Method and system for habitat suitability prediction for marine ecological management

Abstract

The present application relates to the technical field of marine ecological information, and discloses a habitat suitability prediction method and system for marine ecological management, wherein the habitat suitability prediction method for marine ecological management comprises the following steps: obtaining fisherman traditional knowledge data, converting audio data into text format through natural language processing technology, and generating a traditional knowledge graph containing the corresponding relationship between observation characteristics and fishing results; performing canonical correlation analysis on the surface observation characteristics in the traditional knowledge graph and vortex physical parameters, and converting the traditional knowledge into quantitative expression in the vortex parameter space based on a mapping matrix; and integrating multiple source prediction results by using a random forest, taking the traditional knowledge data and the physical model calculation results as input characteristics, and outputting a middle-layer fish habitat probability distribution map. The present application solves the technical problem that traditional knowledge and vortex physical models are difficult to effectively integrate for middle-layer fish habitat prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine ecological information science, more particularly, it relates to a habitat suitability prediction method and system for marine ecological management. BACKGROUND

[0002] Mesopelagic fish (200-1000m depth) is an important marine fishery resource, and its habitat distribution is closely related to the dynamic process of ocean mesoscale eddies. Fishermen have long been observing surface signs (such as water color change, foam line distribution, seabird aggregation, etc.) to determine the location of deep fish schools. These surface signs are often related to the vertical transport process of mesoscale eddies. However, fishermen's experience and knowledge are mostly qualitative descriptions and lack depth information, while eddy physical models can describe three-dimensional flow fields but lack biological behavior response mechanisms. There is a semantic gap between the two knowledge systems, making it difficult to effectively integrate traditional knowledge and scientific observation data and accurately predict the habitat distribution of mesopelagic fish. SUMMARY

[0003] The present application provides a habitat suitability prediction method and system for marine ecological management, which solves the technical problem of the semantic gap between the two knowledge systems in related technologies, which makes it difficult to effectively integrate traditional knowledge and scientific observation data and accurately predict the habitat distribution of mesopelagic fish.

[0004] The present application provides a habitat suitability prediction method for marine ecological management, comprising:

[0005] Obtaining fishermen's traditional knowledge data, including: collecting oral information about surface observation rules and corresponding catch records from fishermen. Transcribe audio data into text format through natural language processing technology, extract environmental clue words and spatial orientation descriptions using long short-term memory network, and generate a traditional knowledge graph containing the correspondence between observation features and catch results;

[0006] Obtaining satellite eddy observation data, using eddy automatic recognition algorithm to detect eddy boundary and center position, calculating three-dimensional velocity field and density field of eddy based on quasi-geostrophic theory, outputting eddy feature data set containing eddy intensity, radius, vertical velocity;

[0007] Calculate the mapping relationship between traditional knowledge and eddy parameters, perform canonical correlation analysis on surface observation features in the traditional knowledge graph and eddy physical parameters, calculate the semantic-physical mapping matrix, and convert the traditional knowledge to quantitative expression in the eddy parameter space based on the mapping matrix;

[0008] The mapping results, fuzzy membership of traditional knowledge, environmental suitability index and transport flux distribution are input into the Bayesian inference network based on Bayesian inference fusion of multi-source prediction information, and the posterior probability distribution is calculated through conditional probability updating;

[0009] The steps of simulating the vertical material transport process by using the particle tracking algorithm are as follows:

[0010] Based on the three-dimensional velocity field of the vortex, the Lagrangian particle tracking method is used to simulate the transport path of the nutrient from the sea surface to the middle layer, the motion equation is solved by numerical integration, the particle position update is calculated by using the fourth-order Runge-Kutta method, the time distribution of the particles from the surface to different depths is calculated, and the space-time distribution diagram of the vertical transport flux is generated;

[0011] The prediction results are integrated by using ensemble learning, and the random forest is used to integrate the multi-source prediction results, and the traditional knowledge data and the calculation results of the physical model are used as input features, and the output is the probability distribution diagram of the middle layer fish habitat.

[0012] Further, the traditional knowledge graph is constructed by using the entity-relation-property triple representation method, the graph nodes include environmental observation entities and fishing result entities, and the edges represent the causal relationship between the observation features and the fishing results; the environmental feature words and the fishing amount values are extracted by named entity recognition, the syntactic relationship between the words is identified by dependency syntax analysis, the mapping relationship between the observation features and the fishing results is established by the relationship extraction algorithm, and the confidence weight is calculated by the fisherman description frequency and consistency.

[0013] Further, the extraction of environmental clue words is realized by word frequency statistics and semantic clustering algorithm, the semantic clustering algorithm adopts K-means clustering method, the input is the vector set of fisherman description text processed by word embedding, the word embedding processing adopts the pre-trained Word2Vec model, and the vectorization of the text segment adopts the weighted average method of word vector, and the weight is weighted by term frequency-inverse document frequency value to highlight the importance of key words.

[0014] Further, the vortex automatic recognition algorithm includes the following sub-steps:

[0015] The input sea surface height anomaly data field is spatially filtered to remove noise less than twice the grid resolution;

[0016] The sea surface height gradient of each grid point is calculated, and the local extreme point with a gradient less than a threshold value is identified as a potential vortex center;

[0017] Starting from each potential center, the closed contour is searched along the radial direction, and the area and equivalent radius within the contour are calculated;

[0018] The closed contour meeting the radius constraint and shape parameter constraint is selected as the vortex boundary;

[0019] Output the coordinates of the vortex center, the vortex radius and the vortex intensity.

