An information gain driven active assimilation method for rice field ecological processes

Abstract

This invention discloses an information gain-driven active assimilation method for paddy field ecological processes. This method acquires multi-source time-series observation data of the study area and constructs a paddy field ecological process model. It establishes a hierarchical Bayesian constraint network at the pixel, field, and regional levels. Based on posterior uncertainty, crop growth process, platform availability, observation cost, and execution risk, it calculates the expected information gain of candidate observation actions and determines the optimal observation action. Using asynchronous multi-source observation operators, new observations are mapped to the model state space, and state variables and process parameters are jointly updated. The method outputs the ecological state at the field or pixel scale, greenhouse gas emission time series, peak time, cumulative emissions at each stage, stage contribution rate, and uncertainty, and updates the observation value library. This invention transforms fixed-frequency passive observation into information gain-driven closed-loop active assimilation, improving the accuracy of paddy field ecological process reconstruction, stage structure identification ability, and cross-field migration ability under limited observation conditions.

Description

Technical Field

[0001] This invention belongs to the field of agricultural remote sensing, ecological process modeling and data assimilation technology. Specifically, it relates to an information gain-driven active assimilation method for paddy field ecological processes, and more particularly to a closed-loop active assimilation method that integrates satellite remote sensing, UAV remote sensing, ground observation and field management information to achieve dynamic reconstruction of the state of paddy field ecological processes, greenhouse gas emission stage structure and their uncertainties at the field scale or pixel scale. Background Technology

[0002] Paddy field ecological processes are influenced by changes in surface water layer, soil hydrothermal conditions, crop growth processes, field management practices, and biogeochemical reactions, exhibiting significant stage-specific, time-varying, and spatial heterogeneity. Particularly during the tillering, jointing-booting, heading-filling, and maturity stages, changes in water and canopy states alter the soil redox environment, substrate supply, gas production rates, oxidation rates, and plant transport capacity, thereby affecting the timing of peak greenhouse gas emissions, cumulative emissions at each stage, and stage contribution rates.

[0003] Existing technologies utilize satellite remote sensing, ground sensors, UAV observations, or ecological process models to estimate paddy field moisture status, crop growth status, or greenhouse gas emissions. Other methods employ ensemble Kalman filtering, particle filtering, and variational assimilation for model state updates. However, existing solutions still suffer from the following shortcomings: First, observation schemes often employ passive observations at fixed times, locations, and frequencies, lacking the ability to proactively determine high-value observation actions based on model posterior uncertainty. Second, model parameters are often independently calibrated within single plots or sample points, lacking hierarchical constraints and parameter sharing mechanisms between pixel layers, plot layers, and regional layers. Third, multi-source observations are often used only as model input or validation data, failing to form a closed-loop proactive assimilation chain of "candidate observation action generation—information gain evaluation—new observation acquisition—model posterior update—observation value accumulation." Fourth, output results often focus on single-moment status or seasonal cumulative emissions, making it difficult to simultaneously express emission time series, peak times, stage cumulative emissions, stage contribution rates, and their uncertainties.

[0004] Therefore, it is necessary to propose a new active assimilation method for paddy field ecological processes, which integrates information gain-driven observation decision-making, hierarchical Bayesian constraints, asynchronous multi-source observation operators, joint updating of process models, and observation value database to improve the accuracy of paddy field ecological process state reconstruction, greenhouse gas emission stage structure identification, and cross-field migration capabilities under limited observation resources. Summary of the Invention

[0005] The purpose of this invention is to provide an information gain-driven active assimilation method for paddy field ecological processes, in order to solve the problems in the prior art such as fixed observation frequency, static observation location, isolated parameter updates, insufficient coordination between remote sensing and ground observation, weak cross-field knowledge transfer ability, and difficulty in outputting stage structure and uncertain results.

[0006] To address the aforementioned technical problems, this invention provides an information gain-driven active assimilation method for paddy field ecological processes, comprising the following steps:

[0007] S1: Acquire multi-source time-series observation data of the study area and perform time registration, spatial registration, quality control and standardization preprocessing. The multi-source time-series observation data includes at least satellite remote sensing data, UAV remote sensing data, ground observation data, field boundary and field management record data.

[0008] S2: Construct a paddy field ecological process model based on the multi-source time-series observation data. The paddy field ecological process model includes water state variables, crop growth state variables, biogeochemical state variables, greenhouse gas flux state variables and corresponding process parameters, and outputs the greenhouse gas emission time series, emission peak time, cumulative emission amount of each stage and stage contribution rate at the field scale or pixel scale.

[0009] S3: Construct a hierarchical Bayesian constraint network with pixel layer, field layer and region layer to impose cross-scale hierarchical constraints on the process parameters, so that the posterior parameter update of the observed field can be transferred to similar fields and unobserved fields.

[0010] S4: Based on the model's posterior uncertainty at the current moment, the availability of the observation platform, weather conditions, crop growth process, and historical observation value, generate a set of candidate observation actions, where each candidate observation action includes at least the observation time, observation location, observation platform, observation variable type, and sampling intensity;

[0011] S5: For each candidate observation action, calculate its expected information gain on the posterior distribution of the target variable, and conduct a comprehensive evaluation by combining phenological sensitivity weight, observation cost, execution risk and migration gain to determine the optimal observation action;

[0012] S6: Based on the optimal observation action, new observation data is obtained, and an asynchronous multi-source observation operator is constructed to map new observations with different time resolutions, different spatial resolutions, and different error structures to the state space of the paddy field ecological process model;

[0013] S7: Under the hierarchical Bayesian constraint network, the state variables and process parameters of the paddy field ecological process model are jointly and actively assimilated and updated to obtain the posterior state, posterior parameters and posterior uncertainty of the target variable;

[0014] S8: Output the updated state of paddy field ecological processes at the field or pixel scale, greenhouse gas emission time series, emission peak time, cumulative emissions of each stage, stage contribution rate and its uncertainty results.

