Fishing ground prediction method and device based on multi-environmental factor deep learning

Abstract

The invention relates to a fishery prediction method based on multi-environmental factor deep learning, and belongs to the ocean fishing condition forecast and ocean information technology, and the method comprises the following steps: obtaining fishery data and multi-source environmental data; calculating the fishing catch per unit fishing effort, and defining the center area of the fishing ground by adopting a quartile method; data sets of different time scales are established, and environment data and fishery center definition are processed for each time scale; constructing a deep learning prediction model; establishing a prediction model by taking the sea surface temperature as a main part and combining other environmental factors; dividing the data set into a training set, a verification set and a test set, and training each prediction model by using the training set and the verification set; performing validity analysis on each prediction model; selecting an optimal prediction model; and obtaining a fishery center prediction result, and forming a fishery center probability distribution diagram. According to the invention, the technical problems of overlarge prediction area and scattered distribution of a single environment factor model are solved, and high-precision prediction of the central area of the fishing ground is realized.

Description

Technical Field

[0001] This invention belongs to the field of deep-sea fishing forecasting and marine information technology, specifically involving a fishing ground prediction method and a fishing ground prediction device based on deep learning of multiple environmental factors. Background Technology

[0002] Fishing ground forecasting is a crucial technical aspect of fisheries resource development and management. Traditional fishing ground forecasting methods primarily rely on the statistical relationship between single environmental factors (such as sea surface temperature, SST) and catch yield, employing methods like the Habitat Suitability Index (HSI) or Generalized Additive Models (GAM). However, these methods suffer from the following technical limitations: First, significant bottlenecks exist in data processing. Marine environmental data exhibits typical multi-source heterogeneity, including various environmental factors such as sea surface temperature (SST), sea surface height (SSH), sea surface salinity (SSS), and chlorophyll a (Chl-a). These data differ considerably in spatiotemporal resolution and acquisition methods, making effective fusion difficult using traditional statistical methods. Simultaneously, fisheries observation data often suffers from insufficient sample size and incomplete spatiotemporal coverage, posing challenges to model construction.

[0003] Secondly, existing prediction models have obvious limitations. Traditional statistical models (such as the generalized additive model GAM and the habitat suitability index HSI) have limited ability to express complex nonlinear relationships; single environmental factor models (such as those based solely on SST) are easy to calculate, but their prediction results are too broad and lack practicality, making it difficult to provide accurate guidance for fishery production.

[0004] Finally, there is a lack of standardized criteria for defining the central area of ​​fishing grounds. Traditional methods often rely on expert experience or simple statistical thresholds, which fail to objectively reflect the concentrated characteristics of resource distribution, resulting in poor operability of the prediction results.

[0005] In recent years, deep learning technology has made breakthrough progress in the field of image processing, and architectures such as U-Net have performed well in semantic segmentation tasks. However, directly applying existing deep learning models to fishery prediction still faces the following technical challenges: (1) Marine environmental data has the characteristics of being multi-source, multi-dimensional, and multi-scale, making it difficult for existing models to effectively extract key features; (2) Fishery data is prone to the problem of "only appearing data", that is, only catch records are available but negative samples without catches are lacking; (3) Models are prone to overfitting and have insufficient generalization ability; (4) The spatial concentration of prediction results is insufficient, which is not conducive to practical fishery production applications. Therefore, developing a deep learning prediction method that can integrate multiple environmental factors, adapt to different time scales, and improve the concentration of the central area of ​​the fishery has important theoretical value and practical application significance. Summary of the Invention

[0006] The purpose of this invention is to address the problems existing in the prior art by providing a fishery prediction method based on deep learning of multiple environmental factors. This method overcomes the problems of excessively large prediction areas and scattered distribution caused by the use of a single environmental factor in existing fishery prediction methods; breaks through the bottleneck that the choice of time scale has a significant impact on the prediction results; and solves the technical difficulty of lacking a systematic method for optimizing the combination of multiple environmental factors.