[0020] Further, the step of quantifying the uncertainty of traditional knowledge based on fuzzy logic is included: converting the qualitative description in traditional knowledge data into membership functions of vortex features, defining the corresponding fuzzy set for each observed feature, quantifying the degree to which the feature value belongs to the fuzzy set through triangular or trapezoidal membership functions, and generating knowledge representation containing uncertainty range.

[0021] Further, the step of obtaining the environmental suitability of the middle layer water body in Argo buoy data analysis is included: obtaining the vertical profile data of temperature, salinity and dissolved oxygen from the Argo buoy database, interpolating the profile data into the cylindrical coordinate system with the vortex center as the origin, and evaluating the environmental suitability index of each depth layer based on the physiological demand threshold of fish. The environmental suitability index is calculated by the product of the temperature suitability index, the salinity suitability index and the dissolved oxygen suitability index.

[0022] Further, the input of the canonical correlation analysis method is two groups of variable matrices: the traditional observed feature matrix and the vortex physical parameter matrix. Due to the dimensional difference, the two groups of matrices need to be preprocessed by Z-score standardization before calculating the correlation. The mapping relationship is obtained by solving the eigenvalue problem, and the output is the canonical correlation coefficient and the corresponding canonical variable coefficient matrix.

[0023] Further, the random forest is composed of multiple decision trees, each of which is constructed by recursively splitting nodes, and the splitting criterion adopts Gini impurity. The input is the fused feature vector, which contains multi-dimensional features such as traditional knowledge mapping value, environmental suitability index and transport flux. The feature vector needs to be preprocessed by standardization before inputting into the random forest, and the output is the habitat existence probability.

[0024] The present application provides a habitat suitability prediction system for marine ecological management, comprising:

[0025] A traditional knowledge acquisition module for acquiring fisherman's traditional knowledge data and generating structured knowledge representation;

[0026] A marine data processing module for acquiring and processing satellite vortex data and marine environment data;

[0027] A semantic mapping module for establishing the mapping relationship between traditional knowledge and physical parameters;

[0028] A multi-source fusion module for fusing multiple prediction information sources;

[0029] A prediction output module for generating habitat distribution prediction results

[0030] The present application has the advantages of:

[0031] The present application establishes the quantitative connection between the traditional observation characteristics of fishermen and the physical parameters of eddies by calculating the semantic-physical mapping matrix, overcomes the semantic gap between traditional knowledge and quantitative scientific models; the qualitative description is converted into quantitative expression with uncertainty range through fuzzy logic method, which overcomes the application difficulties caused by the fuzziness of traditional knowledge; the limitations of surface observation lacking depth information are overcome by mapping traditional knowledge to eddy parameter space and combining with the vertical transport calculation of quasi-geostrophic dynamics, thereby solving the technical problem that traditional knowledge and eddy physical model are difficult to effectively integrate for mid-layer fish habitat prediction. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 is a flow chart of a habitat suitability prediction method for marine ecological management of the present application;

[0033] Figure 2 is a traditional knowledge graph relationship network diagram of the present application;

[0034] Figure 3 is an eddy physical parameter distribution scatter plot of the present application;

[0035] Figure 4 is a semantic-physical mapping relationship heat map of the present application;

[0036] Figure 5 is an environmental suitability depth profile broken line diagram of the present application;

[0037] Figure 6 is a particle vertical transport process Sang base diagram of the present application. DETAILED DESCRIPTION

[0038] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to enable those skilled in the art to better understand and thus implement the subject matter described herein, and the functions and arrangements described can be modified as appropriate without departing from the scope of the disclosure. Various examples can omit, substitute, or add various procedures or components as appropriate. Also, it will be understood that some of the demonstrations described herein can be implemented by software, hardware, firmware, or a combination of these elements.

[0039] A habitat suitability prediction method for marine ecological management is disclosed in at least one embodiment of the present application, as shown in Figure 1 comprising the following steps:

[0040] Step 1: Obtain traditional knowledge data of fishermen, and generate structured knowledge representation;

[0041] Oral information about sea surface observation rules and corresponding catch records was collected from fishermen. The audio data was transcribed into text format, and the qualitative ratings (such as "high yield", "medium yield", "low yield") in the fishermen's descriptions were numerically encoded to convert the categorical data into numerical data: high yield = 3, medium yield = 2, low yield = 1. Environmental cue words and spatial orientation descriptions were extracted using a Long Short-Term Memory (LSTM) network to generate a traditional knowledge graph containing the correspondence between observation features and catch results.

[0042] Furthermore, the construction of traditional knowledge graphs adopts the entity-relation-attribute triple representation method. Graph nodes include environmental observation entities (such as "dark water color" and "seabirds gathering") and catch result entities (such as "high yield" and "low yield"), with edges representing the causal relationship between observation features and catch results. The specific construction process is as follows: first, environmental feature words and catch quantity values ​​are extracted through named entity recognition; then, dependency parsing is used to identify the grammatical relationships between words; finally, a relation extraction algorithm is used to establish the mapping relationship between observation features and catch results. The confidence weight is calculated by the frequency and consistency of the fishermen's descriptions.

[0043] Furthermore, the specific implementation method of named entity recognition adopts a sequence labeling model based on conditional random fields (CRF), with the input being the segmented text sequence. The output is the corresponding label sequence. The tags use the BIO annotation system:

[0044] B-ENV represents the beginning of an environmental feature entity, I-ENV represents the interior of an environmental feature entity, B-CATCH represents the beginning of a catch result entity, I-CATCH represents the interior of a catch result entity, and O represents a non-entity term; the CRF model maximizes conditional probabilities. The training process involves feature functions including lexical features, part-of-speech features, and contextual features.

