present invention, the step of retrieving semantic descriptions of growth states that match the growth evolution characteristics from a rice growth knowledge graph database and generating a set of semantic vectors of growth states is accomplished through the following steps: Receive the growth and evolution characteristics of the rice canopy; The rice canopy growth and evolution characteristics are input into a preset semantic encoder and converted into an initial semantic query vector. In the rice growth knowledge graph database, the cosine similarity between the initial semantic query vector and the feature vectors of all entity nodes in the knowledge graph is calculated; The entity nodes are sorted from high to low based on the cosine similarity, and the top-ranked candidate entity node set is selected. Extract the attribute triples associated with each candidate entity node from the rice growth knowledge graph database. The attribute triples include entity, relation, and attribute value. Semantic parsing is performed on the attribute triples to convert them into natural language descriptions, generating preliminary semantic descriptions corresponding to the candidate entity nodes; The initial semantic description is input into the language model for semantic refinement and structuring, generating semantic description entries of the growth state with a uniform format. The growth state semantic description entries corresponding to all candidate entity nodes are aggregated to form the growth state semantic vector set.

[0010] As a further aspect of the present invention, the set of semantic vectors representing the growth state is spatially matched with an externally input soil moisture dataset, and feature fusion is performed within a spatial grid to generate a fused feature field containing the rice growth state and the soil environment. Specifically, this includes the following operations: Receive an externally input soil moisture dataset, which contains spatial distribution values ​​of soil moisture content corresponding to the spatial location of the target paddy field; The geographic spatial range of the target paddy field is divided into regular spatial grids, with each grid serving as an independent processing unit; Within each spatial grid, the portion of the growth state semantic vector set belonging to the spatial grid is obtained to acquire the grid semantic features; Within the same spatial grid, the soil moisture content values ​​belonging to the spatial grid in the soil moisture dataset are obtained to obtain the grid moisture characteristics; Establish a feature association graph between grid semantic features and grid soil moisture features, wherein the feature association graph uses elements of grid semantic features as nodes and grid soil moisture features as edge weights; The graph neural network performs message propagation and aggregation operations on the feature association graph, enabling the grid semantic features and grid moisture features to be deeply integrated during the iterative update of the graph. After a preset number of graph update iterations, the fused feature representations of all nodes in the feature association graph are extracted and used as the fused feature vectors of the spatial grid. The fused feature vectors of all spatial grids in the target paddy field are arranged according to their spatial coordinates to generate the spatially continuous fused feature field.

[0011] As a further aspect of the present invention, the fused feature field is subjected to growth state decoding processing to output the quantitative value of rice growth state corresponding to the standard time scale, which is achieved through the following process: The quantitative values ​​of rice growth status include chlorophyll density and water stress coefficient; The fused feature field is input into a pre-constructed state decoding network, which includes a multilayer perceptron structure. The first layer of the state decoding network increases the feature dimension of the input fused feature field, mapping the low-dimensional fused features to a high-dimensional feature space to obtain a high-dimensional feature field. The intermediate layer of the state decoding network performs nonlinear transformation and feature extraction on the high-dimensional feature field, captures deep patterns in the fused feature field that are strongly correlated with the rice growth state, and generates deep state features. The output layer of the state decoding network performs regression calculations on the deep state features, mapping the high-dimensional deep state features to the quantized values ​​of the corresponding specific growth parameters, and generating preliminary quantized values ​​of rice growth status. Spatial consistency correction is performed on the preliminary quantitative values ​​of rice growth status to eliminate numerical abrupt changes between adjacent grids caused by spatial grid division; The spatial consistency correction step includes: calculating the average of the preliminary quantization values ​​of each spatial grid and the preliminary quantization values ​​of all its adjacent spatial grids; calculating the difference between the preliminary quantization value and the average value; if the difference is greater than a preset tolerance threshold, replacing the preliminary quantization value of the spatial grid with the average value, otherwise retaining the original value; the corrected value is the final output quantization value of the rice growth status.

[0012] As a further aspect of the present invention, the quantitative value of the rice growth status is input into the growth process simulation model, and the growth process simulation model is driven to calculate the predicted value of the growth status at the next monitoring time, including the following detailed steps: Initialize the internal state of the growth process simulation model, which includes the phenological stage code, cumulative biomass, and nitrogen uptake of rice; The quantitative value of the rice growth status corresponding to the current standard time scale is used as the driving input of the model and loaded into the growth process simulation model; The growth process simulation model is based on internal physiological and ecological process equations to simulate the rice growth process from the current standard time scale to the next standard time scale. The rice growth process includes photosynthesis, respiration consumption, and dry matter distribution. During the simulation, the spatial distribution values ​​of soil moisture content in the soil moisture dataset are used as environmental stress factors to dynamically adjust the stress coefficient in the physiological and ecological process equations. After the simulation is completed, the growth process simulation model outputs the simulated growth state value at the next standard time scale, which is used as the predicted growth state value. The growth process simulation model simultaneously records intermediate state variables during the simulation process, forming a state simulation trajectory, which is used for process correction in subsequent steps.

[0013] As a further aspect of the present invention, the predicted growth state value is returned to the time-series synchronization processing step to correct the processing flow of the subsequently acquired multispectral remote sensing sequence data. This is specifically achieved through the following mechanism: After completing the calculation of the predicted growth state value for the next monitoring time, the simulation intensity parameters of key physiological processes are extracted from the state simulation trajectory. The simulated intensity parameters are converted into expected requirements for the quality and frequency of remote sensing data acquisition, and data acquisition correction instructions are generated. The data acquisition correction command is sent to the control system of the UAV platform responsible for acquiring the multispectral remote sensing sequence data; The control system adjusts the flight path planning, spectral sensor operating parameters, or data acquisition time interval of the UAV according to the data acquisition correction command. When the UAV platform collects new multispectral remote sensing sequence data according to the adjusted parameters, the time synchronization processing step will dynamically adjust the time alignment strategy and weight based on the acquisition parameters that generated the new multispectral remote sensing sequence data, thereby achieving closed-loop correction of the processing flow.

