Tuna longline fishing activity prediction method and system

Abstract

The invention discloses a tuna longline fishing activity prediction method and a tuna longline fishing activity prediction system. The tuna longline fishing activity prediction method comprises the following steps: step S1, obtaining tuna fishing effort data and marine environment data corresponding to the area fishing effort data; s2, constructing space-time sequence tensor data of different space-time scales according to different time resolutions and spatial resolutions; s3, a ConvLSTM-Attention U-Net network model is constructed and trained, an optimal space-time scale model is determined, and the time step length M is larger than or equal to 3; s4, based on the optimal spatial-temporal scale model, visualizing a model prediction result and comparing the model prediction result with real fishing effort distribution; and S5, calculating average absolute gradient values of the variables under different spatial and temporal scales, and evaluating and visualizing the importance of the variables under different spatial and temporal scales. According to the method, the position distribution and the numerical value of the fishing effort can be predicted, and the tuna longline fishing activity in the Chinese and western Pacific Ocean area can be accurately predicted.

Description

Technical Field

[0001] The present invention relates to the field of fisheries and fishing grounds, and in particular to a method and system for predicting tuna longline fishing activities. Background Art

[0003] Based on AIS data, historical fishing intensity can be quantified, but future trends cannot be predicted. Mid- and long-term forecasts of fishing effort are particularly important for fishery resource management. Studies have shown that the distribution of fishing effort of longline fishing vessels in the high seas is affected by environmental factors. By combining environmental data with fishery observations, the prediction model can reveal the predictability of fishing intensity.

[0004] Compared with fishing forecast, the prediction of fishing intensity faces greater challenges. Fishing activities of fishing vessels are a dynamic and continuous process. The choice of fishing points is not only affected by environmental factors, but also related to previous fishing locations. Fishing activities have obvious spatiotemporal correlation. Traditional prediction methods have limitations in processing spatiotemporal data. Most of them only select single-point environmental data for modeling, which makes it difficult to fully reveal the spatiotemporal variation trend of fishing intensity.

[0005] In recent years, deep learning methods have shown potential in fishing ground prediction. Convolutional long short-term memory network (ConvLSTM) combines the advantages of CNN and LSTM and can process spatiotemporal data. Although ConvLSTM can capture local spatial dependencies, it is weak in processing long-distance spatial dependencies. U-Net has outstanding performance in the field of spatiotemporal prediction due to its advantage in capturing multi-scale spatial features. Attention U-Net, which combines U-Net and attention mechanism, further improves the prediction accuracy of the model. The combination of ConvLSTM and Attention U-Net can simultaneously analyze spatial and temporal features and improve the accuracy of fishing intensity prediction.

[0006] Existing studies usually only consider environmental factors and ignore the spatiotemporal evolution of fishery data. However, historical fishing information will affect future fishing activities, so it is important to incorporate it into the model. At the same time, although deep neural networks have high prediction accuracy, they have poor interpretability. Understanding the importance of variables in the network is as important as prediction accuracy. Most of the variable selection methods of existing deep learning models are only applicable to shallow networks or have high computational complexity, while the variable importance assessment method based on backpropagation has better feasibility and scientificity.

[0007] The Western and Central Pacific Ocean is the world's largest tuna fishing ground, with a catch of about 2.7 million tons in 2022, accounting for 54% of the global total. In the past 50 years, the global tuna population has decreased by 60%, making its protection and sustainable development a focus of international attention. Tuna longline fishing covers a large area of ​​sea and exerts the greatest pressure on the marine ecology. Therefore, achieving refined predictions of tuna longline fishing activities in the Western and Central Pacific Ocean and clarifying the importance of variables in predicting fishing activities are of great significance to ensuring the sustainable development of marine fishery resources. Summary of the invention

[0008] The object of the present invention is to provide a method and system for predicting tuna longline fishing activities to solve the problems raised in the above background technology.