[0045] Furthermore, the specific implementation method of dependency parsing adopts a graph-based dependency parsing algorithm, with the input being a sequence of words. The output is a dependency tree. ,in For a set of word nodes, It is a set of dependency arcs; the algorithm calculates the dependency relationship score matrix between words. ,in Words As a word The scores of the governing words are calculated, and then the maximum spanning tree algorithm (such as Edmonds' algorithm) is used to find the optimal dependency tree, ensuring that each word (except the root node) has one and only one governing word.

[0046] Further, the specific implementation method of the relation extraction algorithm adopts a rule method based on pattern matching combined with a machine learning classifier. First, a set of language patterns of causal relationship is defined , such as causal conjunction word patterns like “cause”, “result in”, “lead to”, etc., and then paths matching these patterns are searched in the dependency syntax tree. For the matched entity pairs , a support vector machine (SVM) classifier is used to determine whether a causal relationship exists. The input features of the classifier include the syntactic distance between entities, the dependency path type, and the context word features.

[0047] Further, the specific calculation method of the confidence weight is as follows:

[0048]

[0049] wherein is the relationship frequency ratio, is the number of times the relationship appears in all fisherman descriptions, is the total number of relationships; is the consistency coefficient, and are the mean and standard deviation of the scores of different fishermen for the same relationship, respectively; the weight coefficient , ensures the comprehensive consideration of frequency and consistency.

[0050] The aforementioned semantic clustering algorithm adopts a K-means clustering method, and the input is a set of fisherman description text vectors processed by word embedding , wherein is the vector representation of the th text segment;

[0051] Further, the word embedding processing adopts a pre-trained Word2Vec model, and the input is the segmented fisherman description text sequence, and the output is a dimensional word vector representation; for the vectorization of the text segment, a weighted average method of word vectors is adopted: , wherein is the number of words in the th text segment, is the TF-IDF weight value of the th word, , is the corresponding word vector, and the weight is weighted by the TF-IDF value to highlight the importance of key words.

[0052] L2 normalization preprocessing is performed on the word embedding vector before clustering, i.e. , to eliminate the influence of vector length difference on clustering effect; and the output is Cluster centers and category labels for each text segment;

[0053] The algorithm optimizes the objective function iteratively. To determine the optimal clustering result.

[0054] Furthermore, the convergence condition of the K-means algorithm is set as follows: when the distance the cluster centers move in two consecutive iterations satisfies... The algorithm converges in time, where The convergence threshold, This represents the number of iterations.

[0055] Simultaneously set the maximum number of iterations. To avoid infinite loops, the cluster center update formula is: ,in For the first In the nth iteration, it is assigned to the ... A set of samples for each cluster.

[0056] Furthermore, cluster numbers The contour coefficient method was used to determine the value.

[0057] Calculations are different Profile coefficient at value ,in For the sample The average distance to other samples of the same type

[0058] For the sample The average distance to the nearest outlier sample; select the sample that maximizes the silhouette coefficient. The value is used as the optimal cluster number, and its range is constrained. Ensure that each cluster contains at least 5 samples.

[0059] Furthermore, the profile coefficient The constraint range is ,in A value close to 1 indicates good clustering performance (samples are close to their own kind and large to their dissimilar kind). A value close to 0 indicates that the cluster boundaries are ambiguous (the sample is near the cluster boundary). A value close to -1 indicates poor clustering performance (samples may be assigned to incorrect clusters), and is typically chosen as a low-value cluster. of The value is used as the number of effective clusters.

[0060] The aforementioned LSTM network is used for text sequence feature extraction, and its input layer receives a sequence of word embedding vectors. ,in For word embedding dimension, For each time step in the sequence, the output layer outputs the hidden state at each time step. ,in The hidden layer dimension is used; supervised learning mode is used during training, employing the cross-entropy loss function from the sequence labeling task.

[0061]

[0062] in For the sample size, For the number of categories, For real labels, To predict probabilities, the Adam optimizer is used as the optimization strategy, with a learning rate set to 0.001. During training, a linear mapping layer is used to map the hidden states of the LSTM to the label space, i.e. ,in This is the weight matrix. This is the bias vector; the softmax probability distribution output by the LSTM network needs to be decoded into specific feature labels using the argmax function, i.e. Then, the numerical labels are converted into corresponding environmental feature words through a predefined label mapping table (such as label 1 corresponding to "dark water color", label 2 corresponding to "dense foam zone", etc.).

[0063] Step 2: Acquire satellite vortex observation data and calculate vortex physical parameters;

[0064] Sea surface height anomaly data are obtained from satellite altimeters. The vortex boundary and center position are detected using an automatic vortex identification algorithm. The three-dimensional velocity field and density field of the vortex are calculated based on the quasi-geostrophic theory. The output is a vortex feature dataset containing vortex intensity, radius, and vertical velocity.

[0065] Furthermore, the three-dimensional velocity field calculation method based on quasi-geostrophic theory is as follows: the horizontal geostrophic velocity component is obtained through the sea surface height anomaly gradient calculation.

[0066] ,

[0067] in This represents an anomaly in sea level height (unit: m).

[0068] It is the acceleration due to gravity. For Coriolis parameters, The angular velocity of Earth's rotation. Latitude; Vertical velocity Calculated using the quasi-geostrophic vorticity equation: ,in Stream function (unit: ),

[0069] By solving the elliptic equation Obtain, The geostrophic vorticity (unit: ).

[0070] Further, the time derivative term The calculation uses the finite difference method, and the time window is set to Day,

[0071] By the flow function field at two consecutive times:

[0072] , where The sampling interval of the time series data is daily, which ensures the capture of the quasi-steady characteristics of mesoscale eddies; the vertical velocity is constrained to , and the outliers beyond this range are processed by the median filtering method.

