Method for predicting biomass of charybdis japonica

Abstract

The invention discloses a method for predicting the biomass of charybdis japonica. The method comprises the following steps: step a, collecting, cleaning and processing ecological environment monitoring indexes and biomass data of the charybdis japonica; b, screening ecological environment monitoring indexes based on the magnitude of a variance expansion factor VIF; c, constructing a BP neural network model, and optimizing parameters of the BP neural network model; and d, predicting the growth trend of the charybdis japonica, and outputting a result. According to the method, the problem that the charybdis japonica prediction model is learned and constructed from limited supervision data in a marine ranch is solved, the accuracy of biomass prediction of the charybdis japonica is effectively improved, meanwhile, the prediction cost is reduced, and the method has sustainable use significance.

Description

Technical Field

[0001] This invention belongs to the field of biomass prediction technology, and particularly relates to a method for predicting the biomass of Japanese crab. Background Technology

[0002] As an important marine economic crab species, the study of biomass prediction methods for the Japanese crab is of great significance. The marine ecological environment is a highly complex and dynamic system formed by physical, chemical, and biological processes. Different ecological environment monitoring indicators, such as temperature, salinity, dissolved oxygen, and food availability, affect the Japanese crab's metabolism, feeding, and digestion, thereby influencing its growth rate, growth quality, and biomass. Methods for predicting Japanese crab biomass can be developed in two ways: firstly, by setting up different environmental factor gradients in the laboratory and observing the growth of the Japanese crab under different conditions to determine the threshold and optimal range of each environmental factor's impact on growth; secondly, by conducting long-term monitoring in natural marine areas or aquaculture environments, collecting ecological environment monitoring indicator data and Japanese crab biomass predictions, and using numerical simulation and other methods to establish a model of the relationship between ecological environment monitoring indicators and biomass.

[0003] In recent years, with the rapid development of artificial intelligence, using machine learning technology to study the relationship between marine ecological environment monitoring indicators and marine organisms has become an emerging trend in interdisciplinary research. Advanced algorithms and models have provided unprecedented capabilities for the monitoring, analysis, and prediction of Japanese crab biomass. In this application field, few-shot learning is an important research direction, as it can learn from limited supervised data and build excellent predictive models. Therefore, using machine learning technology, based on few-shot learning algorithms, and combined with field-collected data on Japanese crabs and ecological environment monitoring indicators, to establish a model of the relationship between Japanese crab biomass and ecological environment monitoring indicators, in order to analyze and predict the dynamic changes of Japanese crabs, has become a major need in marine ecological research.

[0004] Currently, the biomass prediction technology for Japanese crabs faces the following problems and challenges: (1) Due to the vastness and inaccessibility of marine areas, traditional research and monitoring methods often have limited resources. Data on Japanese crabs are usually limited, and it is very difficult to explore the mapping relationship between Japanese crabs and ecological environment monitoring indicators using small sample data.

[0005] (2) Establishing an accurate growth prediction model requires long-term, continuous and accurate data support. However, in the actual monitoring process, data collection may be affected by various factors, such as the accuracy of monitoring equipment and human error. The quality and reliability of ecological environment monitoring index data are usually not ideal.

[0006] (3) In the actual environment, multiple factors interact and influence each other, jointly affecting the growth of Japanese crab. Although current research considers multiple ecological and environmental monitoring indicators, it lacks a deep understanding of the complex interaction mechanisms between these indicators, making it difficult to accurately quantify this comprehensive impact. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method for predicting the biomass of Japanese crabs. This invention improves the accuracy of predicting future trends of marine organisms, better fits the changing trends of Japanese crab biomass, solves the problems of inaccurate and low-precision prediction of marine organism biomass in the past, and reduces the cost required for prediction, thus having sustainable application significance.

[0008] The technical problem solved by this invention is achieved through the following technical solution: A method for predicting the biomass of Japanese crab, characterized by comprising the following steps: Step a. Collect, clean, and process ecological environment monitoring indicators and Japanese crab biomass data; Step b. Screen ecological and environmental monitoring indicators based on the variance inflation factor (VIF). Step c. Construct a BP neural network model and optimize its parameters; Step d. Predict the growth trend of Japanese crab and output the results.