[0014] As a further aspect of the present invention, the method further includes the step of constructing and updating the rice growth knowledge graph database, the step of which operates independently of the monitoring process, specifically including: From agricultural databases, entities, attributes, and relationships between entities related to rice growth are automatically extracted to form initial knowledge triples. The initial knowledge triples are subjected to conflict detection and resolution. Entities with the same semantics but different forms are merged to eliminate contradictory or redundant relationships and form clean knowledge triples. Using knowledge representation learning techniques, the entities and relations in the clean knowledge triples are mapped to a low-dimensional continuous vector space to obtain entity vectors and relation vectors; Based on the entity vectors and relation vectors, a graph embedding algorithm is used to optimize the overall structure of the rice growth knowledge graph database, so that entities with similar semantics are closer in the vector space. Periodically receive new growth state semantic descriptions that are not covered in the existing knowledge graph during the generation process of the growth state semantic vector set; The newly generated semantic description of the growth state is transformed into new knowledge triples through entity linking and relation extraction; After the new knowledge triples are quality-verified, they are incrementally merged into the rice growth knowledge graph database to achieve continuous evolution and expansion of the knowledge graph.

[0015] As a further aspect of the present invention, the present invention also includes a rice growth status monitoring system based on multispectral remote sensing. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the rice growth status monitoring method based on multispectral remote sensing described above.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Temporal synchronization processing was performed on multispectral remote sensing sequence data to unify raw data from different acquisition times to a standard time scale, forming a time-aligned remote sensing sequence set. Spectral feature evolution analysis was used to extract rice canopy growth evolution characteristics at each time point. Matching semantic descriptions were retrieved from a rice growth knowledge graph database to generate a set of semantic vectors representing growth states. Temporal unification eliminates time deviations between different observation times, and spectral evolution characteristics can fully reflect the continuous changes in rice growth. The introduction of the knowledge graph transforms numerical features into semantic information with growth mechanisms, ensuring a high degree of correspondence between state representations and actual growth stages.

[0017] The semantic vector set of growth status and the soil moisture dataset are spatially matched and fused within a unified spatial grid to generate a fused feature field containing rice growth status and soil environment. Decoding the fused feature field yields quantified values ​​of growth status, driving the growth process simulation model to output predicted values ​​for the next time step. These predicted values ​​are then returned to the time-series synchronization processing stage. Multi-dimensional features are coupled and superimposed within the spatial grid, resulting in a more complete state representation. The predicted values ​​form a closed-loop feedback path, allowing time-series synchronization and subsequent data processing to be dynamically adjusted based on the prediction results. The monitoring and simulation processes are mutually adapted, and the overall processing flow possesses continuous iteration and self-optimization characteristics. Attached Figure Description

[0018] Figure 1 This is a state diagram of a rice growth status monitoring method based on multispectral remote sensing as described in this invention. Figure 2 A flowchart for time-series synchronization processing of multispectral remote sensing sequences; Figure 3 A flowchart for generating semantic descriptions of growth states. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1 This invention provides a method for monitoring the growth status of rice based on multispectral remote sensing. The overall implementation scheme is as follows: Multispectral remote sensing sequence data of the target paddy field within a specific monitoring period is acquired. This multispectral remote sensing sequence data includes spatial location encoding, temporal labels, and spectral reflectance values. The multispectral remote sensing sequence data undergoes temporal synchronization processing to construct a standard time scale axis. Raw data from different acquisition times are unified to the standard time scale through interpolation and averaging operations, generating a time-aligned remote sensing sequence set. Spectral feature evolution analysis is performed on the time-aligned remote sensing sequence set. Spectral data from consecutive standard time scales are selected to form evolutionary segments. Principal component transformation is performed, and the rate of change is calculated. Trend pattern features of sensitive evolutionary principal components are extracted to generate rice canopy growth evolution features corresponding to each standard time scale. Entity nodes and their attribute triples matching the growth evolution features are retrieved from a rice growth knowledge graph database. Semantic descriptions of the growth state are generated through semantic encoding and refinement, forming a set of growth state semantic vectors. The set of semantic vectors representing the growth state is spatially gridded with the externally input soil moisture dataset. A correlation graph between semantic features and moisture features is established within the same grid, and then fused using a graph neural network to generate a spatially continuous fused feature field. The fused feature field undergoes growth state decoding processing; deep state features are extracted using a multilayer perceptron, and chlorophyll density and water stress coefficient are regressed and quantified. After spatial consistency correction, the quantified values ​​of rice growth state are output. These quantified values ​​are input into a growth process simulation model, driving the model to simulate the predicted growth state at the next monitoring time based on physiological and ecological equations and soil moisture stress factors. The predicted values ​​and key parameters in the simulated trajectory are converted into data acquisition correction instructions, which are fed back to the remote sensing data acquisition platform to adjust the acquisition strategy. This dynamically optimizes the time alignment process in subsequent time-series synchronization processing, achieving closed-loop monitoring.

[0022] In one embodiment of the present invention, see [reference] Figure 2The system receives raw multispectral remote sensing sequence data, which carries timestamp information from multiple acquisition times. It constructs a standard time scale axis composed of pre-defined, equally spaced monitoring points. For each acquisition time, it calculates the time difference between the acquisition timestamp and the two nearest monitoring points on the standard time scale axis. Based on this time difference, it resamples the original spectral curves in the time dimension using a cubic spline interpolation algorithm, generating a transitional spectral value curve with the time difference as the interpolation weight, and extracts the spectral values ​​located at the standard monitoring point locations. For each standard monitoring point, it collects all resampled spectral values ​​assigned to that point, calculates the average spectral value, and generates a standardized spectral data unit. Finally, it arranges all the standardized spectral data units from the standard monitoring points in chronological order and assembles them into a time-aligned remote sensing sequence set.