[0015] S9: Based on the aforementioned uncertainty results, the consistency differences in target variable estimation results under different observation modes, and the posterior uncertainty at the emission peak time point, determine whether to trigger local intensive observation; the trigger threshold for local intensive observation is a dynamic trigger threshold. When the posterior uncertainty index of the target variable group exceeds the uncertainty trigger threshold Or, the difference between the target variable estimation results under different observation modes exceeds the consistency trigger threshold. Or, the posterior uncertainty at the peak emission point exceeds the peak emission point trigger threshold. At that time, local encrypted observation is triggered, and the newly obtained local encrypted observation data is fed back to the asynchronous multi-source observation operator and the joint active assimilation update step;

[0016] S10: Establish and update the observation value library, record the fertility period category, field category, observation platform, observation variable type, environmental background, information gain value, posterior improvement magnitude, observation cost and execution risk corresponding to each round of observation actions, and generate a priori strategies for subsequent active observations based on the observation value library.

[0017] Preferably, in step S1, the multi-source time-series observation data includes at least one of the following: satellite synthetic aperture radar remote sensing data, satellite optical remote sensing data, UAV multispectral imagery, UAV thermal infrared imagery, ground water level observation data, soil hydrothermal observation data, meteorological observation data, greenhouse gas flux verification data, and data on field boundaries, transplanting dates, irrigation and drainage records, fertilization records, variety attributes, and soil attributes. Furthermore, the multi-source time-series observation data undergoes unified time indexing, field boundary clipping, pixel resampling, outlier removal, and observation error calibration. The advantage of this preferred scheme is that it can unify remote sensing observations, ground process observations, and field management records into an assimilable data stream, avoiding interference from inconsistencies in time scales, spatial scales, and error structures between different data sources that could lead to subsequent active assimilation.

[0018] Preferably, in step S2, the state variables of the paddy field ecological process model include water state variables, crop growth state variables, and biogeochemical state variables; wherein, the water state variables include at least one of surface water depth, soil water content, and drainage and seepage status; the crop growth state variables include at least one of leaf area index, canopy cover, biomass, and developmental progress indicators; and the biogeochemical state variables include implicit state quantities related to greenhouse gas production, oxidation, transport, and release processes. The advantage of this preferred scheme is that it incorporates paddy field water processes, crop growth processes, and greenhouse gas release processes into the same state space, enabling the model to express the process coupling relationship between water, crops, and gases, rather than being limited to single flux estimation.

[0019] Preferably, in step S2, the target variable set output by the paddy field ecological process model includes the time series of greenhouse gas emissions. Peak emission time , No. Cumulative emissions during each reproductive stage and contribution rate The advantage of this preferred approach is that it expands greenhouse gas emissions results from seasonal cumulative totals to dynamic indicators with phased structural significance, thereby more accurately identifying critical emission windows and major contributing periods.

[0020] Preferably, in step S3, the hierarchical Bayesian constraint network includes a pixel layer, a field layer, and a region layer; the parameters of the pixel layer are constrained by the parameters of the field layer, and the parameters of the field layer are constrained by the parameters of the region layer. The parameter sharing weights are adaptively adjusted based on the similarity of soil properties, management systems, variety attributes, and growth processes among the fields. The advantage of this preferred scheme is that it can transfer posterior parameter information obtained from observed fields to insufficiently observed fields, improving the stability of model parameter updates and cross-field extrapolation ability under sparse observation conditions.

[0021] Preferably, in step S4, each candidate observation action in the candidate observation action set includes one or more of the following: observation time, observation location, observation platform, observation variable type, sampling density, sampling depth, and number of repetitions; the observation platform includes at least one of satellite remote sensing platform, UAV remote sensing platform, and ground observation platform; and the candidate observation actions are screened for feasibility based on weather conditions, platform accessibility, crop growth process, and field management events. The advantage of this preferred scheme is that it uniformly expresses when to observe, where to observe, what platform to use, and what variables to observe as calculable observation actions, providing a clear optimization target for proactive observation decision-making.

[0022] Preferably, in step S5, the expected information gain is characterized by the divergence, posterior entropy reduction, or posterior variance contraction between the current posterior distribution of the target variable group and the predicted posterior distribution after executing the candidate observation action. The comprehensive utility of the candidate observation action is calculated by combining phenological sensitivity weights, cross-scale migration gain, observation cost, cloud cover risk, platform accessibility, and execution failure risk. The advantage of this preferred scheme is that it prioritizes the observation actions that have the most constraining value for key target variables such as emission peak time, cumulative emissions over a period, and contribution rate over a period, thereby achieving efficient allocation of limited observation resources.

[0023] Preferably, in step S5, the phenological sensitivity weights are dynamically adjusted according to the crop growth process. Specifically, during the tillering stage, the observation weights for water status and changes in field surface water layer are increased; during the jointing and booting stage, the observation weights for canopy structure and crop growth status are increased; and during the heading and grain-filling stage, the observation weights for peak emission time, cumulative emissions at each stage, and stage contribution rate are increased. The advantage of this preferred scheme is that it allows the active observation strategy to be adjusted according to the growth stage, avoiding the waste of high-cost observation resources during non-sensitive periods, and enhancing the ability to identify key emission stages.

