A method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage based on underlying surface type.
By using a method based on the waveform fingerprint of receding water based on underlying surface type, and utilizing flow and pollutant concentration monitoring data, combined with a response kernel library and a joint inversion objective function, the problem of tracing the source of underlying surface type load in existing technologies has been solved, achieving efficient, low-cost, and reliable source tracing of non-point source pollution events.
Patent Information
- Application Number
- CN202611323779.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-28
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies cannot quantitatively decompose the load of a pollution event into specific underlying surface categories using cross-sectional flow and pollutant concentration monitoring data without additional chemical testing. Furthermore, the decomposition results lack physical evidence and cannot accurately trace the source of non-point source pollution.
An agricultural non-point source pollution event tracing method based on the waveform fingerprint of receding water based on underlying surface type is adopted. By acquiring the flow time series and pollutant concentration time series, combined with the pre-established underlying surface type response kernel library, the contribution ratio of each underlying surface is determined by using a correction function and a joint inversion objective function.
It enables the tracing of underlying surface load categories based solely on routine monitoring data, shortening detection time, reducing costs, and providing decomposition results with traceable mechanisms. It avoids physically unreasonable decomposition results and provides reliable estimates of contribution ratios and spatial distribution.
Smart Images

Figure CN123118329A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water environment pollution source tracing technology, specifically to a method for tracing agricultural non-point source pollution events based on the waveform fingerprint of receding water under underlying surface type. Background Technology
[0002] Non-point source pollution in watersheds is typically contributed by a combination of underlying surface types, including paddy fields, dry land, forests and grasslands, hardened surfaces, and retention areas such as ditches and wetlands. These underlying surface types have fundamentally different water storage mechanisms and outflow paths, and the recession response patterns of the same rainfall event differ on different underlying surfaces. Identifying the load contribution ratio of various underlying surfaces in a pollution event is of practical significance for accurate source tracing and zoned prevention and control of non-point source pollution.
[0003] Existing methods for tracing pollution loads mainly fall into two categories. The first is chemical isotope or compositional fingerprinting, which determines the source category by detecting the chemical composition or isotopic ratio of pollutants. This method requires multi-site, multi-time-period chemical testing, resulting in long testing cycles, high costs, and limitations in the number of sampling points, making it difficult to perform continuous and precise load decomposition for a single drainage process. The second category is drainage curve analysis in hydrology. This method decomposes the total runoff into contributions from different underlying surface types or different runoff generation mechanisms based on the flow decay law during the drainage process. However, this method only involves water volume decomposition, not pollutant load, and does not consider situations where the load intensity per unit area of a certain underlying surface may be significantly higher, resulting in a disproportionate contribution to the total load.
[0004] In the absence of a dedicated source tracing method, engineering practice often directly calculates pollution load based on the area ratio of various underlying surfaces. This approach essentially assumes that the pollutant release intensity per unit area of various underlying surfaces is equal. However, the runoff generation mechanisms, pollutant accumulation, and release processes of different underlying surfaces vary significantly. The results of calculation based on area ratio lack physical basis and are prone to deviation from the actual load distribution.
[0005] Therefore, existing technologies lack a method that can quantitatively decompose the load of a pollution event to a specific underlying surface category based solely on cross-sectional flow and pollutant concentration monitoring data, without additional chemical detection, and whose decomposition results can be explained by the underlying surface's own flow generation and outflow mechanisms. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the prior art and to adopt a method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage under the underlying surface type, comprising the following steps:
[0007] Obtain the flow time series of the target monitoring section during the receding stage of the pollution event. With pollutant concentration time series ;
[0008] Obtain a pre-established underlying surface type response kernel library, which includes unit drainage response kernel benchmark parameters for five types of underlying surfaces: paddy fields, dry land, forest and grassland, hardened ground, and ditch and pond wetland retention areas. , of which Unit drainage response core of the underlying surface Inflection point To divide the water level into a rapid receding phase and a slow receding phase:
[0009] ;
[0010] The attenuation constant of the rapid receding section and the turning point The value range is determined by the physical constraints of this type of underlying surface. Constraints, slow receding water section morphology function Based on the outflow mechanism of this type of underlying surface, it is classified as a linear reservoir exponential decay form. Or power-law decay form ;
[0011] Based on the rainfall intensity of this pollution event and the previous soil moisture index Through the correction function The response kernel reference parameters are corrected to obtain the corrected response kernel parameter center. ;
[0012] Weighting based on the contribution of various underlying surfaces and response kernel parameter correction amount The unknown quantity is solved using the following formula as the joint inversion objective function:
[0013] ;
[0014] in:
[0015] ;
[0016] ;
[0017] and These are the concentration-flow hysteresis loop characteristics calculated from the reconstructed sequence and the measured sequence, respectively. These are the weighting coefficients for each item. For the first Permissible variance of the kernel parameter correction for the underlying surface response;
[0018] The constraints are , ,and ;
[0019] Based on the contribution weights obtained from the solution The proportion of the contribution of various underlying surfaces to the pollutant load in this pollution incident will be output as the source tracing result.
[0020] The slow receding section morphology function The mechanism attribution rules are as follows: Paddy fields and wetland retention areas are classified as linear reservoirs with exponential decay. Its parameters Corresponding to the water storage capacity and outlet flow coefficient of this type of underlying surface; dryland and forest / grassland belong to the power-law decay form. ,in The power-law form directly corresponds to the nonlinear decay relationship of the drainage rate with the remaining storage in Horton's runoff theory; the decay constant of the rapid drainage section of hardened ground. The lower limit of its value range is higher than that of the other four types of underlying surfaces, and the inflection point time is... The upper limit of the value range is lower than that of the other four categories, in order to characterize the characteristic that the drainage of hardened ground has basically no tail.
[0021] The concentration-flow hysteresis loop characteristic Defined as:
[0022] ;
[0023] Integral along the receding water stage - Trajectory detour, This indicates that the hysteresis direction is clockwise, corresponding to a flushing-limited release. This indicates counterclockwise rotation, corresponding to dissolution-restricted release. The hysteresis direction and normalized area are both incorporated into the constraints of the joint inversion objective function.
[0024] The first Pollutant release concentration process of underlying surface Based on the release mechanism, it can be classified as: scour-restricted type. Characterizing the release of surface-mobile pollutants by runoff and the exponential decay of their concentrations as water recedes; or dissolved-limited pollutants. This characterizes the approximately equal concentration of pollutants carried out of the reservoir by the outflow; the mechanism and parameters are described. , The concentration was determined by measured concentration curves during the controlled flow test phase.
[0025] The response kernel benchmark parameters The calibration steps include: conducting small-scale controlled runoff generation tests repeatedly for each type of underlying surface under no less than three gradients of artificially simulated rainfall intensity, and measuring the discharge phase flow curves and pollutant concentration curves for each set of replicates; solving the following fitting problem for each set of data to obtain single-test parameter estimates:
[0026] ;
[0027] By summing up the estimated values of each group, the 10th to 90th percentile of their distribution is taken as the baseline parameter range for this type of underlying surface under the current watershed climate conditions. .
[0028] The preceding condition correction function Attenuation constant of rapid water receding section The correction satisfies: That is, the lower the soil moisture index in the early stage, the better the corrected soil moisture index. The smaller the value, the more pronounced the slow receding tail; the monotonic relationship corresponds to the physical mechanism in Horton's formula where infiltration capacity increases as soil moisture decreases, leading to a decrease in the proportion of slope runoff and thus prolonging the receding time; correction function coefficients The controlled runoff generation experiment was repeated under at least two prior soil moisture conditions and the results were determined by fitting.
[0029] It also includes: after the contribution weight is solved, obtaining the composition ratio of the underlying surface type of each sub-region in the upstream area of the target monitoring section, and estimating the spatial distribution of pollutant load among the sub-regions based on the linear proportional relationship between the area ratio of underlying surface type and contribution weight, in conjunction with the contribution weight.
