Deep learning fishery prediction method based on mixed spatial-temporal scale
By constructing a U-Net neural network model with hybrid spatiotemporal scales, the shortcomings of existing fishing ground prediction methods in handling complex spatiotemporal relationships are addressed, achieving high-precision and stable fishing ground prediction, which is suitable for marine fishery resource management.
Patent Information
- Application Number
- CN202511102833.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-07
- Publication Date
- 2025-11-14
AI Technical Summary
Existing fishing ground prediction methods, when dealing with remote sensing big data with nonlinear, multi-scale, and strong spatiotemporal coupling characteristics, have symmetrical spatiotemporal scale designs, which cannot effectively capture irregular dynamic features such as ocean eddies and fronts. Furthermore, they lack efficiency and accuracy in processing complex spatiotemporal relationships and are difficult to handle small sample data.
A deep learning approach based on hybrid spatiotemporal scales is adopted. By constructing a U-Net neural network model, setting an asymmetric combination of temporal and spatial scales, a sample dataset is generated. The optimal prediction model is selected using overall accuracy, precision, and F1 score evaluation indicators. Fishing ground prediction is performed using environmental factor data such as sea surface temperature.
It significantly improves the accuracy and stability of fishing ground forecasting, effectively captures complex spatiotemporal relationships, is highly adaptable, and is suitable for large-scale marine fishery resource management, providing a scientific basis.
Smart Images
Figure CN120952247A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of marine fishing ground prediction technology, specifically involving a deep learning method for predicting fishing grounds based on mixed spatiotemporal scales. Background Technology
[0002] In the era of big data in marine fisheries, the demand for real-time and refined spatiotemporal scale fishery forecasting is increasing. Fishing ground prediction is the most important research content in fishery forecasting. Accurate prediction of fishing ground location is crucial for improving fishing efficiency and saving fuel costs. Fishing ground prediction is also an important foundation for the sustainable utilization and scientific management of fishery resources. Currently, commonly used fishing ground prediction methods include generalized additive models (GAM), habitat suitability index models (HSI), and artificial neural network models (ANN). These methods can reveal the relationship between resource abundance and environmental factors to some extent. However, when processing remote sensing big data with nonlinear, multi-scale, and strongly spatiotemporally coupled characteristics, they still have the following shortcomings: 1. The spatiotemporal scale design often adopts a symmetrical distribution, which cannot effectively capture irregular dynamic features such as ocean eddies and fronts; 2. Insufficient efficiency and accuracy in processing complex spatiotemporal relationships; 3. High sample size requirement, making it difficult to handle small sample data.
[0003] Deep learning, as a cutting-edge technology in artificial intelligence, can accurately and efficiently extract nonlinear information from complex data, providing a new direction for intelligent fisheries forecasting. In particular, the U-Net convolutional neural network, due to its end-to-end training capability, good image segmentation performance, and spatial feature extraction ability, has been gradually applied in the fields of marine ecology and fisheries. However, most current studies use symmetrical spatial grids and fixed time windows for model input scale settings, ignoring the significant directional and scale heterogeneity of real ocean processes such as ocean currents, fronts, and eddies, thus limiting the improvement of model prediction accuracy.
[0004] Therefore, there is an urgent need for a fishing ground prediction method that integrates asymmetric spatiotemporal scale design and deep neural network structure to fully extract high-value information from remote sensing data and improve the accuracy and stability of fishing ground distribution prediction.
[0005] Technical content The purpose of this invention is to overcome the shortcomings of existing fishing ground prediction methods, such as poor adaptability to complex environmental data, low prediction accuracy, and difficulty in handling multi-scale spatiotemporal features, and to provide a deep learning fishing ground prediction method based on mixed spatiotemporal scales with high prediction accuracy, strong adaptability, and good stability.
