A method and system for predicting relative chlorophyll content based on multispectral images

Abstract

The application discloses a kind of chlorophyll relative content prediction method and system based on multispectral image, the method includes: obtaining multi-source data;According to environmental data, determine phenological period category and environmental stress category, and determine comprehensive scene mode accordingly;Environment data and multispectral image are extracted, and vegetation index feature, texture feature and environmental feature are obtained;According to comprehensive scene mode, scene self-adapting interpolation is carried out;According to environmental feature, vegetation index feature correction parameter is generated, and vegetation index feature is dynamically corrected;The corrected feature is constructed as multi-source time series feature sequence, multi-source time series feature sequence is weighted and fused, and input prediction model, and the chlorophyll relative content prediction value is obtained.The system includes data acquisition module, scene mode determination module, feature extraction module, feature processing module and prediction module.The application can be widely applied in the field of agricultural information technology.

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 relative chlorophyll content based on multispectral images. Background Technology

[0002] Relative chlorophyll content (SPAD) is a key physiological indicator characterizing plant photosynthetic capacity and nitrogen nutrition status. Rapid, non-destructive, and accurate acquisition of SPAD is a crucial prerequisite for modern smart agriculture to achieve precision fertilization, stress diagnosis, and yield prediction. Currently, field chlorophyll acquisition still heavily relies on manual, single-point measurements using handheld chlorophyll meters (SPAD-502). While this method provides accurate readings, it is time-consuming, labor-intensive, and has limited sample representativeness, failing to achieve rapid, comprehensive, and continuous monitoring at the orchard or field scale. This fundamentally contradicts the demands of modern precision agriculture for large-scale, high-resolution, and timely physiological information. Summary of the Invention

[0003] In view of this, in order to solve the technical problems of time-consuming, labor-intensive, and limited sample representativeness in existing methods for measuring relative chlorophyll content, the present invention proposes a method for predicting relative chlorophyll content based on multispectral imagery, which includes the following steps: First, multi-source data is acquired, including at least environmental data and multispectral imagery. Phenological stage categories and environmental stress categories are determined based on the environmental data, and a comprehensive scenario model is determined based on these categories. Feature extraction is performed on the environmental data and multispectral imagery to obtain vegetation index features, texture features, and environmental features. According to the comprehensive scenario model, scenario-adaptive interpolation is performed on the vegetation index features, texture features, and environmental features to obtain preprocessed features. Vegetation index feature correction parameters are generated based on the environmental features, and the vegetation index features are dynamically corrected based on these parameters to obtain corrected vegetation index features. The corrected vegetation index features, texture features, and environmental features are constructed into a multi-source time-series feature sequence. Weighted fusion is performed on each time step and feature dimension of the multi-source time-series feature sequence, and the fused features are input into a prediction model to obtain a predicted value for relative chlorophyll content.

[0004] Furthermore, based on the predicted relative chlorophyll content and the dynamic health threshold corresponding to the comprehensive scenario model, a spatial distribution map of canopy chlorophyll health is generated. Subsequently, spatial cluster analysis is performed on this health distribution map to identify regions with different health levels. Finally, based on the results of the cluster analysis, a targeted variable-based work prescription map is generated to guide precise work in different zones.

[0005] Furthermore, the step of determining the phenological period category and environmental stress category based on the environmental data, and determining the comprehensive scenario mode based on the phenological period category and the environmental stress category, includes: determining the corresponding phenological period category based on the environmental data; determining the environmental stress category based on the environmental data and a preset rule base; and determining the comprehensive scenario mode based on the combination relationship between the phenological period category and the environmental stress category.

[0006] Furthermore, the step of performing scenario-adaptive interpolation on the vegetation index features, texture features, and environmental features according to the comprehensive scenario mode includes: using exponential smoothing interpolation combined with accumulated temperature when the comprehensive scenario mode corresponds to the vigorous growth period; and using linear interpolation when the comprehensive scenario mode corresponds to the slow growth period.