[0009] Preferably, records with empty data collection results for Japanese crab biomass in step a are removed, and valid data are retained; ecological and environmental monitoring indicators with more than 50% missing values ​​are eliminated, and ecological and environmental monitoring indicators with less than or equal to 50% missing values ​​are retained; the average value of each retained ecological and environmental monitoring indicator is calculated, and the missing values ​​are filled with the average value of the indicator, thus achieving data cleaning.

[0010] Preferably, step a normalizes the cleaned data, mapping all data to the range [0,1].

[0011] Preferably, if the VIF in step b is greater than or equal to the preset standard, it is determined that there is serious multicollinearity and is removed, while ecological and environmental monitoring indicators that are determined not to have serious multicollinearity are retained.

[0012] Preferably, the GMDH self-organizing algorithm is used to screen ecological and environmental monitoring indicators that are determined not to have serious multicollinearity.

[0013] Preferably, the BP neural network structure constructed in step c includes three layers: an input layer, a hidden layer, and an output layer.

[0014] Preferably, the optimization of the BP neural network model parameters in step c includes adding the sum of squares of the model weights to the loss function and using a genetic algorithm to adjust the connection weights and bias matrix of the BP neural network.

[0015] Preferably, the genetic algorithm uses binary encoding for its parameters, fitness values ​​for its individual performance indicators, and roulette wheel selection for its chromosomes.

[0016] Preferably, the output results in step d are quantitatively evaluated, and the quantitative evaluation methods can include mean squared error, coefficient of determination, and explained variance fraction.

[0017] Compared with the prior art, the present invention has the following advantages: 1. This invention uses methods such as VIF (variance inflation factor) and GMDH (data processing grouping method) to process and screen key ecological and environmental monitoring indicators, which improves the utilization rate of key data and is of great significance for the small sample learning biomass prediction model of Japanese crab. 2. By introducing L2 regularization and genetic algorithm, this invention initially solves the problems of low learning efficiency, slow convergence speed and model getting stuck in local optima in the optimization process of the Japanese crab biomass prediction model, thereby improving the prediction accuracy of Japanese crab biomass. 3. The Japanese crab biomass prediction technology proposed in this invention can provide data support for optimizing the stock enhancement and release program, guide relevant departments and enterprises to formulate more detailed marine biological stock enhancement and release and growth management plans, and adjust physical, chemical and biological conditions based on this technology to ensure the health of biological populations and the suitability of the growth environment, which is conducive to the protection and reproduction of species and provides an effective means to protect biodiversity and improve fishery production. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a biomass prediction method for Japanese crabs combining a genetic algorithm and a backpropagation neural network; Figure 2 This is a schematic diagram of the BP neural network structure; Figure 3 This is a schematic diagram of the chromosome structure in a genetic algorithm. Figure 4 This is a schematic diagram illustrating the operation of a genetic algorithm; Figure 5 The simulation fit plot is shown for the BP model using L2 regularization (L2BP). Figure 6 This is a simulation fitting graph of the BP model based on genetic algorithm (GABP); Figure 7 The simulation fitting plot is for the basic BP model; Figure 8The simulated fit plot for Ridge Regression; Figure 9 A simulated fit plot of a BP model (deep regression) using the Adam optimizer; Figure 10 A schematic diagram showing the extreme points of biological resource density.

[0019] Figure 11 A quantitative comparison chart of the effectiveness of methods for predicting the biomass of Japanese crabs. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0021] To verify the effectiveness of the method of the present invention, this embodiment of the invention relies on the Tianjin Dashentang National Marine Ranch Demonstration Zone (Tianjin Marine Ranch) to simulate and predict the Japanese crab in the Tianjin Marine Ranch. The specific implementation process is as follows: Step a. Collect and process ecological and environmental monitoring data to ensure the reliability of the prediction results. 1) Data Source This case study is based on real-world marine ecological environment monitoring data, collected by a professional offshore team through routine monitoring of the Tianjin sea area using cruise monitoring and information infrastructure. The survey area was located between 117.6° and 118.0° east longitude and 38.6° and 39.2° north latitude, with a total of 15 survey stations established. The data collection period was primarily from May 2015 to September 2022, but also included some historical data from 2006 to 2007. Monitoring was concentrated from April to October each year. The collected information includes data in five dimensions: spatial information, temporal information, physical indicators, chemical indicators, and biological indicators, as well as the biomass data of *Hemiberlesia lataniae* (unit: kg / km2) as the dependent variable.

