A self-organizing neural network-based similar watershed identification and parameter transplantation method

Abstract

The present application relates to the technical field of unsupervised machine learning, and discloses a similar basin identification and parameter transplantation method based on a self-organizing neural network, which comprises the following steps: obtaining the basin characteristic data of a target basin and a reference basin, and standardizing the data; using self-organizing neural network competition learning to form a neuron mapping relationship; when there is no reference basin in the target optimal neuron, forming a neighborhood expansion branch along the adjacent neurons, and determining the end point characteristic difference and local incremental difference of the candidate reference basin; stopping the branch from continuing to expand when the local incremental difference exceeds a threshold value, and retaining the candidate reference basin when the local incremental difference does not exceed the threshold value and the end point characteristic difference meets the requirements; forming a similar reference basin set according to the retained candidate reference basin, and generating the transplantation hydrological parameters of the target basin. The present application reduces the probability of a hydrological property mutation basin entering the parameter source set, and improves the stability of similar basin identification and the reliability of hydrological parameter transplantation.

Description

Technical Field

[0001] This invention relates to the field of unsupervised machine learning technology, and more specifically, to a method for similar watershed identification and parameter transfer based on self-organizing neural networks. Background Technology

[0002] In hydrological simulation and forecasting of watersheds with little or no data, due to the lack of long-term runoff observation data that can be directly used for hydrological model calibration, it is usually necessary to identify reference watersheds with similar hydrological characteristics by utilizing watershed attributes such as topography, meteorology, and underlying surface of watersheds with available data. Then, the parameters of the target watershed are derived using the hydrological model parameters that have already been calibrated in the reference watersheds. With the increasing availability of watershed attribute data, relying solely on geographical proximity or manual selection of a few attributes is no longer sufficient to meet the needs of multidimensional watershed feature processing. Using data analysis methods such as clustering and machine learning to classify the features of multiple watersheds and identify similar watersheds accordingly has become an important technical approach for regionalizing hydrological parameters in watersheds without available data.

[0003] In the prior art, Chinese invention patent application CN103488871A, "A Flood Forecasting Method for Areas Without Runoff Data," clusters watersheds within a candidate range based on watershed characteristics related to hydrological model parameters. If the number of similar watersheds is insufficient, the search range is expanded until the required number is met. Under this approach, the suitability of candidate watersheds as parameter sources depends primarily on the final comprehensive characteristic relationships. When watershed attributes are unevenly distributed within the candidate range, and different attribute combinations produce similar comprehensive evaluation results, the set of similar watersheds is easily affected by changes in the search range and candidate sample composition. This can lead to watersheds with significantly different hydrological response characteristics entering the parameter source set, thereby increasing the uncertainty of the hydrological model parameter transfer results.

[0004] Therefore, it is necessary to design a similar watershed identification and parameter transfer method based on self-organizing neural networks to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes a similar watershed identification and parameter transfer method based on self-organizing neural networks, which aims to solve the problem that watersheds lacking data are easily mixed with watersheds with abrupt changes in hydrological attributes when expanding the search range of reference watersheds, resulting in unstable parameter sources.

[0006] This invention proposes a method for similar watershed identification and parameter transfer based on a self-organizing neural network, comprising: Obtain watershed feature data for the target watershed and reference watersheds, and standardize them to obtain watershed feature vectors; The watershed feature vector is input into a self-organizing neural network for machine learning, and competitive learning is used to determine the winning neurons and form a neuron mapping relationship. The target winning neuron is determined based on the neuron mapping relationship; when there is no reference flow domain in the target winning neuron, an expansion level is determined outward along the adjacent neurons and a neighborhood expansion branch is formed, and the reference flow domain mapped by the neurons in the neighborhood expansion branch is determined as the candidate reference flow domain. The endpoint feature differences between the candidate reference watershed and the target watershed, as well as their local incremental differences with the preceding reference watersheds in the same neighborhood extension branch, are determined. If there is no preceding reference watershed, the target watershed is used as the comparison object. If the local incremental difference is greater than the local incremental threshold, the neighborhood extension branch is stopped from extending outward from the current candidate reference watershed level. If the local incremental difference is not greater than the local incremental threshold and the endpoint feature difference is not greater than the endpoint difference threshold, the current candidate reference watershed is retained. A set of similar reference watersheds is formed based on the retained candidate reference watersheds, and the transplanted hydrological parameters of the target watershed are generated using the hydrological model parameters of the set of similar reference watersheds.

[0007] Furthermore, when forming neighborhood expansion branches, the process includes: neurons directly adjacent to the target winning neuron forming the first expansion layer; neurons directly adjacent to the previous expansion layer but not assigned to a determined expansion layer forming the next expansion layer; and connecting adjacent neurons in an ascending order of layers.