[0023] In practice, a 30-day monitoring cycle was set for the target rice paddy. Within this cycle, 12:00 noon each day was designated as the standard monitoring point, forming a standard time scale. The raw multispectral remote sensing sequence data includes timestamps for three acquisition times: 11:45 on day 5, 12:15 on day 10, and 12:05 on day 15. The spectral data for each acquisition time includes reflectance values ​​for the blue, green, red, and near-infrared bands. Taking the blue band as an example, the raw spectral reflectance at day 5 was 0.082, at day 10 it was 0.095, and at day 15 it was 0.090.

[0024] In the specific implementation, for the raw data collected on the 10th day, the timestamp is 12:15 on the 10th day. The time difference between this timestamp and the two adjacent monitoring points on the standard time scale (12:00 on the 10th day and 12:00 on the 11th day) is 15 minutes and 1425 minutes, respectively. Resampling is performed based on this time difference using a cubic spline interpolation algorithm. The time difference is used as the interpolation weight to smoothly interpolate the raw spectral curve collected on the 10th day between the two standard monitoring points. The interpolation formula is as follows: , in: This represents the interpolated spectral reflectance. Indicates the time of the standard monitoring point. A timestamp indicating the time of data collection. , , , The interpolation coefficients are determined by fitting spectral data from adjacent acquisition times. Spectral values ​​at the standard monitoring point at 12:00 on day 10 are extracted from the generated transition spectral numerical curve, with an interpolation result of 0.094 for the blue light band.

[0025] In practice, all spectral values ​​allocated to each standard monitoring point after resampling are collected, and the average spectral value is calculated. Taking the standard monitoring point at 12:00 on day 10 as an example, the collected spectral values ​​include the resampling results at the acquisition time on day 10 and the interpolation contributions at the acquisition times on days 9 and 11. The three sets of values ​​for the blue band are 0.094, 0.092, and 0.096, respectively, and the average spectral value is calculated to be 0.094. After generating standardized spectral data units, the standardized spectral data units of all standard monitoring points are arranged in chronological order and assembled into a time-aligned remote sensing sequence set. In some embodiments, the standardized blue band spectral data of the standard monitoring point at 12:00 on day 10 in the time-aligned remote sensing sequence set is 0.094, while the reflectance at the original acquisition time is 0.095. The difference between the two reflects the adjustment of data temporal consistency by the time synchronization processing.

[0026] In some embodiments, the deviation between the timestamp of the original multispectral remote sensing sequence data and the standard time scale axis may lead to adjustments in the interpolation results. For example, if the acquisition time on day 5 is 11:45, which is 15 minutes earlier than the standard monitoring point, the original blue light band reflectance of 0.082 is interpolated to 0.083 at the standard monitoring point at 12:00 on day 5. This demonstrates the calibration effect of time-dimensional resampling on the original data. It can be understood that time-series synchronization processing, by unifying the original data from different acquisition times to the standard time scale, eliminates spectral data fluctuations caused by inconsistent acquisition times, providing a time-aligned data foundation for subsequent spectral feature evolution analysis.

[0027] In one embodiment of the present invention, see [reference] Figure 3 Spectral data corresponding to consecutive standard time scales in a time-aligned remote sensing sequence set are selected to form spectral evolution data segments. Principal component transformation is performed on the segments to eliminate correlations between spectral bands, resulting in multiple independent evolutionary principal components. The rate of change of each evolutionary principal component in the time dimension is calculated to generate a set of rate of change vectors. Evolutionary principal components whose change amplitude exceeds a preset threshold are identified and marked as sensitive evolutionary principal components. Temporal pattern mining is performed on the sensitive evolutionary principal components to extract trend pattern features such as linear growth and periodic fluctuations, and these features are encoded into fixed-length feature vectors as rice canopy growth evolution features corresponding to the standard time scales. The growth and evolution features are input into a preset semantic encoder and converted into an initial semantic query vector. The cosine similarity between this vector and the feature vectors of all entity nodes is calculated in the rice growth knowledge graph database. Candidate entity nodes are selected by sorting them from high to low similarity. The attribute triples of each candidate node are extracted from the knowledge graph, converted into natural language descriptions through semantic parsing, and then input into a language model for semantic refinement and structuring to generate a uniformly formatted semantic description entry for growth status. The semantic description entries of all candidate nodes are aggregated to form a set of semantic vectors for growth status.

[0028] In practice, a spectral evolution data segment is formed by selecting spectral data corresponding to five consecutive standard time scales from a time-aligned remote sensing sequence set. This segment contains spectral reflectance values ​​for four bands: blue, green, red, and near-infrared. Principal component transformation is performed on the spectral evolution data segment to eliminate the correlation between the four spectral bands, resulting in four independent evolutionary principal components. The rate of change of each evolutionary principal component along the time dimension of the spectral evolution data segment is calculated, generating a set of rate of change vectors. The formula for calculating the rate of change is as follows: , in: Indicates the first The rate of change of each principal component of evolution This indicates the number of time points in a spectral evolution data segment. Indicates the first The principal evolutionary component in the th... The values ​​at each time point are analyzed. Evolutionary principal components whose change amplitude exceeds a preset threshold of 0.02 in the set of rate of change vectors are identified and marked as sensitive evolutionary principal components. Temporal pattern mining is performed on these sensitive evolutionary principal components to extract trend pattern features of their spectral values ​​over time. These trend pattern features include linear growth patterns and periodic fluctuation patterns. The linear growth pattern is encoded as value 1, and the periodic fluctuation pattern is encoded as value 2, combined into a fixed-length four-dimensional feature vector, which serves as the rice canopy growth evolution feature corresponding to the fifth standard time scale.

[0029] In the specific implementation, the growth and evolution characteristics of the rice canopy are input into a pre-defined semantic encoder, which is a three-layer fully connected neural network that maps the four-dimensional growth and evolution characteristics into an eight-dimensional initial semantic query vector. In the rice growth knowledge graph, the cosine similarity between the initial semantic query vector and the feature vectors of all entity nodes in the knowledge graph is calculated. The rice growth knowledge graph contains entity nodes such as "tillering stage," "jointing stage," and "booting stage," each with a 256-dimensional feature vector representation. The entity nodes are sorted from high to low based on cosine similarity, and the top five candidate entity nodes are selected. These candidate entity nodes include "tillering stage," "rapid growth stage," and "dark green leaf color."

