A method and device for predicting the probability distribution of fishery production based on CNN-LSTM model

Abstract

The present invention provides a method and device for predicting the probability distribution of fishery yield based on a CNN-LSTM model, and relates to the technical field of fishery data processing; the method first collects historical fishing logs and a variety of marine environmental initial data, obtains environmental factors through preprocessing, calculates the unit fishing effort, calculates the correlation coefficient between the environmental factors and the factors, screens out the target correlation coefficient and the corresponding factors, and constructs a training sample set after formatting. Then, CNN and LSTM are used to construct an initial model, and a fishery prediction model is obtained through training with the sample set. The data to be analyzed is imported into the model, and the prediction result of the yield probability distribution is output. The invention integrates multi-source data and preprocesses it to enhance the adaptability of the model; screens features to reduce redundancy and interference, and improves model performance; the constructed CNN-LSTM fishery prediction model combines the advantages of the two, can fully explore the complex characteristics and laws of marine environmental data in the temporal and spatial dimensions, better model the relationship between fisheries and marine environmental factors, and significantly improves the prediction accuracy.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of fishery data processing, and specifically to a method and device for predicting the probability distribution of fishery output based on a CNN-LSTM model. Background Art

[0002] In the field of fishery resource management, for example, tuna fishing ground prediction is of great significance for rationally planning fishing operations, protecting the marine ecological balance and improving the economic benefits of fisheries.

[0003] Previous studies have tried various methods to forecast tuna fishing grounds, including artificial neural network model (ANN), habitat suitability index (HSI) model, Bayesian probability (BA) model, generalized additive model (GAM) and support vector machine (SVM), etc. These methods mainly use a single model combined with environmental and fishery data for forecasting, but there are some limitations.

[0004] Most of the existing integrated forecasting models rely on statistical principles, and have poor fitting effects on high-dimensional marine environmental data, making it difficult to effectively capture the spatiotemporal distribution characteristics of fishing grounds. For example, marine environmental data contains many types (such as vertical profile data, chlorophyll concentration, sea surface height anomalies, etc.) and has spatiotemporal dimensional characteristics. Traditional integrated learning models are difficult to fully mine the complex information contained therein and cannot meet the needs of accurate prediction of tuna fishing grounds. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method and device for predicting the probability distribution of fishery yield based on a CNN-LSTM model in view of the shortcomings of the prior art.

[0006] The technical solution of the present invention to solve the above technical problem is as follows: a method for predicting the probability distribution of fishery yield based on a CNN-LSTM model, comprising the following steps:

[0007] S1. Collect historical fishing logs and multiple types of initial marine environmental data;

[0008] S2. Preprocessing the initial data of each marine environment to obtain multiple marine environmental factors, and calculating the unit fishing effort based on historical fishing logs;

[0009] S3, respectively calculating the correlation coefficients between a plurality of marine environmental factors and the unit fishing effort, and selecting a plurality of target correlation coefficients related to the unit fishing effort from the calculated plurality of correlation coefficients based on a correlation threshold;

[0010] S4, formatting the target marine environmental factors corresponding to the multiple target correlation coefficients, and constructing a training sample set through the formatted target marine environmental factors and unit fishing effort;

[0011] S5. Constructing a CNN-LSTM fishing ground prediction initial model based on a CNN convolutional neural network and an LSTM long short-term memory network, and training the CNN-LSTM fishing ground prediction initial model on the correlation between fishing grounds and marine environmental factors through the training sample set to obtain a fishing ground prediction model;

[0012] S6. Import the data to be analyzed into the fishing ground capture prediction model, and output the fishing ground yield probability distribution prediction result.

[0013] Another technical solution of the present invention to solve the above technical problem is as follows: a device for predicting the probability distribution of fishery yield based on a CNN-LSTM model, comprising:

[0014] A data collection module for collecting historical fishing logs and multiple types of initial marine environmental data;

[0015] A preprocessing module is used to preprocess the initial data of each marine environment to obtain multiple marine environmental factors and calculate the unit fishing effort based on the historical fishing log;

[0016] a calculation module, used to respectively calculate the correlation coefficients between a plurality of marine environmental factors and the unit fishing effort, and screen out a plurality of target correlation coefficients related to the unit fishing effort from the calculated plurality of correlation coefficients based on a correlation threshold;

[0017] A formatting module is used to format the target marine environmental factors corresponding to the multiple target correlation coefficients, and to construct a training sample set through the formatted target marine environmental factors and unit fishing effort;

[0018] A model building module, for building a CNN-LSTM fishery prediction initial model based on a CNN convolutional neural network and an LSTM long short-term memory network, and training the CNN-LSTM fishery prediction initial model on the correlation between the fishery and the marine environmental factors through the training sample set to obtain a fishery prediction model;

[0019] The prediction module is used to import the data to be analyzed into the fishing ground capture prediction model and output the fishing ground yield probability distribution prediction result.

[0020] Another technical solution of the present invention to solve the above-mentioned technical problem is as follows: a device for predicting the probability distribution of fishery yield based on the CNN-LSTM model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that when the processor executes the computer program, the method for predicting the probability distribution of fishery yield based on the CNN-LSTM model as described above is implemented.