[0008] Furthermore, the determination of the preceding reference domain includes: selecting the reference domain whose expansion level of the neuron in the same neighborhood expansion branch is smaller than the expansion level of the neuron in the current candidate reference domain and whose level difference is the smallest; when there are more than two reference domains, determining the endpoint feature differences between each reference domain and the target domain, and determining the reference domain with the smallest endpoint feature difference as the preceding reference domain.

[0009] Furthermore, when determining local incremental differences, the following are included: The watershed feature vector includes topographic feature components, meteorological feature components, and underlying surface feature components; Three types of local incremental differences are determined based on three types of feature components; if any type of local incremental difference is greater than the local incremental threshold of that type, it is determined to exceed the limit; if none of the three types of local incremental differences are greater than their respective local incremental thresholds, it is determined not to exceed the limit.

[0010] Furthermore, the determination of the local increment threshold includes: classifying the pattern clustering categories according to the neuron mapping relationship, forming a reference flow domain pair by forming two reference flow domains that are mapped to adjacent neurons and belong to the same pattern clustering category; and determining the maximum value of each of the three types of local increment differences as the local increment threshold of that type.

[0011] Furthermore, the blocking method of the neighborhood expansion branch is as follows: when the local incremental difference exceeds the limit, the neuron where the current candidate reference watershed is located is determined as the blocking start point, the search along the neighborhood expansion branch from the blocking start point is stopped, and the selection of candidate reference watersheds from the blocking start point in the neighborhood expansion branch and neurons whose expansion level is greater than the expansion level where the blocking start point is located is stopped.

[0012] Furthermore, the handling method when the endpoint feature difference exceeds the limit is as follows: when the endpoint feature difference is greater than the endpoint difference threshold and the local incremental difference does not exceed the limit, the current candidate reference watershed is excluded, and the neighboring extended branch is kept to expand outward.

[0013] Furthermore, the filtering method when the number of retained candidate reference watersheds exceeds a preset number is as follows: For each retained candidate reference watershed, the ratio of various local increment differences of its neighboring extended branches up to the extended level of the candidate reference watershed to the threshold of that local increment is determined, and the maximum value of the obtained ratio is determined as the maximum increment ratio of the path. The paths are selected from smallest to largest based on the maximum increment ratio. If the maximum increment ratios of the paths are the same, they are selected from smallest to largest based on the differences in endpoint characteristics.

[0014] Furthermore, the processing method for multi-branch neurons is as follows: When the same neuron is located in more than two neighborhood expansion branches, determine whether each neighborhood expansion branch segment from the target superior neuron to the neuron contains a blocking origin. When all neighboring extended branch segments contain the blocking origin, the reference watershed in that neuron is not selected; If at least one neighboring extended branch segment does not contain the blocking origin, the reference watershed in that neuron is identified as a candidate reference watershed.

[0015] Furthermore, the expansion method when the number of retained candidate reference watersheds has not reached the preset number is as follows: The expansion level is increased along the unblocked neighborhood expansion branch; expansion stops when the number of retained candidate reference regions reaches a preset number; expansion stops when the preset number is not reached and there are no adjacent neurons in each unblocked neighborhood expansion branch that can be included in the next expansion level, and a set of similar reference regions is formed with the currently retained candidate reference regions.

[0016] Compared with existing technologies, the advantages of this invention are as follows: By standardizing the watershed feature data of the target watershed and the reference watersheds and inputting them into a self-organizing neural network, competitive learning is used to form a neuron mapping relationship. When the target winning neuron lacks a reference watershed, the parameter source is no longer directly selected based solely on the overall similarity result obtained after expanding the search range. Instead, a neighborhood expansion branch with hierarchical relationship is formed along adjacent neurons. At the same time, the endpoint feature differences between the candidate reference watershed and the target watershed, as well as the local incremental differences between the candidate reference watershed and the preceding reference watershed, are used to distinguish and judge the similarity of the candidate reference watershed itself and the continuity of hydrological attributes in the search direction. When the local incremental difference exceeds a threshold, the neighborhood is stopped in time. The extended branch continues to expand outwards, avoiding the inclusion of outer reference basins in the parameter source even after crossing regions of abrupt changes in hydrological attributes during the search range expansion. Candidate reference basins are retained only when local changes remain continuous and endpoint feature differences meet the requirements. This ensures that the set of similar reference basins not only possesses overall similarity at the target basin level but also maintains continuity of local hydrological attributes along the search path. This reduces the interference of similar comprehensive evaluation results formed by different attribute combinations and changes in candidate sample composition on the identification results of similar basins, improving the stability and consistency of hydrological attributes in reference basin selection. It also reduces parameter bias caused by unsuitable reference basins participating in hydrological model parameter transplantation, enhancing the reliability and applicability of transplanting hydrological parameters to basins lacking data. Attached Figure Description

