A Deep Learning-Based Method for Dynamic Evaluation of Gulf Ecosystems

Abstract

This invention relates to the field of deep learning technology, specifically to a deep learning-based method for dynamic evaluation of bay ecosystems. In this invention, a graph attention network participates in the node weight update process. Nodes of water quality, biological community, and ecological pressure are no longer entered into the scoring calculation with uniform intensity. Node weights can be differentiated based on adjacency association strength. A convolutional neural network participates in the node adjacency matrix calculation process. The cumulative results of connection weights, difference results, product results, and local neighborhood convolutional responses are simultaneously entered into the node relationship table. The promoting, inhibiting, synergistic, and lagging effects between indicators can unfold along the neighborhood connection chain. The ecological relationship propagation graph structure is no longer limited to static connection records. The bay ecosystem dynamic evaluation result set simultaneously retains score results, relationship results, and path results, and the sources, propagation directions, and key action chains of ecological state changes have stronger correspondence.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and in particular to a deep learning-based method for dynamic evaluation of bay ecosystems. Background Technology

[0002] The field of deep learning technology aims to build multi-layer neural network models. Through nonlinear mapping, parameter learning, and error backpropagation mechanisms, these models perform representation learning, pattern recognition, state discrimination, relationship modeling, and result prediction on high-dimensional input data. They establish a mapping relationship between input variables and output targets from the original data, extract discriminative representation vectors through multi-layer network structures, and output computational results for classification, regression, detection, segmentation, prediction, and decision-making tasks.

[0003] The purpose of this deep learning-based method for dynamic evaluation of bay ecosystems is to construct a mapping model between state variables of the bay ecosystem and the results of comprehensive ecological evaluation. By performing modeling and calculation on the nonlinear relationships between multiple indicators of the bay ecosystem and continuously evaluating the ecological status based on time series changes, a dynamic change assessment mechanism for the bay ecosystem is formed. This provides quantitative basis for the formulation of bay ecological protection plans, the evaluation of the effectiveness of ecological restoration measures, and ecological management decisions, and improves the ability to model multivariate relationships in complex ecosystems and identify ecological change trends.

[0004] The shortcomings of existing technologies lie in their operational methods, which tend to establish an overall mapping between input variables and output targets. While they can handle high-dimensional data and output classification, regression, and prediction results, for multi-source, multi-temporal, and multi-relational coupled scenarios such as bay ecological assessment, the common calculation process tends to focus on one-time input and one-time output. The continuity between time points, historical memory of scoring, and propagation paths between nodes are often compressed into a unified representation space, leading to evaluation conclusions that rely on overall fitting results. The process basis is not easily expanded, and when multiple indicators work together, fixed weights or overall parameter learning methods tend to mix tidal changes, water salinity fluctuations, biological community changes, and ecological pressure disturbances. The sudden changes of a single indicator and... Differences between multiple indicators undergoing coordinated changes are difficult to separate, and existing technologies for time series processing are mostly focused on continuous input discrimination, lacking specific organization for differences between adjacent nodes, level boundary updates, and path cumulative relationships. Samples in the level critical interval are easily mixed with adjacent states, and continuous fluctuations at multiple time points are easily averaged. In scenarios such as seasonal alternation, sudden pollution, and restoration transition periods, the state transition position is not easy to identify stably. Once multiple indicators fluctuate synchronously, different sea areas interfere with each other, and historical states continue to affect the current state, the conclusions are likely to be characterized by existing results, ambiguous sources, coarse boundaries, and unclear paths, affecting ecological protection planning, restoration timing judgment, and identification of key control points. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a deep learning-based method for dynamic evaluation of bay ecosystems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a deep learning-based method for dynamic evaluation of bay ecology, comprising the following steps:

[0007] S1: Based on the bay ecological index sequence data, read the tide level, salinity, dissolved oxygen, plankton density and chlorophyll concentration sequences, align them by time index and calculate the difference between adjacent nodes, write the historical scores to the tail of the matrix, and generate an ecological state vector matrix.

[0008] S2: Based on the ecological state vector matrix, extract the node numbers of water quality, biological community and ecological pressure and establish an index. Use graph attention network to read the node weights and perform incremental superposition, deduction and original value copying. Multiply the updated weights with the ecological state vector matrix row by row and accumulate the score to generate a score weight configuration rule table.

[0009] S3: Based on the ecological state vector matrix and the scoring weight configuration rule table, extract the phytoplankton, benthic organisms and nutrient index vectors, perform sample pairing and distance calculation, update the level center vector and perform distance comparison, write the time difference into the change sequence, and obtain the ecological state embedded vector sequence.

[0010] S4: Based on the ecological state embedding vector sequence and the scoring weight configuration rule table, establish the node numbers for dissolved oxygen, salinity, phytoplankton, and ecological pressure, construct the adjacency index matrix, and use a convolutional neural network to perform weight accumulation, difference, and product propagation calculations and filter neighboring nodes to obtain the ecological relationship propagation graph structure.

[0011] S5: Based on the ecological relationship propagation graph structure, extract the path of tidal level, nutrients, and biological community nodes and calculate the cumulative weight. Sort the paths according to the weight, read the state vector and perform scoring mapping with the scoring weight configuration rule table to obtain the dynamic evaluation result set of the bay ecology.