[0007] Furthermore, the prediction model includes an environmental correction attention module, a spatiotemporal importance attention module, and a long short-term memory network. The environmental correction attention module is used to generate vegetation index feature correction parameters based on the environmental features and to dynamically correct the vegetation index features. The spatiotemporal importance attention module is used to calculate the weight coefficients of the multi-source time-series feature sequences at each time step and each feature dimension, and to perform weighted fusion of the multi-source time-series feature sequences based on the weight coefficients. The long short-term memory network is used to output the predicted value of the relative chlorophyll content based on the weighted fused features.

[0008] Based on the above method, in a second aspect, the present invention also proposes a chlorophyll relative content prediction system based on multispectral images. The system includes a data acquisition module, a scenario mode determination module, a feature extraction module, a feature processing module, and a prediction module.

[0009] Based on the above scheme, this invention provides a method and system for predicting the relative chlorophyll content based on multispectral imagery. First, it acquires environmental and canopy multispectral data synchronously via the Internet of Things and drones. It innovatively introduces a phenological-environmental scenario discrimination mechanism to automatically identify the current growth stage and stress state of the litchi orchard. Based on the discriminated scenario, the system adaptively selects a feature subset, employs a matching interpolation algorithm, and calls the corresponding expert prediction model. The environmental correction channel explicitly counteracts environmental interference with the spectrum, while the spatiotemporal fusion channel deeply mines multi-source temporal features. Finally, the predicted values, combined with scenario thresholds dynamically optimized based on reinforcement learning, automatically generate a precise fertilization prescription map to guide variable fertilization. This achieves rapid prediction of the relative chlorophyll content under a wide range of environments, representing a leap from static prediction to scenario-adaptive prediction and then to closed-loop decision-making, providing a directly executable technical solution for precision nutrition management in litchi orchards. Attached Figure Description

[0010] Figure 1This is a flowchart of the steps of a method for predicting the relative chlorophyll content based on multispectral images according to the present invention. Figure 2 This is a schematic diagram of the data flow in the method of the present invention. Detailed Implementation

[0011] In addition to the issues mentioned in the background, existing methods for predicting relative chlorophyll content also suffer from the following problems: static models based on remote sensing have poor robustness and insufficient ability to model dynamic environmental disturbances and temporal changes in crops. Multispectral remote sensing technology, using unmanned aerial vehicles (UAVs) as a platform, has become an important technological direction for crop physiological parameter inversion due to its advantages of high efficiency, non-destructive testing, and wide coverage. Existing technologies typically rely on single or multiple independent aerial images, calculating vegetation indices (such as NDVI and GNDVI) and establishing statistical regression or machine learning models (such as random forests) with ground-measured SPAD values ​​to achieve spatial estimation. However, such methods have significant limitations: First, spectral reflectance signals are highly susceptible to transient interference from rapidly changing field environmental factors, resulting in weak generalization ability and large fluctuations in prediction accuracy of inversion models built based on the assumption of stable light intensity under actual variable weather conditions; Second, most models rely solely on spectral information and fail to fully utilize the spatial texture features contained in high-resolution images that can reflect canopy structure and growth status; Third, most models are static analyses of independent time periods, failing to effectively model the dynamic cumulative effect of chlorophyll content with phenological processes and its long-term temporal dependence on historical environmental stresses, resulting in insufficient "dynamic" prediction.

[0012] Secondly, multi-source data often exists in isolated silos, lacking in-depth cross-modal collaborative analysis and adaptive decision-making mechanisms. While IoT technology has been widely applied in agricultural environmental monitoring, enabling continuous collection of multi-dimensional time-series data such as soil temperature and humidity, atmospheric temperature and humidity, and light intensity, current applications largely remain at the level of data visualization and simple threshold alarms. Effective in-depth fusion and causal correlation analysis are lacking between the time-series environmental data collected by IoT and the plant physiological spectral / texture data acquired by remote sensing. The core coupling mechanism of how environmental stress dynamically affects canopy spectral response and thus alters chlorophyll content has not yet been explicitly modeled in existing technologies. Although more advanced cross-domain technologies have demonstrated the enormous potential for intelligent fusion and decision optimization of multi-source data, these solutions are complex, geared towards macro-level ecological environmental protection, and have not yet been deeply adapted and transformed for specific crop growth and management scenarios, especially for the phenological stages and stress responses of specialty economic fruit trees like lychee.