[0007] The technical solution is as follows: A method for predicting fishing grounds based on deep learning of multiple environmental factors includes the following steps: Step 1: For fishery production data of a specified time period, a specified sea area, and a specified fish species, collect corresponding multi-source marine environmental data. The marine environmental data includes sea surface temperature (SST), sea surface height (SSH), sea surface salinity (SSS), and chlorophyll a concentration (Chl-a), with a spatial resolution of not less than 0.25°×0.25° and a temporal resolution of daily. Step 2: Calculate the catch per unit effort (CPUE) of the fisheries data from Step 1, and define the central area of ​​the fishing ground using the quartile method; Step 3: Establish datasets at different time scales, and process environmental data and fishery center definitions separately for each time scale; Step 4: Build a deep learning prediction model based on the U-Net architecture; Step 5: Establish a prediction model based primarily on sea surface temperature, combined with other environmental factors; Step 6: Divide the dataset from Step 3 into a training set, a validation set, and a test set, and use the training set and validation set to train each prediction model from Step 5.

[0008] Step 7: Perform effectiveness analysis on the prediction models obtained in Step 6. The effectiveness analysis includes: overall accuracy (OA) analysis, F1 score analysis, catch coverage (CCR) analysis, fishery center area ratio (APCFG) analysis, and fishery application effect index (AEIFG) analysis. Step 8: Select the prediction model with the highest overall accuracy, the largest F1 score, and the highest AEIFG as the optimal prediction model; Step 9: Apply the optimal prediction model to predict the center of the fishing grounds in the target sea area, obtain the prediction results of the fishing ground center, and form a probability distribution map of the fishing ground center.

[0009] Furthermore, in step 1, the specified time period is from 2002 to the year to be predicted, the specified sea area is the coordinate range of 36°~48°N, 145°~165°E, and the specified fish species is Northwest Pacific squid.

[0010] Furthermore, in step 2, the central area of ​​the fishing ground is defined as the area where the CPUE is above the upper quartile and the label is set to 1, while the rest are non-central areas and the label is set to 0.

[0011] Furthermore, in step 3, the time scale is to divide each month from July to November of the year into one period, and the data of each period is segmented at intervals of 3 days, 6 days, 10 days, 15 days or 30 days.

[0012] Furthermore, the features of the deep learning prediction model in step 4 are as follows: two SpatialDropout2D layers are added to prevent overfitting, and the dropout rate is set to 0.75; the encoding path adopts a three-layer downsampling structure, each layer containing convolution and max pooling operations; the decoding path adopts a three-layer upsampling structure, fusing shallow features through skip connections; the last layer uses the sigmoid activation function for binary classification prediction; and the loss function is binary cross-entropy.

[0013] Furthermore, the prediction model in step 5 has the following 7 schemes: Scheme I, which includes 4 environmental factors: sea surface temperature, sea surface height, sea surface salinity, and chlorophyll a concentration; Scheme II, which includes 3 environmental factors: sea surface temperature, sea surface height, and sea surface salinity; Scheme III, which includes 3 environmental factors: sea surface temperature, sea surface salinity, and chlorophyll a concentration; Scheme IV, which includes 3 environmental factors: sea surface temperature, sea surface salinity, and chlorophyll a concentration; Scheme V, which includes 2 environmental factors: sea surface temperature and sea surface height; Scheme VI, which includes 2 environmental factors: sea surface temperature and sea surface salinity; and Scheme VII, which includes 2 environmental factors: sea surface temperature and chlorophyll a concentration.

[0014] Furthermore, in step 6, 80% of the dataset from 2002 to the previous year is used as the training set, 80% of the dataset from 2002 to the previous year is used as the validation set, and the data of the year to be predicted is used as the test set. The model training parameters are set as follows: batch size 32, learning rate 0.001, 300 iterations, and an early stopping strategy is adopted. The early stopping strategy is to terminate training when the validation set loss does not decrease for 10 consecutive iterations and retain the optimal model parameters.