[0012] As a further embodiment of the present invention, the ecological state vector matrix includes tidal level change index, salinity change index, dissolved oxygen change index, phytoplankton density index, chlorophyll concentration index, and historical score identifiers; the score weight configuration rule table includes water quality index weight items, biological community index weight items, ecological pressure index weight items, and weight adjustment identifier items; the ecological state embedded vector sequence includes phytoplankton index vector, benthic organism index vector, nutrient index vector, and ecological level center vector; the ecological relationship propagation graph structure includes dissolved oxygen nodes, salinity nodes, phytoplankton nodes, ecological pressure nodes, and node relationship adjacency matrix; and the bay ecological dynamic evaluation result set includes ecological score sequence, key relationship path sequence, and ecological state expression vector.

[0013] As a further aspect of the present invention, the specific steps for generating the ecological state vector matrix are as follows:

[0014] Based on the bay ecological indicator sequence data, the sequences of tide level, salinity, dissolved oxygen, plankton density and chlorophyll concentration are sorted according to a unified time index, missing time points are filled in the sequence positions, the difference between adjacent nodes is calculated and written into the change column, and an indicator time series matrix is ​​generated.

[0015] Based on the indicator time series matrix, historical scores are written into the last column in chronological order, indicator category labels are written into the identifier column, and the matrix positions are rearranged according to the time rows and indicator columns to generate an ecological state vector matrix.

[0016] As a further aspect of the present invention, the specific steps for generating the scoring weight configuration rule table are as follows:

[0017] Based on the ecological state vector matrix, water quality node numbers, biological community node numbers, and ecological pressure node numbers are extracted, written into the node index table according to the node category, duplicate numbers are merged, placeholders are written for missing numbers, and node records are rearranged according to category order to generate a node index set.

[0018] Based on the node index set, a graph attention network is used to read the weight values ​​and adjustment flags of each node. The values ​​corresponding to the boost flags are written into the increment column, the values ​​corresponding to the decrement flags are written into the decrease column, and the values ​​corresponding to the keep flags are written into the retain column. Then, the updated weights are summarized in the order of node numbers and written into a unified weight table to obtain the node weight matrix.

[0019] Based on the node weight matrix, the corresponding row data in the ecological state vector matrix is ​​called, and the node weight value is multiplied with the vector value of each row one by one. The product results under the same node category are accumulated in the row, the accumulated results are written into the scoring column, and the scoring records are arranged in chronological order to generate a scoring weight configuration rule table.

[0020] As a further aspect of the present invention, the graph attention network first reads the corresponding weight values, adjustment identifiers, node category numbers, and adjacency association records of water quality nodes, biological community nodes, and ecological pressure nodes based on the node index set. It then concatenates the weight values, node category numbers, and node connection direction codes of individual nodes into a node input vector. For any pair of node input vectors with a connection relationship, it performs a linear mapping to obtain source node representation vectors and target node representation vectors. The source node representation vectors, target node representation vectors, and connection direction codes are then concatenated sequentially and input into the attention scoring unit, outputting the attention score for each node pair. Normalization is performed on the attention scores of all source nodes associated with the same target node to obtain the attention coefficients of neighboring nodes. These neighboring node attention coefficients are multiplied by the corresponding source node weight values ​​to obtain weighted node weight values. The weighted node weight values ​​are then allocated to the incremental, decremental, and retained columns according to the adjustment identifier. The incremental column values ​​are added to the original weight values, the decremental column values ​​are subtracted from the original weight values, and the retained column values ​​are copied and written to the original weight values. Finally, the weights are summarized and updated according to the node number order and written into a unified weight table to form a node weight matrix.

[0021] As a further aspect of the present invention, the specific steps for generating the ecological state embedding vector sequence are as follows:

[0022] Based on the ecological state vector matrix and the scoring weight configuration rule table, phytoplankton, benthic organisms and nutrient index vectors are extracted, samples are paired according to the grade labels, the distance between samples is calculated, and a grade distance table is generated according to the grade set center value.

[0023] Based on the aforementioned distance table, the sample vectors are compared with the level centers item by item. The level corresponding to the first distance value is written into the identifier, and the difference between adjacent time points is written into the change column to obtain the ecological state embedded vector sequence.

[0024] As a further aspect of the present invention, the calculation of the distance between samples involves reading the phytoplankton index vector, benthic organism index vector, and nutrient index vector from the ecological state vector matrix. The three indices are combined into a multidimensional sample vector at the same time point. A dimension-wise difference operation is performed on two sample vectors to obtain the difference between the phytoplankton index, the difference between the benthic organism index, and the difference between the nutrient index. The three differences are squared and summed. Then, the square root operation is performed on the summed value to obtain the distance between samples. The distance value is recorded in the sample pairing record row and written into the grade distance table along with the sample number and grade label.

[0025] As a further aspect of the present invention, the specific steps for generating the ecological relationship propagation graph structure are as follows:

[0026] Based on the ecological state embedding vector sequence and the scoring weight configuration rule table, the dissolved oxygen node number, salinity node number, phytoplankton node number and ecological pressure node number are extracted, written into the node list according to the node source, the same type of nodes are grouped and arranged, the cross-type nodes are sequentially numbered, and the corresponding relationship markers between nodes are retained to generate an ecological node list.

[0027] Based on the ecological node list, connection records are established in pairs according to the node number. The starting node, target node, connection direction and connection weight are written into the adjacency list. Empty value markers are filled in for positions where no connection is established. Multiple connection records of the same node are written in parallel. The matrix structure is arranged according to the connection direction to obtain the node adjacency matrix.

[0028] Based on the node adjacency matrix, a convolutional neural network is used to read the connection weights between nodes, calculate the cumulative weights of adjacent nodes, the weight difference between nodes, and the weight product of nodes, and write the three types of results into the node relationship table. Then, according to the weight size rule, low-association nodes are filtered out and effective neighboring nodes are retained to generate an ecological relationship propagation graph structure.