[0006] To achieve the above objectives, the present invention proposes the following technical solution: A deep learning-based method for predicting fishing grounds based on hybrid spatiotemporal scales includes the following steps: Step 1: Obtain historical catch data and corresponding environmental factor data for a specific species B fishery resource within the target sea area A over a specific time period C, and determine the location of the central fishing ground in the target sea area A based on the historical catch data. Step 2: Set different time scales and different spatial scales in the longitude and latitude directions to construct multiple hybrid spatiotemporal scale combinations. Reconstruct the environmental factor data on the hybrid spatiotemporal combinations to generate a sample dataset. Step 3: Using each sample in the sample dataset and the location of the central fishing ground of the corresponding target sea area A as input and output data respectively, train the deep learning prediction model, select the deep learning prediction model with the best evaluation index as the optimal prediction model, and its corresponding hybrid spatiotemporal combination is the optimal spatiotemporal scale combination. Step 4: Collect environmental factor data of the target sea area corresponding to the optimal combination of spatiotemporal scales as input data, and use the optimal prediction model to predict the fishing grounds of the target sea area.
[0007] Furthermore, in step two, the time period C is divided into m different time scales, and n different scales are set in both the longitude and latitude directions to construct n×n spatial scales, thereby constructing a combination of m×n×n mixed spatial scales.
[0008] Further, generate the sample dataset by following these steps: Step I: Divide the time period C into t time intervals based on a certain time scale. Each time interval contains environmental factor data for multiple days. Step II: For each time interval, identify the environmental factor data of each day within it according to a certain spatial scale, then normalize the pixel values, and then calculate the average value of the environmental factor data corresponding to each day within the time interval, thereby obtaining the sample corresponding to the time interval at that spatial scale. Step III: Repeat Step II to obtain the corresponding samples at all spatial scales at this time scale, for a total of t×n×n samples; Step IV: Repeat steps I-III to obtain the corresponding samples at all time scales and all spatial scales, resulting in a total of t×n×n×m samples, thus generating a sample dataset containing t×n×n×m samples.
[0009] Furthermore, the time scale is set with m=5 values, namely 3 days, 6 days, 10 days, 15 days, and 30 days. In both the longitude and latitude directions, n=4 scales are set, namely 0.05°, 0.1°, 0.25°, and 0.5°, to construct a 4×4 spatial scale, as detailed below: 0.05°*0.05°, 0.05°*0.1°, 0.05°*0.25°, 0.05°*0.5°; 0.1°*0.05°, 0.1°*0.1°, 0.1°*0.25°, 0.1°*0.5°; 0.25°*0.05°, 0.25°*0.1°, 0.25°*0.25°, 0.25°*0.5°; 0.5°*0.05°, 0.5°*0.1°, 0.5°*0.25°, 0.5°*0.5°.
[0010] Furthermore, the deep learning prediction model is set as a U-Net neural network model, including an encoder and a decoder. The encoder includes convolutional layers, ReLU activation function and max pooling operation, and the decoder includes upsampling and feature concatenation operations, and SpatialDropout2D regularization layers are added to the fourth and fifth convolutional layers. To compare the model performance under different combinations of spatiotemporal scales, we use overall accuracy (OA), precision, recall, and F1 score as evaluation metrics. The deep learning prediction model with the best evaluation metrics is selected as the optimal prediction model, and its corresponding spatiotemporal combination is the optimal spatiotemporal scale combination.
[0011] Furthermore, the upper quartile of historical catch data is used as the threshold for determining the central fishing ground. Spatial locations where the historical catch data of a certain day is greater than or equal to its corresponding upper quartile are marked as central fishing grounds, and the rest are non-central fishing grounds.
[0012] Furthermore, the environmental factor data is set as one or more combinations of sea surface temperature image (SST), sea surface height image (SSH), sea surface salinity image (SSS), and chlorophyll image (Chla).
[0013] An electronic device, comprising: One or more processors; one or more memories; one or more programs; a data acquisition device; The data acquisition device is used to acquire fish catch data and marine environmental factor data. The one or more programs are stored in the memory. When the one or more programs are executed by the processor, the electronic device performs the deep learning fishery prediction method based on hybrid spatiotemporal scales as described above.
[0014] Beneficial effects: (1) This invention provides a fishing ground prediction method based on deep learning and hybrid spatiotemporal scales. By constructing a U-Net neural network model and using sea surface temperature SST image data as input, the spatiotemporal scale selection strategy is optimized. By comparing multiple spatiotemporal scale combinations, the optimal spatiotemporal scale (15-day time scale and 0.25°×0.25° spatial scale) is found, which significantly improves the prediction accuracy and stability of the model, effectively captures complex spatiotemporal relationships, and improves the reliability of fishing ground prediction.