[0021] The beneficial effects of the present invention are: collecting and integrating various types of initial marine environmental data and historical fishing logs, performing preprocessing operations on the initial marine environmental data, enhancing the adaptability of the model to different types of data, calculating the correlation coefficient between the marine environmental factor and the unit fishing effort CPUE, and screening the target correlation coefficient and the corresponding target marine environmental factor based on the correlation threshold, effectively eliminating the features with low correlation with CPUE, reducing data redundancy and noise interference, reducing the calculation burden of the model, enabling the model to focus more on key features, thereby improving the prediction performance and efficiency of the model;

[0022] The constructed CNN-LSTM fishery prediction model combines the advantages of CNN convolutional neural network and LSTM long short-term memory network. The combination of the two can fully explore the complex characteristics and laws of marine environmental data in the temporal and spatial dimensions, better model the relationship between fisheries and marine environmental factors, and significantly improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A schematic diagram of a flow chart of a method for predicting the probability distribution of fishery production provided by an embodiment of the present invention;

[0024] Figure 2 A schematic diagram of a flow chart of a method for predicting the probability distribution of tuna fishery production provided by an embodiment of the present invention;

[0025] Figure 3 A schematic diagram of a multi-channel embodiment of the present invention;

[0026] Figure 4 A schematic diagram of the structure of a CNN-LSTM model provided in an embodiment of the present invention;

[0027] Figure 5 A schematic diagram of k-cross validation provided by an embodiment of the present invention;

[0028] Figure 6 A module block diagram of a device for predicting the probability distribution of fishery production provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0030] The purpose of the present invention is to provide a method for predicting the probability distribution of fishery yield based on a CNN-LSTM model, so as to effectively solve the problems of single models in the prior art being prone to overfitting, insufficient generalization ability, and difficulty in processing high-dimensional complex marine environmental data. The following is a detailed description through multiple specific embodiments.

[0031] Example 1: Figure 1 As shown, the embodiment of the present invention provides a method for predicting the probability distribution of fishery yield based on the CNN-LSTM model, comprising the following steps:

[0032] S1. Collect historical fishing logs and multiple types of initial marine environmental data;

[0033] S2. Preprocessing the initial data of each marine environment to obtain multiple marine environmental factors, and calculating the unit fishing effort based on historical fishing logs;

[0034] S3, respectively calculating the correlation coefficients between a plurality of marine environmental factors and the unit fishing effort, and selecting a plurality of target correlation coefficients related to the unit fishing effort from the calculated plurality of correlation coefficients based on a correlation threshold;

[0035] S4, formatting the target marine environmental factors corresponding to the multiple target correlation coefficients, and constructing a training sample set through the formatted target marine environmental factors and unit fishing effort;

[0036] S5. Constructing a CNN-LSTM fishing ground prediction initial model based on a CNN convolutional neural network and an LSTM long short-term memory network, and training the CNN-LSTM fishing ground prediction initial model on the correlation between fishing grounds and marine environmental factors through the training sample set to obtain a fishing ground prediction model;

[0037] S6. Import the data to be analyzed into the fishing ground capture prediction model, and output the fishing ground yield probability distribution prediction result.

[0038] In the above embodiment, various types of initial marine environmental data and historical fishing logs are collected and integrated, and the initial marine environmental data are preprocessed to enhance the adaptability of the model to different types of data. The correlation coefficient between the marine environmental factor and the unit fishing effort CPUE is calculated, and the target correlation coefficient and the corresponding target marine environmental factor are screened based on the correlation threshold, which effectively eliminates the features with low correlation with CPUE, reduces data redundancy and noise interference, reduces the computational burden of the model, and enables the model to focus more on key features, thereby improving the prediction performance and efficiency of the model.

[0039] The constructed CNN-LSTM fishery prediction model combines the advantages of CNN convolutional neural network and LSTM long short-term memory network. The combination of the two can fully explore the complex characteristics and laws of marine environmental data in the temporal and spatial dimensions, better model the relationship between fisheries and marine environmental factors, and significantly improve the prediction accuracy.

[0040] Preferably, the multiple types of marine environment initial data include a vertical profile data set, a chlorophyll concentration data set and a sea level height data set;

[0041] In step S2, each marine environment initial data is preprocessed to obtain multiple marine environment factors, including:

[0042] Setting a unified time resolution and spatial resolution for the data in the vertical profile dataset, the chlorophyll concentration dataset, and the sea level height dataset;

[0043] The missing values ​​in each data set are filled by the Euclidean distance interpolation formula, which is:

[0044]

[0045] Among them, S0 is the position where the missing value of vertical profile data, chlorophyll concentration data or sea level height data needs to be interpolated, is the predicted value corresponding to the missing value (such as temperature, salinity, dissolved oxygen concentration, etc. at a specific depth and location), ω k is the weight of the surrounding data points (such as adjacent depth layers or similar latitude and longitude points) of the missing value location. It is determined by the distance from the target point S0. Generally speaking, the closer the distance, the greater the weight. k is the serial number corresponding to the missing value from 1 to n. Z(S k ) is the actual attribute value of the surrounding sampling points;

[0046] The filled vertical profile data, chlorophyll concentration data and sea level height data are normalized by a normalization formula to obtain a plurality of marine environmental factors. The normalization formula is:

[0047]

[0048] Among them, x min and x max are the minimum and maximum values ​​in the vertical profile dataset, chlorophyll concentration dataset, or sea level height dataset, respectively. i is one of the vertical profile datasets, chlorophyll concentration datasets, or sea level height datasets, x Nor is the normalized value.