[0073] The aforementioned eddy automatic identification algorithm includes the following sub-steps:

[0074] Sub-step 2.1: Spatial filtering is performed on the input sea surface height anomaly data field Remove noise less than , where is the grid resolution;

[0075] Further, the noise judgment uses a Gaussian low-pass filter to suppress high-frequency signals with a spatial scale less than The cutoff frequency of the filter is set to , and the amplitude of the signal retained after filtering is at least of the original signal, ensuring that the effective signal is not excessively suppressed.

[0076] Further, the specific implementation method of the Gaussian low-pass filter is:

[0077] Filter kernel function:

[0078]

[0079] Where the standard deviation , and the filter window size is set to .

[0080] The filtering operation is realized by convolution operation:

[0081]

[0082] Where is the window radius.

[0083] Sub-step 2.2: Calculate the sea surface height gradient of each grid point Identify gradients less than a threshold The local extreme points are taken as potential vortex centers;

[0084] Furthermore, gradient threshold The value of is determined using statistical methods, and the gradient field of the entire sea area is calculated. Standard deviation ,set up Gradient thresholding can effectively identify local extrema where the gradient is close to zero, while avoiding misidentifying noise points as vortex centers.

[0085] Furthermore, gradient threshold The constraint function is in the form of: The lower bound open interval ensures the threshold is positive, and the upper bound closed interval restricts the threshold from exceeding the standard deviation of the gradient field. When the value is too small, too many noise points will be mistaken for the vortex center. When the value is too large, the true vortex center will be missed, so set it to... As the empirically optimal value.

[0086] Sub-step 2.3: Starting from each potential center, search for closed contour lines radially and calculate the area within the contour lines. and equivalent radius ;

[0087] Sub-step 2.4:

[0088] Select the one that satisfies And shape parameters The closed contour lines serve as the vortex boundary, where The perimeter of the contour lines;

[0089] Furthermore, shape parameters The constraint function is in the form of: The lower boundary Corresponding to a perfect circle ( ), Upper Realm Ensure that the contour lines are not excessively distorted, when A value close to 1 indicates that the vortex shape is close to circular. A value close to 2 indicates that the vortex shape is somewhat distorted but still acceptable. Contours exceeding 2 are judged to be non-vortex structures and are discarded.

[0090] Furthermore, the determination of the vortex radius constraint parameters is based on the deformation radius of the geostrophic theory.

[0091] calculate ,in It is the acceleration due to gravity. For Coriolis parameters, To set the vortex intensity; and The vortex radius constraint parameter range can effectively filter out small noise structures and large non-vortex structures.

[0092] Further, the deformation radius The specific calculation method is:

[0093] First, calculate the Coriolis parameter ,

[0094] Where is the angular velocity of the earth's rotation, is the average latitude of the study sea area (in radians); Then calculate the absolute value of the vortex intensity , Finally, calculate the deformation radius by the formula , Where , the unit of the calculation result is meter, which needs to be converted to kilometer for radius constraint.

[0095] Sub-step 2.5: output vortex center coordinates , vortex radius and vortex intensity

[0096] , Where is the average sea surface height on the boundary.

[0097] Further, the vortex intensity The unit is meter (m), which represents the difference in sea surface height between the vortex center and the boundary, and the value range is constrained to , Where a positive value represents an anticyclonic vortex (warm vortex), and a negative value represents a cyclonic vortex (cold vortex);

[0098] The unit of vortex radius is kilometer (km), and the value range is ; When or , the vortex structure is determined as noise rather than a real vortex and is removed.

[0099] Step 3: Calculate the mapping relationship between traditional knowledge and vortex parameters to generate habitat prediction results;

[0100] Correlation analysis is performed between the surface observation features in the traditional knowledge graph obtained in step 1 and the vortex physical parameters obtained in step 2 to calculate the semantic-physical mapping matrix , Where represents the correlation between the th traditional observation feature and the ​correlation coefficient between the two groups of variables; based on the mapping matrix, the traditional knowledge is converted into quantitative expression in the vortex parameter space, combined with the vertical transport characteristics of the vortex, the probability distribution map of the middle-layer fish habitat is output.

[0101] The input of the aforementioned canonical correlation analysis method is two groups of variable matrices:

[0102] Traditional observation feature matrix and vortex physical parameter matrix , wherein is the number of samples, is the dimension of the observation feature, is the dimension of the physical parameter; since there is a significant dimensional difference and numerical range difference between the traditional observation feature (qualitative score, dimensionless value 1-3) and the vortex physical parameter (intensity unit: meter, radius unit: kilometer, velocity unit: meter / second, etc.), direct correlation calculation will cause dimensional mismatch problem, and the two groups of matrices must be preprocessed by Z-score standardization before calculating the correlation:

[0103] , , wherein and are the mean and standard deviation vectors respectively, the standardized data are all dimensionless with mean 0 and standard deviation 1, eliminating the influence of dimensional difference on correlation calculation; the output is the canonical correlation coefficient and the corresponding canonical variable coefficient matrix , , wherein .

[0104] The mapping relationship is obtained by solving the eigenvalue problem , wherein is the covariance matrix.

[0105] Further, the calculation method of the covariance matrix is:

[0106]

[0107]

[0108]

[0109]

[0110] Before solving the eigenvalue problem, the covariance matrix needs to be regularized by adding a small regularization term ( ) to avoid numerical instability.

[0111] Step 4: Quantify the range of uncertainty of traditional knowledge based on fuzzy logic;

[0112] The qualitative descriptions in traditional knowledge data are transformed into membership functions of vortex features. For each observed feature, a corresponding fuzzy set is defined, and triangular or trapezoidal membership functions are used. Quantization eigenvalues The degree to which it belongs to a fuzzy set, where This generates a knowledge representation that includes a range of uncertainties.