[0022] 2) Data cleaning In this example, the ecological environment monitoring indicator data was first cleaned. Specifically, all records with empty data for *Hemiberlesia lataniae* were removed, leaving 247 valid records. Ecological environment monitoring indicators with more than 123 missing values ​​were removed, and the missing values ​​of the remaining ecological environment monitoring indicator data were filled with their own average values. This maximized data retention and avoided excessive impact on the data distribution.

[0023] 3) Normalization processing The Min-Max normalization method was adopted to map all Japanese crab biomass density data and ecological environment monitoring index data to the range [0,1]. This was done to remove the influence of different dimensions and units of measurement, and also to improve the computational speed of the algorithm.

[0024] Step b. Screening ecological and environmental monitoring indicators based on the variance inflation factor (VIF). 1) Remove multicollinearity Multicollinearity refers to a high degree of correlation among explanatory variables, which can affect the stability and accuracy of a regression model. The Variance Inflation Factor (VIF) is a measure of the severity of multicollinearity in a multiple linear regression model. It represents the ratio of the variance of the regression coefficient estimates to the variance assumed to be non-linearly correlated among the independent variables. Its calculation formula is:

[0025] in R 2 Is the first j The coefficient of determination (VIF) is the result of linear regression of each explanatory variable against other environmental variables. Generally, when VIF < 10, there is no multicollinearity; when 10 <= VIF < 100, there is strong multicollinearity; and when VIF >= 100, there is severe multicollinearity.

[0026] In this example, VIF>=100 is used as the criterion for rejection.

[0027] 2) Reduce model complexity When there are numerous environmental variables, using them all as input to a BP network increases network complexity, reduces performance, significantly increases computation time, and impacts accuracy. To address this issue, this example uses the GMDH (Group Method of Data Handling) self-organizing algorithm. GMDH is a heuristic feedforward algorithm for studying relationships between variables. Its core idea is to construct a predictive model for the data through progressive model combination and selection. Based on adaptive polynomial function combinations, it effectively handles nonlinear relationships in the input data and continuously improves model accuracy through iteration, thereby selecting the variables with the greatest influence on the dependent variable as input nodes for the BP network.

[0028] By taking into account the biological habits of Japanese crabs through two steps of "removing multicollinearity" and "reducing model complexity", key factors and cleaned data were obtained. The final results include 13 ecological environment monitoring indicators, including: season, offshore distance, nitrate, nitrite, reactive nitrogen, reactive phosphate, phytoplankton biomass, phytoplankton diversity, chlorophyll a, abundance of larvae and juvenile fish, abundance of fish eggs, benthic biomass, and abundance of benthic organisms.

[0029] Step c. Construct a BP neural network model and optimize its parameters. 1) Determining the structure of the BP neural network model The back propagation algorithm (BP) is a very common machine learning algorithm used to adjust the weights and biases of neural networks to achieve efficient learning and pattern recognition. The basic idea of ​​the BP algorithm is to compare the network's output with the desired target output, and then propagate the error back through the network, adjusting the weights and biases layer by layer. It uses gradient descent to minimize the error function, allowing the neural network to gradually approach the optimal solution of the objective function.

[0030] refer to Figure 2 The steps of the BP algorithm are briefly described below: Initialization: Randomly initialize the network's weights and biases.

[0031] Forward propagation: Input samples are passed through the network, and the output value of each neuron is calculated.

[0032] Error calculation: The network's output value is compared with the desired target output to calculate the error of the output layer. Here, MSE is used as the error function.