[0017] The accompanying drawings are used to illustrate the technical solutions of the present invention and do not constitute a limitation on the scope of protection of the present invention. The process nodes, regional relationships, and functional block diagrams in the drawings are used to illustrate the implementation methods of the embodiments of the present invention. In the drawings: Figure 1 A flowchart illustrating the method for similar watershed identification and parameter transfer based on a self-organizing neural network provided in this embodiment of the invention; Figure 2 A flowchart for candidate reference watershed determination and neighborhood expansion branch control based on local incremental differences is provided for embodiments of the present invention. Figure 3 A schematic diagram illustrating the mapping and neighborhood expansion of the target watershed and the reference watershed in a self-organizing neural network, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the determination and blocking of local incremental differences in a neighboring extended branch, as provided in an embodiment of the present invention. Detailed Implementation

[0018] The present application will be described below with reference to specific embodiments. It should be noted that the target watershed in this application is a watershed lacking long-term complete runoff observation data and requiring hydrological model parameter transplantation; the reference watershed is a watershed with complete observation data and whose hydrological model parameters have been calibrated; the target winning neuron is the winning neuron of the target watershed after mapping through a self-organizing neural network. The expansion level is determined from the inside out according to the adjacency distance of each neuron relative to the target winning neuron; the neighborhood expansion branch is the search path formed by connecting adjacent neurons in the increasing direction of the expansion level. The preceding reference watershed is the reference watershed located inside the current candidate reference watershed and whose expansion level is closest to the current candidate reference watershed's expansion level within the same neighborhood expansion branch. Endpoint feature differences are used to characterize the overall watershed attribute differences between the candidate reference watershed and the target watershed; local incremental differences are used to characterize the degree of local hydrological attribute changes between adjacent reference watersheds along the expansion direction of the same neighborhood expansion branch. The above terms have the same meaning in the following embodiments.

[0019] In some embodiments of this application, see Figure 1-4 As shown, this application proposes a method for similar watershed identification and parameter transfer based on a self-organizing neural network, including: S100: Obtain watershed feature data for the target watershed and reference watersheds, and standardize them to obtain watershed feature vectors.

[0020] S200: Input the watershed feature vector into the self-organizing neural network for machine learning, use competitive learning to determine the winning neurons and form a neuron mapping relationship.

[0021] S300: Determine the target winning neuron based on the neuron mapping relationship. When there is no reference flow domain within the target winning neuron, determine the expansion level outward along the adjacent neurons and form a neighborhood expansion branch. The reference flow domain mapped by the neurons in the neighborhood expansion branch is determined as the candidate reference flow domain.

[0022] S400: Determine the endpoint feature differences between the candidate reference watershed and the target watershed, as well as the local incremental differences between the candidate watershed and its preceding reference watershed in the same neighborhood extension branch. If there is no preceding reference watershed, the target watershed is used as the comparison object. If the local incremental difference is greater than the local incremental threshold, stop the neighborhood extension branch from the current candidate reference watershed's extension level outwards. If the local incremental difference is not greater than the local incremental threshold and the endpoint feature differences are not greater than the endpoint difference threshold, retain the current candidate reference watershed.

[0023] S500: Based on the retained candidate reference watersheds, a set of similar reference watersheds is formed, and the hydrological model parameters of the similar reference watershed set are used to generate the transplanted hydrological parameters of the target watershed.

[0024] Specifically, watershed characteristic data for the target and reference watersheds are acquired. This watershed characteristic data includes at least topographic features, meteorological features, and underlying surface features. Topographic features describe the watershed's geomorphology and runoff conditions; meteorological features describe the watershed's water input and energy conditions; and underlying surface features describe precipitation infiltration, runoff generation, and evapotranspiration conditions. In one specific implementation, topographic features may include average slope, watershed area, shape coefficient, and channel gradient; meteorological features may include average annual precipitation, average annual evaporation, average annual temperature, and the concentration of rainfall during the flood season; and underlying surface features may include vegetation cover, soil composition, land use composition, soil saturated hydraulic conductivity, and soil water holding capacity. Meteorological features and underlying surface features with temporal variability are statistically analyzed using the same statistical period. Topographic features utilize basic geographic data consistent with the spatial scope of the target and reference watersheds to reduce additional discrepancies introduced by inconsistencies in the statistical period and spatial baseline.

[0025] When preprocessing watershed feature data, missing and outlier data are processed first, and then each feature is standardized. Specifically, range standardization can be used, where for the same watershed feature, the minimum value among the watershed samples used in training is used as the lower bound, and the difference between the maximum and minimum values ​​is used as the scale to transform the values ​​of each watershed for that feature to a unified numerical range. The target watershed and the reference watershed use the same standardization benchmark, without establishing separate scales. In this embodiment, the watershed feature vectors formed after standardization have the same component order. Subsequent endpoint feature differences and local incremental differences are directly calculated based on the watershed feature vectors, without replacing the watershed feature vectors with neuron weight vectors for watershed difference evaluation. If the maximum and minimum values ​​of a certain watershed feature are the same across all training watersheds, that watershed feature is not included in the construction of subsequent watershed feature vectors.