[0029] As a further aspect of the present invention, the convolutional neural network first reads the node connection weights based on the node adjacency matrix, extracts the corresponding row vector and column vector of each node in the order of node number, concatenates the row vector, column vector, connection direction encoding, and connection distance encoding into a relational input tensor, and uses a sliding window to scan the relational input tensor grid by grid along the node number direction and the adjacency position direction, performs convolution kernel multiplication and addition operations on the connection weights within each window, outputs local relational response values, and performs addition operations on the connection weights of the previous node and the next node within the same window. After obtaining the weight accumulation result, the weight difference result is obtained by subtracting the weight of the next node from the weight of the previous node connection. The weight product result is obtained by multiplying the weight of the previous node connection and the weight of the next node connection. The local relation response value, weight accumulation result, weight difference result and weight product result are then written into the node relation table. The relation response values ​​of all neighbors of each node in the node relation table are sorted. Low-association node records are deleted according to the preset retention quantity and response threshold. The remaining node connection records are retained and the neighborhood connection table is reconstructed to generate the ecological relation propagation graph structure.

[0030] As a further aspect of the present invention, the specific steps for generating the bay ecological dynamic evaluation result set are as follows:

[0031] Based on the ecological relationship propagation graph structure, the connection paths between tidal level, nutrients and biological community nodes are extracted, the path nodes are unfolded in hierarchical order, and multiple connection weights are accumulated to generate a path weight sequence.

[0032] Based on the path weight sequence, the path weights are arranged in descending order, the state vectors are read and matched with the scoring weight configuration rule table item by item, and the mapping scores and path records are written to obtain the dynamic evaluation result set of the bay ecosystem.

[0033] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0034] In this invention, the graph attention network participates in the node weight update process. The three types of nodes, namely water quality, biological community and ecological pressure, no longer enter the scoring calculation according to the same intensity. The node weight can be differentiated by combining the adjacency association intensity. The incremental superposition, reduction deduction and original value replication have more fine-grained adjustment capabilities. The scoring weight configuration rule table corresponds more closely with the changes in ecological state.

[0035] In this invention, after sample pairing, distance calculation, rank center update and distance comparison work together, the three indices of phytoplankton, benthic organisms and nutrients no longer remain directly mapped to the original values. The rank boundary is constrained by the distance relationship between vectors and the center aggregation relationship. The critical state has higher distinguishability. After the difference value of continuous time points is written into the change sequence, the interpretation of the state transition direction is more coherent.

[0036] In this invention, the convolutional neural network participates in the node adjacency matrix calculation process. The cumulative result of connection weights, the difference result, the product result, and the local neighborhood convolution response are simultaneously entered into the node relationship table. The promoting, inhibiting, synergistic, and lagging effects between indicators can be unfolded along the neighborhood connection chain, and the structure of the ecological relationship propagation graph is no longer limited to static connection records.

[0037] In this invention, path extraction, weight accumulation, descending sorting, and score mapping continue to write the influence links between tide level, nutrients, and biological communities into the result record. The bay ecological dynamic evaluation result set simultaneously retains the score results, relationship results, and path results, and the sources, propagation directions, and key action chains of ecological status changes have stronger correspondence. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the main steps of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0040] Example 1

[0041] Please see Figure 1 This invention provides a technical solution: a deep learning-based method for dynamic evaluation of bay ecosystems, comprising the following steps:

[0042] S1: Based on the bay ecological index sequence data, read the tide level, salinity, dissolved oxygen, plankton density and chlorophyll concentration sequences, align them by time index and calculate the difference between adjacent nodes, write the historical scores to the tail of the matrix, and generate an ecological state vector matrix.

[0043] S2: Based on the ecological state vector matrix, extract the node numbers of water quality, biological community and ecological pressure and establish an index. Use graph attention network to read the node weights and perform incremental superposition, deduction and original value copying. Multiply the updated weights with the ecological state vector matrix row by row and accumulate the score to generate a score weight configuration rule table.

[0044] S3: Based on the ecological state vector matrix and the scoring weight configuration rule table, extract the vectors of phytoplankton, benthic organisms and nutrient index, perform sample pairing and distance calculation, update the level center vector and perform distance comparison, write the time difference into the change sequence, and obtain the ecological state embedded vector sequence.

[0045] S4: Based on the ecological state embedding vector sequence and the scoring weight configuration rule table, establish the node numbers for dissolved oxygen, salinity, phytoplankton, and ecological pressure, construct the adjacency index matrix, and use a convolutional neural network to perform weight accumulation, difference, and product propagation calculations and filter neighboring nodes to obtain the ecological relationship propagation graph structure.

[0046] S5: Based on the ecological relationship propagation graph structure, extract the path of tidal level, nutrients, and biological community nodes and calculate the cumulative weight. Sort the paths according to the weight, read the state vector and perform scoring mapping with the scoring weight configuration rule table to obtain the dynamic evaluation result set of the bay ecology.

[0047] The ecological state vector matrix includes tidal level change indicators, salinity change indicators, dissolved oxygen change indicators, phytoplankton density indicators, chlorophyll concentration indicators, and historical score identifiers. The score weight configuration rule table includes water quality index weight items, biological community index weight items, ecological pressure index weight items, and weight adjustment identifier items. The ecological state embedded vector sequence includes phytoplankton index vectors, benthic organism index vectors, nutrient index vectors, and ecological level center vectors. The ecological relationship propagation graph structure includes dissolved oxygen nodes, salinity nodes, phytoplankton nodes, ecological pressure nodes, and node relationship adjacency matrix. The bay ecological dynamic evaluation result set includes ecological score sequences, key relationship path sequences, and ecological state expression vectors.