[0030] It also includes: calculating the underlying surfaces for any two types before solving the joint inversion objective function. , The corresponding correlation coefficients between the unit retreat response kernels after correction are as follows:
[0031] ;
[0032] in , These represent the time mean of the corresponding response kernel sequence; with all underlying surface categories as nodes and correlation coefficients as parameters... Exceeding the preset threshold An undirected graph is constructed with the category pairs as edges. All surface categories in the same connected component of the undirected graph are merged into a joint category to participate in the objective function solution. The output of the merged joint contribution weight is the sum of the contribution weights of each category in the connected component. The contribution weights of categories in the same connected component are no longer solved separately. The contribution weights of surface categories that do not fall into any connected component are still solved separately.
[0033] When the pollutant concentration time series When missing or marked as unreliable, the degenerate joint inversion objective function is used for solving:
[0034] ;
[0035] In the degraded objective function, only those functions satisfying the correlation coefficient are considered. The contribution weights of the underlying surface categories are calculated separately, while the remaining categories are processed according to the merging rules.
[0036] It also includes: obtaining the optimal solution by solving the joint inversion objective function. Then, based on the local curvature of the objective function at the optimal solution, the confidence interval of the contribution weight is estimated:
[0037] ;
[0038] in Let the joint inversion objective function be... for about The second derivative matrix at the optimal solution The standard normal quantile corresponds to the confidence level; the output source tracing results include the contribution ratio of various underlying surfaces and their confidence intervals.
[0039] The present invention has the following beneficial effects:
[0040] I. This invention can complete the source tracing of underlying surface load categories based solely on the time series of flow and pollutant concentrations routinely collected from monitoring sections, without requiring additional chemical isotope or component fingerprinting for this event, thus shortening the time required for source tracing and reducing detection costs.
[0041] Second, the response kernel library established in this invention has a mathematical form that is directly derived from the runoff generation and outflow mechanisms of various underlying surfaces. For example, the linear reservoir assumption of paddy fields corresponds to the gap flow characteristics, and the power-law decay of dry land and forest grassland corresponds to the nonlinear process of infiltration rate changing with soil moisture content. The response kernel morphology has a clear mechanistic explanation, rather than being simply obtained by fitting historical data. The decomposition results obtained have mechanistic traceability.
[0042] Third, this invention reconstructs the pollutant concentration process using a mass flux weighted approach and introduces the concentration-flow hysteresis loop feature as a solution constraint. This allows for the selection of solutions that are consistent with the calibrated underlying surface release mechanism from multiple candidate weighted solutions with similar numerical fitting accuracy. This avoids the possibility of numerically feasible but physically unreasonable decomposition results that may occur when solving based solely on the point-by-point fitting error of flow rate and concentration.
[0043] Fourth, this invention provides an identification and merging mechanism for situations where different underlying surface response patterns are highly similar and it is difficult to identify the contribution ratio individually, thus avoiding giving unreliable independent estimates; for situations where pollutant concentration data is missing or unreliable, it provides a degradation solution mode based solely on flow data, ensuring that the method can still operate when monitoring conditions are incomplete; and further provides confidence intervals for contribution weights, so that the source tracing results come with a clear explanation of the degree of credibility, making it easier for users to judge the reliability of the results.
[0044] Fifth, the contribution weight of the underlying surface category obtained by this invention can be further combined with the composition ratio of the underlying surface in the sub-region to estimate the spatial distribution of pollutant load in different sub-regions within the watershed, providing a more targeted quantitative basis for the zoning and classification prevention and control of non-point source pollution. Attached Figure Description
[0045] Figure 1 A flowchart illustrating the overall architecture of a method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage based on underlying surface type;
[0046] Figure 2 This is a schematic diagram of the response kernel morphology of a source tracing method for agricultural non-point source pollution events based on the waveform fingerprint of drainage in the underlying surface type.
[0047] Figure 3 This is a schematic diagram of the decomposition of the drainage curve, which is a method for tracing agricultural non-point source pollution events based on the drainage waveform fingerprint of the underlying surface type. Detailed Implementation
[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0049] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0050] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.
[0051] Example 1
[0052] The downstream control section of a small agricultural watershed in a hilly area is designated as S07, covering an area of 12.6 square kilometers. The upstream underlying surface consists of 32% paddy fields, 28% dry land, 25% forest and grassland, 8% hardened ground, and 7% ditch, pond, and wetland retention areas. Under current technological conditions, the measured discharge flow and total phosphorus concentration curves at section S07 are single signals formed by the superposition and mixing of the discharge processes of the five types of underlying surfaces within the watershed. Due to the fundamental physical differences in the water storage mechanisms and outflow paths of different underlying surfaces, the discharge response patterns produced by the same rainfall on different underlying surfaces are not the same. Simply calculating the pollution load based solely on the area ratio of the underlying surfaces essentially assumes that the pollutant release intensity per unit area of each type of underlying surface is equal, an assumption that lacks physical basis. The technical problem to be solved by this invention is to quantitatively decompose the load of a pollution event into specific underlying surface categories based solely on measured data during the receding period of a single cross section, and the decomposition results should be explainable by the runoff generation and outflow mechanisms of the underlying surface itself, rather than merely a result that looks good on the numerical scale.
[0053] Following a rainfall event in July of a certain year, the S07 section entered the receding flood stage. The system first identified the receding flood stage using continuous monitoring flow data at the section: the time interval where the flow sequence peaks and then begins a monotonically decreasing trend, with the second-order difference of the decreasing slope turning from negative to approaching zero, was defined as the start and end of the receding flood stage. The identified receding flood stage was defined by the peak flood time. Observation window to hour, sampling step size Hours, to obtain the flow time series Time series of total phosphorus concentration ; exist It reaches a peak of approximately 180 L / s per hour, and then gradually declines until... The current drops to approximately 5 L / s per hour. The peak concentration was approximately 45 mg / L, occurring about 0.5 hours earlier than the peak flow rate. This phenomenon, where the concentration peak precedes the flow rate peak, is a direct manifestation of hysteresis characteristics and will be quantitatively characterized in subsequent steps.
[0054] The system retrieves a pre-established response database. The baseline parameters for the five types of underlying surfaces are not determined by direct fitting, but rather calibrated during the database construction phase after conducting controlled runoff generation tests with at least three gradients of artificially simulated rainfall intensity for each type of underlying surface. Taking dryland as an example, three sets of tests were repeated at rainfall intensity gradients of 20, 35, and 50 mm / h. The measured runoff flow curves for each set were calculated using the formula... The fitting yielded nine groups. The estimated values range from 0.28 to 0.42. The 10th to 90th percentile values, rather than the maximum and minimum values of all samples, were used as the interval boundaries because the number of trials in small-scale controlled experiments is limited. Individual trials may produce outlier estimates due to accidental factors such as localized water accumulation or instantaneous sensor fluctuations during the runoff test. Directly using extreme values as interval boundaries would lead to problems in subsequent joint inversion. The allowable range is expanded by individual outlier groups; taking the 10th to 90th percentile can retain the main distribution characteristics while excluding the influence of outliers at both ends by about 10%, resulting in... of Components , Components , Components Hour.
[0055] The formula needs to be explained here. Medium proportion coefficient Meaning and value selection method: Due to As a unit drainage response core, its physical meaning is the drainage process morphology of this type of underlying surface under unit runoff input conditions. Therefore, for each type of underlying surface... Apply normalization constraints , Combined with the normalization constraint , And the parameters in the slow range are uniquely determined; they are not independently calibrated or freely fitted quantities. Without this normalization, This will simultaneously mix two types of information: "morphological characteristics of the underlying surface drainage process" and "the magnitude of artificially simulated rainfall input in the reservoir construction experiment." After normalization... The contribution weights in the joint inversion are determined solely by the recession pattern itself. It can purely characterize the relative contribution of this type of underlying surface to this event without being confused with the scale of rainfall input in the reservoir construction experiment.