[0026] The feature vectors of all watersheds are input into a two-dimensional self-organizing neural network for training. Each neuron in the self-organizing neural network has a neuron weight vector with the same dimension as the watershed feature vector. In one specific implementation, when the number of watersheds participating in the training is 15, the competition layer uses a 4×4 neuron grid, with an initial learning rate set to 0.5, gradually decreasing to 0.05 during training. The initial neighborhood radius is set to 2, gradually decreasing to 1. The number of training iterations is set to 1000, and training ends after 1000 iterations.

[0027] For each input watershed feature vector, the distance between the feature vector and the weight vectors of each neuron is calculated. The neuron with the smallest distance is identified as the winning neuron, and the weight vectors of the winning neuron and its neighboring neurons are adjusted synchronously according to the neighborhood range set during training. As the number of training iterations increases, the learning amplitude and neighborhood range are gradually reduced until the set number of training iterations is reached, at which point training ends. After training, each target watershed and reference watershed is mapped to the winning neuron closest to its respective watershed feature vector, thus establishing a neuron mapping relationship.

[0028] For a target watershed, the target winning neuron is determined based on the neuron mapping relationship. When a reference watershed is already mapped within the target winning neuron, candidate screening can be directly performed on that reference watershed. When there is no reference watershed within the target winning neuron, neighborhood expansion is initiated. Preferably, the eight-neighborhood relationship is determined using a two-dimensional neuron grid. In this embodiment, the same eight-neighborhood relationship is used for determining the expansion level, forming neighborhood expansion branches, classifying pattern clusters, and forming reference watershed pairs, without changing the neuron adjacency criteria in subsequent processing. Neurons directly adjacent to the target winning neuron constitute the first expansion level. After excluding the target winning neuron and neurons already assigned to smaller expansion levels, the remaining neurons directly adjacent to the first expansion level constitute the second expansion level, and subsequent expansion levels are determined accordingly. Thus, the expansion level of a neuron actually reflects the minimum number of expansions required to reach the neuron from the target winning neuron through direct adjacency relationships.

[0029] Figure 3 The adjacency relationship between the first and second expansion layers is illustrated using a local neuron grid centered on the target winning neuron. The number of neurons in the illustration is only used to illustrate the neighborhood expansion relationship and does not represent the actual grid size of the competitive layer of the self-organizing neural network. In the actual expansion process, when a neuron located at the boundary of the competitive layer does not have a complete neighborhood, the corresponding expansion layer is determined by the actual adjacent neurons in the competitive layer.

[0030] After determining the expansion levels, the directly adjacent neurons of the target winning neuron are used as the starting positions of different neighborhood expansion branches. Connections are only allowed from neurons at smaller expansion levels to directly adjacent neurons at larger expansion levels, thus forming neighborhood expansion branches extending from the inside out. When an outer layer neuron can be directly adjacent to different inner layer neurons, all neighborhood expansion branch relationships to which that outer layer neuron belongs are retained, without pre-defining it as a unique branch, so that the continuity of local hydrological attributes in different arrival directions can be determined separately later. Each reference watershed mapped by neurons in the neighborhood expansion branches is determined as a candidate reference watershed. When multiple reference watersheds exist in the same neuron, each reference watershed participates in subsequent judgments as an independent candidate reference watershed.

[0031] Specifically, this application does not directly use all reference watersheds close to the target winning neuron in the self-organizing neural network as parameter sources. Instead, when the target winning neuron is empty and the search must be expanded outwards, it separately judges whether the candidate reference watershed itself is similar to the target watershed and whether the expansion direction to the candidate reference watershed still maintains the continuity of local hydrological attributes. Endpoint feature differences are used to determine whether the current candidate reference watershed has the conditions to be directly used as a parameter source for the target watershed, and local incremental differences are used to determine whether the current neighborhood expansion branch has crossed the position of abrupt change in hydrological attributes. The two differences generate candidate reference watershed selection and neighborhood expansion branch control results respectively, thereby avoiding confusing "the current candidate is not similar enough" with "the entire search direction has lost continuity" into the same treatment.

[0032] The endpoint feature difference is determined based on the standardized watershed feature vectors of the target watershed and the current candidate reference watershed. In one specific calculation method, the feature components at the same position in the feature vectors of the two watersheds are obtained sequentially, the difference between each component is calculated, the squares of each difference are summed, and the square root is taken as the endpoint feature difference. Since all features have been standardized in the aforementioned steps, different physical dimensions will not directly dominate the distance. The endpoint difference threshold is determined based on the training samples of the reference watershed. Reference watersheds belonging to the same pattern cluster category and containing two or more reference watersheds within that pattern cluster category are paired to form sample pairs. The endpoint feature difference of each sample pair is calculated, and the maximum value of the obtained endpoint feature difference is determined as the endpoint difference threshold. The endpoint difference threshold is recalculated after the reference watershed training samples are increased or the self-organizing neural network is retrained. If the current training samples cannot form any sample pairs, it is determined that the current reference watershed training samples are insufficient, and the self-organizing neural network is retrained after supplementing the reference watersheds.