[0030] The attribute triples associated with each candidate entity node are extracted from the rice growth knowledge graph. For the "tillering stage" entity, the attribute triples include (tillering stage, light requirement, long day length), (tillering stage, suitable temperature, 20-25℃), and (tillering stage, water requirement, moderate). Semantic parsing is performed on the attribute triples, converting them into natural language descriptions to generate preliminary semantic descriptions for the candidate entity nodes. The preliminary semantic description for "tillering stage" is "The tillering stage requires long day lengths, the suitable temperature is 20-25 degrees Celsius, and the water requirement is moderate." These preliminary semantic descriptions are then input into the BERT-base language model for semantic refinement and structuring, generating standardized growth state semantic description entries. The refined semantic description for "tillering stage" is "Growth stage: tillering stage; Environmental requirements: light - long day length, temperature -20-25℃, water - moderate." All growth state semantic description entries corresponding to the candidate entity nodes are aggregated to form a growth state semantic vector set containing five semantic description entries.

[0031] In some embodiments, the number of bands in the spectral evolution data segment can be adjusted to six, adding red-edge bands and short-wave infrared bands, correspondingly increasing the number of evolutionary principal components to six, setting the rate of change threshold to 0.015, and potentially reducing the number of sensitive evolutionary principal components to two. It is understood that through principal component transformation and temporal pattern mining, sensitive features reflecting rice growth dynamics can be extracted from multi-band spectral data, reducing redundant information interference. In some embodiments, the output dimension of the semantic encoder can be set to 16 dimensions, the number of candidate entity nodes can be set to 10, and the language model can use a GPT-2 structure for semantic refinement. Optionally, the semantic parsing of attribute triples can be achieved by combining rule templates with the model to improve parsing accuracy. Optionally, the aggregation of growth state semantic description entries can be achieved using a weighted average method, assigning weights based on the similarity scores of candidate entity nodes. It is understood that through knowledge graph retrieval and semantic refinement, numerical spectral features can be mapped to interpretable growth state descriptions, enhancing the interpretability of monitoring results.

[0032] In one embodiment of the present invention, a soil moisture dataset is received from an external input. The dataset contains spatial distribution values ​​of soil moisture content corresponding to the spatial location of a target paddy field. The geographic space of the target paddy field is divided into regular spatial grids, with each grid serving as an independent processing unit. Within a single grid, the portion of the growth state semantic vector set belonging to that grid is obtained to acquire the grid semantic features. Simultaneously, the soil moisture content values ​​within the same grid are acquired to acquire the grid moisture features. A feature association graph is established with grid semantic feature elements as nodes and grid moisture features as edge weights. Message propagation and aggregation operations of a graph neural network are performed on the association graph, and the two types of features are deeply fused through iterative updates. After a preset number of iterations, the fused feature representations of all nodes are extracted as the fused feature vectors of the spatial grid. The fused feature vectors of all grids are arranged according to spatial coordinates to generate a spatially continuous fused feature field. The fused feature field is input into a pre-constructed multilayer perceptron state decoding network. The first layer performs dimensionality upscaling mapping on the fused features to obtain a high-dimensional feature field. The intermediate layer extracts deep patterns strongly correlated with the growth state through nonlinear transformation. The output layer regresses the deep state features into preliminary quantified values ​​of chlorophyll density and water stress coefficient. Spatial consistency correction is performed on the preliminary values. The difference between each grid value and the average value of adjacent grids is calculated. If the difference is greater than a preset tolerance threshold, the average value is used to replace it. Otherwise, the original value is retained, and the corrected quantified values ​​of rice growth state are output.

[0033] In the specific implementation, an externally input soil moisture dataset is received. This dataset contains spatial distribution values ​​of soil moisture content corresponding to the spatial location of the target paddy field. The geographic spatial range of the target paddy field is a rectangular area, 200 meters long east-west and 100 meters wide north-south. The geographic spatial range of the target paddy field is divided into regular 10m × 10m spatial grids, forming a total of 200 spatial grids, each serving as an independent processing unit. Within each spatial grid, the portion of the growth state semantic vector set belonging to the spatial grid is extracted to obtain the grid semantic feature, which is an eight-dimensional vector. Within the same spatial grid, the soil moisture content values ​​belonging to the spatial grid in the soil moisture dataset are extracted to obtain the grid moisture feature, which is a scalar value. A feature association graph is established between the grid semantic feature and the grid moisture feature. The feature association graph uses the eight elements of the grid semantic feature as nodes and the grid moisture feature as edge weights, where the edge weights represent the correlation strength between the semantic feature elements and the soil moisture.

[0034] In the specific implementation, message propagation and aggregation operations of a graph neural network are performed on the feature association graph. The message propagation function uses the graph convolution operator, and the aggregation function uses summation pooling. The hidden layer dimension of the graph neural network is set to 16 dimensions, and the number of graph update iterations is set to 3. In each iteration, node features and edge weights are interactively updated, so that the grid semantic features and grid soil moisture features are deeply integrated during the graph iteration update process. After completing 3 graph update iterations, the fused feature representations of all nodes in the feature association graph are extracted and used as the fused feature vectors of the spatial grid. The fused feature vectors are 16-dimensional vectors. The fused feature vectors of all spatial grids in the target paddy field are arranged according to their spatial coordinates to generate a spatially continuous fused feature field. The spatial resolution of the fused feature field is consistent with the grid size, which is 10 meters × 10 meters (see Table 1).