[0056] The reference parameters of the five types of underlying surfaces and their corresponding mechanisms are described below.
[0057] paddy fields ( ) Calibration obtained , The slow-speed segment belongs to the linear reservoir index decay. The drainage process in paddy fields is controlled by the flow capacity of surface gaps or drainage outlets. The flow rate through these gaps is approximately linearly related to the water depth at the surface, directly corresponding to the linear reservoir assumption, which states that the drainage rate is proportional to the remaining water storage. Measured concentration data during the drainage phase of this type of underlying surface show an approximately exponential decay, indicating a scour-limited release. , The soluble phosphorus retained on the paddy field surface is gradually depleted with drainage. Furthermore, the linear reservoir attenuation constant for this type of paddy field... It can be decomposed into the ratio of the outlet flow coefficient to the effective storage capacity: ,in Characterizes the flow capacity of field surface gaps or drainage outlets under standard water head conditions. The effective water storage volume above the field outlet elevation that can participate in drainage is represented by both the outflow velocity and the water storage volume. Both are engineering quantities that can be obtained directly through gap geometry and field surface measurements, independent of drainage curve fitting. The physical basis of this decomposition is that, under the linear reservoir assumption, the outflow velocity is proportional to the water storage capacity, and the proportionality coefficient is the ratio of flow capacity to water storage volume. This is consistent with... The definition of the linear reservoir attenuation constant is completely consistent with the definition itself, and is not an additional approximation. It is estimated by referring to the gap width and surface elevation difference of this type of paddy field. Estimate based on field area and average water depth above outlet The ratio of the two The results obtained from the controlled flow test calibration The magnitudes are consistent. It should be noted that in this example... , The values are estimated based on the general scale of field surface gap structures, used to demonstrate the agreement between the decomposition and the experimental calibration values in terms of scale. No specific engineering structure measurements were conducted on section S07 this time; the values were still calibrated using controlled runoff tests. As the final basis; the practical significance of this decomposition lies in the fact that, if engineering structural measurement conditions are available, , It can be used as an alternative to controlled flow testing. Perform cross-validation.
[0058] dry land( ) Calibration obtained , The slow segment belongs to power-law decay. , In drylands, when rainfall intensity exceeds soil infiltration capacity, overflow occurs on infiltrated slopes. The infiltration rate decreases with increasing soil moisture content. This coupling relationship causes the drainage rate to exhibit a nonlinear decay as described by Horton's runoff generation theory. The characteristic time scale required to characterize the manifestation of nonlinear decay effects Characterize the steepness of decay after manifestation. Calibrate the scour-limited release. , This corresponds to the gradual depletion of surface-mobile phosphorus by runoff, with an initial concentration significantly higher than that in paddy fields. This corresponds to the characteristic of dry land where the surface is exposed and mobile phosphorus is more easily carried directly by runoff.
[0059] Forest and grassland ( ) Calibration obtained , It also belongs to the power-law decay category. , Both parameters are significantly greater than the corresponding values for dryland. Corresponding components are , , Hourly, this difference corresponds to the additional water-holding buffer structure formed by the surface litter layer and the loose root soil layer in forest and grassland areas, causing the nonlinear decay effect to appear later and be steeper after it appears. Calibration yielded scour-limited release, but... , The initial concentration was significantly lower than that in dry land, and the decay constant was significantly higher than that in dry land. This corresponds to the interception effect of forest and grassland vegetation on surface phosphorus, resulting in a lower initial amount of mobile phosphorus, which is depleted faster as the water recedes.
[0060] Hardened ground ( ) Calibration obtained , The value deviates significantly from the other four categories, corresponding to the characteristics of this category: impermeable underlying surface, water storage consisting only of a thin surface water layer, and near-instantaneous drainage process. The calibration yields a scour-limited release. , Its attenuation constant is much higher than the other four types, which corresponds to the characteristic that pollutants on hardened ground surfaces are concentrated and flushed out in the early stage of water receding and then rapidly depleted.
[0061] Ditch and pond wetland retention area ( ) Calibration obtained , The slow-speed segment belongs to the linear reservoir index decay. The drainage is the slowest. The concentration curve of this type of underlying surface consistently remained above 85% of the initial value during the controlled runoff test, indicating a solution-limited release. This corresponds to the physical image where pollutants within the pond / wetland reservoir are fully mixed, and the outflow concentration is close to the overall concentration level of the reservoir. The linear reservoir attenuation constant of the ditch / pond / wetland retention area. Similarly, it can be decomposed into the ratio of the control structure's overcurrent coefficient to the storage volume: , Characterizes the flow capacity of the outlet or gate control structure of a pond or weir. Characterizes the storage capacity of ponds and weirs. Estimated by referring to similar structures. , , The results obtained from the controlled flow test calibration Consistent. This type of underlying surface Significantly smaller than paddy fields After decomposition, the source of this difference becomes clearer: it is not the current carrying capacity. Significantly smaller ( and (Similar in magnitude), but rather in terms of storage capacity. Significantly larger—the design purpose of the reservoir's storage capacity is to reduce flood peaks and extend the storage time, which is consistent with... The calibration result of taking the minimum value among all five categories corresponds perfectly to the mechanism, and it yields a consistent order of magnitude with similar estimation methods when applied to paddy fields, indicating that this decomposition method is not merely for finding numbers for a single category. Similar to the aforementioned explanation of paddy fields, in this example... , The values were estimated based on the general structural magnitude of ponds and weirs. No specific engineering structural measurements were conducted on section S07 this time; the values were still determined using controlled flow generation tests. As the final basis, the significance of this decomposition lies in... Provides a cross-validation pathway independent of controlled flow testing.
[0062] The intensity of rainfall in this event early soil moisture index (The reference conditions for benchmark library calibration are as follows) Dryland correction factor After correction Forest and grassland correction coefficient After correction The correction for paddy fields, ditches, ponds, and wetland retention areas is within 5%, remaining approximately unchanged; hardened surfaces are unaffected. .
[0063] With the corrected parameter center Using the initial baseline, an alternating optimization approach is employed to solve the joint inversion objective function, with weighting coefficients set to... , , , .
[0064] The values of the weighting coefficients are not determined by independent empirical judgments, among which... , and This can be explained within a unified statistical framework. Let... , ,in , The noise is observed independently by the flow meter and the water quality sensor; at the same time Obey the modified parameter center Mean and variance The Gaussian prior—this prior hypothesis is not an additional assumption—" The design of limiting the allowable fluctuation range to one-tenth of the width of the database construction parameter distribution itself describes the basis for the variance value of the Gaussian prior. Under this setting, , The negative logarithmic posterior (neglecting constant terms irrelevant to the unknown) is exactly:
[0065] ;
[0066] By comparing it item by item with the joint inversion objective function, it can be seen that... , , In the context of Bayesian maximum a posteriori estimation, the empirical weights are not freely adjustable, but should be directly taken as the reciprocals of the corresponding observations or prior variances: , The estimated value was obtained by repeating 30 measurements of the S07 cross-sectional flow meter under steady flow conditions. The total phosphorus was estimated by repeated measurements from the online sensor. The ratio of the inverse variances of the two measurements relative to their respective measurement ranges is given directly. , This value relationship no longer relies on the qualitative judgment that "concentration observation noise is usually higher", but is directly determined by repeated measurement calibration experiments of the sensor.
[0067] The same set of derivations does not apply: hysteresis loop characteristic It is a diagnostic quantity obtained by nonlinear calculation of the reconstructed sequence, not a direct observation of a certain original physical quantity, and therefore does not have an independently calibrated observation noise variance. This cannot be directly derived using the aforementioned Bayesian framework. Instead, it serves as a morphological constraint imposed outside of the numerical fitting error, taking the maximum value among the five coefficients. While this selection still has its technical basis, it is essentially a design choice and should not be directly correlated with... , , They are confused into the same type of derivation.