[0113] The aforementioned membership function Using the triangular membership function form:

[0114] ,in , , The left, peak, and right boundary parameters of the triangular membership function are determined based on the eigenvalue range described by the fishermen; for the trapezoidal membership function form:

[0115] ,in , , , For the four vertices of the trapezoid, when hour ,when or hour .

[0116] Furthermore, the method for determining the membership function parameters is as follows: First, perform statistical analysis on the qualitative scores in the fishermen's descriptions, and calculate the log-normal distribution parameters corresponding to each feature. and For the membership function of a triangle, set... , , For trapezoidal membership functions, set , , , The membership function parameter setting method can ensure that 95% of the observations fall within the support interval of the membership function.

[0117] It should be noted that the input vortex eigenvalues Before applying membership functions, min-max normalization preprocessing is required to scale physical parameters of different dimensions uniformly to a uniform scale. Interval: ,in and These are the minimum and maximum values ​​of the vortex characteristic, respectively, to ensure the effectiveness of the membership function calculation.

[0118] Step 5: Obtain Argo buoy data, analyze the environmental suitability of the middle layer water body;

[0119] Obtain the vertical profile data of temperature, salinity, and dissolved oxygen from the Argo buoy database, interpolate the profile data into the cylindrical coordinate system with the eddy center as the origin, calculate the environmental parameter distribution in the depth range of 200-1000 meters, and evaluate the environmental suitability index of each depth layer based on the physiological demand threshold of fish.

[0120] Further, the cylindrical coordinate interpolation method uses the inverse distance weighted interpolation algorithm. For the environmental parameter value at the eddy center coordinate , the calculation formula is: ,

[0121] Where is the environmental parameter observation value of the th Argo station, is the weight, is the three-dimensional Euclidean distance from the target point to the th station, where is the radial distance, is the azimuth angle, is the depth, is the smoothing term to prevent division by zero error, and the interpolation radius is set to 50km to ensure that each grid point has at least 3 observation stations involved in the interpolation.

[0122] Further, the constraint range of the azimuth angle is , which means starting from the east direction ( ), measuring counterclockwise to radian (not including ), covering the full range of 360 degrees. When the value exceeds this range, periodic constraint is performed through modulo operation: , to ensure that the azimuth angle is always within the valid range.

[0123] Further, the calculation of the environmental suitability index is based on the physiological demand threshold of the middle layer fish. The temperature suitability index

[0124] Where °C is the optimal temperature, °C is the temperature tolerance;

[0125] Salinity suitability index:

[0126]

[0127] Where psu, psu;

[0128] dissolved oxygen suitability index wherein mg / L, mg / L is the minimum requirement threshold; the comprehensive suitability index .

[0129] It should be noted that due to the large differences in the dimensions and numerical ranges of environmental parameters such as temperature (unit: °C, numerical range about 2-8 °C), salinity (unit: psu, numerical range about 33-36 psu), dissolved oxygen (unit: mg / L, numerical range about 2-8 mg / L),

[0130] directly multiplying the environmental parameters to calculate the comprehensive suitability index will cause dimension mismatch problems, so it is necessary to perform Z-score standardization preprocessing on each parameter before calculating the environmental suitability index: , , wherein and are the mean and standard deviation of each parameter, respectively. The standardized parameters are dimensionless, eliminating the influence of dimensional differences on suitability assessment.

[0131] Step 6: Simulate the vertical material transport process using the particle tracking algorithm;

[0132] Based on the three-dimensional velocity field of the vortex calculated in step 2, the Lagrangian particle tracking method is used to simulate the transport path of nutrients from the sea surface to the middle layer, and the time distribution of particles from the surface to different depths is calculated wherein is the depth, and the spatiotemporal distribution of vertical transport flux is generated.

[0133] The input of the aforementioned Lagrangian particle tracking method is a three-dimensional velocity field:

[0134]

[0135] and an initial particle position set ,

[0136] wherein is the number of particles; the output is the particle trajectory at each time and the arrival depth time ; the algorithm solves the motion equation by numerical integration , and uses the fourth-order Runge-Kutta method to calculate the particle position update: wherein is the intermediate velocity estimate.

[0137] Further, the intermediate velocity estimate in the fourth-order Runge-Kutta method is calculated as:

[0138]

[0139]

[0140]

[0141]

[0142] The time step size must satisfy the CFL condition, i.e. to ensure numerical stability.

[0143] Further, the constraint function of the CFL condition is in the form of

[0144] where , is the CFL number, which is usually taken as to ensure numerical stability. When exceeds this upper bound, numerical instability and computational divergence will occur.

[0145] The aforementioned time distribution function is defined as:

[0146] where is the total number of particles reaching depth , is the time for the th particle to reach depth from the sea surface, is determined by the condition to determine whether the particle has reached the target depth, where is the depth tolerance threshold.

[0147] Further, the time integration interval is set to days,

[0148] The particle tracking starts at and ends at , with a time step size of hours and a total number of steps of steps;

[0149] For particles that do not reach the target depth within the time window, set as the truncation processing;

[0150] The effective range constraint for the arrival time is , and the average transport time is in units of days, with a value range of Day.

[0151] Further, the depth tolerance threshold value is determined according to the vertical grid resolution, and is set to , wherein is the vertical grid spacing, and the depth tolerance threshold value is set to ensure that the accuracy of the particle reaching judgment matches the vertical resolution of the model, avoiding particle omission due to too small tolerance or inaccurate depth judgment due to too large tolerance.