[0015] (2) The present invention adopts a hybrid spatiotemporal scale design, including symmetrical spatiotemporal scale and asymmetrical spatiotemporal scale, which can overcome the limitations of traditional symmetrical spatiotemporal scale methods, accurately capture the influence of different latitudes and longitudes on fishing ground prediction, and at the same time, by optimizing the spatiotemporal scale settings of latitude and longitude, the model can better reflect the marine climate characteristics, especially in complex marine environments, such as the distribution of fishing grounds in the confluence area of the Kuroshio and Oyashio, thereby improving the spatial consistency and accuracy of the prediction results.
[0016] (3) The deep learning U-Net model used in this invention has a simple and efficient structure, can automatically learn complex features in spatiotemporal data, and does not rely on a large amount of domain knowledge. Its rapid training and high accuracy make it suitable for large-scale marine fishery resource prediction, can provide a scientific basis for fishery resource management and sustainable development, and has strong potential for promotion and application.
[0017] Through these innovative designs, this invention provides an efficient and accurate fishing ground prediction tool, which is not only applicable to the North Pacific squid fishing grounds, but can also be extended to the prediction of fishing grounds for other sea areas and species, and has broad scientific research and practical value. Attached Figure Description
[0018] Figure 1 This is a flowchart illustrating the deep learning-based fishery prediction method based on hybrid spatiotemporal scales of the present invention. Figure 2 This is a comparison chart of the F1 scores of the model under different combinations of time and space scales in this invention; Figure 3 The flowchart for constructing the central fishing ground tag of the present invention illustrates the fishing ground identification process based on the CPUE upper quartile method. Figure 4 This is a schematic diagram of the structure of the U-Net neural network model used in this invention; Figure 5 The diagram shows the loss curves of the training and validation sets of the fishery prediction model of the present invention at different temporal and spatial scales. Figure 6This is a performance evaluation graph of the present invention on test sets at different time scales; Figure 7 This is a comparison diagram showing the spatial overlap between the output of the optimal prediction model of this invention and the actual central fishing ground label map. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only illustrative of one feasible mode of the present invention and do not constitute a limitation on the scope of protection of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] like Figure 1 As shown, this invention provides a deep learning-based method for predicting fishing grounds based on asymmetric spatiotemporal scales. The specific steps are as follows: Step 1: Obtain historical catch data and corresponding environmental factor data for a specific species B fishery resource within the target sea area A over a time period C. The catch data includes, but is not limited to, the catch per unit catch effort (CPUE) sequence, and the environmental factor data includes, but is not limited to, sea surface temperature (SST) image data, etc., to construct a fishery resource abundance dataset. Then, the upper quartile method can be used to define the central fishing ground. Regions with a catch-per-unit effort (CPUE) value greater than the upper quartile are marked as central fishing grounds; that is, the upper quartile of CPUE is used as the threshold for determining central fishing grounds. Specifically, within each time slice (e.g., a day), spatial locations with CPUE greater than or equal to the quartile of that period are marked as central fishing grounds (value 1), and the rest are non-central fishing grounds (value 0). This central fishing ground label image is used as the output of the deep learning model, and the spatial region image corresponding to the SST is used as the input to the deep learning model, constructing an input-output pair for model training.
[0021] Step 2: Set different time scales and different spatial scales in the longitude and latitude directions to construct multiple hybrid spatiotemporal scale combinations. Reconstruct the environmental factor data on the hybrid spatiotemporal combinations to generate a sample dataset.