[0009] To achieve the above-mentioned object of the invention, one aspect of the present invention provides a method for predicting tuna longline fishing activities, comprising the following steps:

[0010] Step S1, obtaining the fishing effort data of tuna in the central and western Pacific region from 2017 to 2023 and the marine environment data corresponding to the fishing effort data in the region;

[0011] Step S2, constructing spatiotemporal sequence tensor data of different spatiotemporal scales according to different time resolutions and spatial resolutions;

[0012] Step S3, constructing a ConvLSTM-Attention U-Net network model, using the fishing effort data and marine environment data of the first M time steps to predict the distribution of fishing activities in the next time step at different spatiotemporal scales, training the ConvLSTM-Attention U-Net network, learning the spatiotemporal information in the historical fishing effort data and environmental variable data and the complex mapping relationship between environmental variables and fishing effort distribution, and determining the optimal spatiotemporal scale model by comparing the prediction performance of the model at different spatiotemporal scales, wherein the time step M is greater than or equal to 3;

[0013] Step S4, based on the optimal spatiotemporal scale model, the model prediction results are visualized and compared with the actual fishing effort distribution;

[0014] Step S5, calculate the average absolute gradient value of the variables at different spatiotemporal scales, evaluate and visualize the importance of the variables at different spatiotemporal scales.

[0015] Furthermore, the fishing effort data in step S1 include operation date, operation location, and fishing effort; the environmental factor data include sea surface temperature, dissolved oxygen concentration in 100-meter deep seawater, sea water temperature in 400-meter deep seawater, chlorophyll a concentration, dissolved oxygen concentration in 400-meter deep seawater, and salinity in 100-meter deep seawater.

[0016] Furthermore, the processing method of fishing effort and marine environmental variables at different spatiotemporal scales in step S2 is to calculate the daily average fishing effort of each grid at each spatiotemporal scale as the fishing effort value of the grid, and calculate the overall mean of environmental data in each grid as the environmental variable value of the grid.

[0017] Furthermore, the model structure of the ConvLSTM-Attention U-Net combines the ConvLSTM model and the Attention U-Net model to capture the long-term temporal characteristics of the data and process spatial features of different scales to achieve refined prediction of fishing activities.

[0018] Furthermore, the ConvLSTM model is good at learning long-term spatiotemporal features in data, which is calculated as follows:

[0019] i t =σ(W xi *X t +W hi *H t-1 +W ci ⊙C t-1 +b i )

[0020] f t =σ(W xf *X t +W hf *H t-1 +W cf ⊙C t-1 +b f )

[0021] C t =f t ⊙C t-1 +i t ⊙tanh(W xc *X t +W hc *H t-1 +b c )

[0022] o t =σ(W xo *X t +W ho *H t-1 +W co ⊙C t +b o )

[0023] H t =o t ⊙tanh(Ct )

[0024] Among them, i t 、f t , C t , o t and H t They represent input gate, forget gate, memory unit, output gate and hidden state respectively, X t is the input of the current time step, H t-1 is the hidden state of the previous time step, C t-1 is the memory unit of the previous time step, σ represents the sigmoid activation function, tanh represents the hyperbolic tangent activation function, * represents the convolution operation, ⊙ represents the Hadamard product of the matrix, W xi ,W hi and W ci represents the weight matrix of the input gate, W xf ,W hf and W cf Represents the weight matrix of the forget gate, W xc and W hc Represents the weight matrix of the memory unit, W xo ,W ho and W co represents the weight matrix of the output gate, b i ,b f ,b c and b o Indicates bias.

[0025] Furthermore, the time step M is set to 5.

[0026] Furthermore, at different spatiotemporal scales, the fishing effort data and marine environment data of the first M time steps are used to predict the distribution of fishing activities in the next time step, the ConvLSTM-Attention U-Net network is trained, the spatiotemporal features in the data are extracted, and the optimal spatiotemporal scale model is selected according to the general evaluation index F1 score and the length of the time series. The calculation formula of the F1 score is as follows:

[0027]

[0028] Among them, N TP The number of grids that actually have fishing effort and are predicted to have fishing effort, N FP The number of grids where there is actual fishing effort but is predicted to be without fishing effort, N FN Indicates the number of grid cells where there is actually no fishing effort but is predicted to be fishing effort.