[0049] In this embodiment, data standardization and integrity are improved, and a unified time resolution and spatial resolution are set for initial marine environmental data such as vertical profile data sets, chlorophyll concentration data sets, and sea level height data sets, ensuring the consistency of data in time and space scales, which is convenient for subsequent comprehensive analysis and processing. Filling missing values ​​with the Euclidean distance interpolation formula can effectively utilize the information of surrounding data points, reasonably estimate missing data, make data more complete, reduce the deviation and uncertainty caused by missing data, and provide more comprehensive data support for subsequent accurate model training.

[0050] Normalization maps data to intervals, making environmental factor data of different magnitudes and ranges comparable, accelerating the convergence of the model and improving the efficiency of model training. In the subsequent calculation of the correlation between marine environmental factors and unit fishing effort (CPUE) and model training, the normalized data can make the calculation more stable and accurate, helping the model to better learn data features and improve model performance.

[0051] Preferably, in step S2, calculating the unit fishing effort based on the historical fishing log includes:

[0052] The catch and the number of days of operation are obtained from the historical fishing log, and the catch is divided by the number of days of operation to obtain the unit fishing effort:

[0053]

[0054] Among them, CPUE is unit fishing effort.

[0055] In this embodiment, the unit fishing effort (CPUE), that is, the ratio of the catch to the number of operating days, is calculated based on the historical fishing logs. This indicator comprehensively considers the fishing results (catch) and inputs (operating days), and can accurately reflect the fishing efficiency within a certain operating time. It provides an important quantitative basis for the subsequent analysis of the relationship between marine environmental factors and fishery output, enabling the model to more accurately capture the factors affecting fishery output, thereby improving the accuracy of the prediction.

[0056] Preferably, in step S3, the correlation coefficients between a plurality of marine environmental factors and the unit fishing effort are calculated respectively, and a plurality of target correlation coefficients related to the unit fishing effort are screened out from the calculated plurality of correlation coefficients based on a correlation threshold, including:

[0057] The correlation coefficients between multiple marine environmental factors and the unit fishing effort are calculated using the Pearson formula, which is:

[0058]

[0059] Among them, r is the correlation coefficient, x and y are the values ​​of any two variables in the marine environmental factors and unit fishing effort;

[0060] The calculated r values ​​are compared with the correlation threshold. If the r value is greater than the correlation threshold, the r value is the target correlation coefficient.

[0061] In this embodiment, the correlation coefficient between the marine environmental factor and CPUE is calculated by the Pearson formula, and the linear relationship between the two can be quantified. Based on the correlation threshold, the target correlation coefficient and the corresponding target marine environmental factor are screened, and the features with low correlation with CPUE are effectively eliminated, which avoids the interference of irrelevant or weakly correlated features on the model and reduces data redundancy and computational burden. This enables the model to focus on key environmental factors closely related to fishery output, improves the model's ability to capture important information, thereby improving the model's prediction performance and efficiency, and making the prediction results more reliable and interpretable.

[0062] Preferably, in step S4, the target marine environmental factors corresponding to the multiple target correlation coefficients are formatted, and a training sample set is constructed by the formatted target marine environmental factors and unit fishing effort, including:

[0063] The vertical profile data in the target marine environmental factor is divided according to the set height levels, and the vertical profile data obtained by the division is generated into three dimensions: time step number, height layer number and feature number;

[0064] The chlorophyll concentration data, sea level height data and unit fishing effort in the target marine environmental factors are dimensionally expanded to obtain three dimensions: time step number, height layer number and feature number;

[0065] A four-dimensional form of the number of samples, the number of channels, the number of height layers, and the width is constructed based on the three dimensions of the number of time steps, the number of height layers, and the number of features, wherein the number of time steps is used as the number of samples, and the number of features is used as the number of channels.

[0066] Specifically: integrate the various data into various data sets, and divide the vertical profile data (temperature, salinity and dissolved oxygen concentration) in the vertical profile data set into four height layers of 0m, 150m, 300m and 450m according to the depth level (0 to 450m), and then organize them into a three-dimensional input tensor, and the data format is set to (time step number, height layer number, feature number).

[0067] For features without height layers (SLA sea surface height data, Chl-a chlorophyll concentration data, and CPUE unit fishing effort), their dimensions are expanded to (number of time steps, number of height layers, number of features), and the number of height layers is repeated 4 times (corresponding to 4 depth layers) to make them conform to the data shape. For example, assuming that the index of chlorophyll concentration in the feature dimension is 4 (counting from 0), then X[:,:,4] means accessing every chlorophyll concentration data sample.

[0068] The shape of the data before entering the CNN is (time steps, depth layers, features), specifically (120, 4, 6). In order for the CNN to recognize channels, the tensor needs to be arranged into a four-dimensional form (number of samples, number of channels, number of height layers, width). With this adjustment, the convolution operation in the CNN will perform independent convolution operations on each feature, and the relationship between each feature in the depth layer can be learned.