[0015] Furthermore, the overall accuracy in step 7 F1 score: Catch Coverage Ratio (CCR) = the proportion of actual catches falling within the predicted central area of ​​the fishing ground, where AP is the central area of ​​the fishing ground. CFG = Predict the proportion of the central fishing ground area to the total sea area, application effect index In this model, NTP (TP stands for true positive) represents the number of true central fishing grounds predicted as central fishing grounds, i.e., the number of true samples; NTN (TN stands for true negative) represents the number of true non-central fishing grounds predicted as non-central fishing grounds, i.e., the number of true negative samples; NFP (FP stands for false positive) represents the number of true non-central fishing grounds predicted as central fishing grounds, i.e., the number of false positive samples; NFN (FN stands for false negative) represents the number of true central fishing grounds predicted as non-central fishing grounds, i.e., the number of false negative samples; Precision represents precision, Recall represents recall; AEIFG is used to quantify the concentration of prediction results, and a higher value indicates a better model performance.

[0016] A fishing ground prediction device includes one or more processors, one or more memories, one or more programs, and a data acquisition device. The data acquisition device is used to acquire fishery production data and corresponding multi-source marine environmental data. The one or more programs are stored in the memory. When the one or more programs are executed by the processor, the fishing ground prediction device gradually executes the above-mentioned fishing ground prediction method based on deep learning of multiple environmental factors.

[0017] Furthermore, it also includes a display device for visually displaying the prediction results of the fishing ground prediction equipment and outputting them in the form of a probability distribution map, with areas having a probability > 0.7 marked as the core area of ​​the fishing ground.

[0018] Beneficial effects: 1) This invention integrates multi-source environmental data such as sea surface temperature and chlorophyll concentration to construct an improved U-Net deep learning model, which can more accurately capture the complex nonlinear relationship between fishing ground distribution and environmental factors, significantly improving the prediction accuracy of the central area of ​​the fishing ground. Compared with traditional single environmental factor models, the prediction results have higher spatial resolution and lower false alarm rate.

[0019] 2) By innovatively introducing the fishing ground application effect index, a quantitative assessment of the concentration of fishing ground prediction has been achieved. This index effectively balances the relationship between catch coverage and predicted area, making the prediction results more focused on actual high-yield fishing areas and providing more operational guidance for fishery production.

[0020] 3) The improved U-Net architecture employs regularization techniques such as SpatialDropout2D and batch normalization, which effectively prevents model overfitting; the multi-timescale design enables the model to adapt to changes in fishery distribution under different seasons and environmental conditions, ensuring the interannual stability of the prediction results.

[0021] 4) Through standardized data preprocessing and lightweight model design, efficient operation on ordinary computing devices is achieved; the model supports real-time data input and rapid inference, which can meet the timeliness requirements of fishery production and provide timely fishing ground forecasting services for ocean-going fishing vessels.

[0022] 5) Fishing ground forecasting equipment has a simple structure, low cost, and can quickly forecast fishing grounds for marine fishery resources, thus having a wide market demand. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of a deep learning prediction model based on the U-Net architecture. Figure 3 The loss curves for the training and validation sets at different time scales are shown. Figure 4 A graph showing the overall accuracy (OA) for the training and validation sets; Figure 5 Graphs showing the performance evaluation (OA and F1 scores) of different solutions on the test set; Figure 6 A comparison chart of the actual catch coverage effects of single-factor and multi-factor models; Figure 7 A visual comparison chart of the best fishing ground prediction schemes; Figure 8 A schematic diagram of the fishery prediction device of the present invention; Wherein: 1 is the processor, 2 is the memory, 3 is the program and data acquisition device, and 4 is the display device. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit it. Terms such as "upper," "lower," "front," "rear," "left," "right," "bottom," "inner," and "outer" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the purpose of describing the invention. They do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention.