[0029] Furthermore, the spatiotemporal resolution of the ConvLSTM-Attention U-Net model is set to 1°×1° and 30 days, and the spatiotemporal scale prediction results are visualized to be closest to the actual distribution of fishing effort.

[0030] Furthermore, the ConvLSTM-Attention U-Net model has the highest importance of historical fishing effort data at all spatiotemporal scales, and the importance of the six marine environmental variables is relatively close, with no obvious difference.

[0031] Another aspect of the present invention provides a prediction system for tuna longline fishing activities in the Central and Western Pacific Ocean based on claims 1 to 9, comprising a data acquisition module, a data construction module, a model construction module, a prediction comparison module, and a prediction evaluation module, wherein:

[0032] A data acquisition module is used to obtain the fishing effort data of tuna in the Western and Central Pacific region from 2017 to 2023 and the marine environment data corresponding to the fishing effort data in the region;

[0033] Data construction module, used to construct spatiotemporal sequence tensor data of different spatiotemporal scales according to different temporal resolutions and spatial resolutions;

[0034] A model building module is used to build a ConvLSTM-Attention U-Net network model. At different spatiotemporal scales, the fishing effort data and marine environment data of the first M time steps are used to predict the distribution of fishing activities in the next time step. The ConvLSTM-Attention U-Net network is trained to learn the spatiotemporal information in the historical fishing effort data and environmental variable data and the complex mapping relationship between environmental variables and fishing effort distribution. The optimal spatiotemporal scale model is determined by comparing the prediction performance of the model at different spatiotemporal scales, wherein the time step M is greater than or equal to 3.

[0035] A prediction comparison module, which is used to visualize the model's prediction results and compare them with the actual fishing effort distribution based on the optimal spatiotemporal scale model;

[0036] The prediction evaluation module is used to calculate the average absolute gradient value of variables at different spatiotemporal scales, evaluate and visualize the importance of variables at different spatiotemporal scales

[0037] Compared with the prior art, the present system and method have the following advantages:

[0038] The present invention can construct a prediction model for the fishing effort of tuna longline fishing in the Western and Central Pacific Ocean at different time and space scales, and can predict the geographical distribution and numerical value of fishing effort; by comparing the model performance, the optimal time and space scale is determined, so that it can more accurately predict the fishing activities of tuna longline fishing in the Western and Central Pacific Ocean. This accurate prediction can not only provide a scientific basis for the formulation of fishery policies, but also support the effective implementation of policies, thereby promoting the long-term sustainability of marine resources and fishing activities. In addition, by calculating the average absolute gradient value of each environmental variable, the impact of each variable on the model prediction effect can be quantitatively evaluated, thereby identifying key variables. This process helps to improve the overall efficiency of the model, reduce the need to collect non-critical data, and reduce the cost and time investment of data acquisition. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 A flow chart for forecasting tuna longline fishing activity in the Western and Central Pacific Ocean.

[0040] Figure 2 Schematic diagram of the ConvLSTM-Attention U-Net model framework.

[0041] Figure 3 This is a visual comparison of the prediction results of the ConvLSTM-Attention U-Net model and the actual fishing effort distribution at the optimal spatiotemporal scale (1°×1°, 30 days).