[0069] In this embodiment, the vertical profile data in the target marine environmental factor is divided according to the set height level, and a tensor form of a specific dimension is generated, and other related data are dimensionally expanded and integrated to construct a four-dimensional form (number of samples, number of channels, number of height layers, width) that meets the model input requirements. This formatting processing method makes the data structure match the input requirements of the convolutional neural network (CNN). In CNN, each feature can be independently convolved, and the relationship between each feature in the depth layer can be fully learned, which improves the data processing efficiency and feature extraction ability in the model, and lays a good data foundation for accurately predicting the probability distribution of fishery yields.

[0070] Preferably, the CNN-LSTM fishery prediction initial model is constructed based on the CNN convolutional neural network and the LSTM long short-term memory network, and the CNN-LSTM fishery prediction initial model is trained on the correlation between the fishery and the marine environmental factors through the training sample set to obtain the fishery prediction model, including:

[0071] The output of the CNN convolutional neural network is connected to the input of the LSTM long short-term memory network, and the output of the LSTM long short-term memory network is connected to the fully connected layer and the output layer in turn.

[0072] The CNN convolutional neural network includes a plurality of sequentially connected 2D convolutional layers, a maximum pooling layer, and a Dropout layer. Each of the 2D convolutional layers uses a 3×3 convolution kernel and a ReLU activation function to obtain local spatial features in the training sample set. Each 2D convolutional layer is connected to the maximum pooling layer to downsample the output local spatial features.

[0073] The local spatial features after downsampling are input into the LSTM long short-term memory network, and the time dependency information of each local spatial feature is obtained through the LSTM long short-term memory network.

[0074] The time dependency relationship information of each of the local spatial features is input into the fully connected layer for feature conversion, the converted local spatial features are input into the output layer, the binary classification task prediction is performed through the Sigmoid activation function in the output layer, and the prediction result of the probability distribution of the fishery yield is output, thereby completing the training of the fishery prediction model.

[0075] like Figure 4 As shown in the figure, the CNN layer structure: The figure shows the structure of the CNN part in detail, including multiple 2D convolutional layers, maximum pooling layers (MaxPool) and Dropout layers connected in sequence. Each 2D convolutional layer uses a 3×3 convolution kernel and a ReLU activation function to extract local spatial features in the training sample set. For example, the first convolutional layer (Conv2d(3x3,64,ReLU)) means that the input data is convolved using a 3×3 convolution kernel, 64 feature maps are output, and nonlinear transformation is performed through the ReLU activation function. The maximum pooling layer (such as MaxPool(2x2,2Stride)) connected after each convolutional layer downsamples the output local spatial features, reduces the spatial dimension of the data by selecting the maximum value in the local area, and retains the most significant spatial features, reducing the amount of calculation and preventing overfitting. A Dropout layer (Dropout(0.5)) is added at the end of the CNN part. During the training process, 50% of the neuron connections are randomly discarded to avoid the network from being overly dependent on certain specific neurons, further preventing overfitting and improving the generalization ability of the model.

[0076] Model connection relationship: The output of the CNN layer will be used as the input of the LSTM long short-term memory network, showing the connection relationship between CNN and LSTM in the model, and reflecting how the model combines spatial feature extraction with time series processing. The output of the LSTM layer is connected to the fully connected layer in turn, which is used to further transform and integrate the features of the LSTM output. Finally, the binary classification task prediction is performed through the Sigmoid activation function in the output layer, and the prediction result of the probability distribution of the fishery yield is output. For example, the first fully connected layer (Fully connected layer 1) contains 1024 neurons, receives the final output from the LSTM (the input shape is (batch size, number of hidden units)), and transforms the high-dimensional features of the LSTM output; the second fully connected layer (Fully connected layer 2) is also used to further process the output of the previous layer, and contains 512 neurons; the last output layer (Layer 3 (1, Sigmoid)) converts the output into a probability value between 0 and 1 through the Sigmoid activation function, indicating the probability of the predicted result belonging to a high-yield fishery or a low-yield fishery.

[0077] In this embodiment, the advantages of model structure fusion improve prediction accuracy, and the initial model of fishing ground prediction built based on CNN and LSTM gives full play to the advantages of both. CNN can effectively extract local spatial features in marine environmental data and capture the changing laws of environmental factors in the spatial dimension through multiple 2D convolutional layers, maximum pooling layers and Dropout layers; LSTM is good at processing time series data and obtaining time dependency information of local spatial features. The combination of the two realizes the full mining of the complex characteristics and laws of marine environmental data in the time and space dimensions, and better models the dynamic relationship between fishing grounds and marine environmental factors.

[0078] Model training and prediction achieve accurate output. After feature conversion of the fully connected layer and Sigmoid activation function of the output layer for binary classification task prediction, the model can output the prediction results of the probability distribution of fishery production, which intuitively reflects the possibility of high or low production of the fishery. During the training process, by learning the training sample set and continuously adjusting the model parameters, the model can accurately capture the intrinsic connection between marine environmental factors and fishery production, thereby significantly improving the prediction accuracy, providing fishery practitioners with more valuable decision-making basis, and helping to rationally plan fishing operations and achieve sustainable utilization of fishery resources.

[0079] The method for predicting the probability distribution of fishery yields in the present invention is applicable to marine ecosystem research and marine fishery management, especially in the fields of fish distribution analysis, fishery prediction and fish dynamic monitoring based on marine environmental factors. By integrating multi-source marine environmental data and fishery data, and using the CNN-LSTM model for spatial distribution prediction and time series prediction, the distribution trend of target fish can be predicted more accurately, thereby improving the efficiency of fishery resource utilization.