[0033] Pattern clustering categories are determined based on neuron mapping relationships and neuron adjacency relationships. Neurons mapped to the reference watershed are used as neurons to be classified, and the aforementioned eight-neighbor relationship is used to determine whether neurons to be classified are directly adjacent. Neurons to be classified that can reach each other through one or more direct adjacency relationships are assigned to the same pattern clustering category, while neurons to be classified that cannot reach each other through direct adjacency relationships are assigned to different pattern clustering categories. Pattern clustering categories are only used for grouping training samples of the reference watershed and do not change the neuron mapping relationships already formed in the self-organizing neural network.

[0034] Local incremental differences are used to reflect changes in adjacent hydrological attributes along the expansion direction of the same neighborhood extension branch. For the current candidate reference watershed, reference watersheds that have already appeared in the same neighborhood extension branch are searched towards the target winning neuron. The reference watershed with the smallest expansion level difference (lower than the current candidate reference watershed's expansion level) is selected as the preceding reference watershed. If there are more than two reference watersheds in this expansion level, the endpoint feature differences between each reference watershed and the target watershed are determined, and the reference watershed with the smallest endpoint feature difference is selected as the preceding reference watershed. When the endpoint feature differences are the same, the preceding reference watershed is determined according to the pre-recorded reference watershed data numbering order. If the current candidate reference watershed is the first reference watershed to appear in this neighborhood extension branch, there is no preceding reference watershed, and the target watershed is directly used as the comparison object. This ensures that each local incremental difference reflects the actual attribute changes from the known position inside the current search branch to the current candidate position, rather than using reference watersheds in other neighborhood extension branches as comparison objects.

[0035] When calculating local incremental differences, the standardized watershed feature vector is divided into three categories based on topographic, meteorological, and underlying surface features. For each category, the difference between the current candidate watershed and the previous candidate watershed is calculated for each feature component. These differences are squared, summed, and their square roots are taken to obtain the local incremental differences in topography, meteorology, and underlying surface. These three categories of local incremental differences are not weighted and canceled out; instead, they are compared separately with their respective local incremental thresholds. This is because a significant change in any of the topographic, meteorological, or underlying surface attributes can alter the hydrological response. If the three categories of differences are first combined into a single average result, the smaller differences in the other two categories might reduce the influence of the abruptly changed feature in the overall result.

[0036] The local increment thresholds are determined by the reference watershed samples after the self-organizing neural network training is completed. Specifically, two reference watersheds that are mapped to directly adjacent neurons and belong to the same pattern cluster category are selected from all pattern cluster categories to form reference watershed pairs. When adjacent neurons each have multiple reference watersheds, the reference watersheds in the two neurons are paired one by one. The topographic local increment difference, meteorological local increment difference, and underlying surface local increment difference of each reference watershed pair are calculated separately, and the maximum value among the calculation results is taken as the topographic local increment threshold, meteorological local increment threshold, and underlying surface local increment threshold, respectively. Since these reference watershed pairs are located in the same pattern cluster category and are in directly adjacent neurons, their maximum local change is used to represent the upper bound of the allowable change between adjacent watersheds that are still considered to have the same hydrological attribute pattern in the training samples. After the reference watershed samples are updated or the self-organizing neural network is retrained, reference watershed pairs are re-formed and the three types of local increment thresholds are recalculated, so that the thresholds are adjusted according to the actual attribute distribution of the current watershed samples, rather than using fixed empirical values ​​that are universal for different study areas.

[0037] When a certain pattern cluster category cannot form a reference watershed pair mapped to a directly adjacent neuron, that pattern cluster category will not participate in the local incremental threshold statistics. When all pattern cluster categories cannot form a reference watershed pair, it is determined that the current reference watershed training samples are insufficient. After supplementing the reference watershed samples, the self-organizing neural network training and neuron mapping are re-performed, and the candidate reference watershed selection for the current target watershed is not continued.

[0038] During the evaluation, if any type of local increment difference in the current candidate reference watershed exceeds the local increment threshold for that type, it is determined that the current neighboring extended branch has experienced a local attribute mutation at the location of the current candidate reference watershed. Only when all three types of local increment differences are not greater than their respective local increment thresholds is it determined that the local increment difference is not exceeded. When the local increment difference equals the local increment threshold, it is classified as not exceeding the limit.