[0042] Figure 4 This is a visualization of the importance of model variables at different spatial and temporal scales. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] like Figure 1 The flowchart of the method of the present invention is shown. The embodiment of the present invention provides a method for short-term and medium-term prediction of tuna longline fishing effort. The specific steps are as follows:

[0045] Step S1, the embodiment of the present invention relates to a method for constructing a prediction model for tuna longline fishing activities in the Western and Central Pacific Ocean based on deep learning, obtaining the fishing effort data of tuna in the Western and Central Pacific Ocean from 2017 to 2023 and the marine environment data corresponding to the fishing effort data in the region (sea surface temperature (SST), dissolved oxygen concentration in seawater at a depth of 100 meters (O2100), seawater temperature at a depth of 400 meters (Tem400), chlorophyll a concentration (Chl-a), dissolved oxygen concentration in seawater at a depth of 400 meters (O2400), and salinity in seawater at a depth of 100 meters (Salt200));

[0046] Step S2, the area is divided into spatial grids according to the spatial resolution of 0.5°×0.5° and 1°×1°, and the temporal resolution is set to 1 day, 3 days, 7 days, 15 days and 30 days. Different temporal resolutions and spatial resolutions are combined into 10 spatiotemporal scales. The processing method of fishing effort and marine environmental variables at different spatiotemporal scales is as follows: the daily average value of fishing effort of each grid at each spatiotemporal scale is calculated as the fishing effort value of the grid; the overall mean value of environmental data in each grid is calculated as the environmental variable value of the grid.

[0047] Step S3, at different spatiotemporal scales, the fishing effort data and marine environment data of the first M (M ≥ 3) time steps are used to predict the distribution of fishing activities in the next time step, and the ConvLSTM-Attention U-Net network is trained to capture the spatiotemporal information in the fishing effort data and environmental variable data and the complex mapping relationship between environmental variables and fishing effort distribution, including:

[0048] The ConvLSTM model is good at learning long-term spatiotemporal features in the data, which is calculated as follows:

[0049] i t =σ(W xi *X t +W hi *H t-1 +W ci ⊙C t-1 +b i )

[0050] f t =σ(W xf *X t +W hf *H t-1 +W cf ⊙C t-1 +b f )

[0051] C t =f t ⊙Ct-1 +i t ⊙tanh(W xc *X t +W hc *H t-1 +b c )

[0052] o t =σ(W xo *X t +W ho *H t-1 +W co ⊙C t +b o )

[0053] H t =o t ⊙tanh(C t )

[0054] Among them, i t 、f t , C t , o t and H t They represent input gate, forget gate, memory unit, output gate and hidden state respectively, X t is the input of the current time step, H t-1 is the hidden state of the previous time step, C t-1 is the memory unit of the previous time step, σ represents the sigmoid activation function, tanh represents the hyperbolic tangent activation function, * represents the convolution operation, ⊙ represents the Hadamard product of the matrix (multiplication of elements at corresponding positions), W xi ,W hi and W ci represents the weight matrix of the input gate,

[0055] W xf ,W hf and W cf Represents the weight matrix of the forget gate, W xc and W hc Represents the weight matrix of the memory unit, W xo ,W ho and W co represents the weight matrix of the output gate, b i ,b f ,b c and b o Indicates bias.

[0056] The ConvLSTM model relies on convolutional layers to capture spatial dependencies, but these dependencies are local and inefficient. The Attention U-Net can capture multi-scale spatial features through the encoder-decoder structure, and the Attention U-Net introduces an attention mechanism that can enhance the model's ability to focus on important areas.

[0057] The model structure of ConvLSTM-Attention U-Net is as follows Figure 2 As shown in the figure, it combines the advantages of the ConvLSTM model and the Attention U-Net, which can not only capture the long-term temporal characteristics of the data, but also process the spatial characteristics of different scales, and can effectively realize the refined prediction of fishing activities.

[0058] Determine M=5, at which point the model has a better effect and high efficiency. Then train the ConvLSTM-Attention U-Net model on these 10 data sets respectively; after the model training is completed, set the true distribution of fishing effort and the positions where the fishing effort is greater than 0 in the predicted results to 1, and the other positions to 0, and then calculate the F1 score of the model results. Use the F1 score to evaluate the prediction performance of the ConvLSTM-Attention U-Net model at different spatiotemporal scales, and select the spatiotemporal scale with the highest F1 score as the optimal spatiotemporal scale. The calculation formula for the F1 score is as follows:

[0059]

[0060] Among them, N TP The number of grids that actually have fishing effort and are predicted to have fishing effort, N FP The number of grids where there is actual fishing effort but is predicted to be without fishing effort, N FN Indicates the number of grid cells where there is actually no fishing effort but is predicted to be fishing effort.