[0080] like Figure 2 As shown, the implementation process of the prediction method of the present invention is explained below through a specific example of tuna fishing ground prediction.

[0081] 1. Data acquisition and preprocessing ( Figure 3 )

[0082] Vertical profile data and two-dimensional data acquisition: Data related to tuna fishing grounds are collected from multiple data sources, including vertical profile data (such as temperature, salinity, dissolved oxygen concentration) and two-dimensional data (such as chlorophyll concentration Chl-a, sea surface height anomaly SLA).

[0083] Fishing log acquisition: Get historical fishing logs, which contain information such as date, longitude and latitude, number of operating days, and catch.

[0084] Temporal and spatial alignment: Adjust the temporal resolution (align the time step to the monthly resolution) and spatial resolution (unify the spatial resolution to 100) of the collected data to ensure the consistency of the data in time and space for subsequent analysis.

[0085] Outlier processing (missing value processing): For the missing values ​​in the data of each environmental factor, the Euclidean distance interpolation method is used to fill them in to make the data complete and improve the data quality.

[0086] Correlation test: Calculate the Pearson correlation coefficient between each environmental factor and the unit fishing effort (CPUE, calculated from the fishing log data). By analyzing the correlation, screen out the features with a high correlation with the target variable (CPUE), and remove the features with a low correlation to optimize the data characteristics.

[0087] Formatting: Organize the processed vertical profile data by depth level (0m, 150m, 300m, 450m), and integrate it with other data into a specific format (such as a three-dimensional tensor form, and then adjust it to a four-dimensional input format) to obtain the data to be trained and prepare for model input.

[0088] 2. Model training and evaluation ( Figure 3 The second box in

[0089] Model training: The formatted training data is input into a deep learning model consisting of a convolutional neural network (CNN) and a long short-term memory network (LSTM) for training. The CNN part is used to extract the spatial features of the data, and the LSTM part is used to capture the time series dependencies. Through multiple iterations of training, the model learns the complex relationship between tuna fishing grounds and environmental factors.

[0090] Calculate F1-score: During model training, calculate F1-score to evaluate model performance. F1-score is the harmonic mean of precision and recall, which can comprehensively measure the accuracy of the model in positive and negative class predictions, especially for unbalanced data.

[0091] Cross-validation: Use the cross-validation method to further evaluate the generalization ability of the model. Divide the data set into multiple subsets, perform training and validation multiple times, and calculate the average of multiple validation results to reduce the impact of data segmentation on the evaluation results, so as to more accurately evaluate the performance of the model on different data subsets.

[0092] Accuracy judgment and parameter tuning: Based on the calculated F1-score and cross-validation results, determine whether the model's prediction accuracy meets the requirements. If not, adjust and optimize the model parameters (such as adjusting the learning rate, optimizer parameters, network structure, etc.), and then train and evaluate again until the model accuracy meets the requirements.

[0093] 3. Result processing

[0094] Prediction result export: When the model accuracy meets the requirements, the trained model is used to predict the tuna fishing grounds and obtain the prediction results (such as the probability distribution of high-yield fishing grounds or low-yield fishing grounds).

[0095] Archiving and further research: The prediction results will be archived and saved, and data support will be provided for further research, such as analyzing the differences between the prediction results and the actual distribution of fishing grounds, and exploring ways to improve the model.

[0096] Specifically, during the data acquisition process:

[0097] Vertical profile data (temperature, salinity, dissolved oxygen concentration) are obtained from the global marine environment database provided by the National Oceanic and Atmospheric Administration (NOAA). The Ocean Color Web website provides a chlorophyll concentration (Chl-a) data set based on satellite observations. The SLA data provided by CMEMS covers the global ocean and can be downloaded from its official website database. The Inter-American Tropical Tuna Commission (IATTC) provides global fishing vessel activity information and various historical tuna fishing logs can be obtained from its official website, including date, longitude and latitude, operating days, and catch (tons).

[0098] During data preprocessing:

[0099] The first step is to adjust the temporal resolution. Align the time steps of all data to a consistent monthly resolution to ensure that data from different sources can be analyzed at the same time scale. Use resampling in ArcGIS to standardize data with different temporal resolutions. For example, if SLA data is taken once a week, the SLA data can be averaged monthly to match the temporal resolution of other data.

[0100] For the adjustment of spatial resolution, a gridding method was used to adjust each data set to the same spatial resolution of 0.5°×0.5°. In the face of different spatial resolutions of data from different sources (chlorophyll data has a resolution of 0.1° and temperature has a resolution of 1°), ArcGIS was also used to downsample the high-resolution data or upsample the low-resolution data to match the consistent resolution.

[0101] In the second step, the missing values ​​of each environmental factor data are filled using the Euclidean distance interpolation formula 1, and then the data are normalized according to formula 2.

[0102] Formula 1 is:

[0103]

[0104] Formula 2 is:

[0105]

[0106] The third step is to calculate CPUE using historical fishing log data according to Formula 3, so that the data can be divided into different categories according to the CPUE value, and labels for high-yield fishing areas and low-yield fishing areas can be set. The CPUE values ​​in high-yield fishing areas are higher than the specified threshold (1.5 times the average value), and these data points are labeled with 1. For areas below the threshold (lower than 0.5 times the average value), a label of 0 is set.