[0024] Preferably, in step S6, the asynchronous multi-source observation operator includes a radar observation operator, an optical observation operator, a thermal infrared observation operator, and a ground observation operator. The radar observation operator maps the field surface water layer state and canopy structure state to radar backscattering observations. The optical observation operator maps the canopy state to vegetation indices, canopy coverage, or leaf area index surrogates. The thermal infrared observation operator maps the hydrothermal state to a canopy temperature field. The ground observation operator maps the local state to water level, soil hydrothermal, and greenhouse gas flux verification observations. The advantage of this preferred scheme is that it can uniformly map observations with different temporal resolutions, spatial resolutions, and error structures to the model state space, improving the collaborative constraint capability among satellite, UAV, and ground observations.

[0025] Preferably, in step S7, the joint active assimilation update is implemented using ensemble Kalman filtering, particle filtering, sequential Monte Carlo methods, variational assimilation methods, or a combination of the above methods with hierarchical Bayesian inference, and the state variables and process parameters of the paddy field ecological process model are updated simultaneously. The advantage of this preferred scheme is that it avoids structural biases caused by only updating state variables while keeping parameters fixed, allowing the model to simultaneously correct for uncertainties in process state and parameters after continuous observations.

[0026] Preferably, in step S8, the uncertainty results include at least one of the posterior mean, posterior variance, confidence interval, or spatial uncertainty distribution map of the target variable group; the output results include the state sequence of paddy field ecological processes at the field scale or pixel scale, greenhouse gas emission time series, emission peak time, cumulative emissions of each stage, stage contribution rate, and their uncertainty results. The advantage of this preferred scheme is that it can simultaneously provide emission results and confidence level expressions, providing clear uncertainty basis for subsequent observation refinement, model diagnosis, and field management decisions.

[0027] Preferably, in step S9, the local encryption observation trigger threshold is a dynamic trigger threshold. The dynamic trigger threshold Including uncertainty trigger threshold Consistency trigger threshold and peak time trigger threshold When the posterior uncertainty index of the target variable group exceeds Or the difference between the estimation results of the target variable under different observation modes exceeds Or the posterior uncertainty at the peak emission point exceeds At that time, localized intensive observation is triggered; the localized intensive observation includes at least one of the following: UAV multispectral supplementary observation, UAV thermal infrared supplementary observation, ground water level intensive sampling, soil hydrothermal state supplementary observation, and greenhouse gas flux verification supplementary observation. The advantage of this preferred scheme is that it can concentrate high-cost observation resources on fields and pixels with the highest posterior uncertainty or the greatest differences in multi-source observations, realizing the transformation from uniform supplementary observation across the entire area to localized high-value supplementary observation.

[0028] Preferably, in step S9, the new observation data obtained from the locally encrypted observation is fed back to the asynchronous multi-source observation operator and the joint active assimilation update step, and the state variables, process parameters, and posterior distributions of the target variable group for the corresponding field or pixel are updated again. The advantage of this preferred scheme is that it enables the present invention to form a closed-loop mechanism of "posterior diagnosis—locally encrypted observation—multi-source observation mapping—re-assimilation update," which differs from the passive update method that only performs a one-time data assimilation.

[0029] Preferably, in step S10, the observation value database records the fertility period category, field category, observation platform, observation variable type, environmental background, information gain value, reduction in posterior uncertainty of the target variable group, reduction in target variable estimation error, observation cost, and execution risk corresponding to each round of observation actions; and calculates the observation value score based on the information gain value, reduction in posterior uncertainty of the target variable group, reduction in target variable estimation error, observation cost, and execution risk. The advantage of this preferred scheme is that it can quantify and preserve the actual contribution of each round of observation actions, making observation behaviors no longer isolated events, but evaluable, comparable, and accumulative strategy samples.

[0030] Preferably, in step S10, a priori strategy for subsequent active observations is generated based on the observation value database. This priori strategy is used to assign prior selection weights to candidate observation actions in subsequent assimilation cycles, subsequent growth periods, or similar fields, and is used in conjunction with the information gain evaluation results of the current round to determine the optimal observation action. The advantage of this preferred scheme is that it enables the transfer of observation experience across rounds, growth periods, and fields, giving the active assimilation system continuous learning and regional adaptability.

[0031] Compared with related technologies, the information gain-driven active assimilation method for paddy field ecological processes provided by this invention has the following beneficial effects:

[0032] (1) This invention integrates information gain-driven observation action optimization, hierarchical Bayesian cross-scale parameter constraints, asynchronous multi-source observation operators, joint active assimilation update, local densification observation triggering and observation value library learning mechanism into the same closed-loop framework, realizing the transformation of paddy field ecological process observation from passive observation with fixed frequency and fixed location to active observation based on posterior uncertainty and expected information gain.

[0033] (2) The present invention can prioritize the constraint of key target variables such as emission peak time, cumulative emissions of each stage and stage contribution rate under limited observation budget conditions, thereby improving the reconstruction accuracy of paddy field ecological processes and greenhouse gas emission stage structure at the field scale or pixel scale.

[0034] (3) This invention achieves the migration of observed field information to insufficiently observed fields and the reuse of single-season observation experience to subsequent seasons through the hierarchical Bayesian constraints of the pixel layer, field layer and region layer and the continuous updating of the observation value library, thereby improving the stability, transferability and regional application capability of the model. Attached Figure Description

[0035] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention.