[0025] like Figure 1 The method for predicting fishing grounds based on deep learning of multiple environmental factors, as shown, includes the following steps: Step 1: For fishery production data of a specified time period, a specified sea area, and a specified fish species, collect corresponding multi-source marine environmental data. The marine environmental data includes sea surface temperature (SST), sea surface height (SSH), sea surface salinity (SSS), and chlorophyll a concentration (Chl-a), with a spatial resolution of not less than 0.25°×0.25° and a temporal resolution of daily. Step 2: Calculate the catch per unit effort (CPUE) of the fisheries data from Step 1, and define the central area of ​​the fishing ground using the quartile method; Step 3: Establish datasets at different time scales, and process environmental data and fishery center definitions separately for each time scale; Step 4: Build a deep learning prediction model based on the U-Net architecture; Step 5: Establish a prediction model based primarily on sea surface temperature, combined with other environmental factors; Step 6: Divide the dataset from Step 3 into a training set, a validation set, and a test set, and use the training set and validation set to train each prediction model from Step 5.

[0026] Step 7: Perform effectiveness analysis on the prediction models obtained in Step 6. The effectiveness analysis includes: overall accuracy (OA) analysis, F1 score analysis, catch coverage (CCR) analysis, fishery center area ratio (APCFG) analysis, and fishery application effect index (AEIFG) analysis. Step 8: Select the prediction model with the highest overall accuracy, the largest F1 score, and the highest AEIFG as the optimal prediction model; Step 9: Apply the optimal prediction model to predict the center of the fishing grounds in the target sea area, obtain the prediction results of the fishing ground center, and form a probability distribution map of the fishing ground center.

[0027] In step 1, the specified time period is from 2002 to the year to be predicted, the specified sea area is the coordinate range of 36°~48°N, 145°~165°E, and the specified fish species is Northwest Pacific squid.

[0028] In step 2, the central area of ​​the fishing ground is defined as the area where the CPUE is above the upper quartile and the label is set to 1. The rest are non-central areas and the label is set to 0.

[0029] The time scale in step 3 is to divide each month from July to November of the year into one period, and the data of each period is segmented at intervals of 3 days, 6 days, 10 days, 15 days or 30 days.

[0030] The features of the deep learning prediction model in step 4 are as follows: two SpatialDropout2D layers are added to prevent overfitting, and the dropout rate is set to 0.75; the encoding path adopts a three-layer downsampling structure, each layer containing convolution and max pooling operations; the decoding path adopts a three-layer upsampling structure, fusing shallow features through skip connections; the last layer uses the sigmoid activation function for binary classification prediction; and the loss function is binary cross-entropy.

[0031] In step 5, the prediction model has the following 7 options: Option I, which includes 4 environmental factors: sea surface temperature, sea surface height, sea surface salinity, and chlorophyll a concentration; Option II, which includes 3 environmental factors: sea surface temperature, sea surface height, and sea surface salinity; Option III, which includes 3 environmental factors: sea surface temperature, sea surface salinity, and chlorophyll a concentration; Option IV, which includes 3 environmental factors: sea surface temperature, sea surface salinity, and chlorophyll a concentration; Option V, which includes 2 environmental factors: sea surface temperature and sea surface height; Option VI, which includes 2 environmental factors: sea surface temperature and sea surface salinity; and Option VII, which includes 2 environmental factors: sea surface temperature and chlorophyll a concentration.

[0032] In step 6, 80% of the dataset from 2002 to the previous year is used as the training set, 80% of the dataset from 2002 to the previous year is used as the validation set, and the data of the year to be predicted is used as the test set. The model training parameters are set as follows: batch size 32, learning rate 0.001, 300 iterations, and an early stopping strategy is adopted. The early stopping strategy is to terminate training when the validation set loss does not decrease for 10 consecutive iterations and retain the optimal model parameters.