[0013] In summary, existing technical solutions either suffer from efficiency and coverage issues, or lack robustness and dynamism in their models, or while possessing advanced data fusion frameworks, they fail to deeply integrate with specific agronomic management scenarios. In particular, there is a lack of a closed-loop system capable of automatically sensing crop growth conditions and dynamically adjusting data strategies and predictive models accordingly, ultimately forming precise agricultural decision-making. Therefore, there is an urgent need to develop an intelligent prediction and decision-making method tailored to specific crops, possessing situational awareness and adaptive capabilities, to achieve a leap from "passive state monitoring" to "proactive management intervention."

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] It should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0016] It should be understood that the terms "system," "apparatus," "unit," and / or "module" used in this application are a method of distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0017] In the description of the embodiments of this application, "a plurality of" refers to two or more. The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature.

[0018] Furthermore, flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Additionally, other operations can be added to these processes, or one or more steps can be removed from them.

[0019] Reference Figure 1 This is a flowchart illustrating an optional example of the chlorophyll relative content prediction method based on multispectral imagery proposed in this invention. The method can be applied to computer equipment. The chlorophyll relative content prediction method proposed in this embodiment may include, but is not limited to, the following steps: Step S1: Acquire multi-source data, which includes at least environmental data and multispectral imagery; Step S2: Determine the phenological period category and environmental stress category based on the environmental data, and determine the comprehensive scenario model based on the phenological period category and the environmental stress category; Step S3: Extract features from the environmental data and the multispectral image to obtain vegetation index features, texture features, and environmental features; Step S4: According to the comprehensive scenario mode, perform scenario-adaptive interpolation on the vegetation index features, the texture features, and the environmental features to obtain the preprocessed features; Step S5: Generate vegetation index feature correction parameters based on the environmental features, and dynamically correct the vegetation index features based on the vegetation index feature correction parameters to obtain corrected vegetation index features; construct a multi-source time-series feature sequence from the corrected vegetation index features, the texture features, and the environmental features; perform weighted fusion on each time step and each feature dimension of the multi-source time-series feature sequence, and input the fused features into the prediction model to obtain the predicted value of relative chlorophyll content.

[0020] Step S6: Generate a spatial distribution map of canopy chlorophyll health based on the predicted relative chlorophyll content and the dynamic health threshold corresponding to the comprehensive scenario model. Step S7: Perform spatial cluster analysis based on the spatial distribution map of canopy chlorophyll health, and generate a variable job prescription map based on the cluster analysis results.

[0021] Data flow reference of this method Figure 2 .

[0022] In some feasible embodiments, step S1 specifically includes: By deploying an IoT sensor network in the litchi orchard, environmental time-series monitoring data (such as temperature, humidity, soil moisture, soil EC value, etc.) are collected periodically; multispectral images of the litchi canopy are collected periodically through a drone platform; and ground stratified sampling is carried out simultaneously to obtain measured values ​​of leaf chlorophyll relative content (SPAD) as a benchmark for model training and validation.

[0023] In some feasible embodiments, step S2 specifically includes: Phenological Stage Determination: The system has a built-in litchi phenological knowledge base based on an accumulated temperature model or a calendar model. Based on the current date and historical temperature accumulation data, it automatically determines the current major phenological stage. Phenological stages include at least: flower bud differentiation stage, flowering and fruit setting stage, young fruit enlargement stage, and autumn shoot growth stage. For example, when the accumulated temperature reaches a preset threshold and the date is in May or June, it is determined to be the "young fruit enlargement stage." Environmental stress category identification: Read IoT data in real time and identify the main types of environmental stress through preset rules.