[0107] Formula 3 is:

[0108]

[0109] In the fourth step, the Pearson correlation coefficient was calculated according to Formula 4. By calculating the correlation between each environmental factor and CPUE, the features with correlations with the target variable higher than the threshold were screened out, and the features with correlations lower than the threshold were eliminated.

[0110] Formula 4 is:

[0111]

[0112] During data formatting:

[0113] According to the data formatting steps, the final data is a three-dimensional tensor consisting of 120 time steps, 4 height layers, and 6 feature numbers, expressed in the following form:

[0114]

[0115] Adjust (120,4,6) to (120,6,4,1). After the adjustment, each time step is a sample, 6 features correspond to 6 channels, and each channel contains 4 depth layer values. For the input data of a certain time step, such as the Xth sample, it can be represented as a (6,4,1) tensor. Figure 3 As shown, Figure 3 By showing the formatted tensor structure of the data and the partial structure of the model, it clearly shows how the data is input into the model and how the model processes the data. This helps to understand the interaction between data and model in the entire prediction method, and how the model can achieve accurate prediction of the probability distribution of fishery yield through the collaborative work of various layers of structure.

[0116] An example of organized data is as follows:

[0117]

[0118] During model training:

[0119] In the first step, the processed and aligned three-dimensional tensor data with time steps and spatial resolutions are used as model input. The dataset is divided into training set and validation set in a ratio of 7.5:2.5 to ensure the generalization ability of the model.

[0120] In the second step, in the setting of training parameters, select the cross-entropy loss function (Cross-Entropy Loss) suitable for classification problems; the optimizer uses the Adam optimizer, the initial learning rate is set to 0.01, and the learning rate decay is set to reduce the learning rate every 10 epochs;

[0121] The batch size is set according to the computer performance to achieve a balance between training stability and memory consumption; the model is trained for 200 epochs, and the loss and accuracy are calculated on the validation set after each epoch to observe the model convergence.

[0122] The third step is to load the input data in batches. Each batch of data is input into the CNN part to extract spatial features, and then input into the LSTM to capture time series dependencies. The output of the last layer of LSTM passes through the fully connected layer to output the classification result.

[0123] During model evaluation:

[0124] F1-score is the harmonic mean of precision and recall, which can effectively balance these two indicators. In the case of class imbalance, F1-score can better reflect the performance of the model on the minority class, and is particularly suitable for tasks that focus on classification accuracy.

[0125] After calculating the confusion matrix and obtaining the model's judgment on positive and negative predictions, the F1-score is obtained according to Formula 5-Formula 7. By comparing the F1-scores of multiple other models, the performance of the model on a given data set is judged, and the F1-scores of multiple models are calculated using cross-validation to observe whether the prediction accuracy of the model is improved relative to other models.

[0126] Formula 5 is:

[0127] Formula 6 is:

[0128] Formula 7 is:

[0129] in:

[0130] TP (True Positives): True positives, the number of samples correctly predicted as positive.

[0131] FP (False Positives): False positive examples, the number of samples incorrectly predicted as positive.

[0132] FN (False Negatives): False negative examples, the number of samples incorrectly predicted as negative.

[0133] like Figure 5As shown in the figure, k-Cross Validation is used to evaluate the performance of the model and reduce the variance caused by data segmentation. The specific method is to divide the data set into k subsets, train the model with K-1 subsets respectively, and then evaluate the model on the remaining subset, repeat k times, and finally calculate the average evaluation index.

[0134] Embodiment 2:

[0135] like Figure 6 As shown, the embodiment of the present invention also provides a device for predicting the probability distribution of fishery production based on the CNN-LSTM model, including:

[0136] A data collection module for collecting historical fishing logs and multiple types of initial marine environmental data;

[0137] A preprocessing module is used to preprocess the initial data of each marine environment to obtain multiple marine environmental factors and calculate the unit fishing effort based on the historical fishing log;

[0138] a calculation module, used to respectively calculate the correlation coefficients between a plurality of marine environmental factors and the unit fishing effort, and screen out a plurality of target correlation coefficients related to the unit fishing effort from the calculated plurality of correlation coefficients based on a correlation threshold;

[0139] A formatting module is used to format the target marine environmental factors corresponding to the multiple target correlation coefficients, and to construct a training sample set through the formatted target marine environmental factors and unit fishing effort;

[0140] A model building module, for building a CNN-LSTM fishery prediction initial model based on a CNN convolutional neural network and an LSTM long short-term memory network, and training the CNN-LSTM fishery prediction initial model on the correlation between the fishery and the marine environmental factors through the training sample set to obtain a fishery prediction model;

[0141] The prediction module is used to import the data to be analyzed into the fishing ground capture prediction model and output the fishing ground yield probability distribution prediction result.