[0033] Backpropagation: Starting from the output layer, the error is propagated back to the neurons in the previous layer using the chain rule, and the error of each neuron is calculated.

[0034] Update weights and biases: Using the calculated error, update the weights and biases in the network according to the gradient descent method, so that the error function gradually decreases.

[0035] Repeat steps 2 to 5: repeat forward propagation, error calculation, back propagation and weight update until the preset stopping condition is met (such as reaching the maximum number of iterations or the error is reduced to an acceptable range).

[0036] To avoid overfitting caused by too many model parameters in few-shot learning, this example uses a three-layer BP network structure, including an input layer, a hidden layer, and an output layer.

[0037] 2) Improve predictive ability and ensure stable results To enhance the model's generalization ability and improve prediction accuracy, this example optimizes the original BP model by introducing mini-batch training and L2 regularization. This optimization method does not change the basic training steps. L2 regularization is a commonly used technique to prevent overfitting in machine learning models. Overfitting refers to a model that performs well on training data but poorly on unseen new data. This phenomenon is usually due to the model being too complex, overfitting to noise in the training data, and failing to capture the true underlying patterns.

[0038] L2 regularization is achieved by adding an extra term to the model's loss function, which consists of the sum of squared model weights. This extra term penalizes models with excessively large weights, as this can lead to overcomplication. This extra term is also called the regularization term, and its preceding coefficient (usually denoted as λ) is a hyperparameter, often referred to as the weight decay coefficient, used to control the strength of regularization. The formula is as follows:

[0039] Recorded as

[0040] After adding the L2 regularization term, the formula for calculating the derivative and the implementation method of backpropagation also change accordingly, as follows:

[0041] The addition of mini-batch training and L2 regularization to the BP model enhances its ability to predict extreme values. The weight decay coefficient λ is adjusted in the original experimental method parameters, but the accuracy of the model's prediction still has problems. Although the model can predict values ​​close to the maximum, the prediction time point still needs to be adjusted to achieve more accurate predictions.

[0042] 3) Optimize the BP model parameters to accurately locate the search space. Backpropagation (BP) models often face challenges during optimization, including low learning efficiency, slow convergence, and susceptibility to local optima. To address these issues, this example employs a genetic algorithm (GA) for heuristic optimization of the BP model's parameters. By using a genetic algorithm to determine the initial values ​​of the connection weights, the overall performance of the model can be improved, further optimizing the BP model's training process.

[0043] The Genetic Algorithm (GA) is an optimization search algorithm that simulates natural selection and genetic mechanisms. It exhibits excellent performance in solving complex optimization problems, especially when no explicit solution exists. As a non-derivative optimization stochastic method, it is suitable for global searches of complex, multimodal, nonlinear, and non-differentiable functions. In contrast, the Backpropagation (BP) algorithm performs better in local searches. To quickly find a satisfactory solution, the initial weights can be optimized using a genetic algorithm to locate a better search space in the solution space. Then, the BP algorithm is used to search within these smaller solution spaces to find the optimal solution. This example uses the connection weights and bias matrices between every two layers of the BP model structure as the objects of genetic algorithm optimization.

[0044] Encoding method refer to Figure 3 To optimize the parameters of the connection weight matrix and biases, their upper and lower bounds are used as decision variables. Simultaneously, a precision value (delta) needs to be chosen to determine the number of decimal places the result can be accurate to. There are two ways to encode the parameters: binary encoding and real-number encoding; binary encoding is chosen here. Then, based on the range of the decision variables, the precision requirements, and the neural network structure, the binary encoding length of each decision variable is calculated using the following formula: 2 length-1 <(upper-lower) / delta<=2 length-1 By solving nonlinear equations ((upper-lower) / delta-2 x +1)=0 The final encoded length L is obtained by rounding up. Correspondingly, the binary code can be converted to the corresponding floating-point number (decoding process) using the following formula: Genetic algorithms simulate the evolutionary process of organisms under natural conditions, including the survival of the fittest and the ability to adapt to the environment. In a genetic algorithm, the population is abstracted as a group of multiple individuals, each represented as a chromosome, with the parameters on the chromosome corresponding to genes. The evolutionary process in a genetic algorithm involves chromosome crossover and mutation.