[0024] Water stress identification: If the average soil volumetric water content (VWC) for three consecutive days is lower than the preset stress threshold for this phenological period (e.g., 50% of the field level), it is marked as "water stress".

[0025] High-temperature light suppression stress identification: If the highest temperature of the day exceeds 35°C and the photosynthetically active radiation (PAR) remains at a high value, it is marked as "high-temperature light suppression stress".

[0026] Based on historical phenological models (such as flower bud differentiation period, flowering and fruit setting period, young fruit enlargement period, and autumn shoot growth period divided by date) and real-time IoT environmental data (such as soil moisture being below the stress threshold for several consecutive days), the system automatically determines the current "comprehensive scenario mode" through a preset rule base or a lightweight machine learning model (scenario discriminator). This mode is composed of phenological period categories (such as "young fruit enlargement period") and major environmental stress categories (such as "water stress"), for example, "Scenario S3: Young fruit enlargement period - water stress".

[0027] In some feasible embodiments, step S3 specifically includes: S3.1 Perform integrated processing on the original multispectral images.

[0028] This was achieved using DJI Terra, DJI's official software. This software combines high-precision RTK positioning data and solar radiation sensor data to automatically complete multispectral band registration, radiometric correction (converting the original DN values ​​to surface reflectance), and orthorectification, generating a high-precision four-band (G, R, RE, NIR) orthophoto map.

[0029] S3.2 Based on the corrected multispectral images, nine vegetation indices sensitive to the physiological state of litchi were calculated for each region of interest (ROI) in litchi canopy using professional software such as ENVI 5.6.

[0030] The specific indices and calculation formulas are shown in the table below: Table 1 Vegetation Index S3.3 To further quantify the spatial structure information of the canopy, based on the Gray-Level Co-occurrence Matrix (GLCM) method, eight types of second-order texture features were extracted from images in four bands: green (G), red (R), red edge (RE), and near-infrared (NIR): mean, variance, homogeneity, contrast, dissimilarity, entropy, energy, and correlation. Thus, each canopy ROI can obtain a 32-dimensional texture feature vector of 4 bands × 8 textures per flight day.

[0031] S3.4. Aggregate the raw IoT time-series data (collected every 30 minutes) stored in the cloud on a daily basis. Preferably, use the daily average value as the environmental characteristic value for the day, including: average temperature, average humidity, average soil temperature, average soil moisture, average soil electrical conductivity, etc. For parameters such as wind speed and precipitation, the daily maximum value or cumulative value can be used.

[0032] In some feasible embodiments, step S4 specifically includes: First, data normalization is performed: To ensure the effective fusion of feature data (environmental features, 9 vegetation indices, and 32-dimensional texture features) from different sensors with varying dimensions, an improved one-dimensional convolutional autoencoder method is used to detect and repair outliers in the original IoT time-series data. If an outlier is detected (corresponding to 99.9% confidence level), it is determined to be an anomaly, and KNN interpolation is performed using normal sensor data that are in the same spatiotemporal proximity to repair the outlier, which significantly improves the reliability of the data. Secondly, time anchor alignment: Using the UAV flight day as the time anchor, the alignment of ground SPAD values ​​uses the flight date of each UAV aerial survey as the core time anchor. An interpolation model is selected based on the currently identified phenological stage. For this anchor day: if it is during a period of vigorous growth (such as the young fruit enlargement stage), SPAD changes rapidly, so exponential smoothing interpolation combined with effective accumulated temperature (GDD) is used. The formula can be expressed as: This method is more in line with physiological patterns; if the tree is in a slow growth or dormant period, linear interpolation is used. This method constructs a spatiotemporally strictly matched dataset, where each sample point represents the complete feature vector (including environment, vegetation index, and texture) of a sample tree on a specific flight day and the corresponding target SPAD value.