[0068] This solution converged in the 14th round, ultimately yielding the contribution weights. , , , , At convergence, the objective function takes the following values: flow fitting term. Concentration fitting term Hysteresis loop deviation term Regularization term The total objective function value is approximately 182.1. The hysteresis loop bias term accounts for a very small proportion of the total objective function value, which is consistent with the fact that the measured hysteresis direction is already very close to the reconstruction result. Its main role is not to contribute to the numerical reduction space, but to eliminate those solutions that do not satisfy the hysteresis loop shape from among the many candidate weighted solutions that can also reduce the fitting error of flow and concentration to a low value.
[0069] The convergence values of the response kernel parameter correction are as follows:
[0070] , ;
[0071] The other three categories are close to zero. This is because the prior condition correction function only applies to... ,check When this constraint applies, middle The corresponding components should be shifted synchronously by the same correction amount, that is... , , The corresponding components use the original calibration area. Substitute these into the previously calibrated values. , And the corresponding correction amount, verifying both dry land and forest / grassland. All of them fall within the translated interval, the other three categories Since it is close to zero, it naturally falls within its respective interval, and all constraints are satisfied.
[0072] Based on this convergence result, the application of the flow and concentration reconstruction formula at a specific time point is demonstrated. The aforementioned calibrated response kernel parameters Substitution The unit response kernel values for the five types of underlying surfaces at that moment were obtained as follows: (Paddy fields, fast section) (Dryland, slow section) (Forestry and grassland, slow section) (On the hardened ground, the slow-speed attenuation has been basically completed.) (Ditch and wetland retention area, fast section). Substitute the above response values and the calculated contribution weights into... ,have to Based on the total flow generated in this event, the corresponding actual flow rate is approximately 115 L / s, which is consistent with the measured flow rate at that moment. The L / s values are close, with a deviation of approximately 3%.
[0073] The previously marked , Substituting the respective release concentration processes, we obtain , , , , (Dissolution-limited, approximately constant). Substitute the above values into the mass flux weighted reconstruction formula. ,have to ;
[0074] In contrast, if the concentration is simply weighted and averaged according to the contribution weights of various categories... Without considering the differences in actual outflow intensity, we get The difference between the two is approximately 17%. This difference stems from the fact that dry land... The outflow intensity at time 10 is significantly higher than the assumption of "equal outflow intensities for all categories" implied by simple weighted averaging. At this moment, the dryland happens to be the category with higher concentrations, thus mass flux weighting correctly amplifies the actual contribution of the dryland to the overall concentration at that time. Compared with the measured concentration at that time In comparison, the deviation is about 2%, which is significantly lower than the deviation of about 13% between the simple weighted average and the measured value. This comparison verifies the necessity of the mass flux weighted concentration reconstruction formula relative to the simple weighted average in restoring the measured concentration.
[0075] Based on the aforementioned definition of the concentration-flow hysteresis loop characteristic quantity Reconstruction along the receding phase of this event - Trajectory and Actual Measurement - The trajectories are calculated separately, and the results are obtained. , The deviation is within the convergence tolerance; the peak value of the reconstructed concentration appears about 0.4 hours earlier than the peak value of the reconstructed flow rate, and the direction is clockwise, which is consistent with the measured hysteresis direction, indicating that the release of pollutants in this event is dominated by the scour-limited mechanism, which is consistent with the solution result that the dry land and hardened ground account for 53% of the contribution weight.
[0076] In contrast, if the objective function is... Removing this constraint term and solving only for the point-by-point fitting error and regularization term of flow rate and concentration yields another set of weights: , , , , This set of solutions differs from the aforementioned solutions by no more than 3% in the root mean square error of point-by-point fitting of the flow rate and concentration sequences, belonging to the category of solutions that are also "feasible" at the numerical fitting level, but their corresponding... It dropped to 0.06, compared to the measured value. The difference is as high as 0.18, and this set of solutions significantly increases the weight of forest and grassland from 12% to 31%, which contradicts the established mechanism attribute that forest and grassland have the longest drainage time and the weakest rapid erosion characteristics. This comparison shows that the point-by-point fitting error itself constitutes an under-constrained objective function, and the role of the hysteresis loop feature constraint term is to select the set of solutions that are consistent with the established underlying surface release mechanism from multiple numerically equivalent solutions.
[0077] The final source tracing results for this pollution incident are as follows: dry land contributed 35%, paddy fields contributed 28%, hardened ground contributed 18%, forest and grassland contributed 12%, and ditch and wetland retention areas contributed 7%, which serve as the quantitative basis for tracing the underlying surface categories of this total phosphorus pollution incident.
[0078] In some embodiments, if the general blind source decomposition method is directly applied to the historical receding events at section S07 to decompose the contribution of the underlying surface type, the specific steps are as follows: collect the flow curves of 40 historical receding events at section S07, and resample each event to a unified time grid to form a matrix. ,right Perform nonnegative matrix decomposition with rank 5 The solution employs a multiplication update rule, initializing... , Each element is sampled independently. The convergence criterion is the reconstruction error between two adjacent iterations. The relative change is less than Or it can reach a maximum of 200 iterations.
[0079] The solution obtained after convergence The five lines represent the five basic recession patterns extracted, denoted as follows: to Calculate these five components and their correlation with the five types of response kernels defined in this invention. (Paddy field) to The complete Pearson correlation coefficient matrix between (ditch, pond, wetland, and water storage areas) is shown in Table 1 below:
[0080] Table 1. Pearson Correlation Coefficient Matrix:
[0081] ;
[0082] The complete matrix shows: and The correlation coefficient is 0.89, which is significantly higher than the other values in the same row (the second highest being 0.31), and can be regarded as clearly corresponding to the single category of paddy fields; and The correlation coefficient is 0.85, clearly corresponding to the water retention area of ditches and wetlands. However... and (0.61) (0.58) The two values are close and cannot be classified into a single category; and (0.55) (0.47) is also close, and there is also ambiguity in its attribution; The correlation coefficients with the five types of response kernels are all in the range of 0.24 to 0.33. They do not form a clear correspondence with any one type and belong to the intermediate component that neither clearly corresponds to a certain type of underlying surface nor is a negligible residual.
[0083] Baseline parameters of five types of response kernels Independent calibration was performed for each individual underlying surface type during the controlled runoff test phase. The input signal during calibration originated solely from that single type, without involving mixed signal decomposition, thus eliminating rotational and scalar uncertainties. Repeated calibration tests on the same type of underlying surface (e.g., 9 sets of tests in dryland) yielded... The estimated values are distributed in Within the interval, the consistency of the calibration results is directly guaranteed by the repeatability of the experiment, rather than by the numerical decomposition of the mixed signal.
[0084] If we consider the power-law decay form of the slow segment of dryland and grassland... The same exponential decay form as that used in paddy fields, ditches, and wetlands is uniformly replaced. This means eliminating the mechanism-differentiated design of response nuclei and only retaining the range of values for the baseline parameters of the classification, requiring recalibration for dryland and forest / grassland separately.
[0085] For the slow-speed segment of the measured drainage curves in nine controlled experiments in dryland areas, an exponential decay model was fitted, and the goodness of fit was determined. The mean value is 0.86, which is significantly lower than the original power-law form fit. Mean 0.97; Fit residuals at the end of the slow segment ( The residual mean (hours) is systematically positively skewed, with the mean residual being 11% of the mean measured concentration, indicating that the exponential decay form has a continuous and cumulative systematic underestimation in the long-tail stage of the dryland water receding process.