[0039] When a local attribute mutation occurs, the neuron containing the current candidate reference watershed is designated as the blocking origin of the neighborhood expansion branch. Reference watersheds within the blocking origin are not selected, and the search along the neighborhood expansion branch from the blocking origin to larger expansion levels is stopped. Candidate reference watersheds located inside the blocking origin and already retained are not deleted due to subsequent local attribute mutations. Therefore, the blocking result only applies to the mutation location and the search range outside the mutation location, and does not negate watershed attribute relationships that were previously continuous at the mutation location.

[0040] When the local incremental difference is within the limit, the determination of whether to retain the current candidate reference watershed is based on the endpoint feature difference. If the endpoint feature difference is not greater than the endpoint difference threshold, the current candidate reference watershed is retained, and the current neighboring extension branch is allowed to continue expanding outward. If the endpoint feature difference is greater than the endpoint difference threshold, only the current candidate reference watershed is excluded, but the neighboring extension branch is still allowed to continue expanding outward. This is because the overall attribute difference between the current candidate reference watershed and the target watershed is large, which only indicates that the current candidate reference watershed is not suitable for directly providing parameters, and cannot be used to conclude that there must be a sudden change in local attributes in the outer layer in this direction. Conversely, regardless of whether the endpoint feature difference exceeds the limit, as long as any type of local incremental difference exceeds the limit, the neighboring extension branch is blocked first, and the search for the outermost reference watershed in this direction is no longer conducted.

[0041] When the same neuron is located in more than two neighborhood expansion branches, the arrival relationship from the target superior neuron to the neuron is preserved for each neighborhood expansion branch, and local increment judgment is performed separately. If all neighborhood expansion branches have formed blocking origins before reaching the neuron, the neuron cannot be reached through the locally continuous neighborhood expansion branches, and no candidate reference domain is selected from the neuron. If at least one neighborhood expansion branch has not yet formed a blocking origin, the search qualification of the neighborhood expansion branch is preserved, and the reference domain in the neuron is further judged through the unblocked branch. When the same candidate reference domain has more than two unblocked neighborhood expansion branches, the maximum path increment ratio is determined according to the ratio of the local increment difference of each neighborhood expansion branch up to the current expansion level of the candidate reference domain to its respective local increment threshold. The neighborhood expansion branch with the smallest maximum path increment ratio is determined as the effective neighborhood expansion branch, and subsequent path continuity evaluation is performed along the effective neighborhood expansion branch. When the maximum increment ratio of paths is the same, the effective neighbor expansion branch is determined according to the pre-fixed traversal order of the neighbor expansion branch.

[0042] When the number of retained candidate reference watersheds exceeds a preset number, instead of directly selecting the top few based on endpoint feature differences, the entire expansion process from the target winning neuron to each candidate reference watershed is further evaluated. For an effective neighborhood expansion branch, the local incremental difference calculated up to the current candidate reference watershed's expansion level is divided by the local incremental threshold for that type. The resulting ratio represents the degree of closeness of the local change to the upper bound of the allowed change. The maximum value among all ratios is taken as the maximum incremental ratio of the path. When the local incremental threshold for a certain type is 0, if the current local incremental difference for that type is greater than 0, then according to the aforementioned local attribute mutation criterion, the local incremental difference for that type is directly determined to be out of bounds and is not included in the maximum incremental ratio of the path. If the current local incremental difference for that type is equal to 0, then the ratio of the local incremental difference for that type to the local incremental threshold is recorded as 0.

[0043] A smaller path maximum increment ratio indicates that the most significant local attribute change in the extended branch of that neighborhood is still far from the abrupt change boundary. A predetermined number of candidate reference watersheds are selected based on the path maximum increment ratio from smallest to largest. If the path maximum increment ratios are the same, then selection is based on the endpoint feature differences from smallest to largest. When both the path maximum increment ratio and endpoint feature differences are the same, candidate reference watersheds with smaller extension levels are prioritized. If the extension levels are still the same, selection is based on a pre-fixed reference watershed data numbering order. This avoids a candidate reference watershed from masking a significant local attribute change that occurred during its journey to the target watershed simply because it is close to the final endpoint of the target watershed.

[0044] The preset number is set based on the number of reference samples required for hydrological parameter transplantation. In a preferred embodiment, the preset number is set to 5. When the number of retained candidate reference basins is less than the preset number, the expansion level is increased only along the neighboring expansion branches that have not yet formed a blocking point. The expansion search stops after the preset number is reached. If none of the unblocked neighboring expansion branches have adjacent neurons that can be included in the next expansion level, and the number of retained candidate reference basins is still less than the preset number, the expansion stops, and a similar reference basin set is formed with all currently retained candidate reference basins. This prevents the reintroduction of outer reference basins across already identified local attribute mutation locations in order to force a fixed number to be reached.