[0142] Preferably, in the preprocessing module, each marine environment initial data is preprocessed to obtain a plurality of marine environment factors, including:

[0143] Multiple types of initial marine environmental data including vertical profile datasets, chlorophyll concentration datasets, and sea surface height datasets;

[0144] Setting a unified time resolution and spatial resolution for the data in the vertical profile dataset, the chlorophyll concentration dataset, and the sea level height dataset;

[0145] The missing values ​​in each data set are filled by the Euclidean distance interpolation formula, which is:

[0146]

[0147] Among them, S0 is the position where the missing value of vertical profile data, chlorophyll concentration data or sea level height data needs to be interpolated, is the predicted value corresponding to the missing value, ω k is the weight of the data points around the missing value location, k is the serial number corresponding to the missing value from 1 to n, Z(S k ) is the actual attribute value of the surrounding sampling points;

[0148] The filled vertical profile data, chlorophyll concentration data and sea level height data are normalized by a normalization formula to obtain a plurality of marine environmental factors. The normalization formula is:

[0149]

[0150] Among them, x min and x max are the minimum and maximum values ​​in the vertical profile dataset, chlorophyll concentration dataset, or sea level height dataset, respectively. i is one of the vertical profile datasets, chlorophyll concentration datasets, or sea level height datasets, x Nor is the normalized value.

[0151] Preferably, in step S3, the correlation coefficients between a plurality of marine environmental factors and the unit fishing effort are calculated respectively, and a plurality of target correlation coefficients related to the unit fishing effort are screened out from the calculated plurality of correlation coefficients based on a correlation threshold, including:

[0152] The correlation coefficients between multiple marine environmental factors and the unit fishing effort are calculated using the Pearson formula, which is:

[0153]

[0154] Among them, r is the correlation coefficient, x and y are the values ​​of any two variables in the marine environmental factors and unit fishing effort;

[0155] The calculated r values ​​are compared with the correlation threshold. If the r value is greater than the correlation threshold, the r value is the target correlation coefficient.

[0156] Embodiment 3:

[0157] An embodiment of the present invention also provides a device for predicting the probability distribution of fishery yield based on the CNN-LSTM model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for predicting the probability distribution of fishery yield based on the CNN-LSTM model as described above is implemented.

[0158] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0159] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0160] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the embodiments of the present invention.

[0161] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting the probability distribution of fishery yield based on the CNN-LSTM model, characterized in that: The steps include: S1. Collect historical fishing logs and multiple types of initial marine environmental data; S2. Preprocessing the initial data of each marine environment to obtain multiple marine environmental factors, and calculating the unit fishing effort based on historical fishing logs; S3, respectively calculating the correlation coefficients between a plurality of marine environmental factors and the unit fishing effort, and selecting a plurality of target correlation coefficients related to the unit fishing effort from the calculated plurality of correlation coefficients based on a correlation threshold; S4, formatting the target marine environmental factors corresponding to the multiple target correlation coefficients, and constructing a training sample set through the formatted target marine environmental factors and unit fishing effort; S5. Constructing a CNN-LSTM fishing ground prediction initial model based on a CNN convolutional neural network and an LSTM long short-term memory network, and training the CNN-LSTM fishing ground prediction initial model on the correlation between fishing grounds and marine environmental factors through the training sample set to obtain a fishing ground prediction model; S6. Import the data to be analyzed into the fishing ground capture prediction model, and output the fishing ground yield probability distribution prediction result.

2. The method for predicting the probability distribution of fishery yield according to claim 1, characterized in that: The multiple types of marine environment initial data include a vertical profile data set, a chlorophyll concentration data set and a sea surface height data set; In step S2, each marine environment initial data is preprocessed to obtain multiple marine environment factors, including: Setting a unified time resolution and spatial resolution for the data in the vertical profile dataset, the chlorophyll concentration dataset, and the sea level height dataset; The missing values ​​in each data set are filled by the Euclidean distance interpolation formula, which is: Among them, S0 is the position where the missing value of vertical profile data, chlorophyll concentration data or sea level height data needs to be interpolated, is the predicted value corresponding to the missing value, ω k is the weight of the data points around the missing value location, k is the serial number corresponding to the missing value from 1 to n, Z(S k ) is the actual attribute value of the surrounding sampling points; The filled vertical profile data, chlorophyll concentration data and sea level height data are normalized by a normalization formula to obtain a plurality of marine environmental factors. The normalization formula is: Among them, x min and x max are the minimum and maximum values ​​in the vertical profile dataset, chlorophyll concentration dataset, or sea level height dataset, respectively. i is one of the vertical profile datasets, chlorophyll concentration datasets, or sea level height datasets, x Nor is the normalized value.

3. The method for predicting the probability distribution of fishery yield according to claim 1, characterized in that: In step S2, the unit fishing effort is calculated based on the historical fishing log, including: The catch and the number of days of operation are obtained from the historical fishing log, and the catch is divided by the number of days of operation to obtain the unit fishing effort: Among them, CPUE is unit fishing effort.

4. The method for predicting the probability distribution of fishery yield according to claim 1, characterized in that: In step S3, the correlation coefficients between the plurality of marine environmental factors and the unit fishing effort are calculated respectively, and a plurality of target correlation coefficients related to the unit fishing effort are screened out from the plurality of calculated correlation coefficients based on a correlation threshold, including: The correlation coefficients between multiple marine environmental factors and the unit fishing effort are calculated using the Pearson formula, which is: Among them, r is the correlation coefficient, x and y are the values ​​of any two variables in the marine environmental factors and unit fishing effort; The calculated r values ​​are compared with the correlation threshold. If the r value is greater than the correlation threshold, the r value is the target correlation coefficient.