[0086] The data from the nine controlled experiments on forestry and grassland were also fitted with exponential decay models, and the goodness of fit was determined. The mean value is 0.79, which is lower than the original power-law form fit. The mean is 0.98, which is lower than the 0.86 of the fitted dryland forced index; the fitting residual is at the end of the slow range ( The residual mean (hourly) also exhibited a systematic positive skewness, with a mean residual of 16% of the measured concentration mean, higher than the 11% corresponding to dryland. The goodness-of-fit decrease and terminal systematic bias were more pronounced in forest-grassland than in dryland, a difference directly corresponding to the identified mechanism: the power-law parameters of forest-grassland. , All are greater than the corresponding values for dry land , This means that the nonlinear decay effect of forest and grassland appears later and decays more steeply after it appears. This implies that the forest and grassland water receding curve deviates more from the exponential decay form on medium and long time scales than that of dry land. Therefore, forcing the exponential decay form to describe the forest and grassland water receding process will result in a correspondingly larger systematic bias.
[0087] Using the ablation of the dryland and grassland response kernels, the alternating optimization solution process is reused (initialization) ,fixed Solving constrained quadratic programming subproblems, fixed right Using alternating gradient descent techniques, the relative change is less than... Or stop after 30 rounds, weighting coefficient , , , (Unchanged) A joint inversion of the S07 event was performed again. This solution converged in the 16th round, yielding the contribution weights. , , , , , compared with the original result ( , Compared to the previous year, the weight of dryland increased from 35% to 44%, while the weight of forest and grassland decreased from 12% to 9%.
[0088] Because both the index approximations for dryland and forest-grassland exhibit systematic underestimation (11% and 16% respectively), they display a competitive rather than independent compensatory relationship in the joint inversion after ablation. The forest-grassland index approximation has a larger deviation, and its reconstructed curve's contribution to measured flow is underestimated to a greater extent on medium- to long-term scales. Under the constraint of a fixed total flow reconstruction target, the joint inversion algorithm tends to use the relatively smaller deviation of dryland weights to fill this gap, rather than keeping the forest-grassland weights unchanged. This directionality is determined by the relative magnitude of the residuals of the two, rather than the result of either one acting alone. This result indicates that the differentiated selection of slow-speed segment morphology based on the outflow mechanism of the underlying surface is a necessary condition for the accuracy of the inversion results of this invention. This necessity is even more critical for categories with longer characteristic time scales and stronger nonlinearity in the slow-speed segment (such as forest-grassland in this example).
[0089] Example 2
[0090] This embodiment aims to illustrate the specific data structure and its operation mode within the joint inversion solution step in the complete process of obtaining the underlying surface type response kernel library, correcting the response kernel baseline parameters based on the rainfall intensity and previous soil moisture index of this pollution event, and then solving the joint inversion objective function with the contribution weights of various underlying surfaces and the correction amount of the response kernel parameters as the variables to be solved.
[0091] Specifically, in the alternating optimization process of joint inversion, the inner layer is fixed. right Solving, outer layer fixed right The two steps of the solution process alternate. Observe the specific form of the outer subproblem: fix the current... back, right In terms of linearity, this subproblem is already a standard nonnegative least squares problem. However... As itself The function is a nonlinear quantity calculated using exponential or power functions. If in each round of alternating optimization, the inner layer... When performing gradient search, each probe must be considered. The value is recalculated completely once. The computational cost accumulates linearly with the number of iterations and the number of attempts in the gradient search, depending on the value taken at all sampling times.
[0092] Therefore, for the first The underlying surface, at the beginning of each outer iteration, revolves around the current parameter center. Define and cache the following Jacobian structure:
[0093] ;
[0094] in The sequence of response kernel values at the parameter center. In response to the gradient sequence of the check parameter vector, both are based on the sampling time. The entire index is cached, only in the parameter center. When an update occurs (i.e., when the outer iteration enters the next round), the calculation is recalculated once. Based on Response core The first-order expansion within the neighborhood is:
[0095] ;
[0096] It needs to be explained that, Its corresponding inflection point moment The component is always zero in this embodiment, only and slow segment shape parameters ( ,or , The corresponding components participate in the linearization and inner-layer iterative solution described here.
[0097] Substituting the above linear expansion into the flow fitting term in the joint inversion objective function, the inner layer... (at this time The subproblem, fixed at the current outer iteration value, is approximated as:
[0098] ;
[0099] This is about Linear least squares problem (residual terms pair) It is linear (the regularization term is quadratic) and has a closed-form ridge regression solution.
[0100] Now consider the outer sub-problem: fixed ,right Please solve this. already cached The whole is denoted as The outer subproblem is written as:
[0101] ;
[0102] It plays the role of "endmember" or "source component spectrum" in classic mixture models. It plays the role of a mixing ratio, and both are subject to non-negativity and normalization constraints. The only difference between the two is that in the classic mixture model, the endmembers... It is a constant that does not change with the solution process and is directly taken from the measured chemical characteristics of the corresponding source; in this method It is a quantity that is continuously updated by the inner subproblems as the outer layer iterates, and its range of values is... Limited to It is a neighborhood, not a completely free quantity.
[0103] This leads to a precise proposition of degeneration, if we take Shrink into a set of single points (i.e., order) If the response kernel shape is not allowed to be fine-tuned with event conditions, then the inner subproblem no longer has free variables, and alternating optimization degenerates into solving the outer subproblem only once. With fixed endmembers, we directly solve for the linear least squares mixed ratio under non-negative, normalized constraints.
[0104] This method is achieved through... and Introduced, Being The endmember adaptive deformation capability, limited to a small neighborhood, allows the endmember itself to make minor corrections to the baseline shape at the event level, without requiring the baseline response kernel calibrated during the library construction phase to be exactly true under any event conditions.
[0105] In addition, this method will Excluded in Jacobi structure Outside the linearization range, Calculate the gradient only for the smooth parameters within the same segment: classify the slow segment as a linear reservoir. The component is and For categories belonging to the power-law form, the components are: , , ; The value of is always determined solely by the prior condition correction function. It is determined once and does not participate in the iterative solution of inner subproblems. That is... The moment when the mechanism of the receding water process changes is directly given by two observable driving factors: the rainfall intensity of the event and the soil moisture in the preceding period. It describes the residual shape deviation within the same mechanistic stage that is empirically difficult to fully explain by these two driving forces.
[0106] The structure is generated, stored, and invoked during the computation process as follows;
[0107] At the start of each outer iteration, around the current... (i.e., when the previous outer iteration ends) The values are accumulated to the new center after the parameter center is reached, and the calculation is repeated once; within the outer iteration of this round, regardless of how many attempts are needed to solve the gradient search or ridge regression for the inner subproblem. The values are all directly reused from the same set. ,pass This linear operation yields the corresponding response kernel approximation without requiring re-evaluation of the exponential or power function. Because , , , All appear in parsed form In the expression, Each component also has an analytical expression (e.g., fast segment). This eliminates the need to approximate gradients using numerical differencing, thus avoiding the truncation error and additional function evaluations introduced by numerical differencing itself.
[0108] The inner subproblem is transformed from general nonlinear least squares, which requires repeated calls to nonlinear functions for evaluation, into linear least squares (ridge regression) with a one-time gradient cache. This is the specific benefit brought by the Jacobi structure at the computational implementation level, and the accuracy of this transformation is continuously narrowed by the outer iterations. The search neighborhood is used to ensure that, after each round of the outer iteration, the parameter center is... The value is updated to the actual convergence value of the previous round, and the linearization base point moves closer to the true optimal solution. The linear approximation error in the neighborhood decreases as the iteration progresses.
[0109] It is calculated and written into memory at the beginning of each outer iteration, and is read by all calculation steps within that outer iteration. It is replaced with a new value before entering the next outer iteration, and no history is retained across iterations.
[0110] The specific process is as follows. The response kernel value corresponding to each sampling time and gradient components All have analytical expressions, as mentioned before, fast segments During the calculation, all sampling times were considered. Substituting each value into the analytical expression yields a numerical sequence of length equal to the number of sampling points. Store as a one-dimensional array of numbers. Stored as a two-dimensional numerical array, with rows corresponding to sampling times and columns corresponding to parameter components (the slow segment is categorized into power-law forms, with 3 columns). , , For categories belonging to the linear reservoir type, the number of columns is 2, corresponding to... , The five types of underlying surfaces each have their own specific characteristics. and Calculations and storage are performed separately and are not shared between them.