[0036] Figure 2A schematic diagram showing the integrated rainfall-irrigation process of field surface water depth, LAI, CH4 flux and rainfall during the growing season;

[0037] Figure 3 Flowchart for optimizing active observation actions driven by information gain;

[0038] Figure 4 A schematic diagram of the coupling between a hierarchical Bayesian constraint network and an asynchronous multi-source observation operator;

[0039] Figure 5 A schematic diagram comparing the uncertainty and stage structure of key target variables before and after active assimilation; (a) contraction of uncertainty of key target variables; (b) CH4 flux and uncertainty band in representative fields; (c) schematic diagram of stage contribution rate structure;

[0040] Figure 6 A schematic diagram of information gain-driven local encrypted observation triggering and observation value library update; (a) trigger utility value and dynamic threshold; (b) posterior uncertainty and expected information gain evolution; (c) active observation action execution timeline. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] This embodiment uses paddy fields in the experimental area of ​​Taihe County, Ji'an City, Jiangxi Province as the research object, and selects representative continuous paddy fields as research units. These fields have relatively complete irrigation and drainage conditions and field management records. The main rice growth processes include the tillering stage, jointing and booting stage, heading and grain-filling stage, and maturity stage. To achieve joint representation at the field scale and pixel scale, each field is divided into regular pixels with a spatial resolution of 10m × 10m. The model time step is set to 1 day, and the monitoring period covers the complete growth cycle of one rice season.

[0043] To avoid symbol confusion, the following symbols are used uniformly in this embodiment: Indicates the cell number. Indicates the field number. Indicates the order of days, Indicates the reproductive period number. Represents the observation round; the state variable is denoted as The process parameters are denoted as External drive is denoted as The observation data is recorded as The observation action is recorded as The target variable set is denoted as The target variable set includes time-series greenhouse gas emissions. Peak emission time , No. Cumulative emissions during each reproductive stage and stage contribution rate .

[0044] 1. Acquisition and preprocessing of multi-source time-series observation data

[0045] In this embodiment, multi-source time-series observation data of the study area are acquired and preprocessed by time registration, spatial registration, and standardization. The multi-source time-series observation data includes satellite synthetic aperture radar remote sensing data, satellite optical remote sensing data, UAV multispectral imagery, UAV thermal infrared imagery, ground water level observation data, soil hydrothermal observation data, meteorological observation data, greenhouse gas flux verification data, as well as data on field boundaries, transplanting dates, irrigation and drainage records, fertilization records, variety attributes, and soil attributes.

[0046] Among them, satellite synthetic aperture radar remote sensing data is used to obtain information on the state of the field water layer and changes in canopy scattering; satellite optical remote sensing data is used to obtain vegetation index, canopy coverage, and leaf area index proxy; UAV multispectral imagery is used to supplement high spatial resolution canopy information within key growth windows; UAV thermal infrared imagery is used to identify water and heat stress and local moisture anomalies; ground water level observation data is used to constrain the depth of the field water layer; soil water and heat observation data is used to constrain soil moisture and thermal states; and greenhouse gas flux verification data is used to verify the timing and peak time of greenhouse gas emissions.

[0047] The above data underwent unified preprocessing, primarily including projecting remote sensing images onto a unified coordinate system, cropping them according to field boundaries, resampling the remote sensing data to a 10m×10m pixel scale, aggregating ground observation data by daily scale, removing obvious outliers, and establishing the error variances for different observation variables. All observation data were then uniformly organized as follows:

[0048]

[0049] in, Indicates the radar observation sequence. Represents an optical observation sequence. This represents a multispectral observation sequence from a UAV. Indicates a thermal infrared observation sequence. This represents the observation sequence of surface water level and soil hydrothermal data. This represents the greenhouse gas flux verification observation sequence.

[0050] Through the above processing, observation data from different platforms, time scales, and spatial scales can be uniformly transformed into a standardized input data stream that can be invoked by the active assimilation system.

[0051] 2. Construction of a Paddy Field Ecological Process Model

[0052] Based on the aforementioned multi-source time-series observation data, a paddy field ecological process model was constructed. For the first... The first of the plots The pixel, in the... The state variables for a given day are defined as follows:

[0053]

[0054] in, Indicates the depth of the water layer on the field surface. Indicates the soil moisture content. Indicates drainage leakage status. Indicates leaf area index or canopy structure status. Indicates the state of crop biomass. Indicators representing crop development progress This represents the biogeochemical latent states associated with the production, oxidation, transport, and release of greenhouse gases. This represents the greenhouse gas emission flux.

[0055] The process parameter vector is defined as follows:

[0056]

[0057] in, Indicates parameters of the moisture process. Indicates parameters of crop growth process, Represents greenhouse gas process parameters, These represent the model error and observation error parameters.

[0058] The state evolution relationship of the paddy field ecological process model is represented as follows:

[0059]

[0060] in, This represents a model of rice paddy ecological processes. External drivers include weather conditions, irrigation and drainage measures, fertilization measures, and field management events. This indicates process error.

[0061] To demonstrate the invention's ability to reconstruct the structure of greenhouse gas emission stages, a further set of target variables is defined:

[0062]

[0063] in,

[0064]

[0065]

[0066]

[0067] in, Indicates the first The time window corresponding to each reproductive period Indicates the total number of reproductive years. This indicates the model's time step.

[0068] In this embodiment, the tillering stage, jointing and booting stage, heading and grain-filling stage, and maturity stage can be used as calculation windows for the cumulative emissions and contribution rate of each stage. Therefore, the model output is no longer just the seasonal cumulative emissions, but a dynamic result including the emission process, peak time points, and stage contribution structure.

[0069] 3. Construction of Hierarchical Bayesian Constraint Networks

[0070] To achieve information transfer from observed fields to under-observed fields, a hierarchical Bayesian constrained network is constructed, consisting of a pixel layer, a field layer, and a region layer. Let the pixel layer parameters be... Field layer parameters are The regional layer parameters are Then the following hierarchical relationship is established:

[0071]

[0072]

[0073]

[0074] in, , and These represent the covariance matrices of parameters for the cell layer, field layer, and region layer, respectively. This represents the prior mean of the region.