[0045] In a specific application, a sub-basin of a watershed was selected as the research object, setting up 10 reference watersheds with long-term observation data and 5 target watersheds lacking data. For the reference watersheds, watershed characteristic data were formed using meteorological statistics, underlying surface statistics, and topographic data from 2003 to 2018, and hydrological model parameters were calibrated using concurrent runoff observation data. For the target watersheds, watershed characteristic data were formed according to the same characteristic terms and statistical caliber. Short-term runoff observation data from 2019 to 2021 were only used for independent validation after parameter transplantation and did not participate in the determination of endpoint difference thresholds and local increment thresholds.

[0046] In specific implementation, the topographic, meteorological, and underlying surface features of 10 reference watersheds and one target watershed to be processed are first processed according to the method described in the above embodiment to obtain standardized watershed feature vectors with consistent dimensions. Then, a self-organizing neural network is trained to form neuron mapping relationships. Suppose the target watershed is mapped to a target winning neuron that does not contain reference watersheds, then an expansion hierarchy and neighborhood expansion branches are established outward from the target winning neuron according to eight neighborhoods. For each neuron mapped to a reference watershed, the preceding reference watershed is determined according to the above embodiment, and the three types of local incremental differences in topography, meteorology, and underlying surface are calculated respectively, and local attribute mutation judgment is performed.

[0047] For example, when the endpoint feature difference between the outer candidate reference watershed and the target watershed in a certain neighborhood expansion branch is still within the endpoint difference threshold, but the local incremental difference of the underlying surface relative to the preceding reference watershed has exceeded the local incremental threshold, even if the outer candidate reference watershed is still a relatively close sample in terms of endpoint distance from the target watershed, the neuron containing the outer candidate reference watershed will be used as the blocking starting point, and the search along that neighborhood expansion branch will be stopped. At the same time, neighborhood expansion branches where the local incremental differences in topography, meteorology, and underlying surface are all within the limit will continue to expand. This results in the selection of "reference watersheds with closer endpoint distances being excluded, while reference watersheds with slightly larger endpoint distances are retained because their expansion branches remain continuous." This result is the main implementation difference between this application and methods that only rank based on final similarity.

[0048] After forming a set of similar reference watersheds, the parameters of the same hydrological model that have been calibrated in each reference watershed within the set are extracted. In one specific parameter transplantation method, the Xin'anjiang hydrological model can be used to transplant parameters related to runoff generation, confluence, and groundwater. For each parameter of the same name in the hydrological model, the reciprocal of the endpoint feature difference of each similar reference watershed is taken. The proportion of each reciprocal to the sum of all reciprocals is used as the parameter transplantation weight. Then, the parameters of the same name are weighted and summed according to the parameter transplantation weight to obtain the initial transplanted hydrological parameters of the target watershed. If the endpoint feature difference of a certain similar reference watershed is 0, it means that the reference watershed is consistent with the target watershed in the adopted standardized feature space, and the corresponding parameters of the reference watershed can be directly adopted. When there are multiple reference watersheds with endpoint feature differences of 0, the average value of the parameters of the same name in these reference watersheds is taken. Finally, the initial transplanted hydrological parameters are checked according to the parameter physical range specified by the adopted hydrological model to obtain the transplanted hydrological parameters of the target watershed.

[0049] When short-term observed runoff data is available for the target watershed, the transplanted hydrological parameters can be input into the adopted hydrological model for independent runoff simulation. The simulation results are then compared with the short-term observed runoff data to verify the parameter transplantation results. This verification process is preferably used only for evaluating the implementation effect and does not reverse the already determined local incremental thresholds, blocking initiation points, and neighboring extension branches, thereby avoiding the need to readjust the core reference watershed search criteria of this embodiment using the target watershed verification results.

[0050] This embodiment does not simply expand or shrink the uniform search radius, but rather allows different neighboring expansion branches to stop or continue according to changes in hydrological properties along their routes. Therefore, a local mutation in one direction will not force all directions to terminate simultaneously, and a neuron being unreachable via an abnormal branch will not exclude a reference watershed that can still be reached via another continuous branch. Simultaneously, once a blocking point is formed, subsequent expansion of the expansion hierarchy will not allow reference watersheds outside the blocking point to re-enter the candidate set, making the set of similar reference watersheds highly stable to changes in the search hierarchy. When screening for an excessive number of candidates using the maximum path increment ratio, the reference watershed least close to the mutation boundary during the entire expansion process can be prioritized, rather than simply those with smaller final endpoint distances. Therefore, this embodiment reduces jumps in the reference watershed set caused by changes in the search radius and reduces false exclusions due to single path anomalies, making it more effective in data-scarce watershed scenarios with uneven spatial distribution of reference samples.