[0111] Solving inner subproblems At that time, read all five categories. and ,according to The design matrix is constructed using linear combinations, and the closed-form solution of the ridge regression is obtained to obtain the current round. The value of ; this step only reads the stored value throughout the entire process. , Numerical values are used without re-invoking exponents or power functions. Solving outer subproblems. At that time, the inner layer was just obtained Substitution Similarly, only linear combination operations on the arrays are performed to obtain the results of this round. Subsequent non-negative least squares solution .
[0112] After the inner and outer sub-problems of this round of outer iteration are solved, Accumulated to With the new parameter centers obtained, the analytical expression is re-invoked using the new parameter centers to calculate a new round of results. and This overwrites the old values in storage. The decision to continue to the next round is based on the magnitude of the change in the objective function value relative to the previous round. If the magnitude of the change is below a preset threshold or the maximum number of rounds has been reached, the iteration stops, and the stored values are overwritten. That is, the final The value is directly passed to the subsequent calculation of the correlation coefficient. The steps are read and will not be recalculated.
[0113] The calculation reads the same document. Arrays, for any two classes , Retrieve the data stored in their respective final rounds. , Calculate the inner product and modulus after removing the mean, and obtain This step does not re-invoke the parsed generation of response kernel values, but only reads the array left in storage from the last round of alternating optimization.
[0114] The Jacobi structure described in this embodiment is a specific implementation of the aforementioned joint inversion solution step. This structure transforms the solution of the response kernel parameter correction quantity from a search process that relies on repeated evaluation of nonlinear functions into a closed-loop linear solution based on cached data. Furthermore, the cached response kernel values in this structure can be directly read for the calculation of the response kernel correlation coefficient without repeated generation. The entire implementation process does not change the basic setting of using contribution weights and response kernel parameter correction quantities as the variables to be solved and performing joint solution through a unified objective function. It also does not change the composition of the response kernel library, the basis for parameter correction, or the method of finally outputting the source tracing results. It only elaborates on how the solution process is carried out internally. Together with the response kernel mechanism attribution, parameter calibration method, and event condition correction rules explained in the previous embodiments, it constitutes a complete technical solution.
[0115] Example 3
[0116] This embodiment aims to illustrate the meaning of each specific calculation formula, the mathematical relationship between them, and how these mechanisms operate and cooperate in actual events in three stages: determining whether each underlying surface response kernel can be solved separately based on its correlation, adjusting the solution method when pollutant concentration data is missing or unreliable, and estimating the credibility of each type of contribution ratio after the solution is completed.
[0117] First, let's explain the correlation coefficient. The formula itself is as follows:
[0118] ;
[0119] in, , This is the code for the underlying surface category (the value ranges from 1 to 5, corresponding to paddy fields, dry land, forest and grassland, hardened ground, and ditch and pond wetland retention areas, respectively). That is, the first Class and the The correlation coefficients between the underlying surfaces are obtained. The sampling time is defined as the time grid used in the aforementioned response kernel calculation. For the first At the sampling time, the underlying surface of the type Furthermore, the response kernel parameter takes the currently corrected center value. The unit water discharge response value at that time is verified. The meaning is the same, corresponding to the first Subsurface type; here we take Reference parameters obtained by non-calibration This is because the correlation coefficient needs to reflect the actual form of the two types of response kernels under the conditions of this event, rather than the general form during the database construction phase. , They are respectively , The time mean over all sampling times is calculated by summing the values of the response kernel within the observation window and dividing by the total number of sampling points.
[0120] Molecular part This involves subtracting the time mean from each of the two response kernel curves, multiplying them time-by-time, and then summing over all sampling times. This step measures whether the two curves deviate from their respective average levels in the same direction at the same time point—if... At a certain moment, it is higher than its own average. If the deviation is also higher than its own mean, and this same deviation holds true at most sampling times, then the product of each term is mostly positive, and the summation yields a large positive number; if the deviation directions of the two curves are uncorrelated, the positive and negative terms will cancel each other out, and the summation result will approach zero.
[0121] and As a response kernel that has been normalized and modified for event conditions, its absolute value is itself subject to various underlying surface normalization constants. Due to factors such as the magnitude of corrections, the overall heights of the two curves may not be the same. If the mean is not subtracted and the two curves are multiplied and summed directly, two different things will be mixed up: "the difference in the overall amplitude of the two curves" and "whether the shapes of the two curves are similar over time". After subtracting the mean, the object of comparison becomes "the way the two curves fluctuate around their own average level", that is, the shape characteristics of the curves, rather than the absolute magnitude.
[0122] The denominator is the product of the square roots of the sums of the squares of the two mean-reduced sequences. Its function is to compress the unnormalized numerator to a smaller value. Within a fixed range, the calculated values between different categories are... The magnitudes can be directly compared, unaffected by differences in the absolute amplitudes of their respective curves. Since all five types of response kernels exhibit a monotonically decreasing trend from their peak values during the recession phase, In practice, the value of this method is usually positive. The closer the value is to 1, the more similar the decay rhythm and shape of the two curves over time, making it difficult to distinguish the contributions of the two types of underlying surfaces using only their respective time series. The closer the value is to 0, the more significant the difference in the decay rhythm of the two curves, and the shape itself contains information that can be used for differentiation. The merging mechanism uses this value to determine whether the contributions of the two types of underlying surfaces can still be solved separately.
[0123] The judgment mechanism for responding to kernel merging and the estimation mechanism for the contribution weight reset confidence interval actually deal with the different manifestations of the same mathematical quantity in different value intervals, rather than two independently designed functional modules. The first... Class and First The response kernel is viewed as two vectors on the sampling time grid. , Define the mean-removed sequence , Flow fitting term The second derivative submatrix is:
[0124] ;
[0125] Normalization constraint by response kernel Dryland and grassland response kernels, which belong to the power-law decay pattern, are similar in variance scale and can be taken as... At this point, the eigenvectors of the matrix are exactly... and , respectively corresponding direction and Direction, eigenvalues: ;
[0126] This derivation shows hour That is, the curvature in the direction required to distinguish the contribution of the two categories tends to disappear, while Directional curvature Unaffected, this is the precise mathematical reason why the merger mechanism holds. This approximate dependency... This only applies to category pairs with similar shapes; category pairs with significant differences are not applicable. The value is very low and will not approach the merging threshold, so the simplified derivation is always applicable in scenarios where judgment is actually required.
[0127] Merge threshold The value of is inferred from the magnitude (approximately 0.25) of the weights of dryland and grassland separately in historical events: [Requirement] The direction estimation is considered meaningful only when the standard error does not exceed half of this order of magnitude. The corresponding result is obtained by substituting the flow observation noise level into the aforementioned eigenvalue relationship. It is approximately 0.8.
[0128] The above derivation provides a criterion for judging any pair of categories. However, when the number of underlying surface categories exceeds two, a transitive relationship may occur between the pairwise correlation coefficients. A specific algorithm is needed to handle this situation, rather than simply processing isolated pairs. Before solving the joint inversion objective function, the system performs the following judgment process: using the five underlying surface categories as the node set... Calculate all Correlation coefficient between any two nodes For each pair that satisfies node pairs In the node set Add an undirected edge to form a graph. ;right The connected components are identified, and each connected component forms a merge group. All categories within the same merge group are merged into a single joint weight. Different merge groups and isolated nodes not falling into any merge group are solved independently. In this example, if the correlation coefficient between dryland and forest / grassland exceeds the threshold, the correlation coefficient between forest / grassland and ditch / wetland retention area also exceeds the threshold, but the correlation coefficient between dryland and ditch / wetland retention area does not exceed the threshold, all three will still be grouped into the same merge group through connected components, and a single joint weight will be solved together. The necessity of this approach lies in the fact that if only pairwise independent judgments are made and connectivity is not considered, dry land and ditch wetland retention areas will appear simultaneously in two different "merging pairs", resulting in the same category being included in two different joint weight definitions, causing weight assignment conflicts. The connected component algorithm ensures that each category uniquely belongs to one merging group in this solution, and there will be no assignment conflicts.