[0035] Table 1: Numerical Table of Fusion Feature Vectors for Partial Spatial Grids (10,10) [0.21,0.18,0.24,0.19] [0.22,0.26,0.23,0.27] (10,20) [0.28,0.31,0.29,0.33] [0.32,0.34,0.35,0.36] (20,10) [0.41,0.42,0.43,0.44] [0.46,0.47,0.48,0.49] In the specific implementation, the fused feature field is input into a pre-constructed state decoding network. The state decoding network comprises a three-layer perceptron structure: an input layer with 16 dimensions, a hidden layer with 32 dimensions, and an output layer with 2 dimensions. The first layer of the state decoding network increases the dimensionality of the input fused feature field by linearly transforming the 16-dimensional low-dimensional fused features to a 32-dimensional high-dimensional feature space, resulting in a high-dimensional feature field. The intermediate layer of the state decoding network performs nonlinear transformation and feature extraction on the high-dimensional feature field, using the ReLU activation function to capture deep patterns strongly correlated with rice growth status in the fused feature field, generating deep state features. The output layer of the state decoding network performs regression calculations on the deep state features, mapping the 32-dimensional deep state features to corresponding quantified values ​​of chlorophyll density and water stress coefficient, generating preliminary quantitative values ​​of rice growth status. The regression calculation formula for the water stress coefficient is expressed as: , in: Indicates the water stress coefficient. Represents the feature vector of deep states. This represents the weight vector of the output layer. Indicates the bias term. This represents the Sigmoid activation function. Spatial consistency correction is applied to the initial quantized values ​​of rice growth status by calculating the average of the initial quantized values ​​of each spatial grid and its four adjacent spatial grids (east, west, south, and north). The difference between the initial quantized value and the average value is calculated. If the difference is greater than a preset tolerance threshold of 0.03, the average value is used to replace the initial quantized value of the spatial grid; otherwise, the original value is retained. The corrected value is the final output quantized value of rice growth status.

[0036] In some embodiments, the size of the spatial grid can be adjusted to 5m × 5m, the number of grids can be increased to 800, the hidden layer dimension of the graph neural network can be set to 32 dimensions, and the number of graph update iterations can be set to 5. In some embodiments, the hidden layer dimension of the state decoding network can be adjusted to 64 dimensions, and nitrogen content can be added as a third output dimension to the output layer. Optionally, the neighborhood range of spatial consistency correction can be extended to eight directions, including adjacent grids in the diagonal direction. Optionally, the tolerance threshold can be adjusted according to the change of grid size; when the grid size decreases, the tolerance threshold can be set to 0.02. It can be understood that through feature fusion and spatial consistency correction within the spatial grid, remote sensing spectral information and soil environmental data can be effectively integrated, local abnormal fluctuations can be eliminated, and the spatial continuity and reliability of quantitative growth state values ​​can be improved.

[0037] In one embodiment of the present invention, the internal state of the growth process simulation model is initialized, including rice phenological stage codes, cumulative biomass, and nitrogen uptake. The quantified value of the rice growth state corresponding to the current standard time scale is loaded into the model as a driving input. Based on internal physiological and ecological process equations, the model simulates the photosynthesis, respiration consumption, and dry matter distribution processes from the current moment to the next standard time scale, and dynamically adjusts the stress coefficient using the moisture content in the soil moisture dataset as an environmental stress factor. After the simulation is completed, the simulated growth state value for the next standard time scale is output as a predicted value, while simultaneously recording intermediate state variables to form a state simulation trajectory. Simulation intensity parameters of key physiological processes are extracted from the state simulation trajectory and converted into expected requirements for the quality and frequency of remote sensing data acquisition, generating data acquisition correction instructions. These instructions are sent to the UAV platform control system for acquiring multispectral remote sensing sequence data, adjusting its flight path planning, spectral sensor parameters, or acquisition time interval. When new multispectral remote sensing sequence data is obtained, the time synchronization processing dynamically adjusts the time alignment strategy and weights based on the acquisition parameters of the data to achieve closed-loop correction.

[0038] In the specific implementation, the internal state of the rice growth process simulation model is initialized. This internal state includes the phenological stage code, cumulative biomass, and nitrogen uptake. The phenological stage code is set to 3 to represent the peak tillering stage. The initial cumulative biomass value is 500 grams per square meter, and the initial nitrogen uptake value is 8 grams per square meter. The quantified value of the rice growth state corresponding to the current standard time scale is loaded into the growth process simulation model as the driving input. In the current input quantified value of the rice growth state, the chlorophyll density is 2.5 mg per square meter, and the water stress coefficient is 0.15. The growth process simulation model, based on its internal physiological and ecological process equations, simulates the rice growth process from the current standard time scale to the next standard time scale. The rice growth process includes photosynthesis, respiration consumption, and dry matter distribution. The photosynthetic rate is calculated using the light response curve. Respiration consumption is divided into maintenance respiration and growth respiration. The dry matter distribution ratio is adjusted according to the phenological stage (see Table 2).

[0039] Table 2: Comparison of numerical values ​​of the growth process simulation model before and after simulation Phenological Period Codes 3 4 Cumulative biomass (g / m²) 500 620 Nitrogen uptake (g / m²) 8 9.5 Chlorophyll density (mg / m²) 2.5 2.8 Water stress coefficient 0.15 0.12 During the simulation, the spatial distribution of soil moisture content in the soil moisture dataset was used as an environmental stress factor to dynamically adjust the stress coefficient in the physiological and ecological process equations. When the soil moisture content was below 60% of field capacity, the water stress coefficient increased, inhibiting the photosynthetic rate. After the simulation was completed, the growth process simulation model output the simulated growth state values ​​at the next standard time scale as predicted growth state values. In the output values, the chlorophyll density was 2.8 mg / m², the water stress coefficient was 0.12, and the phenological stage code was updated to 4 to represent the jointing stage. The growth process simulation model also recorded intermediate state variables during the simulation process, forming a state simulation trajectory, which included hourly cumulative photosynthesis, respiration consumption, and dry matter distribution sequences.