[0033] In addition, the system pre-configures a "scenario-feature correlation mapping table". Based on the current scenario mode s, it dynamically selects the optimal feature subset. Under the "water stress" scenario, the system prioritizes the NDWI vegetation index and related texture features that are sensitive to water, while reducing the weight of features that are not sensitive to water, thus achieving scenario-adaptive optimization of feature engineering.

[0034] The fusion feature sequences arranged in chronological order, with a length of [missing information], are [missing information]. Continuous sampling is performed within the window, where To ensure temporal continuity between adjacent samples, the time steps were set to 7 to 15 based on the phenological period of litchi, with a sliding step size of 1. Finally, the multivariate feature data within each window were reconstructed into a format of size (…). The fusion tensor of (total number of features) is used as the input tensor of the multivariate long short-term memory network with the embedded attention mechanism.

[0035] Through the above processing, time-aligned sample data is constructed, providing preprocessed features for subsequent prediction model input.

[0036] In some feasible embodiments, step S5 specifically includes: Vegetation index feature correction parameters are generated based on environmental characteristics, and the vegetation index features are dynamically corrected based on the vegetation index feature correction parameters to obtain the corrected vegetation index features. The corrected vegetation index features, texture features and environmental features are constructed into a multi-source time series feature sequence. The time steps and feature dimensions of the multi-source time series feature sequence are weighted and fused, and the fused features are input into the prediction model to obtain the predicted value of chlorophyll relative content.

[0037] The core of the prediction model is a Long Short-Term Memory (LSTM) network, but its key innovation lies in the introduction of a dual-channel adaptive attention fusion module, whose parameters are bound to the contextual pattern s. The specific structure is as follows: Channel 1: Environmental Correction Attention Module. This module takes the current environmental feature vector as input. Through a light quantum network, it outputs a set of dynamic correction parameters (weight matrix) for spectral features (vegetation index). and bias vector ).

[0038] The correction formula is: This structure explicitly and dynamically models the impact of environmental disturbances on spectral signals, going beyond simply using environmental data as input features; the calculation formulas demonstrate increased complexity and specificity.

[0039] Channel 2: Spatiotemporal Importance Attention Module. The corrected spectral features, original texture features, and other features are concatenated and input into a spatiotemporal attention layer. This layer simultaneously calculates the comprehensive importance weights for different historical time steps and different feature dimensions. To achieve more refined feature fusion, the attention weight calculation formula can be implemented based on the scaling dot product attention mechanism.

[0040] Attention weight coefficient The calculation formula is: in, and For trainable parameters, This is the Sigmoid function.

[0041] LSTM layer: The processed temporal fusion feature sequence is input into a multi-layer LSTM network. The number of LSTM units in each layer (64 or 128) can be adjusted according to the data size and complexity. The internal updates of the LSTM units strictly follow the gating mechanism, automatically learning the long-term temporal dependencies between features.

[0042] Output layer: Take the LSTM hidden state at the last time step, perform nonlinear transformation through one or more fully connected layers, and finally output a scalar value, namely the predicted SPAD value.

[0043] In multivariate long short-term memory network prediction models, the update process of long short-term memory units includes the forgetting gate. Input gate Temporary cell state Cell state and output gate The formula for calculating is: In some feasible embodiments, the method further includes training the prediction model: Data partitioning: The constructed 3D time series dataset is divided into training set, validation set and test set according to sample tree ID (or time), with a ratio of 7:2:1 to ensure data independence between different sets.

[0044] Loss function and optimizer: The mean squared error (MSE) is used as the loss function to directly optimize the deviation between the predicted and measured values. The optimizer chosen is Adam, with an initial learning rate of 0.001, and a learning rate decay strategy.

[0045] Training process: Set the batch size (e.g., 32 or 64) and the number of training epochs (e.g., 100-200). After each training epoch, evaluate the model performance on the validation set (using RMSE and R²). Use early stopping to prevent overfitting; that is, stop training and retain the model parameters with the best validation performance when the validation set loss no longer decreases over 10 consecutive epochs.