[0129] In a pollution incident involving a small upstream basin and a large area of abandoned land at the boundary between dryland and forest / grassland, the initial soil moisture index was [missing information]. Substituting into the correction function, we get dry land. Forest and grassland The two responses show similar nuclear morphologies. Calculating all 10 pairs of correlation coefficients according to the above determination process, only... The threshold was exceeded, while the remaining 9 pairs were all below 0.5. (Figure) There is only one edge between nodes 2 and 3, and the connected component is , , , Therefore, only dry land and forest / grassland are combined, while the other three categories are solved independently. To verify the numerical necessity of this determination, measured flow sequences were analyzed. Apply independent and identically distributed perturbations, , , For each perturbated sequence, perform the complete joint inversion solution again, and record the results of 10 solutions. , Values: A separate solution The range fluctuated significantly within the range, with a range of 0.37. The sum is in Within the interval, the range is only 0.04, consistent with the aforementioned eigenvalue relationship. The system combines the two and only solves for the joint weight. After merging, the response kernel parameter corrections for the two categories are no longer solved independently, since they are... Since the direction itself is unidentifiable, continuing to retain separate free correction quantities for the two will reintroduce the same unidentifiable degree of freedom at the correction quantity level; the system should be changed to follow the direction before merging. , The internal proportions of their respective watershed areas (28% for dryland and 25% for forest and grassland in this example, with an internal ratio of approximately 53:47) fix their relative morphologies, and only the single shared correction quantity is solved for the entire equivalent response kernel after merging. .
[0130] Because the merging mechanism, degenerate solution mode, and confidence interval estimation work together, the final output contribution ratio may present various combinations such as single category, merged category, with confidence interval, or without confidence interval. The system uniformly encapsulates the source tracing results of each solution into a structured object containing the following fields: event identifier; category grouping list, where each group records the set of underlying surface category numbers (single category or merged group), the contribution weight of the group, and a flag indicating whether the group is a merged group; confidence interval field, providing a half-width confidence interval for groups that are not merged and pass the curvature test, and setting it to empty with a reliability description for merged groups or groups with questionable curvature test results; data integrity field, recording whether the flow and concentration data used in this solution are complete, and recording the missing time period if missing data exists; the objective function terms actually used in this solution (four complete terms or two degenerate terms); and the pairwise correlation coefficient matrix of all categories calculated before the solution. This structure unifies the result expression methods for data missing, confidence interval, and the aforementioned merging scenarios, avoiding incompatible output formats defined by the three mechanisms.
[0131] Correlation coefficient The calculation depends only on the modified response kernel parameters. Since this is completely unrelated to pollutant concentration observation data, the merging and judgment process is executed as is when concentration data is missing, without the need for a separate criterion. When concentration data is missing or unreliable, the objective function upon which the joint inversion is based degenerates accordingly:
[0132] ;
[0133] The two terms retained in this formula are the cumulative term of the traffic reconstruction deviation ( ,in To measure the actual flow rate, To be in accordance with the current , The reconstructed traffic is obtained by summing the squared differences between the two traffic samples over all sampling times, with weighting coefficients. The values are the same as in the aforementioned complete objective function) and the regularization term for the response kernel parameter correction ( , For the first Class response kernel parameter correction amount, For its allowable variance, the weighting coefficient (The values remain unchanged). The reason for removing only the concentration fitting term and the hysteresis loop bias term while retaining these two terms is that when concentration data is missing or unreliable, only the data that is affected is the missing or unreliable hysteresis loop bias term. and the results obtained from calculations based on the complete concentration trajectory Traffic data The two retained items are not affected by the concentration sensor malfunction and do not depend on concentration observation. They can continue to be calculated normally according to their original meaning. If the two removed items are retained, the deviation brought in by the unreliable concentration data will be regarded as the true fitting signal, which will mislead the weight solution.
[0134] The system adds a time-by-time validity marker to the input concentration time series, denoted as In one incident, the online monitoring equipment for total phosphorus at section S07 produced continuous abnormal readings from hour 18 to 41 due to probe scaling. The summation range of the concentration bias term and hysteresis loop bias term in the complete objective function correspondingly only covers... At any given moment; if the effective concentration coverage of an event is too low, the system directly discards the concentration deviation term and the hysteresis loop deviation term, and solves the problem using the degradation form described above.
[0135] The initial conditions for this event were: , The merging threshold was not exceeded, the judgment process did not form any edges, and the five types of underlying surfaces were solved independently. The degenerate solution yielded... , , , , After the equipment was repaired, complete concentration data for the event was measured again and the solution was recalculated. , , , , The difference between the two sets of results in each category does not exceed 2 percentage points.
[0136] The other event's initial conditions were , The concentration was still below the merging threshold but significantly higher than the previous event, also encountering a lack of concentration data. Complete concentration data (re-measured afterward) was obtained. , Degenerate solution yields , The two solutions differ by approximately 8 percentage points in their distribution across dryland and forest / grassland, and in opposite directions, although the fitting errors of the two solutions on the flow series differ by no more than 3%. This comparison indicates that the reliability of the degradation model varies with... Increased and then systematically decreased: The flow rate pattern itself already carries sufficient distinguishing information, so the lack of concentration dimension has little impact. The flow pattern alone is insufficient for independent differentiation; the true discriminative errors caused by the missing concentration and hysteresis loop dimensions become apparent. The source tracing results for both events are output according to the aforementioned unified structure, with the data integrity field recording the corresponding missing time segments and solution modes, allowing downstream users to assess the reliability of the results.
[0137] The solution results for the July event at section S07 Based on (the event) The weight coefficients of each term in the objective function correspond to the reciprocal of the variance of the observation noise within the Bayesian maximum a posteriori estimation framework. This is the maximum a posteriori estimate within this framework. The confidence interval for the contribution weights is estimated by the local curvature of the objective function at the optimal solution:
[0138] ;
[0139] In the first formula, It is the objective function Regarding the total contribution weight In the optimal solution The second derivative matrix at that point, i.e., in the second section above. , The complete matrix formed by the curvatures in both directions indicates that the greater the curvature in a certain direction, the faster the objective function deviates from the optimal solution in that direction, meaning the estimate in that direction is less susceptible to observation noise. Inverting this matrix yields... It converts the magnitude of curvature into the magnitude of uncertainty in the estimated value—the direction with greater curvature corresponds to... Smaller values correspond to larger values in directions with lower curvature, which is consistent with the second section. lead to The conclusion that the directional curvature approaches zero, and therefore the directional estimate is unreliable, is the same thing expressed from two perspectives. In the second equation, It is the first Point estimation results of the contribution weights of the underlying surface. yes The first in the matrix Line 1 The diagonal element of the column, i.e. The standard error is obtained by taking the square root of the variance of the self-estimated value. To determine the standard normal quantiles corresponding to the confidence level, this embodiment uses the quantiles corresponding to the 95% confidence level. ; That is, point estimation Expand on both sides The interval obtained after multiplying the standard error directly reflects the degree to which this category is distinguishable from other categories in this solution.
[0140] Based on this calculation, the confidence interval for forest and grassland is obtained. Hardened ground To verify the reliability of the curvature approximation itself, a perturbation resampling process with the same structure as the aforementioned merging verification was performed: a random perturbation with the same distribution as described above was applied to the flow sequence, the solution was repeated 30 times, and the results of each solution were recorded. , Take values and calculate the empirical standard deviation of 30 results: The empirical standard deviation is approximately 0.031. Approximately 0.011. The curvature approximation for hardened surfaces (0.0102) deviates from the empirical value (0.011) by less than 8%; the estimate for forest and grassland (0.0255) is approximately 18% lower than the empirical value. This deviation is due to the difference between forest / grassland and dryland. Already made In directions with relatively small curvature, the influence of higher-order terms of the objective function relative to the quadratic term becomes more pronounced, thus reducing the accuracy of the local quadratic approximation. Hardened ground generally has lower correlation coefficients with other categories, corresponding to larger directional curvatures, making the quadratic approximation closer to reality. Based on this, the system adjusts the confidence interval field in the output structure accordingly. For groups whose values are close to the merging threshold, an additional reliability statement is added, indicating that the width of the confidence interval should be considered as a lower limit rather than an exact value, so that downstream users can use this label to determine whether further verification by perturbation resampling is needed.