[0152] Step 7: Fusion of multi-source prediction information based on Bayesian inference;

[0153] The mapping result obtained in step 3, the fuzzy membership obtained in step 4, the environmental suitability index obtained in step 5, and the transport flux distribution obtained in step 6 are input into the Bayesian inference network, and the posterior probability distribution is calculated through conditional probability updating , wherein represents the hypothesis of the existence of the habitat, represents the th evidence, and the fused habitat probability evaluation result is output.

[0154] The aforementioned Bayesian inference network is composed of nodes and directed edges, and the nodes include evidence nodes and hypothesis nodes , and the directed edges represent conditional dependency relationships; the input of the evidence nodes is the characteristic value of each prediction source , wherein is the dimension of the th evidence; since the characteristics (mapping correlation coefficient, fuzzy membership, environmental suitability index, and transport flux) of each prediction source have differences in dimension and numerical range, min-max normalization processing is required before inputting into the Bayesian network, and all characteristics are scaled to the interval ; the network calculates the posterior probability through the Bayesian theorem:

[0155] The training method uses maximum likelihood estimation, and estimates the conditional probability table and the prior probability through historical observation data; the loss function is the negative log-likelihood: , wherein is the number of training samples, and the superscript represents the th sample; the optimization strategy uses the expectation maximization (EM) algorithm, and iteratively updates the network parameters until convergence.

[0156] Further, the specific implementation method of maximum likelihood estimation is:

[0157] For discrete evidence variables​ conditional probability Calculated through frequency statistics:

[0158] ,in In the training data and The number of samples, for The total number of samples; for continuous evidence variables, assume they follow a Gaussian distribution. The parameters are estimated using the sample mean and variance:

[0159] , .

[0160] Furthermore, the convergence criterion for the EM algorithm is set as follows: when the difference between the log-likelihood function values ​​of two consecutive iterations satisfies... The algorithm converges in time, where The convergence threshold, Set the number of EM iterations; also set the maximum number of iterations. To avoid an infinite loop;

[0161] The E-step estimates the posterior distribution of the latent variables using the current parameters, and the M-step updates the parameters by maximizing the expected log-likelihood. ,in For observation data, As a latent variable, These are network parameters.

[0162] It should be noted that the posterior probability output by the Bayesian inference network... For continuous numerical values ​​(range [0,1]), a probability threshold is needed to determine whether the decoding process converts them into discrete habitat prediction results.

[0163] when If the habitat is deemed suitable (label=1), it is deemed unsuitable (label=0). The threshold is used for determination; at the same time, the original probability value is retained as a quantitative assessment of habitat suitability, providing continuous input features for subsequent ensemble learning.

[0164] Furthermore, determine the threshold. The selection is based on the principle of maximizing classification accuracy. The decision threshold minimizes the overall error rate of Type I error (misclassifying suitable habitats as unsuitable) and Type II error (misclassifying unsuitable habitats as suitable), which is in line with the minimization of total risk criterion in Bayesian decision theory.

[0165] When prior probability Furthermore, when the costs of the two types of errors are equal, 0.5 is the theoretically optimal threshold.

[0166] Step 8: Integrate the prediction results using ensemble learning and output the final habitat distribution map;

[0167] Random forest is used to integrate multi-source prediction results. Traditional knowledge data and physical model calculation results are used as input features. The results of each prediction source are integrated through weighted voting to output a probability distribution map of mid-water fish habitat in the depth range of 200-1000 meters, with probability values ​​ranging from [0,1].

[0168] The aforementioned random forest consists of multiple decision trees, each decision tree The splitting criterion is Gini impurity, constructed recursively through splitting nodes. ,in Category in the node The sample ratio; the input is the fused feature vector. It includes traditional knowledge mapping values ​​(dimensionless correlation coefficient), environmental suitability index (dimensionless probability value), and transport flux (unit). )wait For multidimensional features, since each feature comes from a different source and has different dimensions (correlation coefficients and probability values ​​are dimensionless, while transport flux has dimensions), the feature vectors need to be preprocessed using Z-score standardization before being input into the random forest. Eliminate the impact of differences in dimensions and numerical ranges on decision tree splitting;

[0169] The output is the probability of habitat existence. ,in The number of decision trees; the training method uses bootstrap sampling, with each tree using randomly selected... One sample and Training is performed using 10 features; the loss function is the logarithmic loss for the classification task.

[0170] ,in The labels are real; the optimization strategy is a greedy algorithm that selects the optimal splitting feature and threshold at each node.

[0171] Furthermore, the specific implementation method of bootstrap sampling is as follows:

[0172] For the original training set Each decision tree training set Generated through random sampling with replacement, the sampling process is as follows: From Random selection Generate samples (repeated samples allowed) from 100 samples. ,in is a random index; meanwhile, at each node split, a random feature is selected from the set of features as the candidate split feature, ensuring the diversity of the decision tree.

[0173] the aforementioned decision tree function is defined as: starting from the root node, the input feature vector is judged layer by layer, at each internal node decides to proceed to the left or right sub-tree according to the split condition , where is the split feature index of the internal node, and is the split threshold; when reaching the leaf node , the predicted value of the leaf node is output , where and are the number of positive and negative class samples in the leaf node, respectively.

[0174] It should be noted that the habitat existence probability output by the random forest needs to be converted into a complete habitat distribution map by a spatial interpolation method: for each spatial grid point in the study area, where represents the depth of the layer, the habitat probability is calculated according to the feature vector of the point, and then a bilinear interpolation method is used to generate a continuous three-dimensional habitat probability distribution field, finally output as the habitat suitability distribution map of each layer in the depth range of 200-1000 meters.