[0048] The specific steps for generating the ecological state vector matrix are as follows:

[0049] Based on the bay ecological indicator sequence data, the sequences of tide level, salinity, dissolved oxygen, plankton density and chlorophyll concentration are sorted according to a unified time index, missing time points are filled in the sequence positions, the difference between adjacent nodes is calculated and written into the change column, and an indicator time series matrix is ​​generated.

[0050] Based on the indicator time series matrix, historical scores are written into the last column in chronological order, indicator category labels are written into the identifier column, and the matrix positions are rearranged according to the time row and indicator column to generate an ecological state vector matrix.

[0051] Based on the bay ecological indicator sequence data, a dynamic time warping algorithm and cubic spline interpolation method were used to align the time indices of tidal level, salinity, dissolved oxygen, plankton density, and chlorophyll concentration sequences. The baseline sampling interval was set to 1 hour, the time window length to 24, and the path constraint bandwidth to 3. First, five sets of sequence index tables were established in ascending order of timestamps. Then, a distance matrix was constructed using the absolute difference. The first row and first column were written with the cumulative cost value. Alignment was completed by selecting the position with the smaller cumulative value along three search paths: right shift, down shift, and right-down shift. Missing time points were filled in. For single missing points, cubic spline interpolation was used. For two or three consecutive missing points, four adjacent time points were fitted and interpolated. Then, the value of the next time point was subtracted from the value of the previous time point for adjacent nodes. The differences in tidal level, salinity, dissolved oxygen, plankton density, and chlorophyll concentration were written into the change column to generate the indicator time series matrix.

[0052] Based on the index time series matrix, a sliding window writing method and a column-major order matrix rearrangement method are used to write historical scores into the last column in chronological order. The window length is set to 6 and the step size is set to 1. First, the score values ​​corresponding to six consecutive time nodes are read and filled into the last column in chronological order. Then, index category labels are written, with tidal level labeled as 01, salinity labeled as 02, dissolved oxygen labeled as 03, plankton density labeled as 04, and chlorophyll concentration labeled as 05. Subsequently, the rows and columns are rearranged in column-major order, with the time rows remaining in ascending order. The index column order is fixed as tidal level column, salinity column, dissolved oxygen column, plankton density column, chlorophyll concentration column, change column, historical score column, and category label column. The rearranged matrix is ​​then checked for empty spaces and indexed and rewritten column by column to generate an ecological state vector matrix.

[0053] The specific steps for generating the scoring weight configuration rule table are as follows:

[0054] Based on the ecological state vector matrix, water quality node numbers, biological community node numbers, and ecological pressure node numbers are extracted, written into the node index table according to the node category, duplicate numbers are merged, placeholders are written for missing numbers, and node records are rearranged according to category order to generate a node index set.

[0055] Based on the node index set, a graph attention network is used to read the weight values ​​and adjustment flags of each node. The values ​​corresponding to the boost flags are written into the increment column, the values ​​corresponding to the decrement flags are written into the decrease column, and the values ​​corresponding to the keep flags are written into the retain column. Then, the updated weights are summarized in the order of node number and written into a unified weight table to obtain the node weight matrix.

[0056] Based on the node weight matrix, the corresponding row data in the ecological state vector matrix is ​​called, and the node weight value is multiplied with the vector value of each row one by one. The product results under the same node category are accumulated in the row, the accumulated results are written into the scoring column, and the scoring records are arranged in chronological order to generate a scoring weight configuration rule table.

[0057] Based on the ecological state vector matrix, a rule-based numbering mapping algorithm and a hash merging algorithm are used to extract water quality node numbers, biological community node numbers, and ecological pressure node numbers. The water quality node number range is set to 1001 to 1999, the biological community node number range is set to 2001 to 2999, and the ecological pressure node number range is set to 3001 to 3999. First, the category marker column values ​​are read according to the matrix column number. Then, the category marker column values ​​and column numbers are concatenated to generate an eight-bit node code. The node's category is then written into the node index table. The index table fields are fixed in the following order: node code column, category code column, original column number column, duplicate marker column, and placeholder marker column. Duplicate numbers are compared using hash buckets. The number of hash buckets is set to 257. Conflict handling uses a linear probing step size of 1. Duplicate node codes are written to the same bucket and a 1 is written to the duplicate marker column. For missing numbers, the positions are scanned bit by bit according to the number range and a 9 is written to the placeholder marker column. Finally, the node records are rearranged according to the ascending order of the category code and the ascending order of the original column number to generate a node index set.

[0058] Based on a node index set, a graph attention network is used to read the corresponding weight values ​​and adjustment flags for each node. The input dimension is set to 4, and the input items are, in order, node weight value, category code value, connection direction value, and node degree value. The number of attention heads is set to 4, and the output dimension of each head is set to 8. The weight matrix has 4 rows and 8 columns, and the activation slope is set to 0.2. First, the original weight columns and adjustment flag columns of the unified weight table are read in the order of node codes. Then, the adjacent node numbers and connection direction codes are obtained by querying the adjacency record table using the node codes. The current node input vector and the adjacent node input vectors are multiplied by the weight matrix to obtain the mapping direction. The system calculates the attention coefficients for each adjacent node by performing vector concatenation, dot product of the scoring vectors, and exponential normalization. Then, it multiplies the attention coefficients by the weight values ​​of the adjacent nodes and sums them one by one to obtain the weighted node weight values. Subsequently, it performs column-by-column writing based on the adjustment flags. The boost flag is written to the incremental column with three levels of incremental coefficients: 0.05, 0.10, and 0.15. The debuff flag is written to the decrement column with three levels of decrement coefficients: 0.05, 0.10, and 0.15. The retain flag is written to the retain column with a retain coefficient of 1.00. Finally, the updated weights are summarized and written to a unified weight table in the order of node numbers to obtain the node weight matrix.