[0045] In the scenario of optimizing the BP parameter matrix, the parameters of a chromosome are constructed by concatenating four parameter matrices row-wise; therefore, one chromosome represents a set of parameter matrices. Crossover between chromosomes is characterized by a certain probability of exchanging a segment of two chromosomes, while mutation is characterized by a certain probability of replacing a segment of a chromosome with a random binary value.

[0046] All these chromosome operations are performed within a single population. During each generation of evolution, the crossover and mutation rates remain constant, while the population generates a certain number of new chromosomes that are compared with the existing chromosomes. The chromosome with the highest fitness is then selected as part of the next generation, forming a new population. In genetic algorithms, several variables need to be set, including the crossover rate, mutation rate, maximum number of offspring, population size, and the number of newly generated chromosomes. In experiments, these variables are continuously adjusted and the optimal combination is selected.

[0047] Individual performance identification refer to Figure 4 The fitness function plays a crucial role in genetic algorithms. It iterates through each individual in the population, uses the neural network parameters represented by that individual to make predictions, and calculates the error between the predicted value and the actual label. The error is calculated using squared error, which is the sum of squared differences between the neural network output and the dataset labels. The purpose of the fitness function is to measure the performance of an individual; by calculating the prediction error of each individual, its performance in solving a specific problem can be evaluated. A higher fitness value indicates better performance. Ultimately, the fitness values ​​of all individuals are used as important indicators in genetic algorithms for operations such as selection, crossover, and mutation. By comparing the fitness values ​​of individuals, it is possible to determine which individuals should be selected for the next generation of the population during evolution, and which individuals should participate in crossover and mutation operations.

[0048] in, fitness Indicates the fitness value. y represents the label of the dataset, and represents the predicted value of the neural network. By calculating the sum of squared errors, the overall prediction error of the neural network for a given dataset can be obtained.

[0049] Chromosomal selection In genetic algorithms, Roulette Wheel Selection is a probability-based selection method used to select chromosomes in a population. It uses an individual's fitness value to calculate the selection probability, giving individuals with higher fitness a greater selection probability and individuals with lower fitness a smaller selection probability. This selection method works by simulating a roulette game, where the selection probability corresponds to the sector size on the roulette wheel.

[0050] The selection process in roulette wheel selection is as follows: First, the selection probability of each individual is calculated by dividing its fitness value by the sum of the fitness values ​​of all individuals in the population, thus determining its relative weight in the selection process. Then, the cumulative probability is calculated by summing the selection probabilities according to the individuals' order in the population. Next, a random number between 0 and 1 is generated, and this random number is used to determine the selected individual. Specifically, by comparing the random number with the cumulative probability, the first individual with a probability greater than the random number is selected.

[0051] Through roulette wheel selection, individuals with higher fitness have a greater probability of being selected, and are therefore more likely to be chosen. This gives superior individuals a greater chance to pass on their excellent genetic information to the next generation. As the selection process is repeated, the individuals in the population gradually tend towards the best solution, realizing the evolutionary process of the genetic algorithm.

[0052] ① First, randomly initialize n individuals, and the encoding length of each individual corresponds to the required binary encoding length (total number of parameters). (The encoding length of a single value). These individuals form the initial population, each representing a different BP model parameter; ② Decode each chromosome in the population to obtain n sets of connection weights and bias matrices for the BP model. Then, input these parameters into the network for testing to obtain the fitness value of each individual; ③ Next, the next generation population is generated according to the iteration rules (based on content (2) and (4)) to form a new network. That is, new individuals are generated by using genetic algorithm operations such as selection, crossover and mutation; ④ Repeat steps ② and ③ until the maximum number of offspring generations is reached, and so on, until the preset maximum number of offspring generations is reached. In each generation, the chromosomes in the population are decoded, and the individual with the highest fitness value is selected. Finally, the individual with the highest fitness value is obtained, and its decoded connection weights and bias matrix are regarded as the optimal parameters of the BP model; ⑤ Finally, the individual with the highest fitness value is used as the initial parameter of the BP model; In summary, genetic algorithms, by simulating the principles of biological evolution, optimize the solution space of a problem through operations such as selection, crossover, and mutation. Through multiple iterations and selections, they gradually find individuals with higher fitness, thereby obtaining a better solution. In scenarios involving the optimization of backpropagation (BP) models, genetic algorithms can help find optimal connection weights and bias matrices, improving the model's generalization ability and prediction accuracy, and reducing the risk of getting trapped in local optima.