[0022] The time period C can be divided into m different time scales. n different scales are set in both the longitude and latitude directions to construct n×n spatial scales, thus creating m×n×n combinations of mixed spatial scales. Specifically, we define the time period C as a month, with each month consisting of 30 days. Therefore, m=5 time scales can be set, namely 3 days, 6 days, 10 days, 15 days, and 30 days. n=4 scales are set in both the longitude and latitude directions, namely 0.05°, 0.1°, 0.25°, and 0.5°, to construct 4×4 spatial scales, as detailed below: 0.05°*0.05°, 0.05°*0.1°, 0.05°*0.25°, 0.05°*0.5°; 0.1°*0.05°, 0.1°*0.1°, 0.1°*0.25°, 0.1°*0.5°; 0.25°*0.05°, 0.25°*0.1°, 0.25°*0.25°, 0.25°*0.5°; 0.5°*0.05°, 0.5°*0.1°, 0.5°*0.25°, 0.5°*0.5°.
[0023] Furthermore, the environmental factor data were reconstructed at different spatial and temporal scales, resulting in the following 80 schemes: Option 1: Time scale 3 days, spatial scale 0.05° × 0.05° Option 2: Time scale 3 days, spatial scale 0.05° × 0.10° Option 3: Time scale 3 days, spatial scale 0.05° × 0.25° Option 4: Time scale 3 days, spatial scale 0.05° × 0.50° Option 5: Time scale 3 days, spatial scale 0.10° × 0.05° Option 6: Time scale 3 days, spatial scale 0.10° × 0.10° Option 7: Time scale 3 days, spatial scale 0.10° × 0.25° Option 8: Time scale 3 days, spatial scale 0.10° × 0.50° Option 9: Time scale 3 days, spatial scale 0.25° × 0.05° Option 10: Time scale 3 days, spatial scale 0.25° × 0.10° Option 11: Time scale 3 days, spatial scale 0.25° × 0.25° Option 12: Time scale 3 days, spatial scale 0.25° × 0.50° Option 13: Time scale 3 days, spatial scale 0.50° × 0.05° Option 14: Time scale 3 days, spatial scale 0.50° × 0.10° Option 15: Time scale 3 days, spatial scale 0.50° × 0.25° Option 16: Time scale 3 days, spatial scale 0.50° × 0.50° Option 17: Time scale 6 days, spatial scale 0.05° × 0.05° Option 18: Time scale 6 days, spatial scale 0.05° × 0.10° Option 19: Time scale 6 days, spatial scale 0.05° × 0.25° Option 20: Time scale 6 days, spatial scale 0.05° × 0.50° Option 21: Time scale 6 days, spatial scale 0.10° × 0.05° Option 22: Time scale 6 days, spatial scale 0.10° × 0.10° Option 23: Time scale 6 days, spatial scale 0.10° × 0.25° Option 24: Time scale 6 days, spatial scale 0.10° × 0.50° Option 25: Time scale 6 days, spatial scale 0.25° × 0.05° Option 26: Time scale 6 days, spatial scale 0.25° × 0.10° Option 27: Time scale 6 days, spatial scale 0.25° × 0.25° Option 28: Time scale 6 days, spatial scale 0.25° × 0.50° Option 29: Time scale 6 days, spatial scale 0.50° × 0.05° Option 30: Time scale 6 days, spatial scale 0.50° × 0.10° Option 31: Time scale 6 days, spatial scale 0.50° × 0.25° Option 32: Time scale 6 days, spatial scale 0.50° × 0.50° Option 33: Time scale 10 days, spatial scale 0.05° × 0.05° Option 34: Time scale 10 days, spatial scale 0.05° × 0.10° Option 35: Time scale 10 days, spatial scale 0.05° × 0.25° Option 36: Time scale 10 days, spatial scale 0.05° × 0.50° Option 37: Time scale 10 days, spatial scale 0.10° × 0.05° Option 38: Time scale 10 days, spatial scale 0.10° × 0.10° Option 39: Time scale 10 days, spatial scale 0.10° × 0.25° Option 40: Time scale 10 days, spatial scale 0.10° × 0.50° Option 41: Time scale 10 days, spatial scale 0.25° × 0.05° Option 42: Time scale 10 days, spatial scale 0.25° × 0.10° Option 43: Time scale 10 days, spatial scale 0.25° × 0.25° Option 44: Time scale 10 days, spatial scale 0.25° × 0.50° Option 45: Time scale 10 days, spatial scale 0.50° × 0.05° Option 46: Time scale 10 days, spatial scale 0.50° × 0.10° Option 47: Time scale 10 days, spatial scale 0.50° × 0.25° Option 48: Time scale 10 days, spatial scale 