[0061] Step S4, determine the optimal prediction performance at the optimal spatiotemporal scale

[0062] The optimal spatiotemporal scale of the ConvLSTM-Attention U-Net model is 1°×1° and 30 days. The F1 score of the ConvLSTM-Attention U-Net model at the optimal spatiotemporal scale is 0.769. Therefore, the ConvLSTM-Attention U-Net model at the spatiotemporal scale of 1°×1° and 30 days is the optimal prediction method for tuna longline fishing activities in the Western and Central Pacific.

[0063] Step S5: Visualization of prediction results and variable importance. Based on the standard gradient back-propagation algorithm, the average absolute gradient value of each variable is calculated to evaluate and visualize the importance of the variable.

[0064] The model prediction results and the visualization results of the actual fishing effort distribution at the optimal spatiotemporal scale (1°×1°, 30 days) are shown in Figure 2. Figure 3 As shown in the figure, the prediction results of the ConvLSTM-Attention U-Net model are similar to the actual distribution of fishing effort in terms of the location and intensity of fishing activities, and the prediction results of the model are relatively accurate. Figure 4 As shown in the figure, judging from the percentage of the average absolute gradient value of the variable to the total gradient value, the importance of fishing effort data at each temporal and spatial scale is the highest, and the importance of the six marine environmental variables is relatively close, with no obvious distinction.

[0065] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for predicting tuna longline fishing activities, characterized in that: The following steps are involved: Step S1, obtaining the fishing effort data of tuna in the central and western Pacific region from 2017 to 2023 and the marine environment data corresponding to the fishing effort data in the region; Step S2, constructing spatiotemporal sequence tensor data of different spatiotemporal scales according to different time resolutions and spatial resolutions; Step S3, constructing a ConvLSTM-Attention U-Net network model, using the fishing effort data and marine environment data of the first M time steps to predict the distribution of fishing activities in the next time step at different spatiotemporal scales, training the ConvLSTM-Attention U-Net network, learning the spatiotemporal information in the historical fishing effort data and environmental variable data and the complex mapping relationship between environmental variables and fishing effort distribution, and determining the optimal spatiotemporal scale model by comparing the prediction performance of the model at different spatiotemporal scales, wherein the time step M is greater than or equal to 3; Step S4, based on the optimal spatiotemporal scale model, the model prediction results are visualized and compared with the actual fishing effort distribution; Step S5, calculate the average absolute gradient value of the variables at different spatiotemporal scales, evaluate and visualize the importance of the variables at different spatiotemporal scales.

2. A tuna longline fishing activity prediction method according to claim 1, characterized in that: The fishing effort data in step S1 include operation date, operation location, and fishing effort; the environmental factor data include sea surface temperature, dissolved oxygen concentration in 100-meter deep seawater, sea water temperature in 400-meter deep seawater, chlorophyll a concentration, dissolved oxygen concentration in 400-meter deep seawater, and salinity in 100-meter deep seawater.

3. The method for predicting tuna longline fishing activities according to claim 1, characterized in that: The processing method of fishing effort and marine environmental variables at different spatiotemporal scales in step S2 is to calculate the daily average fishing effort of each grid at each spatiotemporal scale as the fishing effort value of the grid, and calculate the overall mean of the environmental data in each grid as the environmental variable value of the grid.

4. The method for predicting tuna longline fishing activities according to claim 1, characterized in that: The ConvLSTM-Attention U-Net model structure combines the ConvLSTM model and the Attention U-Net model to capture the long-term temporal characteristics of the data and process spatial features of different scales to achieve refined prediction of fishing activities.