[0046] Regularization: A Dropout mechanism (with a dropout rate typically set to 0.3-0.5) is introduced into the fully connected layer to randomly drop some neurons in some feasible implementations to enhance the model's generalization ability.

[0047] In some feasible embodiments, step S6 specifically includes: The system maintains a "scenario-dynamic threshold mapping library". Each scenario mode s corresponds to an initial health threshold. The threshold is set based on the statistical distribution of historical SPAD values ​​in this scenario (such as a certain percentile).

[0048] The threshold is dynamically optimized using reinforcement learning algorithms (such as Proximal Policy Optimization, PPO). Define the state. (Current forecast of SPAD, meteorological conditions, phenological period), action (Fine-tuning threshold range), rewards The reward function is designed as follows: , in, To provide accurate early warning rewards, and The penalties are for false alarms and false negatives, respectively. The system continuously optimizes rewards by interacting with the environment (historical data), autonomously learning and adjusting the optimal warning threshold for each scenario. .

[0049] According to the predicted value and optimized dynamic threshold A spatial distribution map of chlorophyll health in the litchi canopy is generated, which visually displays abnormal areas below the threshold.

[0050] In some feasible embodiments, step S7 specifically includes: Spatial clustering analysis is performed on abnormal areas, and the cluster boundaries are precisely mapped to pre-divided work grids in the orchard (corresponding to row numbers / tree numbers). Combining the species and age information of the trees within the grid, an NPK recommended fertilization amount is generated for each grid using a built-in fertilization knowledge base.

[0051] The system outputs a shapefile-based application prescription file. This file contains the geographical location and recommended fertilization amount for each application grid, which can be directly imported into a variable-rate fertilizer applicator or smart agricultural machinery to drive it to perform differentiated precision fertilization operations, truly realizing a complete smart agriculture closed loop of "perception-analysis-decision-execution".

[0052] A chlorophyll relative content prediction system based on multispectral imagery includes: The data acquisition module is used to execute step S1; The scenario mode determination module is used to execute step S2; The feature extraction module is used to perform step S3; The feature processing module is used to execute step S4; The prediction module is used to perform step S5; The decision module is used to execute steps S6 to S7.

[0053] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0054] A device for predicting relative chlorophyll content based on multispectral imagery: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements a method for predicting the relative chlorophyll content based on multispectral images as described above.

[0055] The content of the above method embodiments is applicable to the device embodiments. The specific functions implemented by the device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0056] A storage medium storing processor-executable instructions, which, when executed by a processor, are used to implement a method for predicting relative chlorophyll content based on multispectral images as described above.

[0057] The content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0058] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for predicting the relative chlorophyll content based on multispectral imagery, characterized in that, Includes the following steps: Acquire multi-source data, which includes at least environmental data and multispectral imagery; Based on the environmental data, determine the phenological period category and the environmental stress category, and determine the comprehensive scenario model based on the phenological period category and the environmental stress category; Feature extraction is performed on the environmental data and the multispectral image to obtain vegetation index features, texture features, and environmental features; According to the comprehensive scenario model, scenario-adaptive interpolation is performed on the vegetation index features, the texture features, and the environmental features to obtain the preprocessed features. Based on the environmental characteristics, vegetation index feature correction parameters are generated, and the vegetation index features are dynamically corrected based on the vegetation index feature correction parameters to obtain the corrected vegetation index features. The corrected vegetation index features, texture features, and environmental features are constructed into a multi-source time-series feature sequence. The time steps and feature dimensions of the multi-source time-series feature sequence are weighted and fused, and the fused features are input into the prediction model to obtain the predicted value of chlorophyll relative content.

2. The method for predicting relative chlorophyll content based on multispectral imagery according to claim 1, characterized in that, Also includes: Based on the predicted relative chlorophyll content and the dynamic health threshold corresponding to the comprehensive scenario model, a spatial distribution map of canopy chlorophyll health is generated.