[0059] Based on the node weight matrix, a matrix multiplication-addition calculation method and a grouped cumulative sorting method are used to retrieve the data corresponding to the rows of the ecological state vector matrix. The row numbers of the node weight matrix correspond one-to-one with the column numbers of the ecological state vector matrix. First, the value blocks of the ecological state vector matrix are read in chronological order. Then, the node weight values ​​are multiplied item by item with the vector values ​​in the same column. The product is retained to four decimal places, and the product results are accumulated within the same node category. The cumulative columns for water quality category are set to 1 to 32, for biological community category to 33 to 64, and for ecological pressure category to 65 to 96. Subsequently, the cumulative values ​​of the three categories are multiplied and added using a base scoring coefficient of 0.50, a penalty scoring coefficient of 0.30, and a compensation scoring coefficient of 0.20. The accumulated results are written to the scoring column, and the scoring records are sorted in ascending order of timestamp and descending order of score value to generate a scoring weight configuration rule table.

[0060] The graph attention network first reads the weight values, adjustment identifiers, node category numbers, and adjacency association records of water quality nodes, biological community nodes, and ecological pressure nodes based on the node index set. It then concatenates the weight value, node category number, and node connection direction code of a single node into a node input vector. For any pair of connected node input vectors, it performs a linear mapping to obtain the source node representation vector and the target node representation vector. These vectors are then concatenated sequentially with the connection direction code and input into the attention scoring unit, outputting the attention score for each node pair. Normalization is performed on the attention scores of all source nodes associated with the same target node to obtain the attention coefficients of neighboring nodes. These neighboring node attention coefficients are multiplied by the corresponding source node weight values ​​to obtain weighted node weight values. These weighted node weight values ​​are then allocated to the incremental, decremental, and retention columns according to the adjustment identifier. The incremental column values ​​are added to the original weight values, the decremental column values ​​are subtracted from the original weight values, and the retention column values ​​are copied and written to the original weight values. Finally, the weights are summarized and updated according to the node number order and written to a unified weight table to form a node weight matrix.

[0061] The specific steps for generating the ecological state embedding vector sequence are as follows:

[0062] Based on the ecological state vector matrix and the scoring weight configuration rule table, phytoplankton, benthic organisms and nutrient index vectors are extracted, samples are paired according to the level labels, the distance between samples is calculated, and the level distance table is generated according to the level set center value.

[0063] Based on the rank distance table, the sample vector is compared with the rank center item by item. The rank corresponding to the first rank of the distance value is written into the identifier, and the difference between adjacent time points is written into the change column to obtain the ecological state embedding vector sequence.

[0064] To calculate the distance between samples, the phytoplankton index vector, benthic organism index vector, and nutrient index vector are read from the ecological state vector matrix. The three indices are combined into a multidimensional sample vector at the same time point. The dimensional difference operation is performed on two sample vectors to obtain the difference between the phytoplankton index, the difference between the benthic organism index, and the difference between the nutrient index. The three differences are squared and summed. Then, the square root operation is performed on the summed value to obtain the distance between samples. The distance value is recorded in the sample pairing record row and written into the grade distance table together with the sample number and grade label.

[0065] Based on the ecological state vector matrix and the scoring weight configuration rule table, a dual-branch coding network, Euclidean distance calculation formula, and moving mean aggregation method are used to extract phytoplankton, benthic organism, and nutrient index vectors. The dimensions of the phytoplankton index vector, benthic organism index vector, and nutrient index vector are all set to 16. The three vectors are concatenated column-wise to form a 48-dimensional sample vector. The sample vectors are paired according to the level labels in the scoring weight configuration rule table. The pairing rules are defined as pairing adjacent time points of the same level and pairing across three time points of the same level. The maximum number of pairings is set to 64 pairs per level. Then, the Euclidean distance calculation formula is used to perform dimension-wise subtraction, squaring, summation, and square root operations on each pair of sample vectors. The distance value is retained to six decimal places. The level aggregation center value is calculated. The center value aggregation window is set to 32 consecutive samples of the same level. The center value is calculated by dimension-wise summation and then divided by 32. The sample number, level number, distance value, center value number, and time number are written into the header of the table in a fixed five-column record table to generate a level distance table.

[0066] Based on the rank distance table, a prototype matching method, distance sorting and writing rules, and first-order difference calculation method are used to compare sample vectors with rank centers item by item. The sample vector dimension is maintained at 48 dimensions, and the number of rank centers is set to 4 groups. Each group of rank centers is stored in 48-dimensional order. First, a single sample vector and the corresponding elements of the 4 groups of rank centers are read, and dimension-wise subtraction, square accumulation, and square root operation are performed to obtain 4 center distance values. Then, the distance values ​​are sorted in ascending order, and the rank number corresponding to the first sorted value is written into the identifier field with a length of 2 bits. Subsequently, the first-order difference calculation method is used to read sample vectors at adjacent time points, and the values ​​of the previous time point and the values ​​of the next time point in the same dimension are subtracted. The difference results are written into the change column in 48-dimensional order with a field width of 48 columns. The sample number, rank number, identifier, change column, and time number are rearranged and recorded in ascending order of time to obtain the ecological state embedded vector sequence.