[0040] In practice, after calculating the predicted growth status for the next monitoring time, simulation intensity parameters of key physiological processes are extracted from the state simulation trajectory. These parameters include the maximum photosynthetic rate and the average water stress factor. The maximum photosynthetic rate is 0.5 grams of carbon dioxide per square meter per hour, and the average water stress factor is 0.85. These simulation intensity parameters are then converted into expected requirements for the quality and frequency of remote sensing data acquisition, generating data acquisition correction instructions. When the maximum photosynthetic rate is high and the water stress effect is low, it is required to increase the band resolution of the spectral sensor and shorten the acquisition interval. The data acquisition correction instructions are sent to the control system of the UAV platform responsible for acquiring multispectral remote sensing sequence data. These instructions include an adjusted flight path spacing of 5 meters, an increase in the number of spectral sensor bands to 10, and a shortened acquisition interval to twice per day.

[0041] In practice, the control system adjusts the flight path planning, spectral sensor operating parameters, and data acquisition intervals of the UAV based on data acquisition correction instructions. The flight path planning uses a dense grid pattern to cover the target rice paddy, the spectral sensor is set to a high-gain mode to improve the signal-to-noise ratio, and the acquisition intervals are adjusted to 10:00 AM and 2:00 PM. When the UAV platform acquires new multispectral remote sensing sequence data based on the adjusted parameters, the time synchronization processing step dynamically adjusts the time alignment strategy and weights based on the acquisition parameters that generated the new multispectral remote sensing sequence data, achieving closed-loop correction of the processing flow. The time alignment weights are redistributed based on the proximity of the acquisition time to the standard monitoring point; data points with small deviations from the standard monitoring point have a weight increased to 0.7, while data points with large deviations have a weight decreased to 0.3.

[0042] In some embodiments, the internal states of the growth process simulation model can include root depth and leaf area index, and the phenological stage codes can adopt the classification standards of the International Rice Research Institute, divided into ten sub-stages including the vegetative growth stage and the reproductive growth stage. In some embodiments, the recording frequency of the state simulation trajectory can be adjusted to once every 6 hours, the simulation intensity parameters can increase nitrogen absorption rate and transpiration efficiency, and the data acquisition correction command can specify that the UAV flight altitude is reduced to 50 meters to improve spatial resolution. Optionally, the control system's adjustment of the UAV flight path planning can include increasing the overlap rate to 80% to ensure that no spectral data is missed. Optionally, the weight adjustment of the time-series synchronous processing can adopt an exponential decay function, where the weight is inversely proportional to the square of the time difference, and the larger the time difference, the lower the weight. It can be understood that through the prediction and feedback of the growth process simulation model, the data acquisition strategy can be optimized according to the real-time state of rice growth, making the acquisition of multispectral remote sensing sequence data more targeted and improving monitoring accuracy. It can be understood that the closed-loop correction mechanism enables the entire monitoring system to have adaptive capabilities, dynamically adjusting the data acquisition and processing parameters according to changes in growth status, adapting to the dynamic changes in rice growth.

[0043] In one embodiment of the present invention, entities, attributes, and relationships related to rice growth are automatically extracted from an agronomic database to form initial knowledge triples. The agronomic database is a database established based on existing data records of rice growth knowledge. Conflict detection and resolution are performed on the initial triples, merging synonymous and heterogeneous entities and eliminating contradictory and redundant relationships to form clean knowledge triples. Knowledge representation learning technology is used to map the entities and relationships of the clean triples to a low-dimensional continuous vector space to obtain entity vectors and relationship vectors. Based on the vectors, a graph embedding algorithm is used to optimize the knowledge graph structure, making semantically similar entities closer in the vector space. Newly emerging uncovered semantic descriptions during the generation of the growth state semantic vector set are periodically received and transformed into new knowledge triples through entity linking and relationship extraction. After quality verification, the new triples are incrementally merged into the rice growth knowledge graph library to achieve continuous evolution and expansion of the knowledge graph.

[0044] In practice, entities, attributes, and relationships related to rice growth are automatically extracted from growth model documents published by the International Rice Research Institute and the agricultural database of the Food and Agriculture Organization of the United Nations to form initial knowledge triples. These initial knowledge triples include the entity "tillering stage," the attribute "lower limit of average daily temperature" with a value of "20," the attribute "upper limit of average daily temperature" with a value of "25," and the relationship between the entities "tillering stage" and "nitrogen fertilizer application rate" as "fertilizer requirement stage." Conflict detection and resolution are performed on the initial knowledge triples, merging entities with the same semantic meaning but different forms. For example, "tillering stage" and "tillering phase" are unified into "tillering stage," eliminating contradictory or redundant relationships to form clean knowledge triples.

[0045] In practical implementation, knowledge representation learning technology is used to map the entities and relations in the clean knowledge triples to a low-dimensional continuous vector space, resulting in entity vectors and relation vectors. The entity vector dimension is set to 100 dimensions. Based on the entity vectors and relation vectors, a graph embedding algorithm is used to optimize the overall structure of the rice growth knowledge graph database, making semantically similar entities closer in the vector space. For example, the Euclidean distance between the vectors of "tillering stage" and "jointing stage" is smaller than the distance between "tillering stage" and "harvest stage". During the optimization process, the translation distance model formula is used to evaluate the rationality of the triple combination: , in: Represents the head entity vector. Represents a relation vector. Represents the tail entity vector. This indicates the translation distance of the triplet; the smaller the distance, the more reasonable the triplet is.

[0046] In practice, the system periodically receives newly generated semantic descriptions of growth states that are not covered in the existing knowledge graph during the generation of the semantic vector set of growth states, such as "decreased grain filling during the grain-filling stage under high-temperature stress." These newly generated semantic descriptions are transformed into new knowledge triples through entity linking and relation extraction. Entity linking identifies "grain-filling stage" and "high-temperature stress," while relation extraction extracts the "environmental impact" relation. After quality verification, these new knowledge triples are incrementally merged into the rice growth knowledge graph database, enabling the continuous evolution and expansion of the knowledge graph.