[0075] Furthermore, to distinguish the degree of transferability between different fields, a similarity weight between fields is defined:

[0076]

[0077] in, Indicates the first The plot of land and the first Similarity weights between individual plots Indicates similarity in soil properties. Indicating similarities in water management systems, Indicates similarity in varietal attributes. Indicating similarity in reproductive processes, , , and These are the normalized weighting coefficients.

[0078] After obtaining new posterior parameters for a particular field, it can be based on The parameter updates for this field are propagated to similar fields. This process enables under-observed fields to also receive parameter constraints from similar fields, thereby improving stability when applied at a regional scale.

[0079] 4. Generation of candidate observation action set

[0080] In each assimilation cycle, a set of candidate observation actions is generated based on the current model posterior uncertainty, observation platform availability, weather conditions, crop growth process, and field management events:

[0081]

[0082] Each candidate observation action is defined as:

[0083]

[0084] in, Indicates the candidate observation time. Indicates candidate observation locations, Indicates the observation platform. Indicates the type of observed variable. Indicates sampling density, Indicates the sampling depth. Indicates the number of repetitions.

[0085] For example, during the tillering and greening stage, priority candidate observation actions include field surface water depth, water level changes, and radar backscattering observations; during the jointing and booting stage, priority candidate observation actions include leaf area index, canopy coverage, and crop growth status observations; and during the heading and grain-filling stage, priority candidate observation actions include thermal infrared hydrothermal status, greenhouse gas flux verification, and observations related to peak emission times.

[0086] When implementing the experiment in Taihe County, the following candidate actions were set: perform UAV multispectral supplementary measurements on fields with high post-test uncertainty; perform ground water level verification on fields where radar observations and ground water level observations are inconsistent; perform additional greenhouse gas flux measurements on fields with high uncertainty at the peak emission time during the heading and grain-filling stages; and perform supplementary soil hydrothermal state measurements on areas with abnormal thermal infrared temperatures.

[0087] Therefore, the observation action is no longer a passive sampling at fixed time and fixed location, but is dynamically generated based on the model state and the uncertainty of the target variable.

[0088] 5. Optimal Observation Actions Driven by Information Gain

[0089] For each candidate observation action, calculate its expected information gain with respect to the posterior distribution of the target variable group. Its expected information gain is defined as:

[0090]

[0091] in, Indicates the execution of candidate actions New observations that may be obtained later Indicates the Kullback-Leibler divergence. This represents the posterior distribution of the current target variable group. This represents the predicted posterior distribution after assuming the candidate action is performed.

[0092] Furthermore, taking into account phenological sensitivity weights, cross-scale migration gains, observation costs, and execution risks, the comprehensive utility of candidate observation actions is defined as follows:

[0093]

[0094] in, Indicates the first Phenological sensitivity weights for each reproductive period Indicates candidate observation actions The cross-scale migration gain is used to characterize the value of the information obtained by the observation action in the current field for the migration constraint value of similar fields or similar pixels. Indicates the observation cost of candidate actions. This indicates the execution risk of the candidate action. and These are the weighting coefficients.

[0095] It can be represented as:

[0096]

[0097] in, Indicates the relationship with the first A set of fields that are similar to each other. Indicates the first The plot of land and the first Similarity weights between individual plots To prevent constants with a denominator of zero, Indicates candidate observation actions For the The expected reduction in posterior uncertainty for a group of similar field target variables.

[0098]

[0099] in, Indicates the execution of candidate observation actions The previous chapter Posterior uncertainty index of the target variable set for each field. This indicates the assumption that candidate observation actions will be performed. And through hierarchical Bayes constraints to the first The posterior uncertainty index of the target variable set obtained after propagation to individual fields. This indicates the expectation for the candidate observation results.

[0100] The posterior uncertainty index of the target variable group can be represented by the trace of the posterior covariance matrix, i.e.:

[0101]

[0102]

[0103] in, Indicates the first The field in the The posterior covariance matrix of the target variable group for the day. This represents the matrix trace operation. Therefore, The larger the value, the more likely it is to indicate a candidate observation action. It can not only constrain the current field, but also reduce uncertainty for more similar fields through field similarity relationships.

[0104] Finally, the candidate observation action with the highest overall utility is selected as the optimal observation action for the current round:

[0105]

[0106] In this embodiment, if a certain field is in the heading and grain-filling stage... If the posterior variance is large, and both UAV thermal infrared supplementation and flux measurement can be performed, then the system compares the two types of actions. , and The expected constraint capability is considered, while simultaneously incorporating cross-scale migration gains and deducting costs and execution risks, ultimately selecting the action with the highest overall utility. This avoids averaging measurements across all fields, instead prioritizing observations most valuable to the target variable.

[0107] 6. Construction of Asynchronous Multi-Source Observation Operator

[0108] After acquiring new observation data based on the optimal observation action, an asynchronous multi-source observation operator is constructed to map observations from different platforms to the state space of the paddy field ecological process model. The unified observation equation is expressed as:

[0109]

[0110] in, Indicates the observation mode. Indicates the first Observational operators, This represents the corresponding observation error.

[0111] Specifically, the radar observation operator is expressed as:

[0112]

[0113] It is used to map the state of the field water layer and the state of the canopy structure into radar backscatter observations.

[0114] The optical observation operator is represented as:

[0115]

[0116] Used to map canopy structure and biomass status to vegetation indices, canopy coverage, or leaf area index proxies.

[0117] The thermal infrared observation operator is represented as:

[0118]

[0119] Used to map hydrothermal states to a canopy temperature field.