[0033] Overall accuracy in step 7 F1 score: Catch Coverage Ratio (CCR) = the proportion of actual catches falling within the predicted central area of ​​the fishing ground, where AP is the central area of ​​the fishing ground. CFG = Predict the proportion of the central fishing ground area to the total sea area, application effectiveness index: In this model, NTP (TP stands for true positive) represents the number of true central fishing grounds predicted as central fishing grounds, i.e., the number of true samples; NTN (TN stands for true negative) represents the number of true non-central fishing grounds predicted as non-central fishing grounds, i.e., the number of true negative samples; NFP (FP stands for false positive) represents the number of true non-central fishing grounds predicted as central fishing grounds, i.e., the number of false positive samples; NFN (FN stands for false negative) represents the number of true central fishing grounds predicted as non-central fishing grounds, i.e., the number of false negative samples; Precision represents precision, Recall represents recall; AEIFG is used to quantify the concentration of prediction results, and a higher value indicates a better model performance.

[0034] A fishing ground prediction device includes one or more processors 1, one or more memories 2, one or more programs, and a data acquisition device 3. The data acquisition device is used to acquire fishery production data and corresponding multi-source marine environmental data. The one or more programs are stored in the memories. When the one or more programs are executed by the processors, the fishing ground prediction device progressively executes the aforementioned fishing ground prediction method based on deep learning of multiple environmental factors. It also includes a display device 4 for visually displaying the prediction results of the fishing ground prediction device and outputting them in the form of a probability distribution map, marking areas with a probability > 0.7 as the core area of ​​the fishing ground.

[0035] Example 1: (1) For the fishery production statistics of squid in the Northwest Pacific within the coordinate range of 36°~48°N and 145°~165°E from 2002 to 2020, the unit catch effort (CPUE) sequence was extracted and the catch was used as the resource abundance index. (2) Preprocessing of marine environmental factor data, including: sea surface temperature (SST), sea surface height (SSH), sea surface salinity (SSS), and chlorophyll a (Chl-a); uniform spatial resolution of 0.25° and temporal resolution of 3 / 6 / 10 / 15 / 30 days; normalization of all data to the [0,1] interval, and removal of invalid values; (3) Based on the data processed in steps (1) and (2), construct the central fishing ground label as follows: using the quartile method, extract the region where CPUE is higher than the upper quartile and define it as "central fishing ground" with a label of 1; the rest are 0; take the intersection of the "central fishing ground" regions of the four environmental factors as the final label; (4) Construct an improved U-Net deep learning model, with the following structure: Figure 2 The following are examples: Input layer: a multi-channel image with multiple environmental factors superimposed (e.g., 4 channels × 48 × 80 pixels in Scheme 1); Encoding path: three convolutional layers + max pooling + ReLU activation + SpatialDropout2D (dropout rate = 0.75); Decoding path: three-layer upsampling + skip connections + convolution + sigmoid output; Loss function: Binary cross-entropy; Optimizer: Adam; (5) Seven combinations of environmental factors were designed (Case 1~7), and each combination was tested at five time scales (3 / 6 / 10 / 15 / 30 days), for a total of 35 experiments, see Table 1.

[0036] Table 1 Environmental Factor Combination Design Scheme (6) Verify the application effect of the model through actual fishing data, and define: Overall accuracy: F1 score: Catch Coverage Ratio (CCR): The proportion of the actual catch falling in the central area of ​​the predicted fishing ground; Fishing ground center area ratio: AP CFG = Predict the proportion of the central fishing ground area to the total sea area; Application Effectiveness Index: The AEI FG This is used to quantify the concentration of prediction results; a higher value indicates a better model application effect.

[0037] (7) Figure 3 The loss curves for the training and validation sets at different time scales show the changes in the model's loss function at five time scales (3 / 6 / 10 / 15 / 30 days). All schemes converged after 300 training rounds, and the validation set loss stabilized between 0.09 and 0.27. It can be seen that the loss is lowest at the 30-day time scale, indicating that long-term data integration improves the model's stability. Figure 4 The overall accuracy (OA) curves of the training and validation sets show that the OA of the validation set reaches a maximum of 96.37% on the 30-day time scale. These two figures illustrate the balance between loss and accuracy of the model, thus 30 days was chosen as the optimal time scale.