[0053] Step d. Predict the growth trend of the Japanese crab and output the results. After extracting and selecting key factors using VIF (variance inflation factor) and GMDH (self-organizing method), the BP model retained 13 independent variables for training. Experimental results show that the model achieves the best prediction performance when the number of hidden neurons is set to 8, the number of training epochs is 700, and the weight matrix is ​​set to [-5, 5]. It can be seen that the mean squared error decreases, and the ability to fit changing trends is improved. By using VIF and GMDH methods to select key factors, the independent variables affecting model performance can be more accurately identified, and these key factors can be retained as independent variables in subsequent improvements to enhance model performance.

[0054] The L2BP model showed improved generalization ability and prediction accuracy after adding batch training and regularization. After adding regularization, an additional adjustment parameter, the weight decay coefficient λ, was introduced. Experimental results showed that the model achieved the best prediction performance with 8 hidden neurons, 1000 training epochs, and a weight matrix of [-5, 5]. Experiments were conducted with different values ​​of λ, specifically 0.01, 0.005, and 0.001.

[0055] refer to Figure 5 Observing the experimental results, it can be seen that when λ is 0.005, the model has high prediction accuracy and its ability to predict extreme values ​​is improved, approaching the true value. However, the predicted extreme value occurrence time does not completely match the true value, indicating that there is still room for improvement.

[0056] After introducing the genetic algorithm into the GABP model, the parameters that need to be tuned increase, including crossover rate, mutation rate, maximum number of offspring, population size, and number of newly generated chromosomes. However, since an excessively large population size or too many generations can slow down the experiment, the value range of these parameters needs to be carefully selected. Experiments have shown that this example selects a moderate value range. Specifically, the crossover rate is chosen between 0.7 and 0.9, the mutation rate between 0.01 and 0.2, and the population size between 10 and 20. The number of newly generated chromosomes is kept consistent with the population size, and the maximum number of offspring is chosen between 30 and 80. These parameter selections aim to balance algorithm performance and experimental efficiency. Experimental results show that, under the conditions of 8 hidden neurons, 700 training rounds, and a weight matrix of [-5, 5], the model achieves the best prediction performance when the crossover rate is set to 0.9, the mutation rate to 0.2, the population size to 20, and the maximum number of offspring to 50. refer to Figure 6 Under the above parameter combination, the model can predict extreme values ​​and simulate the trend of data changes relatively well. However, it should be noted that the bias of the model has increased.

[0057] Compared with traditional methods for predicting the biomass of Japanese crabs This study focuses on adjusting parameters and training strategies based on the BP model and its improved methods to obtain a superior method for predicting the biomass of Japanese crabs. It compares and analyzes this method with the basic BP model, ridge regression, and regression algorithms using the Adam optimizer, using mean-squared error and coefficient of determination (R²). 2 The quantitative analysis uses the score and the explained variance score as reference evaluation indicators.

[0058] 1) Evaluation indicators Mean Squared Error (MSE) is one of the most commonly used performance metrics in regression tasks, used to measure the degree of difference between model predictions and actual observations. It calculates the average of the squares of the differences between the predicted and actual values; the smaller the value, the closer the model's prediction is to the true value.

[0059]

[0060] Where n represents the number of samples, y represents the actual observed value, and represents the model's predicted value.