[0120] The ground observation operator is represented as:

[0121]

[0122] Used to map local model states to verification observations of water level, soil hydrothermal energy, and greenhouse gas fluxes.

[0123] For observations arriving at different times, sequential updates are used; for observations with different spatial resolutions, pixel mapping, field aggregation, or local weighted projection are used; for observations with different error structures, the observation error covariance matrix is ​​used for weighting. Thus, satellite, UAV, and ground-based observations can all enter a unified assimilation state space.

[0124] 7. Joint active assimilation update of state variables and process parameters

[0125] Under the hierarchical Bayesian constraint network, the state variables and process parameters of the paddy field ecological process model are jointly and actively assimilated and updated. Its posterior distribution is expressed as:

[0126]

[0127] in, Represents the observation likelihood term. Represents a state transition term. This represents the hierarchical Bayesian prior.

[0128] In this embodiment, a combination of sequential Monte Carlo method and hierarchical Bayesian inference is used for updating. For each particle or set member, a state variable is also carried. and process parameters After new observations are received, the particle weights are updated based on the observation likelihood, and the parameter distributions of the field layer and the region layer are adjusted according to the hierarchical Bayesian constraints.

[0129] After the update, the posterior means of the state variables and process parameters are obtained:

[0130]

[0131]

[0132] And the posterior covariance of the target variable group:

[0133]

[0134] This step enables the synchronous updating of state variables and process parameters, avoiding process deviations caused by updating only the state while keeping the parameters fixed.

[0135] 8. The structural output of the ecological process state and greenhouse gas emission stages in paddy fields

[0136] After completing the joint active assimilation update, the system outputs the state of paddy field ecological processes, greenhouse gas emission time series, emission peak time, cumulative emissions of each stage, stage contribution rate and its uncertainty results at the field or pixel scale.

[0137] Specifically, this includes (1) the state sequence of paddy field ecological processes, including the depth of the water layer on the field surface, soil moisture content, drainage and seepage status, canopy structure status, biomass status, and biogeochemical implicit status; and (2) the time series of greenhouse gas emissions. (3) Peak emission time (4) Cumulative emissions during each reproductive stage (5) Contribution rate of each reproductive stage (6) Posterior variance, confidence interval and spatial uncertainty distribution of the target variable group.

[0138] In the pilot program in Taihe County, the differences in emission contributions across different fields during the tillering, jointing and booting, heading and grain-filling, and maturity stages were observed. For example, some low-lying fields may show higher methane emission contributions during the tillering and jointing and booting stages, while locally dried-up-re-watered fields may experience an earlier peak emission point or increased uncertainty in the peak emission after re-watering. By analyzing the cumulative emissions and stage contribution rates at each stage, major emission contribution windows and high-risk fields can be identified.

[0139] 9. Local Encryption Observation Triggering and Backflow Update

[0140] After outputting the uncertainty results, based on the posterior uncertainty of the target variable group, the consistency difference of the target variable estimation results under different observation modes, and the posterior uncertainty of the emission peak time point, it is determined whether to trigger local intensified observation.

[0141] To characterize the dynamic adjustment of the local intensified observation trigger threshold as the crop growth stage, historical observation value, and observation conditions change, this embodiment defines a dynamic trigger threshold. . This can be represented as a threshold triggered by uncertainty. Consistency trigger threshold and peak time trigger threshold The dynamic threshold set is used to adaptively determine whether local intensive observation needs to be initiated under different growth stages and different field types.

[0142]

[0143] in, Used to determine whether the posterior uncertainty of the target variable group is too high. This is used to determine whether there are significant differences in the estimation results of the target variable under different observation modes. This is used to determine whether the posterior uncertainty at the peak emission point is too high. The dynamic trigger threshold... The system can be adaptively adjusted based on crop growth stage, target variable sensitivity, observation platform availability, weather conditions, field management events, and prior strategy weights in the observation value pool. For example, during periods when emission peaks are likely to occur, such as the heading and grain-filling stages or after drying and rehydration, the emission level can be appropriately reduced. To improve the sensitivity to abnormal changes in the timing of emission peaks; when the UAV platform is unavailable or the risk of rainfall is high, the trigger threshold related to UAV supplementary measurements can be appropriately increased to avoid unexecutable or unreliable encrypted observations.

[0144] For the The first pixel or the For each field, calculate the posterior covariance of the target variable group:

[0145]

[0146] When the posterior uncertainty of the target variable set satisfies:

[0147]

[0148] When this occurs, it is determined that the pixel or field has a high posterior uncertainty.

[0149] Furthermore, the consistency differences in the target variable estimation results under different observation models are calculated:

[0150]

[0151] in, and They represent the first Class and First The target variable estimation results obtained from observation-like pattern constraints. When the following conditions are met:

[0152]

[0153] At that time, it was determined that there were significant differences between different observation modes.

[0154] In addition, for the peak emission time point, a trigger condition for uncertainty at the peak time point is set:

[0155]

[0156] When any of the above conditions are met, local encrypted observation is triggered, and a set of local candidate observation actions is generated:

[0157]

[0158] Localized intensive observation includes at least one of the following: UAV multispectral supplementary measurement, UAV thermal infrared supplementary measurement, intensive sampling of ground water level, supplementary measurement of soil hydrothermal status, and additional measurement for greenhouse gas flux verification.

[0159] Calculate the local comprehensive utility for each local candidate observation action:

[0160]

[0161] And select:

[0162]

[0163] This is a locally optimal encrypted observation action.