[0038] (8) such as Figure 5 By comparing the OA and FI scores of seven environmental factor combinations across five time scales, it was found that the larger the time scale, the better the performance, i.e., 30 days > 15 days > 10 days > 6 days > 3 days. Compared with the single-factor SST scheme, the multi-subject scheme showed better AP. CFG A decrease of 11.82%, AEI FG An improvement of 0.55 was achieved; Option 7 (SST+Chl-a) performed best on the 30-day timescale, with a validation set OA of 88.74% and F1 of 0.8732; Table 2 clearly shows that compared to other combinations of environmental factors, the AP of the multi-environmental factor combination is significantly higher. CFG The average decrease of 11.82% (from 53.37% to 41.55%) indicates a more accurate prediction of fishing ground area; the CCR decreased only slightly by 0.97% (from 98.55% to 97.57%), indicating that while the model reduces the fishing ground area, it still covers the vast majority of actual catches; the AEI of the multi-factor model... FG An average improvement of 28.5% (from 1.93 to 2.48) demonstrates its superior performance in practical applications. Combined with... Figure 6The comparison chart of the actual catch coverage effects of single-factor and multi-factor models clearly shows that the multi-factor model, by introducing factors such as SSH, SSS, and Chl-a, significantly optimizes the spatial concentration of fishing grounds without losing CCR.

[0039] Table 2 Comparison of Actual Fishing Results of Single-Factor and Multi-Factor Combined Models (8) Finally, Scheme 7 (SST+Chl-a) + 30-day scale was selected as the optimal model for predicting the central fishing grounds of North Pacific squid. The prediction results are as follows: Figure 7 The data shows that the central fishing grounds are distributed in a strip shape, with a significantly increased spatial concentration.

[0040] Example 2: An electronic device, such as Figure 5 As shown, it includes one or more processors, one or more memories, one or more programs, and a data collection device; The data collection device is used to acquire fishery production data and environmental factor data of fish species C in sea area B within time period A (such as squid data within 36°~48°N and 145°~165°E from 2002 to 2020 in Example 1). One or more programs are stored in memory. When one or more programs are executed by the processor, the electronic device performs the fishery center area prediction method based on deep learning of multiple environmental factors as described in Example 1.

[0041] 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 and improvements made within the principles and spirit of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting fishing grounds based on deep learning of multiple environmental factors, characterized in that: Includes the following steps: Step 1: For fishery production data of a specified time period, a specified sea area, and a specified fish species, collect corresponding multi-source marine environmental data. The marine environmental data includes sea surface temperature, sea surface height, sea surface salinity, and chlorophyll a concentration, with a spatial resolution of not less than 0.25°×0.25° and a temporal resolution of daily. Step 2: Calculate the catch per unit fishing effort for the fishery data in Step 1, and define the central area of ​​the fishing ground using the quartile method; Step 3: Establish datasets at different time scales, and process environmental data and fishery center definitions separately for each time scale; Step 4: Build a deep learning prediction model based on the U-Net architecture; Step 5: Establish a prediction model based primarily on sea surface temperature, combined with other environmental factors; Step 6: Divide the dataset from Step 3 into a training set, a validation set, and a test set, and use the training set and validation set to train each prediction model from Step 5. Step 7: Perform effectiveness analysis on each prediction model obtained in Step 6. The effectiveness analysis includes: overall accuracy analysis, F1 score analysis, catch coverage analysis, fishery center area ratio analysis, and fishery application effect index analysis. Step 8: Select the prediction model with the highest overall accuracy, the largest F1 score, and the highest application effect index in the fishing grounds as the optimal prediction model; Step 9: Apply the optimal prediction model to predict the center of the fishing grounds in the target sea area, obtain the prediction results of the fishing ground center, and form a probability distribution map of the fishing ground center.

2. The fishing ground prediction method based on deep learning of multiple environmental factors as described in claim 1, characterized in that: In step 1, the specified time period is from 2002 to the year to be predicted, the specified sea area is the coordinate range of 36°~48°N, 145°~165°E, and the specified fish species is Northwest Pacific squid.