[0067] The specific steps for generating the ecological relationship propagation graph structure are as follows:

[0068] Based on the ecological state embedding vector sequence and the scoring weight configuration rule table, the dissolved oxygen node number, salinity node number, phytoplankton node number and ecological pressure node number are extracted, written into the node list according to the node source, the same type of nodes are grouped and arranged, the cross-type nodes are sequentially numbered, and the corresponding relationship markers between nodes are retained to generate an ecological node list.

[0069] Based on the list of ecological nodes, connection records are established in pairs according to the node number. The starting node, target node, connection direction and connection weight are written into the adjacency list. Empty value markers are filled in for positions where no connection is established. Multiple connection records of the same node are written in parallel. The matrix structure is arranged according to the connection direction to obtain the node adjacency matrix.

[0070] Based on the node adjacency matrix, a convolutional neural network is used to read the connection weights between nodes, calculate the cumulative weight of adjacent nodes, the weight difference between nodes, and the weight product of nodes, and write the three types of results into the node relationship table. Then, according to the weight size rule, low-association nodes are filtered out and effective neighboring nodes are retained to generate the ecological relationship propagation graph structure.

[0071] Based on the ecological state embedding vector sequence and the scoring weight configuration rule table, the dissolved oxygen node number, salinity node number, phytoplankton node number, and ecological pressure node number are extracted using the type number mapping method and segmented sequential encoding rule. The starting value of the dissolved oxygen node number is set to 4001, the salinity node number to 5001, the phytoplankton node number to 6001, and the ecological pressure node number to 7001. First, the node source marker is read according to the column order of the ecological state embedding vector sequence. Then, the source marker and column number are used to perform decimal concatenation and write the node encoding column. The node list is then written, and the fields of the node list are set in the following order: node number column, source number column, time number column, group number column, and correspondence column. Nodes of the same type are grouped and arranged in ascending order of source number and time number. Nodes of different types are sequentially numbered according to the source number segment and column number segment. The correspondence marker is written, and the length of the correspondence marker is set to 6 digits. The first two digits are the source number, the middle two digits are the time number, and the last two digits are the group number. This generates the ecological node list.

[0072] Based on the ecological node list, a dual-graph structure table construction method and directional coding writing rules are adopted to establish connection records for each pair of node numbers. The connection table for nodes at the same time records the connections between nodes at the current time, and the connection table for nodes at historical time records the connections from the previous time node to the current time node. The starting node field is set to 8 bits, the target node field is set to 8 bits, and the connection direction code is set to four categories: 01, 02, 03, and 04. The initial value of the connection weight is read according to the node weight corresponding to the scoring weight configuration rule table, and 6 decimal places are retained. The adjacency table is written in a fixed order: starting node column, target node column, connection direction column, connection weight column, and null value identifier column. Multiple connection records for the same node are written in parallel in ascending order of target node number. Null value identifier 99 is filled in for positions where no connection is established. Then, the matrix structure is organized using the row-first writing rule and the directional coding partitioning rule, where the horizontal axis is arranged in ascending order of target node number and the vertical axis is arranged in ascending order of starting node number, resulting in a node adjacency matrix.

[0073] Based on the node adjacency matrix, a convolutional neural network and edge weight update method are used to read the connection weights between nodes. The input channel count is set to 4, and these four channels are sequentially written into the connection weight matrix, direction encoding matrix, time offset matrix, and source number matrix. Three convolutional layers are used: the first layer has a 3x3 kernel size, a stride of 1, padding of 0, and 16 output channels; the second layer has a 3x3 kernel size and 32 output channels; and the third layer has a 1x1 kernel size and 8 output channels. A local window is first read along the node number direction and the time number direction, and then the connection weights within the window are updated. The convolution kernel is multiplied and added, and the results of the cumulative weights of adjacent nodes, the difference between weights of nodes, and the product of weights of nodes are calculated. These three types of results, along with the convolution output value, are written into a node relationship table. The fields of the node relationship table are set in the following order: cumulative column, difference column, product column, convolution response column, node number column, and neighborhood number column. Then, the connection weights are rewritten according to the convolution response value using the edge weight update method. The update coefficients are set to three levels: 0.10, 0.20, and 0.30. Threshold filtering is performed according to the weight size, with the lower threshold set to 0.15 and the number of neighbors retained set to 8 for each node, generating an ecological relationship propagation graph structure.

[0074] The convolutional neural network first reads the node connection weights based on the node adjacency matrix. Then, it extracts the corresponding row and column vectors for each node in order of node number. The row vectors, column vectors, connection direction encoding, and connection distance encoding are concatenated into a relational input tensor. A sliding window scans the relational input tensor grid by grid along the node number direction and the adjacency position direction. For each window, a convolution kernel multiplication-addition operation is performed on the connection weights, outputting the local relational response value. The connection weights of the preceding and following nodes within the same window are added to obtain the weight accumulation result. The weight difference result is obtained by subtracting the weight of the following node from the previous node's connection weight. Finally, the connection weights of the preceding and following nodes are multiplied to obtain the weight product result. The local relational response value, the weight accumulation result, the weight difference result, and the weight product result are then written into the node relational table. All relational response values ​​in the neighborhood of each node in the node relational table are sorted. Low-association node records are deleted according to a preset retention quantity and response threshold. The remaining node connection records are retained, and the neighborhood connection table is reconstructed to generate the ecological relational propagation graph structure.

[0075] The specific steps for generating the dynamic evaluation result set of the bay's ecosystem are as follows:

[0076] Based on the structure of the ecological relationship propagation graph, the connection paths between tidal level, nutrients and biological community nodes are extracted, the path node order is unfolded according to the hierarchy, and multiple connection weights are accumulated to generate a path weight sequence.