[0175] Further, the specific implementation of the bilinear interpolation method is as follows: for the target point , first determine its position in the grid, find the four adjacent grid points , , , surrounding the point and their corresponding probability values , , , ; calculate the interpolation weights: , ; the probability value of the target point is calculated by the bilinear interpolation formula:

[0176] The bilinear interpolation method ensures that the generated probability distribution field is spatially continuous and smooth.

[0177] Further, the constraint range of the bilinear interpolation weight is: where the weights satisfy the normalization condition when the target point is located inside the grid element:​

[0178] , ensuring that the probability values of the interpolation results remain within the valid range corresponding weights are 1 and 0, respectively, when the target point is located exactly on a grid node.

[0179] This implementation example takes the Western Pacific Kuroshio Extension region (35°N-40°N, 140°E-150°E) as the study area, and predicts the habitat suitability of medium-depth tuna (habitat depth 200-1000 meters). This sea area is frequently affected by eddies and is an important fishing area, where local fishermen have accumulated rich traditional observation experience. The research time is April 2023, and the data sources include local fishermen's oral records, Jason-3 satellite altimeter data, and Argo float network observation data.

[0180] Step 1: Structured processing results of traditional knowledge

[0181] Through the LSTM network, the fisherman's description text is processed to extract standardized environmental clue words,

[0182] After K-means clustering analysis, a traditional knowledge graph is constructed, and the observation features and fishing results are represented by entity-relation-property triplets .

[0183] Figure 2 The structured processing results of fishermen's traditional knowledge are displayed, and the causal relationship network between the observation feature entities and the fishing result entities is visualized.

[0184] Step 2: Calculate the eddy physical parameters

[0185] Based on satellite altimeter data, the eddy automatic recognition algorithm and quasi-geostrophic theory are used to calculate the three-dimensional velocity field.

[0186] Figure 3 The distribution characteristics of the eddy physical parameters observed by satellites in Step 2 are shown.

[0187] Step 3: Calculate the mapping relationship between traditional knowledge and eddy parameters, and generate habitat prediction results

[0188] The mapping matrix between traditional observation features and eddy physical parameters is calculated using canonical correlation analysis , and based on the mapping relationship, the probability distribution of medium-depth fish habitat is generated.

[0189] Figure 4 The mapping matrix M between traditional observation features and eddy physical parameters is shown, with color depth representing the strength of the correlation coefficient.

[0190] Step 4: Quantify the uncertainty of traditional knowledge based on fuzzy logic;

[0191] Using triangular membership function:

[0192] Quantify the uncertainty of traditional knowledge.

[0193] Step 5: Obtain Argo buoy data and analyze the environmental suitability of middle water;

[0194] Calculate the environmental suitability index of each depth layer based on Argo buoy data. The temperature suitability index is calculated using .

[0195] Figure 5 Show the changes of environmental suitability index at different depths, including temperature, salinity, dissolved oxygen and comprehensive suitability index. Reveal the environmental preference distribution of middle fish in the depth range of 200-1000 meters.

[0196] Step 6: Simulate the vertical material transport process using particle tracking algorithm;

[0197] Based on the three-dimensional velocity field of vortex calculated in step 2, use Lagrangian particle tracking method to simulate the transport path of nutrients from the sea surface to the middle layer, and calculate the time distribution of particles reaching different depths .

[0198] Figure 6 Show the transport flow direction and flux distribution of Lagrangian particles from the sea surface to different depth layers, and the width of the flow band represents the transport flux size. Reveal the hierarchical characteristics of vortex vertical transport and the efficiency of nutrient transfer.

[0199] Step 7: Integrate multi-source prediction information based on Bayesian inference;

[0200] Input the mapping results, fuzzy membership, environmental suitability index and transport flux distribution into the Bayesian inference network, and calculate the posterior probability distribution through conditional probability update .

[0201] Step 8: Integrate the prediction results using ensemble learning;

[0202] Integrate multi-source prediction results using random forest, and output the final habitat probability distribution map.

[0203] This implementation example successfully converts fishermen's traditional observation experience into quantitative knowledge graph triple representation:

[0204] The confidence of the causal relationship between the “abnormal change of water color” and the “high yield” reaches 0.85, and the confidence of the “warm water temperature phenomenon” reaches 0.82. The semantic-physical mapping matrix is established by the canonical correlation analysis The correlation coefficient is between 0.66 and 0.83, wherein the correlation coefficient of the “warm water temperature phenomenon” and the vortex parameter is the highest (0.83).

[0205] The uncertainty quantification result of the fuzzy logic processing shows that the membership degree of the traditional knowledge ranges from 0.67 to 0.92, wherein the membership degree of the “warm water temperature phenomenon” is the highest (0.92), and the credibility of the experience knowledge is effectively quantified. The particle tracking result shows that the average time of the nutrient substance from the sea surface to the depth of 400 meters is 7.8 days, the transport flux is , and the transport efficiency reaches 0.85, verifying the effectiveness of the vortex vertical transport.

[0206] After the Bayesian inference fuses the multi-source information, the posterior probability ranges from 0.59 to 0.79, wherein the existence probability of the habitat at the position (36.8°N, 147.3°E) is the highest (0.79). The final integrated learning result shows that the high suitable habitat accounts for 23.5% of the research sea area, mainly distributed in the 400-meter depth layer, which is highly consistent with the sinking flow area of the anticyclonic vortex, verifying the internal relationship between the observation rule of the “fish gathering in the dark water color area” in the traditional knowledge and the vortex physical mechanism.

[0207] The above describes the embodiments of the present application, but the embodiments are not limited to the above specific implementation, and the above specific implementation is only illustrative but not restrictive, and the person skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection scope of the embodiments.