[0047] In some embodiments, the sources of knowledge extraction can be expanded to include national agricultural science data centers and local variety trial reports, and the number of initial knowledge triples can be increased from 10,000 to 50,000. In some embodiments, the dimension of entity vectors can be set to 200, and the number of training rounds of the graph embedding algorithm can be increased from 100 to 300 to improve the discriminative power of knowledge representation. Optionally, the quality verification of new knowledge triples can be achieved through a combination of manual review and rule filtering to ensure the correctness of the added content. Optionally, the knowledge graph can be updated weekly to adapt to the rapid changes in environmental conditions during the growing season. It can be understood that through automated knowledge extraction and incremental updates, the rice growth knowledge graph database can continuously incorporate the latest scientific research results and field observation facts, maintaining the timeliness and coverage of knowledge.

[0048] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for monitoring the growth status of rice based on multispectral remote sensing, characterized in that, The method includes: Acquire multispectral remote sensing sequence data of the target paddy field within the monitoring period, perform time-series synchronization processing on the multispectral remote sensing sequence data, unify the original data at different acquisition times to a standard time scale, and generate a time-aligned remote sensing sequence set. Spectral feature evolution analysis was performed on the time-aligned remote sensing sequence set to extract the rice canopy growth evolution features corresponding to each standard time scale; Retrieve semantic descriptions of growth states that match the growth evolution characteristics from the rice growth knowledge graph database, and generate a set of semantic vectors of growth states; The set of semantic vectors representing the growth state is spatially matched with the externally input soil moisture dataset, and feature fusion is performed within the spatial grid to generate a fused feature field containing the rice growth state and the soil environment. The fused feature field is subjected to growth state decoding processing, and the quantitative value of rice growth state corresponding to the standard time scale is output. The quantitative value of rice growth state is input into the growth process simulation model, and the growth process simulation model is driven to calculate the predicted value of growth state at the next monitoring time. The predicted growth status is returned to the time-series synchronization processing step to correct the processing flow of the subsequently acquired multispectral remote sensing sequence data.

2. The method for monitoring rice growth status based on multispectral remote sensing according to claim 1, characterized in that, The multispectral remote sensing sequence data undergoes time-series synchronization processing to unify the raw data from different acquisition times to a standard time scale, generating a time-aligned remote sensing sequence set. This includes the following steps: The multispectral remote sensing sequence data includes spatial location coding, time-series labels, and spectral reflectance values; Receive raw multispectral remote sensing sequence data, which carries timestamp information for multiple acquisition times; A standard time scale axis is constructed, which consists of multiple pre-set equally spaced monitoring points; For each data acquisition moment, the time difference between the acquisition timestamp of the raw data and the two nearest monitoring points on the standard time scale axis is calculated. Based on the time difference, the original data is resampled in the time dimension, and the spectral values ​​of the original data are interpolated to the corresponding two standard monitoring points. For each standard monitoring point, collect all spectral values ​​that have been resampled and allocated to the standard monitoring point, calculate the average spectral value of the standard monitoring point, and generate the standardized spectral data unit corresponding to the standard time scale. The standardized spectral data units of all standard monitoring points are arranged in chronological order and assembled into the time-aligned remote sensing sequence set. The resampling process specifically includes: using a cubic spline interpolation algorithm, with the time difference as the interpolation weight, to smoothly interpolate the original spectral curve between two standard monitoring points to generate a transitional spectral numerical curve; and extracting the spectral values ​​located at the standard monitoring point from the transitional spectral numerical curve.

3. The method for monitoring rice growth status based on multispectral remote sensing according to claim 1, characterized in that, Spectral feature evolution analysis was performed on the time-aligned remote sensing sequence set to extract the rice canopy growth evolution characteristics corresponding to each standard time scale. This was achieved through the following steps: Spectral data corresponding to consecutive standard time scales in the time-aligned remote sensing sequence set are selected to form a spectral evolution data segment; Principal component transformation is performed on the spectral evolution data segment to eliminate the correlation between spectral bands and obtain multiple independent evolutionary principal components; Calculate the rate of change of each evolutionary principal component along the time dimension of the spectral evolution data segment, and generate a set of rate of change vectors; Identify the principal components in the set of rate of change vectors whose change magnitude exceeds a preset threshold, and mark them as sensitive principal components; Temporal pattern mining is performed on the sensitive evolutionary principal components to extract the trend pattern features of their spectral values ​​changing over time. The trend pattern features include linear growth and periodic fluctuations. The trend pattern features are encoded into a fixed-length feature vector, which serves as the rice canopy growth and evolution features corresponding to the standard time scale.

4. The method for monitoring rice growth status based on multispectral remote sensing according to claim 1, characterized in that, The step of retrieving semantic descriptions of growth states that match the growth evolution characteristics from the rice growth knowledge graph database and generating a set of semantic vectors of growth states is accomplished through the following steps: Receive the growth and evolution characteristics of the rice canopy; The rice canopy growth and evolution characteristics are input into a preset semantic encoder and converted into an initial semantic query vector. In the rice growth knowledge graph database, the cosine similarity between the initial semantic query vector and the feature vectors of all entity nodes in the knowledge graph is calculated; The entity nodes are sorted from high to low based on the cosine similarity, and the top-ranked candidate entity node set is selected. Extract the attribute triples associated with each candidate entity node from the rice growth knowledge graph database. The attribute triples include entity, relation, and attribute value. Semantic parsing is performed on the attribute triples to convert them into natural language descriptions, generating preliminary semantic descriptions corresponding to the candidate entity nodes; The initial semantic description is input into the language model for semantic refinement and structuring, generating semantic description entries of the growth state with a uniform format. The growth state semantic description entries corresponding to all candidate entity nodes are aggregated to form the growth state semantic vector set.