3. The fishing ground prediction method based on deep learning of multiple environmental factors as described in claim 1, characterized in that: In step 2, the central area of ​​the fishing ground is defined as the area where the catch per unit fishing effort is higher than the upper quartile and is labeled as 1. The rest are non-central areas and are labeled as 0.

4. The fishing ground prediction method based on deep learning of multiple environmental factors as described in claim 1, characterized in that: The time scale in step 3 is to divide each month from July to November of the year into one period, and the data of each period is segmented at intervals of 3 days, 6 days, 10 days, 15 days or 30 days.

5. The fishing ground prediction method based on deep learning of multiple environmental factors as described in claim 1, characterized in that: The features of the deep learning prediction model in step 4 are as follows: two SpatialDropout2D layers are added to prevent overfitting, and the dropout rate is set to 0.75; the encoding path adopts a three-layer downsampling structure, each layer containing convolution and max pooling operations; the decoding path adopts a three-layer upsampling structure, fusing shallow features through skip connections; the last layer uses the sigmoid activation function for binary classification prediction; and the loss function is binary cross-entropy.

6. The fishing ground prediction method based on deep learning of multiple environmental factors as described in claim 1, characterized in that: The prediction model in step 5 has the following seven schemes: Scheme I, which includes four environmental factors: sea surface temperature, sea surface height, sea surface salinity, and chlorophyll a concentration; Scheme II, which includes three environmental factors: sea surface temperature, sea surface height, and sea surface salinity; Scheme III, which includes three environmental factors: sea surface temperature, sea surface salinity, and chlorophyll a concentration; Scheme IV, which includes three environmental factors: sea surface temperature, sea surface salinity, and chlorophyll a concentration; Scheme V, which includes two environmental factors: sea surface temperature and sea surface height; Scheme VI, which includes two environmental factors: sea surface temperature and sea surface salinity; and Scheme VII, which includes two environmental factors: sea surface temperature and chlorophyll a concentration.

7. The fishing ground prediction method based on deep learning of multiple environmental factors as described in claim 1, characterized in that: In step 6, 80% of the dataset from 2002 to the previous year is used as the training set, 80% of the dataset from 2002 to the previous year is used as the validation set, and the data of the year to be predicted is used as the test set. The model training parameters were set as follows: batch size 32, learning rate 0.001, 300 iterations, and an early stopping strategy. The early stopping strategy is to terminate training when the validation set loss does not decrease for 10 consecutive iterations and retain the optimal model parameters.

8. The fishing ground prediction method based on deep learning of multiple environmental factors as described in claim 1, characterized in that: Overall accuracy in step 7 F1 score: Catch Coverage Ratio (CCR) = the proportion of actual catches falling within the predicted central area of ​​the fishing ground, where AP is the central area of ​​the fishing ground. CFG = Predict the proportion of the central fishing ground area to the total sea area, application effectiveness index: Where NTP represents the number of true central fishing grounds predicted as central fishing grounds, i.e., the number of true samples; NTN represents the number of true non-central fishing grounds predicted as non-central fishing grounds, i.e., the number of true negative samples; NFP represents the number of true non-central fishing grounds predicted as central fishing grounds, i.e., the number of false positive samples; NFN represents the number of true central fishing grounds predicted as non-central fishing grounds, i.e., the number of false negative samples. Precision represents accuracy, Recall represents recall; AEIFG is used to quantify the concentration of prediction results, and a higher value indicates a better model performance.

9. A fishery prediction device, characterized in that: It includes one or more processors (1), one or more memories (2), one or more programs and a data acquisition device (3), wherein the data acquisition device is used to acquire fishery production data and corresponding multi-source marine environmental data, and the one or more programs are stored in the memory. When the one or more programs are executed by the processor, the fishing ground prediction device gradually executes the fishing ground prediction method based on deep learning of multiple environmental factors as described in any one of claims 1-8.

10. The fishery prediction device as described in claim 9, characterized in that: It also includes a display device (4) for visually displaying the prediction results of the fishing ground prediction device and outputting them in the form of a probability distribution map, and marking the core area of ​​the fishing ground with a probability > 0.7.

Images