[0061] The floating accuracy rate is defined as follows: points where the deviation from the true value does not exceed 5% are considered accurately predicted. The proportion of accurately predicted points to the total number of sample points in the extreme value region is used to calculate the floating accuracy rate, as shown in the formula:

[0062]

[0063] in pre , y test These refer to the predicted value and the actual sample value at the corresponding time points, respectively. COUNT stands for the counting function, and ABS stands for the absolute value function. AC n , Sample n This refers to the number of samples that are accurately predicted versus the total number of samples.

[0064] 2) Comparison Model Basic BP model To select the optimal parameter combination for the basic BP model, a series of experiments were designed to adjust the number of hidden layer neurons, the number of training epochs, and the upper and lower bounds of the weight matrix. The ratio of the training set to the test set was set to 7:3, and the mean squared error was used as the loss function, while the fit of the image was considered to measure the model's performance. Unfiltered datasets were used in the experiments for the basic BP model. The number of hidden neurons was adjusted from 4 to 16, the number of training epochs from 300 to 2000, and the upper and lower bounds of the weight matrix were selected between [-1, 1] and [-5, 5]. By observing the learning curves and convergence of the model under different training epochs, the model showed the best prediction performance when the number of hidden neurons was set to 4, the number of training epochs was 600, and the weight matrix was selected within the range of [-5, 5].

[0065] refer to Figure 7 The model can detect relatively small overall errors, but it is less effective in capturing fluctuation trends and simulating extreme values.

[0066] Ridge Regression refer to Figure 8 Ridge regression performs well in fitting data trends, and its predictions are close to the actual values, with a relatively small mean squared error. This result is attributed to the loss function of ridge regression, which consists of the regression squared error and a regularization term. The regularization term constrains the regression coefficients, helping to reduce the impact of multicollinearity on the model.

[0067] BP model using Adam optimizer refer to Figure 9The training results of deep regression models are highly unstable, largely depending on the appropriateness of the random initialization of the weight matrix. Otherwise, the model may output a horizontal line or exceed the expected range of predictions. This indicates that network models implemented using the PyTorch framework and the Adam optimizer are not suitable for tasks with small sample sizes.

[0068] 3) Comparative evaluation results refer to Figure 10 In practical applications, this example focuses more on predicting the location of extreme points, with accuracy being secondary. Therefore, this example quantitatively evaluates the method from the perspective of predicting extreme points. Due to the temporal characteristics of the data, the resource density of organisms will briefly reach a maximum, and will be at the lowest density level most of the time. The regions in the test set data where extreme points exist are divided into two segments, labeled T1 and T2 respectively.

[0069] The MSE was calculated separately for the data in these two time periods, as shown in Table 3.2-6 below. It can be seen that L2BP has a larger error in the T1 time period due to the error in predicting the extreme point time, while the GABP algorithm has the least error in the region with the maximum point and has a better prediction effect for the maximum value.

[0070] refer to Figure 11 The accuracy rates of each method are as follows. Overall, the GABP method proposed in this example is superior in both the prediction of extreme values ​​and the accuracy of the prediction of extreme values. For the topic of Japanese crab biomass prediction, simulating the trend and predicting the time point of the maximum value will have more practical application value and can provide decision support for farmers. Accuracy is a secondary focus, and a small range of fluctuations is allowed. The specific predicted value can only be used as a decision reference.

[0071] The basic backpropagation (BP) model can achieve good predictive results under certain parameter combinations, but it performs poorly in simulating extreme values. Introducing the L2BP and GABP models can improve prediction accuracy and the ability to simulate extreme values ​​to some extent. Among the compared models, ridge regression exhibits good performance, fitting the data trend well and approaching the true values, while the training results of deep regression models are very unstable, largely affected by random initialization, and are not suitable for small sample sizes.