5. A method for predicting tuna longline fishing activities according to claim 4, characterized in that: The ConvLSTM model is good at learning long-term spatiotemporal features in the data, which is calculated as follows: i t =σ(W xi *X t +W hi *H t-1 +W ci ⊙C t-1 +b i ) f t =σ(W xf *X t +W hf *H t-1 +W cf ⊙C t-1 +b f ) C t =f t ⊙C t-1 +i t ⊙tanh(W xc *X t +W hc *H t-1 +b c ) o t =σ(W xo *X t +W ho *H t-1 +W co ⊙C t +b o ) H t =o t ⊙tanh(C t ) Among them, i t 、f t , C t , o t and H t They represent input gate, forget gate, memory unit, output gate and hidden state respectively, X t is the input of the current time step, H t-1 is the hidden state of the previous time step, C t-1 is the memory unit of the previous time step, σ represents the sigmoid activation function, tanh represents the hyperbolic tangent activation function, * represents the convolution operation, ⊙ represents the Hadamard product of the matrix, W xi , W hi and W ci represents the weight matrix of the input gate, W xf , W hf and W cf Represents the weight matrix of the forget gate, W xc and W hc Represents the weight matrix of the memory unit, W xo , W ho and W co represents the weight matrix of the output gate, b i , b f , b c and b o Indicates bias.

6. The method for predicting tuna longline fishing activities according to claim 1, characterized in that: The time step M is set to 5.

7. The method for predicting tuna longline fishing activities according to claim 1, characterized in that: In different spatiotemporal scales, the fishing effort data and marine environment data of the first M time steps are used to predict the distribution of fishing activities in the next time step, and the ConvLSTM-Attention U-Net network is trained to extract the spatiotemporal features in the data. The optimal spatiotemporal scale model is selected according to the general evaluation index F1 score and the length of the time series. The calculation formula of the F1 score is as follows: Among them, N TP The number of grids that actually have fishing effort and are predicted to have fishing effort, N FP The number of grids where there is actual fishing effort but is predicted to be without fishing effort, N FN Indicates the number of grid cells where there is actually no fishing effort but is predicted to be fishing effort.

8. The method for predicting tuna longline fishing activities according to claim 1, characterized in that: The spatiotemporal resolution of the ConvLSTM-Attention U-Net model is set to 1°×1° and 30 days, and the spatiotemporal scale prediction results are visualized to be closest to the actual distribution of fishing effort.

9. The method for predicting tuna longline fishing activities according to claim 1, characterized in that: The ConvLSTM-Attention U-Net model has the highest importance of historical fishing effort data at all spatiotemporal scales, and the importance of the six marine environmental variables is relatively close, with no obvious difference.

10. A tuna longline fishing activity prediction system according to claims 1 to 9, characterized in that: It includes data acquisition module, data construction module, model construction module, prediction comparison module and prediction evaluation module, among which: A data acquisition module is used to obtain the fishing effort data of tuna in the Western and Central Pacific region from 2017 to 2023 and the marine environment data corresponding to the fishing effort data in the region; Data construction module, used to construct spatiotemporal sequence tensor data of different spatiotemporal scales according to different temporal resolutions and spatial resolutions; A model building module is used to build a ConvLSTM-Attention U-Net network model. At different spatiotemporal scales, the fishing effort data and marine environment data of the first M time steps are used to predict the distribution of fishing activities in the next time step. The ConvLSTM-Attention U-Net network is trained to learn the spatiotemporal information in the historical fishing effort data and environmental variable data and the complex mapping relationship between environmental variables and fishing effort distribution. The optimal spatiotemporal scale model is determined by comparing the prediction performance of the model at different spatiotemporal scales, wherein the time step M is greater than or equal to 3. A prediction comparison module, which is used to visualize the model's prediction results and compare them with the actual fishing effort distribution based on the optimal spatiotemporal scale model; The prediction evaluation module is used to calculate the average absolute gradient value of variables at different spatiotemporal scales, evaluate and visualize the importance of variables at different spatiotemporal scales.

Images