[0077] Based on the path weight sequence, the path weights are arranged in descending order, the state vectors are read and matched with the scoring weight configuration rule table item by item, and the mapping scores and path records are written to obtain the dynamic evaluation result set of the bay ecosystem.

[0078] Based on the ecological relationship propagation graph structure, a depth-first traversal algorithm and a path multiplication accumulation rule are used to extract the connection paths of tidal nodes, nutrient nodes, and biological community nodes. The starting node is limited to the tidal node number segment, and the ending node is limited to the biological community node number segment. The maximum expansion level is set to 6, the backtracking stack length is set to 6, and the access flag length is set to the total number of nodes. First, all external edges of tidal nodes are read in the adjacency list order. Then, the next target node number is written layer by layer using a recursive stack, and the path nodes are expanded sequentially. When a visited node is encountered, a backtracking flag is written. When a terminal node is encountered, the expansion stops and the complete path is recorded. Then, the connection weight of each segment is read in the path order. The first connection weight is written to the starting position of the cumulative column, and the subsequent connection weights are multiplied and accumulated in sequence. The cumulative result is retained to six decimal places, and the path number, level number, and cumulative weight are written to generate a path weight sequence.

[0079] Based on the path weight sequence, a quicksort algorithm and a linear mapping calculation formula are used to sort the path weights in descending order. The benchmark value is selected from the tail path weights, the starting position of the left pointer is set as the first row, and the starting position of the right pointer is set as the tail row. The comparison rule is that values ​​greater than the benchmark value are written to the left partition, and values ​​less than the benchmark value are written to the right partition. After the path sorting is completed, the time row corresponding to the state vector matrix and the weight column corresponding to the scoring weight configuration rule table are read. Then, the linear mapping calculation formula is used to perform item-by-item mapping. The state vector value is multiplied by the weight coefficient to obtain the individual mapping value. The individual mapping value is written to the mapping score column in chronological order. After sorting, the path number, path weight, mapping score, and time number are written together into the path record table. The record rows are rearranged and the number is repeated to obtain the dynamic evaluation result set of the bay ecosystem.

[0080] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for evaluating ecological dynamics of a bay based on deep learning, characterized in that, Includes the following steps: S1: Based on the bay ecological index sequence data, read the tide level, salinity, dissolved oxygen, plankton density and chlorophyll concentration sequences, align them by time index and calculate the difference between adjacent nodes, write the historical scores to the tail of the matrix, and generate an ecological state vector matrix. S2: Based on the ecological state vector matrix, extract the node numbers of water quality, biological community and ecological pressure and establish an index. Use graph attention network to read the node weights and perform incremental superposition, deduction and original value copying. Multiply the updated weights with the ecological state vector matrix row by row and accumulate the score to generate a score weight configuration rule table. S3: Based on the ecological state vector matrix and the scoring weight configuration rule table, extract the phytoplankton, benthic organisms and nutrient index vectors, perform sample pairing and distance calculation, update the level center vector and perform distance comparison, write the time difference into the change sequence, and obtain the ecological state embedded vector sequence. S4: Based on the ecological state embedding vector sequence and the scoring weight configuration rule table, establish the node numbers for dissolved oxygen, salinity, phytoplankton, and ecological pressure, construct the adjacency index matrix, and use a convolutional neural network to perform weight accumulation, difference, and product propagation calculations and filter neighboring nodes to obtain the ecological relationship propagation graph structure. S5: Based on the ecological relationship propagation graph structure, extract the path of tidal level, nutrients, and biological community nodes and calculate the cumulative weight. Sort the paths according to the weight, read the state vector and perform scoring mapping with the scoring weight configuration rule table to obtain the dynamic evaluation result set of the bay ecology. 2.The method of claim 1, wherein, The ecological state vector matrix includes tidal level change indicators, salinity change indicators, dissolved oxygen change indicators, phytoplankton density indicators, chlorophyll concentration indicators, and historical score identifiers. The score weight configuration rule table includes water quality index weight items, biological community index weight items, ecological pressure index weight items, and weight adjustment identifier items. The ecological state embedded vector sequence includes phytoplankton index vectors, benthic organism index vectors, nutrient index vectors, and ecological level center vectors. The ecological relationship propagation graph structure includes dissolved oxygen nodes, salinity nodes, phytoplankton nodes, ecological pressure nodes, and node relationship adjacency matrix. The bay ecological dynamic evaluation result set includes ecological score sequences, key relationship path sequences, and ecological state expression vectors.