0.50° × 0.50° Option 49: Time scale 15 days, spatial scale 0.05° × 0.05° Option 50: Time scale 15 days, spatial scale 0.05° × 0.10° Option 51: Time scale 15 days, spatial scale 0.05° × 0.25° Option 52: Time scale 15 days, spatial scale 0.05° × 0.50° Option 53: Time scale 15 days, spatial scale 0.10° × 0.05° Option 54: Time scale 15 days, spatial scale 0.10° × 0.10° Option 55: Time scale 15 days, spatial scale 0.10° × 0.25° Option 56: Time scale 15 days, spatial scale 0.10° × 0.50° Option 57: Time scale 15 days, spatial scale 0.25° × 0.05° Option 58: Time scale 15 days, spatial scale 0.25° × 0.10° Option 59: Time scale 15 days, spatial scale 0.25° × 0.25° Option 60: Time scale 15 days, spatial scale 0.25° × 0.50° Option 61: Time scale 15 days, spatial scale 0.50° × 0.05° Option 62: Time scale 15 days, spatial scale 0.50° × 0.10° Option 63: Time scale 15 days, spatial scale 0.50° × 0.25° Option 64: Time scale 15 days, spatial scale 0.50° × 0.50° Option 65: Time scale 30 days, spatial scale 0.05° × 0.05° Option 66: Time scale 30 days, spatial scale 0.05° × 0.10° Option 67: Time scale 30 days, spatial scale 0.05° × 0.25° Option 68: Time scale 30 days, spatial scale 0.05° × 0.50° Option 69: Time scale 30 days, spatial scale 0.10° × 0.05° Option 70: Time scale 30 days, spatial scale 0.10° × 0.10° Option 71: Time scale 30 days, spatial scale 0.10° × 0.25° Option 72: Time scale 30 days, spatial scale 0.10° × 0.50° Option 73: Time scale 30 days, spatial scale 0.25° × 0.05° Option 74: Time scale 30 days, spatial scale 0.25° × 0.10° Option 75: Time scale 30 days, spatial scale 0.25° × 0.25° Option 76: Time scale 30 days, spatial scale 0.25° × 0.50° Option 77: Time scale 30 days, spatial scale 0.50° × 0.05° Option 78: Time scale 30 days, spatial scale 0.50° × 0.10° Option 79: Time scale 30 days, spatial scale 0.50° × 0.25° Option 80: Time scale 30 days, spatial scale 0.50° × 0.50° Finally, the steps for generating the sample dataset using the above combination of mixed spatial scales are as follows: Step I: Divide the time period C into t time intervals based on a certain time scale. Each time interval contains environmental factor data for multiple days. Step II: For each time interval, identify the environmental factor data of each day within it pixel by pixel according to a certain spatial scale, then normalize the pixel values, replace invalid values with specific label values, and unify the sample dimension; then calculate the average value of the environmental factor data corresponding to each day within the time interval pixel by pixel, so as to obtain the sample corresponding to the time interval at the spatial scale. Step III: Repeat Step II to obtain the corresponding samples at all spatial scales at this time scale, for a total of t×n×n samples; Step IV: Repeat steps I-III to obtain the corresponding samples at all time scales and all spatial scales, resulting in a total of t×n×n×m samples, thus generating a sample dataset containing t×n×n×m samples.
[0024] Of course, considering that time period C includes multiple months, the above method for generating sample datasets is for a sample dataset corresponding to a one-month time period. By repeating this process, we can obtain the sample dataset for the entire time period C.
[0025] Step 3: Using each sample in the sample dataset and the location of the central fishing ground of the target sea area A as input and output data respectively, train the deep learning prediction model, select the deep learning prediction model with the best evaluation index as the optimal prediction model, and its corresponding hybrid spatiotemporal combination is the optimal spatiotemporal scale combination.