[0164] In this embodiment, if a field shows inconsistencies between radar-retrieved water levels and surface water level records during the heading and grain-filling stages, and Exceeding the consistency trigger threshold The system will prioritize triggering ground water level verification or UAV thermal infrared supplementary measurement; if a field experiences unstable peak discharge times after drying and re-watering, and Exceeding the peak trigger threshold This triggers additional greenhouse gas flux verification measurements. The newly acquired locally encrypted observation data is fed back to S6 and S7, and asynchronous observation mapping and joint active assimilation updates are re-executed.

[0165] Through this step, the system forms a closed-loop mechanism of "posterior diagnosis - dynamic threshold judgment - local densification observation - observation operator mapping - re-assimilation update", which enables limited observation resources to be concentrated on the fields, pixels and fertility windows that need the most constraints.

[0166] 10. Observation value base update and prior strategy generation

[0167] The observation value base is used to record each round of observation actions and their contribution to the posterior improvement of the model. For the ... Rotate observation actions to construct observation value record items:

[0168]

[0169] in, Indicates the first Wheel observation action, Indicates the corresponding reproductive period category. Indicates the type of field. Indicates the observation platform. Indicates the type of observed variable. Indicates environmental background label, This represents the information gain value of the observation action in that round. This represents the reduction in posterior uncertainty of the target variable group. This indicates the reduction in the estimation error of the target variable. Indicates the cost of observation. Indicates execution risk, This indicates the score for the observation value.

[0170] The reduction in posterior uncertainty can be expressed as:

[0171]

[0172] The reduction in the estimation error of the target variable can be expressed as:

[0173]

[0174] in, and Let these represent the posterior covariance matrices of the target variable group before and after the observation action, respectively. and These represent the estimation errors of the target variable before and after performing the observation action, respectively.

[0175] The observation value score is defined as follows:

[0176]

[0177] in, , , and These are the weighting coefficients. To prevent constants with a denominator of zero.

[0178] Furthermore, the observation value records are archived according to reproductive period category, field category, observation platform, observation variable type, and environmental background label. For subsequent assimilation cycles, prior strategy weights for candidate observation actions are generated based on the observation value database.

[0179]

[0180] in, Indicates the category of a given reproductive period Field categories Types of observed variables and environmental background Under the conditions, candidate observation actions The prior choice probability.

[0181] In subsequent active observation and decision-making, the prior policy weights are incorporated into the comprehensive utility function:

[0182]

[0183] in, This represents the weighting coefficient of the prior strategy. In this way, historical observation experience and cross-field migration value can jointly participate in the optimization of subsequent observation actions.

[0184] After multiple assimilation cycles in the Taihe County experimental area, the observation value database can gradually identify high-value observation combinations under different growth stages and field types. For example, during the tillering stage, the combination of intensive surface water level sampling and radar observation significantly contributes to reducing the uncertainty of water status; during the jointing and heading stage, UAV multispectral supplementary measurements provide strong constraints on canopy structure and biomass status; and during the heading and grain-filling stage, thermal infrared supplementary measurements and flux verification measurements significantly contribute to reducing the uncertainty of emission peak timing and stage contribution rate. Thus, this invention can transform single-cycle observation experience into a reusable active observation strategy that spans multiple cycles, growth stages, and field types.

[0185] 11. Implementation Results

[0186] Using the method of this embodiment, under the same observation budget constraint, the system can prioritize the allocation of observation resources to fields with high uncertainty, high-contribution fertility windows, and regions with inconsistent multi-source observation results based on the posterior uncertainty of the target variable and the expected information gain. Compared with passive observation methods with fixed frequency and fixed location, this invention can reduce low-value repeated observations and improve the ability to constrain emission peak times, cumulative emissions in stages, and stage contribution rates.

[0187] Meanwhile, by leveraging hierarchical Bayesian constraints at the pixel, field, and region levels, posterior parameters obtained from fully observed fields can be transferred to similar fields, improving the simulation stability of under-observed fields. Asynchronous multi-source observation operators enable satellite remote sensing, UAV remote sensing, and ground observations to be uniformly incorporated into the model's state space. A localized dense observation triggering mechanism allows the system to promptly supplement high-value observations at critical stages. Furthermore, an observation value database transforms historical observation experience into prior strategies for subsequent proactive observations.

[0188] Therefore, this invention can realize the transformation of the paddy field ecological process model from "fixed observation - passive update" to a closed loop of "information gain decision - active observation - hierarchical update - local densification - experience learning", which is applicable to the dynamic reconstruction of paddy field water status, crop growth status, biogeochemical processes and greenhouse gas emission stage structure at the field scale or pixel scale.