3. The deep learning-based dynamic evaluation method for bay ecosystems according to claim 1, characterized in that, The specific steps for generating the ecological state vector matrix are as follows: Based on the bay ecological indicator sequence data, the sequences of tide level, salinity, dissolved oxygen, plankton density and chlorophyll concentration are sorted according to a unified time index, missing time points are filled in the sequence positions, the difference between adjacent nodes is calculated and written into the change column, and an indicator time series matrix is ​​generated. Based on the indicator time series matrix, historical scores are written into the last column in chronological order, indicator category labels are written into the identifier column, and the matrix positions are rearranged according to the time rows and indicator columns to generate an ecological state vector matrix. 4.The method of claim 1, wherein, The specific steps for generating the scoring weight configuration rule table are as follows: Based on the ecological state vector matrix, water quality node numbers, biological community node numbers, and ecological pressure node numbers are extracted, written into the node index table according to the node category, duplicate numbers are merged, placeholders are written for missing numbers, and node records are rearranged according to category order to generate a node index set. Based on the node index set, a graph attention network is used to read the weight values ​​and adjustment flags of each node. The values ​​corresponding to the boost flags are written into the increment column, the values ​​corresponding to the decrement flags are written into the decrease column, and the values ​​corresponding to the keep flags are written into the retain column. Then, the updated weights are summarized in the order of node numbers and written into a unified weight table to obtain the node weight matrix. Based on the node weight matrix, the corresponding row data in the ecological state vector matrix is ​​called, and the node weight value is multiplied with the vector value of each row one by one. The product results under the same node category are accumulated in the row, the accumulated results are written into the scoring column, and the scoring records are arranged in chronological order to generate a scoring weight configuration rule table. 5.The method of claim 4, wherein, The graph attention network first reads the weight values, adjustment identifiers, node category numbers, and adjacency association records of water quality nodes, biological community nodes, and ecological pressure nodes based on the node index set. It then concatenates the weight value, node category number, and node connection direction code of a single node into a node input vector. For any pair of node input vectors with a connection relationship, it performs a linear mapping to obtain the source node representation vector and the target node representation vector. The source node representation vector, target node representation vector, and connection direction code are then concatenated sequentially and input into the attention scoring unit, outputting the attention score for each node pair. Normalization is performed on the attention scores of all source nodes associated with the same target node to obtain the attention coefficient of the neighboring nodes. The attention coefficient of the neighboring nodes is multiplied by the corresponding source node weight value to obtain the weighted node weight value. The weighted node weight value is then allocated to the incremental, decremental, and retained columns according to the adjustment identifier. The incremental column value is added to the original weight value, the decremental column value is subtracted from the original weight value, and the retained column value is copied and written to the original weight value. Finally, the weights are summarized and updated according to the node number order and written into a unified weight table to form a node weight matrix. 6.The method of claim 1, wherein, The specific steps for generating the ecological state embedding vector sequence are as follows: Based on the ecological state vector matrix and the scoring weight configuration rule table, phytoplankton, benthic organisms and nutrient index vectors are extracted, samples are paired according to the grade labels, the distance between samples is calculated, and a grade distance table is generated according to the grade set center value. Based on the aforementioned distance table, the sample vectors are compared with the level centers item by item. The level corresponding to the first distance value is written into the identifier, and the difference between adjacent time points is written into the change column to obtain the ecological state embedded vector sequence. 7.The method of claim 6, wherein, The calculation of the inter-sample distance value involves reading the phytoplankton index vector, benthic organism index vector, and nutrient index vector from the ecological state vector matrix. These three indices are combined into a multi-dimensional sample vector at the same time point. A dimension-wise difference operation is performed on two sample vectors to obtain the differences in the phytoplankton index, benthic organism index, and nutrient index. The three differences are squared and summed. Then, the square root operation is performed on the summed value to obtain the inter-sample distance value. The distance value is recorded in the sample pairing record row and written into the grade distance table along with the sample number and grade label.

8. The deep learning-based dynamic evaluation method for bay ecosystems according to claim 1, characterized in that, The specific steps for generating the ecological relationship propagation graph structure are as follows: Based on the ecological state embedding vector sequence and the scoring weight configuration rule table, the dissolved oxygen node number, salinity node number, phytoplankton node number and ecological pressure node number are extracted, written into the node list according to the node source, the same type of nodes are grouped and arranged, the cross-type nodes are sequentially numbered, and the corresponding relationship markers between nodes are retained to generate an ecological node list. Based on the ecological node list, connection records are established in pairs according to the node number. The starting node, target node, connection direction and connection weight are written into the adjacency list. Empty value markers are filled in for positions where no connection is established. Multiple connection records of the same node are written in parallel. The matrix structure is arranged according to the connection direction to obtain the node adjacency matrix. Based on the node adjacency matrix, a convolutional neural network is used to read the connection weights between nodes, calculate the cumulative weights of adjacent nodes, the weight difference between nodes, and the weight product of nodes, and write the three types of results into the node relationship table. Then, according to the weight size rule, low-association nodes are filtered out and effective neighboring nodes are retained to generate an ecological relationship propagation graph structure. 9.The method of claim 8, wherein, The convolutional neural network first reads the node connection weights based on the node adjacency matrix, extracts the corresponding row and column vectors of each node in the order of node number, and concatenates the row vectors, column vectors, connection direction encoding, and connection distance encoding into a relational input tensor. A sliding window is used to scan the relational input tensor grid by grid along the node number direction and the adjacency position direction. A convolution kernel multiplication-addition operation is performed on the connection weights within each window to output the local relational response value. The connection weights of the preceding and following nodes within the same window are added to obtain the weight accumulation result. The connection weights of the preceding and following nodes are subtracted to obtain the weight difference result. The connection weights of the preceding and following nodes are multiplied to obtain the weight product result. The local relational response value, the weight accumulation result, the weight difference result, and the weight product result are then written into the node relational table. All relational response values ​​in the neighborhood of each node in the node relational table are sorted. Low-association node records are deleted according to a preset retention quantity and response threshold. The remaining node connection records are retained, and the neighborhood connection table is reconstructed to generate the ecological relational propagation graph structure. 10.The method of claim 1, wherein, The specific steps for generating the aforementioned dynamic ecological assessment result set of the bay are as follows: Based on the ecological relationship propagation graph structure, the connection paths between tidal level, nutrients and biological community nodes are extracted, the path nodes are unfolded in hierarchical order, and multiple connection weights are accumulated to generate a path weight sequence. Based on the path weight sequence, the path weights are arranged in descending order, the state vectors are read and matched with the scoring weight configuration rule table item by item, and the mapping scores and path records are written to obtain the dynamic evaluation result set of the bay ecosystem.

Images