[0026] The deep learning prediction model is set as a U-Net neural network model, which has powerful image feature extraction and spatial recognition capabilities and can achieve pixel-level fish farm prediction. It includes an encoder and a decoder. The encoder contains convolutional layers, ReLU activation function and max pooling operation. The decoder contains upsampling and feature concatenation operations. SpatialDropout2D regularization layers are added to the fourth and fifth convolutional layers to prevent overfitting. The model is trained using sigmoid activation function and cross-entropy loss function.
[0027] By comparing the model performance under different combinations of spatiotemporal scales, the overall accuracy (OA), precision, recall, and F1 score are used as evaluation metrics. The deep learning prediction model with the best evaluation metrics is selected as the optimal prediction model, and its corresponding spatiotemporal combination is the optimal spatiotemporal scale combination.
[0028] Step 4: Collect environmental factor data of the target sea area corresponding to the optimal combination of spatiotemporal scales as input data, and use the optimal prediction model to predict the fishing grounds of the target sea area.
[0029] To verify the feasibility of this invention, we used squid from the Northwest Pacific Ocean as an example for testing and illustration: (1) Historical CPUE data (in t•v⁻¹) of Ommastrephes bartramii in the Northwest Pacific (36°N–48°N, 145°E–165°E) from 1998 to 2020, as well as remotely sensed sea surface temperature (SST) data for the corresponding time and spatial location. The CPUE data were aggregated daily, and the SST data were daily-scale Level-4 products.
[0030] (2) For example Figure 2 As shown, multiple time scales (3 days, 6 days, 10 days, 15 days, 30 days) and spatial scales (longitude 0.05°, 0.1°, 0.25°, 0.5°; latitude 0.05°, 0.1°, 0.25°, 0.5°) are set and combined to form a total of 80 mixed time and space scale combination schemes.
[0031] (3) such as Figure 3 As shown, the upper quartile of CPUE is used as the threshold for determining the central fishing ground. Specifically, at each time slice, spatial locations where CPUE is greater than or equal to the quartile of that period are marked as central fishing grounds (value 1), and the rest are non-central fishing grounds (value 0). This central fishing ground label map serves as the supervised output of the deep learning model.
[0032] (4) The sea surface temperature (SST) images are combined according to the set spatiotemporal scales and then extracted to form input samples. Each image is normalized, and missing values are set as fixed identifiers and paired with the corresponding central fishing ground labels to form input-output pairs.
[0033] (5) such as Figure 4 As shown, a U-Net convolutional neural network model is constructed, containing four downsampling layers and four upsampling layers, coupled with a skip connection structure. SpatialDropout2D modules are introduced in the 4th and 5th convolutional layers, with a Dropout rate set to 0.75 to prevent overfitting. The model uses the sigmoid function for output normalization, and the loss function is binary cross-entropy. Data from 1998–2019 is used as the training set, and data from 2020 is used as the test set. The Adam optimizer is used during training, with an initial learning rate of 0.001, a batch size of 8, and 100 training epochs. Figure 5 As shown in the training loss curves, after sufficient training iterations, the loss function values of all models tend to stabilize and reach convergence, verifying the sufficiency of the training process and the effective fitting of the models.
[0034] (6) For example Figure 6 As shown, the overall accuracy (OA) and F1 score of the model were evaluated under different scale combinations, and a performance comparison chart was plotted (see...). Figure 2The results show that the model achieves the best performance on the test set with a time scale of 15 days and a spatial scale of 0.25° × 0.25°, resulting in the highest F1 score and predictions that most closely match the actual overlapping area of the label space. Figure 7 The results show that the outline of the central fishing ground predicted by the optimal scheme (15 days / 0.25°×0.25°) highly coincides with the actual label, with smooth edges and accurate positioning. The practicality of the model is further verified by introducing site coverage (the proportion of the actual fishing ground that is predicted to be covered) and catch coverage (the proportion of the catch in the predicted area).
[0035] In summary, model performance improves with increasing time scale, with 15 days being the optimal critical point. Spatially, finer meshes offer higher accuracy, but excessively fine meshes introduce noise; a combined 0.25° (latitudinal) × 0.25° (longitudinal) mesh is optimal. Further analysis reveals that adjusting the latitudinal resolution has a significantly greater impact on the output than adjusting only the longitude under fixed longitude conditions, indicating that fishing ground distribution is more sensitive to latitudinal ecological changes. Therefore, the asymmetric scale design combining 15 days and 0.25° × 0.25° is significantly superior to traditional symmetric meshes, and can more accurately characterize fishing ground migration trends in complex marine environments.