[0072] Overall, the GABP model introduces a heuristic genetic algorithm compared to other methods. By performing a global search of the solution space through genetic operations, it successfully optimizes the problem of easily getting trapped in local optima and fits extreme points well. After feeding the optimized parameter matrix into the L2BP model for training, the overall simulation effect of fluctuation trends is also improved. This is mainly because a regularization term is introduced to reduce the impact of multicollinearity, enabling the model to better handle collinearity issues in multivariate regression tasks. Furthermore, L2 regularization can be intuitively understood as severely penalizing large-value weight vectors, favoring more dispersed weight vectors. Due to the multiplication operation between input and weights, this has a beneficial property: it makes the network more inclined to use all input features, rather than heavily relying on a small subset of features. The L2 penalty favors smaller, more dispersed weight vectors, which encourages the model to ultimately use features across all dimensions, rather than strongly relying on a few. This improves the model's generalization ability and reduces the risk of overfitting.

[0073] Compared with the prior art, the present invention has the following advantages: 1. This invention uses methods such as VIF (variance inflation factor) and GMDH (data processing grouping method) to process and screen key ecological and environmental monitoring indicators, which improves the utilization rate of key data and is of great significance for the small sample learning biomass prediction model of Japanese crab. 2. By introducing L2 regularization and genetic algorithm, this invention initially solves the problems of low learning efficiency, slow convergence speed and model getting stuck in local optima in the optimization process of the Japanese crab biomass prediction model, thereby improving the prediction accuracy of Japanese crab biomass. 3. The Japanese crab biomass prediction technology proposed in this invention can provide data support for optimizing the stock enhancement and release program, guide relevant departments and enterprises to formulate more detailed marine biological stock enhancement and release and growth management plans, and adjust physical, chemical and biological conditions based on this technology to ensure the health of biological populations and the suitability of the growth environment, which is conducive to the protection and reproduction of species and provides an effective means to protect biodiversity and improve fishery production.

[0074] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A method for predicting the biomass of Charybdis japonica, characterized by, The method comprises the following steps: Step a. Collect, clean and process the ecological environment monitoring indicators and the data of the biomass of Charybdis japonica; Step b. Screen the ecological environment monitoring indicators based on the size of the variance inflation factor (VIF); Step c. Construct a BP neural network model and optimize the parameters thereof; Step d. Predict the growth trend of Charybdis japonica and output the results.

2. The method of predicting biomass of Charybdis japonica according to claim 1, characterized by: In step a, the records with empty data of the biomass of Charybdis japonica are removed, and the valid data are retained; the ecological environment monitoring indicators with more than 50% of missing values are removed, and the ecological environment monitoring indicators with less than or equal to 50% of missing values are retained; the average value of each retained ecological environment monitoring indicator is calculated, and the missing values of the indicator are filled with the average value, so that the data cleaning is achieved.

3. The method of predicting biomass of Charybdis japonica according to claim 2, characterized by: In step a, the normalized processing is performed on the cleaned data, and all the data are mapped to the range of [0, 1].

4. The method of predicting biomass of Charybdis japonica according to claim 1, characterized by: In step b, if the VIF is greater than or equal to a preset standard, it is determined that there is serious multicollinearity, and the ecological environment monitoring indicators are removed; otherwise, the ecological environment monitoring indicators are retained.

5. The method of predicting biomass of Charybdis japonica according to claim 4, characterized in that: The GMDH self-organizing algorithm is performed on the ecological environment monitoring indicators determined to have no serious multicollinearity.

6. The new method of predicting biomass of Charybdis japonica according to claim 1, characterized by: In step c, the BP neural network structure constructed comprises three layers, namely an input layer, a hidden layer and an output layer.

7. The method of predicting biomass of Charybdis japonica according to claim 1, characterized in that: In step c, the parameter optimization of the BP neural network model comprises increasing the sum of squares of model weights in a loss function and adjusting the connection weights and bias matrix of the BP neural network by using a genetic algorithm.

8. The method of predicting biomass of Charybdis japonica according to claim 7, characterized by: In the genetic algorithm, the binary coding is selected as the parameter coding mode, the fitness value is used as the individual performance index, and the roulette selection is used as the chromosome selection mode.

9. The method of predicting biomass of Charybdis japonica according to claim 1, characterized in that: In step d, the output results are quantitatively evaluated, and the quantitative evaluation mode can adopt the mean square error, the determination coefficient and the explained variance score.

Images