[0189] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An information gain-driven active assimilation method for paddy field ecological processes, characterized in that, Includes the following steps: S1: Acquire multi-source time-series observation data of the study area and perform time registration, spatial registration, quality control and standardization preprocessing. The multi-source time-series observation data includes at least satellite remote sensing data, UAV remote sensing data, ground observation data, field boundary and field management record data. S2: Construct a paddy field ecological process model based on the multi-source time-series observation data. The paddy field ecological process model includes water state variables, crop growth state variables, biogeochemical state variables, greenhouse gas flux state variables and corresponding process parameters, and outputs the greenhouse gas emission time series, emission peak time, cumulative emission amount of each stage and stage contribution rate at the field scale or pixel scale. S3: Construct a hierarchical Bayesian constraint network with pixel layer, field layer and region layer to impose cross-scale hierarchical constraints on the process parameters, so that the posterior parameter update of the observed field can be transferred to similar fields and unobserved fields. S4: Based on the model's posterior uncertainty at the current moment, the availability of the observation platform, weather conditions, crop growth process, and historical observation value, generate a set of candidate observation actions, where each candidate observation action includes at least the observation time, observation location, observation platform, observation variable type, and sampling intensity; S5: For each candidate observation action, calculate its expected information gain on the posterior distribution of the target variable, and conduct a comprehensive evaluation by combining phenological sensitivity weight, observation cost, execution risk and migration gain to determine the optimal observation action; S6: Based on the optimal observation action, new observation data is obtained, and an asynchronous multi-source observation operator is constructed to map new observations with different time resolutions, different spatial resolutions, and different error structures to the state space of the paddy field ecological process model; S7: Under the hierarchical Bayesian constraint network, the state variables and process parameters of the paddy field ecological process model are jointly and actively assimilated and updated to obtain the posterior state, posterior parameters and posterior uncertainty of the target variable; S8: Output the updated state of paddy field ecological processes at the field or pixel scale, greenhouse gas emission time series, emission peak time, cumulative emissions of each stage, stage contribution rate and its uncertainty results. S9: Based on the aforementioned uncertainty results, the consistency differences in target variable estimation results under different observation modes, and the posterior uncertainty at the emission peak time point, determine whether to trigger local intensive observation; the trigger threshold for local intensive observation is a dynamic trigger threshold. When the posterior uncertainty index of the target variable group exceeds the uncertainty trigger threshold Or, the difference between the target variable estimation results under different observation modes exceeds the consistency trigger threshold. Or, the posterior uncertainty at the peak emission point exceeds the peak emission point trigger threshold. At that time, local encrypted observation is triggered, and the newly obtained local encrypted observation data is fed back to the asynchronous multi-source observation operator and the joint active assimilation update step; S10: Establish and update the observation value library, record the fertility period category, field category, observation platform, observation variable type, environmental background, information gain value, posterior improvement magnitude, observation cost and execution risk corresponding to each round of observation actions, and generate a priori strategies for subsequent active observations based on the observation value library.

2. The information gain-driven active assimilation method for paddy field ecological processes according to claim 1, characterized in that, In step S1, the satellite remote sensing data includes at least one of synthetic aperture radar remote sensing data and optical remote sensing data; the UAV remote sensing data includes at least one of multispectral imagery and thermal infrared imagery; and the ground observation data includes at least one of field surface water level observation data, soil water and heat observation data, meteorological observation data, and greenhouse gas flux verification data.

3. The information gain-driven active assimilation method for paddy field ecological processes according to claim 1, characterized in that, In S2, the water state variables include at least one of field surface water depth, soil water content, and drainage and seepage status; the crop growth state variables include at least one of leaf area index, canopy coverage, biomass, and developmental progress indicators; and the biogeochemical state variables include implicit state quantities related to the production, oxidation, transport, and release processes of methane, nitrous oxide, or carbon dioxide.

4. The information gain-driven active assimilation method for paddy field ecological processes according to claim 1, characterized in that, In S3, the pixel layer parameters are subject to the field layer parameter constraints, the field layer parameters are subject to the regional layer parameter constraints, and the parameter sharing weights are adaptively adjusted according to the similarity of soil properties, field management system, variety attributes and growth process between fields.

5. The information gain-driven active assimilation method for paddy field ecological processes according to claim 1, characterized in that, In S4, the candidate observation action also includes at least one of sampling density, number of repetitions or sampling depth, and different candidate observation priorities are set according to the sensitivity of the target variable at the tillering stage, jointing and booting stage, heading and grain filling stage and maturity stage.

6. The information gain-driven active assimilation method for paddy field ecological processes according to claim 1, characterized in that, In S5, the expected information gain is characterized by the Kullback-Leibler divergence, posterior entropy reduction, posterior variance contraction, or target variable confidence interval narrowing between the current posterior distribution of the target variable and the posterior distribution after the assumption of performing candidate observation actions. The target variable includes at least one of the following: peak greenhouse gas emission time point, cumulative emissions of a stage, stage contribution rate, and emission time series.

7. The information gain-driven active assimilation method for paddy field ecological processes according to claim 1, characterized in that, In S5, the comprehensive evaluation simultaneously considers phenological sensitivity weight, observation cost, cloud cover risk, rainfall risk, platform accessibility, execution failure risk, and cross-field migration gain, so as to prioritize the allocation of high-value observation resources to fields or pixels with high posterior uncertainty and high migration value.

8. The information gain-driven active assimilation method for paddy field ecological processes according to claim 1, characterized in that, In S6, the asynchronous multi-source observation operator includes a radar observation operator that maps the field surface water layer state and canopy structure state to radar backscattering observations, an optical observation operator that maps the canopy state to vegetation index or leaf area index proxy quantities, a thermal infrared observation operator that maps the hydrothermal state to the canopy temperature field, and a ground observation operator that maps the local state to field surface water level, soil hydrothermal or flux verification values.

9. The information gain-driven active assimilation method for paddy field ecological processes according to claim 1, characterized in that, In S9, the local encryption observation trigger threshold is a dynamic trigger threshold. The dynamic trigger threshold Including uncertainty trigger threshold Consistency trigger threshold and peak time trigger threshold ; When the posterior uncertainty index of the target variable group exceeds Or the difference between the posterior estimates of the target variable obtained from different observation models exceeds Or the posterior uncertainty at the peak emission point exceeds When this occurs, local encrypted observation is triggered; the local encrypted observation includes at least one of the following: UAV supplementary measurement, ground water level encrypted sampling, soil hydrothermal state supplementary measurement, and flux verification additional measurement.

10. The information gain-driven active assimilation method for paddy field ecological processes according to claim 1, characterized in that, In S10, the fertility period category, field category, observation platform, observation variable type, environmental background, information gain value, posterior improvement magnitude, observation cost, and execution risk corresponding to each round of observation actions are recorded, and a priori strategy for subsequent active observation is generated based on the observation value library.

Images