[0036] (7) The method of the present invention can be deployed in electronic devices with data input, model loading and prediction display functions, including but not limited to shipborne fishery navigation systems and fishery information management platforms. After receiving the latest remote sensing SST data, the device can load the pre-trained optimal model to perform fishing ground prediction.
[0037] Notes: This embodiment uses sea surface temperature (SST) as a single input feature, but it can also be extended to multi-factor fusion input (such as SSS, Chl-a, etc.); the fishery label construction method can adapt to different species resource abundance threshold standards; the method provided by this invention has good transferability and can be applied to spatial distribution prediction tasks of various pelagic fish species.
[0038] The present invention also provides an electronic device, comprising: 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 fish catch data and marine environmental factor data. One or more programs are stored in memory. When one or more programs are executed by the processor, the electronic device performs the deep learning fishery prediction method based on hybrid spatiotemporal scales as described above.
[0039] In this embodiment, the electronic device, computer storage medium, computer program product or chip are all used to execute the corresponding method provided above. Therefore, the beneficial effects that can be achieved can be referred to the beneficial effects of the corresponding method provided above, and will not be repeated here.
[0040] Through the above description of the embodiments, those skilled in the art will understand that, for the sake of convenience and brevity, only the division of the above functional modules is used as an example. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.
[0041] In summary, the deep learning-based fishery prediction method based on asymmetric spatiotemporal scales of this invention fully considers the spatiotemporal dynamics of fish distribution before model building, and designs multiple combinations of asymmetric scales to avoid the problem of poor adaptability of traditional fixed symmetric grid models to the changing characteristics of real fishery grounds. During model training, mechanisms such as skip connections and spatial dropout are introduced to effectively improve the stability and noise resistance of the model. Finally, through cross-year evaluation and verification, the generalization ability and prediction accuracy of the model are improved.
[0042] This invention is applicable to fishing ground prediction tasks, using pelagic fish such as the calamari (Ommastrephes bartramii) as examples, and can also be extended to spatial distribution modeling of other economically important fish species. Compared with existing fishing ground prediction models, this invention has the following advantages: (1) Asymmetric spatiotemporal scales are used to construct multi-granularity samples, making full use of the detailed features of remote sensing data and the scale response relationship of fish behavior; (2) Based on the improved U-Net deep learning structure, it has powerful image feature extraction and spatial recognition capabilities, and can achieve pixel-level fishery prediction; (3) By evaluating and selecting the optimal solution through multi-scale and multi-model methods, the prediction model can have both high accuracy and good timeliness and stability. (4) The optimal prediction model proposed in this invention can be dynamically updated based on the latest acquired remote sensing and fishing data, and is suitable for long-term operation or deployment in marine fisheries management systems.
[0043] (5) The final output of the model is in the form of a spatial image, which can intuitively identify the central fishing ground area and can be fused with real-time remote sensing images for fisheries production scheduling. The prediction framework of this invention can be used not only for Northwest Pacific squid, but also for other distant-water economic fish species with significant spatiotemporal mobility or seasonal influence.
[0044] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these are merely illustrative examples. Various changes or modifications can be made to these embodiments without departing from the principles and essence of the present invention. Therefore, the scope of protection of the present invention is defined by the appended claims.
Claims
1. A deep learning-based method for predicting fishing grounds based on hybrid spatiotemporal scales, characterized in that... Includes the following steps: Step 1: Obtain historical catch data and corresponding environmental factor data for a specific species B fishery resource within the target sea area A over a specific time period C, and determine the location of the central fishing ground in the target sea area A based on the historical catch data. Step 2: Set different time scales and different spatial scales in the longitude and latitude directions to construct multiple hybrid spatiotemporal scale combinations. Reconstruct the environmental factor data on the hybrid spatiotemporal combinations to generate a sample dataset. Step 3: Using each sample in the sample dataset and the location of the central fishing ground of the corresponding target sea area A as input and output data respectively, train the deep learning prediction model, select the deep learning prediction model with the best evaluation index as the optimal prediction model, and its corresponding hybrid spatiotemporal combination is the optimal spatiotemporal scale combination. Step 4: Collect environmental factor data of the target sea area corresponding to the optimal combination of spatiotemporal scales as input data, and use the optimal prediction model to predict the fishing grounds of the target sea area.