3. The method for predicting relative chlorophyll content based on multispectral imagery according to claim 2, characterized in that, The dynamic health threshold is obtained through a dynamic adjustment strategy, and the reward function of the dynamic adjustment strategy is: in, Rewards are given for correct early warnings. For false alarms, As a penalty for underreporting, and These represent the false alarm penalty weight coefficient and the false alarm penalty weight coefficient, respectively.

4. The method for predicting relative chlorophyll content based on multispectral imagery according to claim 2, characterized in that, Also includes: Spatial clustering analysis was performed based on the spatial distribution map of canopy chlorophyll health, and a variable job prescription map was generated based on the clustering analysis results.

5. The method for predicting relative chlorophyll content based on multispectral imagery according to claim 1, characterized in that, The step of determining phenological period categories and environmental stress categories based on the environmental data, and determining a comprehensive scenario model based on the phenological period categories and environmental stress categories, specifically includes: The corresponding phenological period category is determined based on the environmental data in the multi-source data. Based on the environmental data in the multi-source data and the preset rule base, environmental stress categories are generated; Based on the combined relationship between the phenological period category and the environmental stress category, a comprehensive scenario pattern is determined.

6. The method for predicting relative chlorophyll content based on multispectral imagery according to claim 5, characterized in that, The phenological period categories include at least the flower bud differentiation period, flowering and fruit setting period, young fruit enlargement period, and autumn shoot growth period; the environmental stress categories include at least water stress and high temperature and light inhibition stress; the comprehensive scenario model includes the vigorous growth period and the slow growth period.

7. The method for predicting relative chlorophyll content based on multispectral imagery according to claim 1, characterized in that, The step of performing scenario-adaptive interpolation on the vegetation index features, texture features, and environmental features according to the comprehensive scenario model includes: During the vigorous growth period corresponding to the comprehensive scenario model, exponential smoothing interpolation combined with accumulated temperature is adopted; Linear interpolation is used during the slow growth period corresponding to the comprehensive scenario mode.

8. The method for predicting relative chlorophyll content based on multispectral imagery according to claim 1, characterized in that, The step of inputting the preprocessed features into the prediction model and outputting the predicted value of the relative chlorophyll content specifically includes: The prediction model includes an environmental correction attention module, a spatiotemporal importance attention module, and a long short-term memory network; The environmental correction attention module is used to generate vegetation index feature correction parameters based on the environmental features, and to dynamically correct the vegetation index features. The spatiotemporal importance attention module is used to calculate the weight coefficients of the multi-source temporal feature sequence at each time step and each feature dimension, and to perform weighted fusion of the multi-source temporal feature sequence based on the weight coefficients; The long short-term memory network is used to output the predicted value of the relative chlorophyll content based on the weighted fused features.

9. A chlorophyll relative content prediction system based on multispectral imagery, characterized in that, include: The data acquisition module is used to acquire multi-source data, which includes at least environmental data and multispectral images; The scenario pattern determination module is used to determine the phenological period category and the environmental stress category based on the environmental data, and to determine the comprehensive scenario pattern based on the phenological period category and the environmental stress category. The feature extraction module is used to extract features from the environmental data and the multispectral image to obtain vegetation index features, texture features and environmental features; The feature processing module is used to perform scenario-adaptive interpolation on the vegetation index features, texture features and environmental features according to the comprehensive scenario mode, and generate vegetation index feature correction parameters according to the environmental features to dynamically correct the vegetation index features. The prediction module is used to construct a multi-source time-series feature sequence from the corrected vegetation index features, the texture features, and the environmental features, perform weighted fusion on the multi-source time-series feature sequence, and output the predicted value of the relative chlorophyll content.

10. A device for predicting the relative chlorophyll content based on multispectral imagery, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method for predicting the relative chlorophyll content based on multispectral images as described in any one of claims 1-8.

Images