Claims

1. A habitat suitability prediction method for marine ecological management, characterized by, The method comprises the following steps: acquiring traditional knowledge data of fishermen, including collecting oral information about sea surface observation rules and corresponding fishing records from fishermen; transcribing audio data into text format through natural language processing technology, extracting environmental clue words and spatial orientation descriptions using long short-term memory networks, and generating a traditional knowledge graph containing the corresponding relationship between observation features and fishing results; acquiring satellite vortex observation data, detecting vortex boundaries and center positions using an automatic vortex recognition algorithm, calculating three-dimensional velocity fields and density fields of the vortex based on quasi-geostrophic theory, and outputting vortex feature data sets containing vortex intensity, radius, and vertical velocity; calculating the mapping relationship between traditional knowledge and vortex parameters, performing canonical correlation analysis on surface observation features in the traditional knowledge graph and vortex physical parameters, calculating a semantic-physical mapping matrix, and converting traditional knowledge into quantitative expressions in the vortex parameter space based on the mapping matrix; fusing multiple source prediction information based on Bayesian inference, inputting the mapping results, fuzzy membership of traditional knowledge, environmental suitability index, and transport flux distribution into the Bayesian inference network, and calculating posterior probability distribution through conditional probability update; using a particle tracking algorithm to simulate the vertical material transport process, which comprises the following steps: based on the three-dimensional velocity field of the vortex, using the Lagrangian particle tracking method to simulate the transport path of nutrients from the sea surface to the middle layer, solving the motion equation by numerical integration, calculating the particle position update using the fourth-order Runge-Kutta method, calculating the time distribution of particles reaching different depths from the surface layer, and generating a spatiotemporal distribution map of vertical transport flux; integrating the prediction results using ensemble learning, integrating multiple source prediction results using random forest, taking traditional knowledge data and physical model calculation results as input features, and outputting a middle layer fish habitat probability distribution map.

2. The method of claim 1, wherein, The traditional knowledge graph is constructed using entity-relation-attribute triple representation, the graph nodes include environmental observation entities and fishing result entities, and the edges represent the causal relationship between observation features and fishing results; environmental feature words and fishing quantity values are extracted through named entity recognition, the syntactic relationship between words is identified using dependency syntax analysis, the mapping relationship between observation features and fishing results is established through a relationship extraction algorithm, and the confidence weight is calculated based on the frequency and consistency of fishermen's descriptions.

3. The method of claim 1, wherein, The extraction of environmental clue words is realized through word frequency statistics and semantic clustering algorithm, the semantic clustering algorithm uses the K-means clustering method, the input is the vector set of fishermen's description text after word embedding processing, the word embedding processing uses the pre-trained Word2Vec model, the vectorization of text fragments uses the weighted average method of word vectors, and the weight is weighted using term frequency-inverse document frequency value to highlight the importance of key words.

4. The method of claim 1, wherein, The automatic vortex recognition algorithm comprises the following sub-steps: spatial filtering is performed on the input sea surface height anomaly data field to remove noise smaller than twice the grid resolution; calculate the sea surface height gradient of each grid point, and identify the local extreme points with a gradient less than a threshold value as potential vortex centers; starting from each potential center, search for closed contour along the radial direction, calculate the area and equivalent radius within the contour; Selecting the closed contour line satisfying the radius constraint and the shape parameter constraint as the vortex boundary; Outputting the vortex center coordinates, vortex radius and vortex intensity.

5. The method of claim 1, wherein, It also includes the step of quantifying the uncertainty of traditional knowledge based on fuzzy logic: converting the qualitative description in traditional knowledge data into membership function of vortex characteristics, defining the corresponding fuzzy set for each observed feature, quantifying the degree to which the feature value belongs to the fuzzy set through triangular or trapezoidal membership function, and generating knowledge representation containing uncertainty range.

6. The method of claim 1, wherein, It also includes the step of obtaining the environmental suitability of the middle layer water body in Argo buoy data analysis: obtaining the vertical profile data of temperature, salinity and dissolved oxygen from Argo buoy database, interpolating the profile data to the cylindrical coordinate system with vortex center as the origin, evaluating the environmental suitability index of each depth layer based on the physiological demand threshold of fish, and calculating the environmental suitability index through the product of temperature suitability index, salinity suitability index and dissolved oxygen suitability index.

7. The method of claim 1, wherein, The input of the canonical correlation analysis method is two groups of variable matrices: traditional observed feature matrix and vortex physical parameter matrix. Due to the difference in dimension, the two groups of matrices need to be preprocessed by Z-score standardization before calculating the correlation. The mapping relationship is obtained by solving the eigenvalue problem, and the output is the canonical correlation coefficient and the corresponding canonical variable coefficient matrix.

8. The method of claim 1, wherein, The random forest is composed of multiple decision trees, each of which is constructed by recursive splitting nodes, and the splitting criterion adopts Gini impurity. The input is the fused feature vector, including traditional knowledge mapping value, environmental suitability index and transport flux. The feature vector needs to be preprocessed by standardization before inputting into the random forest, and the output is the habitat existence probability.

9. A habitat suitability prediction system for marine ecological management for performing the method of any one of claims 1-8, characterized by It includes: Traditional knowledge acquisition module, for acquiring fisherman's traditional knowledge data and generating structured knowledge representation; Marine data processing module, for acquiring and processing satellite vortex data and marine environment data; Semantic mapping module, for establishing the mapping relationship between traditional knowledge and physical parameters; Multi-source fusion module, for fusing multiple prediction information sources; Prediction output module, for generating habitat distribution prediction results.

Images