2. The deep learning-based fishery prediction method based on hybrid spatiotemporal scales according to claim 1, characterized in that: In step two, the time period C is divided into m different time scales, and n different scales are set in both the longitude and latitude directions to construct n×n spatial scales, thereby constructing m×n×n mixed spatial scale combinations.
3. The deep learning-based fishery prediction method based on hybrid spatiotemporal scales according to claim 2, characterized in that... Generate the sample dataset by following these steps: Step I: Divide the time period C into t time intervals based on a certain time scale. Each time interval contains environmental factor data for multiple days. Step II: For each time interval, identify the environmental factor data of each day within it according to a certain spatial scale, then normalize the pixel values, and then calculate the average value of the environmental factor data corresponding to each day within the time interval, thereby obtaining the sample corresponding to the time interval at that spatial scale. Step III: Repeat Step II to obtain the corresponding samples at all spatial scales at this time scale, for a total of t×n×n samples; Step IV: Repeat steps I-III to obtain the corresponding samples at all time scales and all spatial scales, resulting in a total of t×n×n×m samples, thus generating a sample dataset containing t×n×n×m samples.
4. The deep learning-based fishery prediction method based on hybrid spatiotemporal scales according to claim 2, characterized in that: The time scale is set with m=5 values, namely 3 days, 6 days, 10 days, 15 days, and 30 days. In both the longitude and latitude directions, n=4 scales are set, namely 0.05°, 0.1°, 0.25°, and 0.5°, to construct a 4×4 spatial scale, as detailed below: 0.05°*0.05°、0.05°*0.1°、0.05°*0.25°、0.05°*0.5°； 0.1°*0.05°、0.1°*0.1°、0.1°*0.25°、0.1°*0.5°； 0.25°*0.05°、0.25°*0.1°、0.25°*0.25°、0.25°*0.5°； 0.5°*0.05°、0.5°*0.1°、0.5°*0.25°、0.5°*0.5°。 5. The deep learning-based fishery prediction method based on hybrid spatiotemporal scales according to claim 1, characterized in that: The deep learning prediction model is set as a U-Net neural network model, including an encoder and a decoder. The encoder includes convolutional layers, ReLU activation function and max pooling operation. The decoder includes upsampling and feature concatenation operations, and SpatialDropout2D regularization layers are added to the fourth and fifth convolutional layers. To compare the model performance under different combinations of spatiotemporal scales, we use overall accuracy (OA), precision, recall, and F1 score as evaluation metrics. The deep learning prediction model with the best evaluation metrics is selected as the optimal prediction model, and its corresponding spatiotemporal combination is the optimal spatiotemporal scale combination.
6. The deep learning-based fishery prediction method based on hybrid spatiotemporal scales according to claim 1, characterized in that: The upper quartile of historical catch data is used as the threshold for determining the central fishing ground. Spatial locations where the historical catch data of a certain day is greater than or equal to its corresponding upper quartile are marked as central fishing grounds, and the rest are non-central fishing grounds.
7. The deep learning-based fishery prediction method based on hybrid spatiotemporal scales according to claim 1, characterized in that: The environmental factor data are set as one or more combinations of sea surface temperature image (SST), sea surface height image (SSH), sea surface salinity image (SSS), and chlorophyll image (Chla).
8. An electronic device, characterized in that, include: One or more processors; One or more memory units; One or more programs; A data acquisition device; The data acquisition device is used to acquire fish catch data and marine environmental factor data for fishery production. The one or more programs are stored in the memory. When the one or more programs are executed by the processor, the electronic device performs the deep learning fishery prediction method based on hybrid spatiotemporal scales as described in any one of claims 1-7.