[0051] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for similar watershed identification and parameter transfer based on self-organizing neural networks, characterized in that, include: Obtain watershed feature data for the target watershed and reference watersheds, and standardize them to obtain watershed feature vectors; The watershed feature vector is input into a self-organizing neural network for machine learning, and competitive learning is used to determine the winning neurons and form a neuron mapping relationship. The target winning neuron is determined based on the aforementioned neuron mapping relationship; When there is no reference flow domain within the target winning neuron, an expansion level is determined outward along the adjacent neurons and a neighborhood expansion branch is formed. The reference flow domain mapped by the neurons in the neighborhood expansion branch is determined as the candidate reference flow domain. The endpoint feature differences between the candidate reference watershed and the target watershed, as well as their local incremental differences with the preceding reference watersheds in the same neighborhood extension branch, are determined. If there is no preceding reference watershed, the target watershed is used as the comparison object. If the local incremental difference is greater than the local incremental threshold, the neighborhood extension branch is stopped from extending outward from the current candidate reference watershed level. If the local incremental difference is not greater than the local incremental threshold and the endpoint feature difference is not greater than the endpoint difference threshold, the current candidate reference watershed is retained. A set of similar reference watersheds is formed based on the retained candidate reference watersheds, and the transplanted hydrological parameters of the target watershed are generated using the hydrological model parameters of the set of similar reference watersheds.

2. The method according to claim 1, characterized in that, When forming a neighborhood expansion branch, the process includes: neurons directly adjacent to the target winning neuron forming the first expansion layer; neurons directly adjacent to the previous expansion layer but not yet assigned to a determined expansion layer forming the next expansion layer; and connecting adjacent neurons in an ascending order of layers.

3. The method according to claim 1, characterized in that, The determination of the preceding reference watershed includes: The reference region is selected where the expansion level of the neuron in the same neighborhood expansion branch is smaller than the expansion level of the neuron in the current candidate reference region and the level difference is the smallest. When there are more than two reference regions, the endpoint feature differences between each reference region and the target region are determined, and the reference region with the smallest endpoint feature difference is determined as the preceding reference region.

4. The method according to claim 3, characterized in that, Determining local incremental differences includes: The watershed feature vector includes topographic feature components, meteorological feature components, and underlying surface feature components; Three types of local incremental differences are determined based on three types of feature components; if any type of local incremental difference is greater than the local incremental threshold of that type, it is determined to exceed the limit; if none of the three types of local incremental differences are greater than their respective local incremental thresholds, it is determined not to exceed the limit.

5. The method according to claim 4, characterized in that, The determination of the local increment threshold includes: classifying the pattern clustering categories according to the neuron mapping relationship, forming a reference flow field pair by forming two reference flow fields that are mapped to adjacent neurons and belong to the same pattern clustering category; and determining the maximum value of each of the three types of local increment differences as the local increment threshold of that type.

6. The method according to claim 4, characterized in that, The blocking method of the neighborhood expansion branch is as follows: when the local incremental difference exceeds the limit, the neuron where the current candidate reference watershed is located is determined as the blocking start point, the search along the neighborhood expansion branch from the blocking start point is stopped, and the selection of candidate reference watersheds from the blocking start point in the neighborhood expansion branch and neurons whose expansion level is greater than the expansion level of the blocking start point is stopped.

7. The method according to claim 4, characterized in that, The handling method when the endpoint feature difference exceeds the limit is as follows: when the endpoint feature difference is greater than the endpoint difference threshold and the local incremental difference does not exceed the limit, the current candidate reference watershed is excluded, and the neighboring extended branch continues to expand outward.

8. The method according to claim 4, characterized in that, The screening method when the number of retained candidate reference watersheds exceeds a preset number is as follows: For each retained candidate reference watershed, the ratio of various local increment differences of its neighboring extended branches up to the extended level of the candidate reference watershed to the threshold of that local increment is determined, and the maximum value of the obtained ratio is determined as the maximum increment ratio of the path. The paths are selected from smallest to largest based on the maximum increment ratio. If the maximum increment ratios of the paths are the same, they are selected from smallest to largest based on the differences in endpoint characteristics.

9. The method according to claim 6, characterized in that, The processing method for multi-branch neurons is as follows: When the same neuron is located in more than two neighborhood expansion branches, determine whether each neighborhood expansion branch segment from the target superior neuron to the neuron contains a blocking origin. When all neighboring extended branch segments contain the blocking origin, the reference watershed in that neuron is not selected; If at least one neighboring extended branch segment does not contain the blocking origin, the reference watershed in that neuron is identified as a candidate reference watershed.

10. The method according to claim 6, characterized in that, The expansion method when the number of retained candidate reference watersheds has not reached the preset number is as follows: The expansion level is increased along the unblocked neighborhood expansion branch; expansion stops when the number of retained candidate reference regions reaches a preset number; expansion stops when the preset number is not reached and there are no adjacent neurons in each unblocked neighborhood expansion branch that can be included in the next expansion level, and a set of similar reference regions is formed with the currently retained candidate reference regions.

Images