5. The method for monitoring rice growth status based on multispectral remote sensing according to claim 1, characterized in that, The set of semantic vectors representing the growth state is spatially matched with an externally input soil moisture dataset. Feature fusion is then performed within a spatial grid to generate a fused feature field containing both rice growth state and soil environment data. Includes the following operations: Receive an externally input soil moisture dataset, which contains spatial distribution values ​​of soil moisture content corresponding to the spatial location of the target paddy field; The geographic spatial range of the target paddy field is divided into regular spatial grids, with each grid serving as an independent processing unit; Within each spatial grid, the portion of the growth state semantic vector set belonging to the spatial grid is obtained to acquire the grid semantic features; Within the same spatial grid, the soil moisture content values ​​belonging to the spatial grid in the soil moisture dataset are obtained to obtain the grid moisture characteristics; Establish a feature association graph between grid semantic features and grid soil moisture features, wherein the feature association graph uses elements of grid semantic features as nodes and grid soil moisture features as edge weights; The graph neural network performs message propagation and aggregation operations on the feature association graph, enabling the grid semantic features and grid moisture features to be deeply integrated during the iterative update of the graph. After a preset number of graph update iterations, the fused feature representations of all nodes in the feature association graph are extracted and used as the fused feature vectors of the spatial grid. The fused feature vectors of all spatial grids in the target paddy field are arranged according to their spatial coordinates to generate the spatially continuous fused feature field.

6. The method for monitoring rice growth status based on multispectral remote sensing according to claim 1, characterized in that, The fused feature field is subjected to growth state decoding processing to output the quantitative value of rice growth state corresponding to the standard time scale, which is achieved through the following process: The quantitative values ​​of rice growth status include chlorophyll density and water stress coefficient; The fused feature field is input into a pre-constructed state decoding network, which includes a multilayer perceptron structure. The first layer of the state decoding network increases the feature dimension of the input fused feature field, mapping the low-dimensional fused features to a high-dimensional feature space to obtain a high-dimensional feature field. The intermediate layer of the state decoding network performs nonlinear transformation and feature extraction on the high-dimensional feature field, captures deep patterns in the fused feature field that are strongly correlated with the rice growth state, and generates deep state features. The output layer of the state decoding network performs regression calculations on the deep state features, mapping the high-dimensional deep state features to the quantized values ​​of the corresponding specific growth parameters, and generating preliminary quantized values ​​of rice growth status. Spatial consistency correction is performed on the preliminary quantitative values ​​of rice growth status to eliminate numerical abrupt changes between adjacent grids caused by spatial grid division; The spatial consistency correction step includes: calculating the average of the preliminary quantization values ​​of each spatial grid and the preliminary quantization values ​​of all its adjacent spatial grids; calculating the difference between the preliminary quantization value and the average value; if the difference is greater than a preset tolerance threshold, replacing the preliminary quantization value of the spatial grid with the average value, otherwise retaining the original value; the corrected value is the final output quantization value of the rice growth status.

7. The method for monitoring rice growth status based on multispectral remote sensing according to claim 1, characterized in that, The quantitative values ​​of rice growth status are input into the growth process simulation model, which then drives the model to calculate the predicted growth status value at the next monitoring time. This includes the following detailed steps: Initialize the internal state of the growth process simulation model, which includes the phenological stage code, cumulative biomass, and nitrogen uptake of rice; The quantitative value of the rice growth status corresponding to the current standard time scale is used as the driving input of the model and loaded into the growth process simulation model. The growth process simulation model is based on internal physiological and ecological process equations to simulate the rice growth process from the current standard time scale to the next standard time scale. The rice growth process includes photosynthesis, respiration consumption, and dry matter distribution. During the simulation, the spatial distribution values ​​of soil moisture content in the soil moisture dataset are used as environmental stress factors to dynamically adjust the stress coefficient in the physiological and ecological process equations. After the simulation is completed, the growth process simulation model outputs the simulated growth state value at the next standard time scale, which is used as the predicted growth state value. The growth process simulation model simultaneously records intermediate state variables during the simulation process, forming a state simulation trajectory, which is used for process correction in subsequent steps.

8. The method for monitoring rice growth status based on multispectral remote sensing according to claim 7, characterized in that, The predicted growth status value is returned to the time-series synchronization processing step to correct the processing flow of the subsequently acquired multispectral remote sensing sequence data. This is specifically achieved through the following mechanism: After completing the calculation of the predicted growth state value for the next monitoring time, the simulation intensity parameters of key physiological processes are extracted from the state simulation trajectory. The simulated intensity parameters are converted into expected requirements for the quality and frequency of remote sensing data acquisition, and data acquisition correction instructions are generated. The data acquisition correction command is sent to the control system of the UAV platform responsible for acquiring the multispectral remote sensing sequence data; The control system adjusts the flight path planning, spectral sensor operating parameters, or data acquisition time interval of the UAV according to the data acquisition correction command. When the UAV platform collects new multispectral remote sensing sequence data according to the adjusted parameters, the time synchronization processing step will dynamically adjust the time alignment strategy and weight based on the acquisition parameters that generated the new multispectral remote sensing sequence data, thereby achieving closed-loop correction of the processing flow.

9. The method for monitoring rice growth status based on multispectral remote sensing according to claim 1, characterized in that, The method further includes the step of constructing and updating the rice growth knowledge graph database, which operates independently of the monitoring process, and specifically includes: From agricultural databases, entities, attributes, and relationships between entities related to rice growth are automatically extracted to form initial knowledge triples. The initial knowledge triples are subjected to conflict detection and resolution. Entities with the same semantics but different forms are merged to eliminate contradictory or redundant relationships and form clean knowledge triples. Using knowledge representation learning techniques, the entities and relations in the clean knowledge triples are mapped to a low-dimensional continuous vector space to obtain entity vectors and relation vectors; Based on the entity vectors and relation vectors, a graph embedding algorithm is used to optimize the overall structure of the rice growth knowledge graph database, so that entities with similar semantics are closer in the vector space. Periodically receive new growth state semantic descriptions that are not covered in the existing knowledge graph during the generation process of the growth state semantic vector set; The newly generated semantic description of the growth state is transformed into new knowledge triples through entity linking and relation extraction; After the new knowledge triples are quality-verified, they are incrementally merged into the rice growth knowledge graph database to achieve continuous evolution and expansion of the knowledge graph.

10. A rice growth status monitoring system based on multispectral remote sensing, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the rice growth status monitoring method based on multispectral remote sensing as described in any one of claims 1 to 9.

Images