5. The method for predicting the probability distribution of fishery yield according to claim 1, characterized in that: In step S4, the target marine environmental factors corresponding to the multiple target correlation coefficients are formatted, and a training sample set is constructed by the formatted target marine environmental factors and unit fishing effort, including: The vertical profile data in the target marine environmental factor is divided according to the set height levels, and the vertical profile data obtained by the division is generated into three dimensions: time step number, height layer number and feature number; The chlorophyll concentration data, sea level height data and unit fishing effort in the target marine environmental factors are dimensionally expanded to obtain three dimensions: time step number, height layer number and feature number; A four-dimensional form of the number of samples, the number of channels, the number of height layers, and the width is constructed based on the three dimensions of the number of time steps, the number of height layers, and the number of features, wherein the number of time steps is used as the number of samples, and the number of features is used as the number of channels.

6. The method for predicting the probability distribution of fishery yield according to claim 1, characterized in that: The CNN-LSTM fishing ground prediction initial model is constructed based on the CNN convolutional neural network and the LSTM long short-term memory network, and the CNN-LSTM fishing ground prediction initial model is trained on the correlation between the fishing ground and the marine environmental factors through the training sample set to obtain the fishing ground prediction model, including: The output of the CNN convolutional neural network is connected to the input of the LSTM long short-term memory network, and the output of the LSTM long short-term memory network is connected to the fully connected layer and the output layer in turn. The CNN convolutional neural network includes a plurality of sequentially connected 2D convolutional layers, a maximum pooling layer, and a Dropout layer. Each of the 2D convolutional layers uses a 3×3 convolution kernel and a ReLU activation function to obtain local spatial features in the training sample set. Each 2D convolutional layer is connected to the maximum pooling layer to downsample the output local spatial features. The local spatial features after downsampling are input into the LSTM long short-term memory network, and the time dependency information of each local spatial feature is obtained through the LSTM long short-term memory network. The time dependency relationship information of each of the local spatial features is input into the fully connected layer for feature conversion, the converted local spatial features are input into the output layer, the binary classification task prediction is performed through the Sigmoid activation function in the output layer, and the prediction result of the probability distribution of the fishery yield is output, thereby completing the training of the fishery prediction model.

7. A device for predicting the probability distribution of fishery yield based on the CNN-LSTM model, characterized in that: include: A data collection module for collecting historical fishing logs and multiple types of initial marine environmental data; A preprocessing module is used to preprocess the initial data of each marine environment to obtain multiple marine environmental factors and calculate the unit fishing effort based on the historical fishing log; a calculation module, used to respectively calculate the correlation coefficients between a plurality of marine environmental factors and the unit fishing effort, and screen out a plurality of target correlation coefficients related to the unit fishing effort from the calculated plurality of correlation coefficients based on a correlation threshold; A formatting module is used to format the target marine environmental factors corresponding to the multiple target correlation coefficients, and to construct a training sample set through the formatted target marine environmental factors and unit fishing effort; A model building module, for building a CNN-LSTM fishery prediction initial model based on a CNN convolutional neural network and an LSTM long short-term memory network, and training the CNN-LSTM fishery prediction initial model on the correlation between the fishery and the marine environmental factors through the training sample set to obtain a fishery prediction model; The prediction module is used to import the data to be analyzed into the fishing ground capture prediction model and output the fishing ground yield probability distribution prediction result.

8. The device for predicting the probability distribution of fishery yield according to claim 7, characterized in that: In the preprocessing module, each marine environment initial data is preprocessed to obtain multiple marine environment factors, including: Multiple types of initial marine environmental data including vertical profile datasets, chlorophyll concentration datasets, and sea surface height datasets; Setting a unified time resolution and spatial resolution for the data in the vertical profile dataset, the chlorophyll concentration dataset, and the sea level height dataset; The missing values ​​in each data set are filled by the Euclidean distance interpolation formula, which is: Among them, S0 is the position where the missing value of vertical profile data, chlorophyll concentration data or sea level height data needs to be interpolated, is the predicted value corresponding to the missing value, ω k is the weight of the data points around the missing value location, k is the serial number corresponding to the missing value from 1 to n, Z(S k ) is the actual attribute value of the surrounding sampling points; The filled vertical profile data, chlorophyll concentration data and sea level height data are normalized by a normalization formula to obtain a plurality of marine environmental factors. The normalization formula is: Among them, x min and x max are the minimum and maximum values ​​in the vertical profile dataset, chlorophyll concentration dataset, or sea level height dataset, respectively. i is one of the vertical profile datasets, chlorophyll concentration datasets, or sea level height datasets, x Nor is the normalized value.

9. The device for predicting the probability distribution of fishery yield according to claim 7, characterized in that: In step S3, the correlation coefficients between the plurality of marine environmental factors and the unit fishing effort are calculated respectively, and a plurality of target correlation coefficients related to the unit fishing effort are screened out from the plurality of calculated correlation coefficients based on a correlation threshold, including: The correlation coefficients between multiple marine environmental factors and the unit fishing effort are calculated using the Pearson formula, which is: Among them, r is the correlation coefficient, x and y are the values ​​of any two variables in the marine environmental factors and unit fishing effort; The calculated r values ​​are compared with the correlation threshold. If the r value is greater than the correlation threshold, the r value is the target correlation coefficient.

10. A device for predicting the probability distribution of fishery yield based on a CNN-LSTM model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the method for predicting the probability distribution of fishery yield based on the CNN-LSTM model as described in any one of claims 1 to 6.

Images