[0141] The correlation coefficient calculation, connected component determination, degenerate objective function, and confidence interval estimation described in this embodiment constitute a complete verification of the reliability of the contribution weight solution results. The correlation coefficient and confidence interval are essentially two manifestations of the same curvature relationship within different value ranges. The former is used to promptly merge data when the distinguishability is too low, avoiding meaningless independent estimates; the latter is used to accurately label the accuracy of each category's estimation results when the distinguishability is acceptable. The degenerate solution mode demonstrates that even if information in the concentration dimension is missing, the above judgment and solution process can still be fully executed based on the flow data itself, and its reliability also depends on the correlation between response kernels. All three share the same underlying mathematical relationships and a unified result output structure. Together with the response kernel mechanism attribution, parameter calibration methods, event condition correction rules, and joint inversion solution process explained in the preceding embodiments, they constitute a complete technical solution.
Claims
1. A method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage from underlying surface type, characterized in that, Includes the following steps: Obtain the flow time series of the target monitoring section during the receding stage of the pollution event. With pollutant concentration time series ; Obtain a pre-established underlying surface type response kernel library, which includes unit drainage response kernel benchmark parameters for five types of underlying surfaces: paddy fields, dry land, forest and grassland, hardened ground, and ditch and pond wetland retention areas. , of which Unit drainage response core of the underlying surface Inflection point To divide the water level into a rapid receding phase and a slow receding phase: ; The attenuation constant of the rapid receding section and the turning point The value range is determined by the physical constraints of this type of underlying surface. Constraints, slow receding water section morphology function Based on the outflow mechanism of this type of underlying surface, it is classified as a linear reservoir exponential decay form. Or power-law decay form ; Based on the rainfall intensity of this pollution event and the previous soil moisture index Through the correction function The response kernel reference parameters are corrected to obtain the corrected response kernel parameter center. ; Weighting based on the contribution of various underlying surfaces and response kernel parameter correction amount The unknown quantity is solved using the following formula as the joint inversion objective function: ; in: ; ; and These are the concentration-flow hysteresis loop characteristics calculated from the reconstructed sequence and the measured sequence, respectively. These are the weighting coefficients for each item. For the first Permissible variance of the kernel parameter correction for the underlying surface response; The constraints are , ,and ; Based on the contribution weights obtained from the solution The proportion of the contribution of various underlying surfaces to the pollutant load in this pollution incident will be output as the source tracing result.
2. The method for tracing agricultural non-point source pollution events based on the waveform fingerprint of underlying surface type drainage as described in claim 1, characterized in that, The slow receding section morphology function The mechanism attribution rules are as follows: Paddy fields and wetland retention areas are classified as linear reservoirs with exponential decay. Its parameters Corresponding to the water storage capacity and outlet flow coefficient of this type of underlying surface; dryland and forest / grassland belong to the power-law decay form. ,in The power-law form directly corresponds to the nonlinear decay relationship of the drainage rate with the remaining storage in Horton's runoff theory; the decay constant of the rapid drainage section of hardened ground. The lower limit of its value range is higher than that of the other four types of underlying surfaces, and the inflection point time is... The upper limit of the value range is lower than that of the other four categories, in order to characterize the characteristic that the drainage of hardened ground has basically no tail.
3. The method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage based on underlying surface type as described in claim 1, characterized in that, The concentration-flow hysteresis loop characteristic Defined as: ; Integral along the receding water stage - Trajectory detour, This indicates that the hysteresis direction is clockwise, corresponding to a flushing-limited release. This indicates counterclockwise rotation, corresponding to dissolution-restricted release. The hysteresis direction and normalized area are both incorporated into the constraints of the joint inversion objective function.
4. The method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage based on underlying surface type as described in claim 1, characterized in that, The first Pollutant release concentration process of underlying surface Based on the release mechanism, it can be classified as: scour-restricted type. Characterizing the release of surface-mobile pollutants by runoff and the exponential decay of their concentrations as water recedes; or dissolved-limited pollutants. This characterizes the approximately equal concentration of pollutants carried out of the reservoir by the outflow; the mechanism and parameters are described. , The concentration was determined by measured concentration curves during the controlled flow test phase.
5. The method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage based on underlying surface type as described in claim 1, characterized in that, The response kernel benchmark parameters The calibration steps include: conducting small-scale controlled runoff generation tests repeatedly for each type of underlying surface under no less than three gradients of artificially simulated rainfall intensity, and measuring the discharge phase flow curves and pollutant concentration curves for each set of replicates; solving the following fitting problem for each set of data to obtain single-test parameter estimates: ; By summing up the estimated values of each group, the 10th to 90th percentile of their distribution is taken as the baseline parameter range for this type of underlying surface under the current watershed climate conditions. .
6. The method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage based on underlying surface type as described in claim 1, characterized in that, The preceding condition correction function Attenuation constant of rapid water receding section The correction satisfies: That is, the lower the soil moisture index in the early stage, the better the corrected soil moisture index. The smaller the value, the more pronounced the slow receding tail; the monotonic relationship corresponds to the physical mechanism in Horton's formula where the infiltration capacity increases as soil moisture decreases, leading to a decrease in the slope runoff ratio and thus prolonging the receding time. Correction function coefficients The controlled runoff generation experiment was repeated under at least two prior soil moisture conditions and the results were determined by fitting.
7. The method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage based on underlying surface type as described in claim 1, characterized in that, Also includes: After the contribution weight is solved, the composition ratio of underlying surface types in each sub-region within the upstream area of the target monitoring section is obtained. Based on the contribution weight and the linear proportional relationship between the area ratio of underlying surface types and the contribution weight, the spatial distribution of pollutant load in each sub-region is estimated.
8. The method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage based on underlying surface type as described in claim 1, characterized in that, Also includes: Before solving the joint inversion objective function, calculate the underlying surface for any two types of surfaces. , The corresponding correlation coefficients between the unit retreat response kernels after correction are as follows: ; in , These are the time averages of the corresponding response kernel sequences; Using all underlying surface categories as nodes and correlation coefficients as parameters... Exceeding the preset threshold An undirected graph is constructed with the category pairs as edges. All surface categories in the same connected component of the undirected graph are merged into a joint category to participate in the objective function solution. The output of the merged joint contribution weight is the sum of the contribution weights of each category in the connected component. The contribution weights of categories in the same connected component are no longer solved separately. The contribution weights of surface categories that do not fall into any connected component are still solved separately.
9. The method for tracing agricultural non-point source pollution events based on the waveform fingerprint of underlying surface type drainage according to claim 8, characterized in that, When the pollutant concentration time series When missing or marked as unreliable, the degenerate joint inversion objective function is used for solving: ; In the degraded objective function, only those functions satisfying the correlation coefficient are considered. The contribution weights of the underlying surface categories are calculated separately, while the remaining categories are processed according to the merging rules.
10. The method for tracing agricultural non-point source pollution events based on the waveform fingerprint of drainage based on underlying surface type as described in claim 1, characterized in that, Also includes: The optimal solution is obtained by solving the joint inversion objective function. Then, based on the local curvature of the objective function at the optimal solution, the confidence interval of the contribution weight is estimated: ; in Let the joint inversion objective function be... for about The second derivative matrix at the optimal solution The standard normal quantile corresponds to the confidence level; the output source tracing results include the contribution ratio of various underlying surfaces and their confidence intervals.
