Deep learning-based short and medium-term fishery space distribution forecasting method

Abstract

The invention discloses a deep learning-based short and medium-term fishery space distribution forecasting method, and belongs to the technical field of marine fishery resource forecasting and ecological modeling. Comprising the following steps: acquiring historical fishery production data of a specific species in a target sea area and corresponding historical environmental factor data; establishing a plurality of data sets according to different time scales and advance moments; a deep learning network for fishery prediction is constructed on the basis of a U-Net model and trained, a coding part of the deep learning network comprises a plurality of layers of down-sampling, a Spatial Dropout 2D layer is added after at least the last layer of down-sampling, and an attention mechanism is added after at least one layer of down-sampling; and according to the best time scale and advance moment combination, obtaining the environment factor data of the advance moment, inputting the environment factor data into the trained deep learning network, and predicting to obtain the current fishery space distribution. The method effectively captures the nonlinear spatial-temporal characteristics of the environmental data, and improves the accuracy of fishery spatial distribution prediction.

Description

Technical Field

[0001] This invention belongs to the field of deep-sea fishing ground forecasting technology, and relates to a method for forecasting the spatial distribution of fishing grounds in the short and medium term based on deep learning. Background Technology

[0002] The spatiotemporal distribution of pelagic economic species resources is closely related to their marine environmental conditions. Accurate spatial prediction of fishing grounds is crucial for improving fishing efficiency and reducing fishing costs. Traditional fishing ground prediction focuses on real-time forecasting and lacks methods for continuous forecasting of short- and medium-term changes. Current fishing ground prediction mainly relies on statistical or empirical models, which struggle to effectively consider the spatiotemporal autocorrelation and nonlinear characteristics of environmental factors. Deep learning technology, as an effective tool for processing large amounts of marine environmental data, is increasingly being applied to the field of fishing ground prediction. However, existing technologies still lack high-precision spatial distribution prediction methods that can effectively handle the fluctuation information and temporal correlation of environmental factors. Therefore, developing deep learning-based short- and medium-term fishing ground spatial distribution prediction technology has significant theoretical and practical implications. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a deep learning-based method for predicting the spatial distribution of fishing grounds in the short to medium term. This method overcomes the limitations of traditional methods in handling temporal fluctuations and nonlinear relationships of environmental factors, thereby improving the timeliness and stability of fishing ground prediction.

[0004] To achieve the above objectives, this invention provides a short- to medium-term fishing ground spatial distribution prediction method based on deep learning, comprising the following steps: (1) Obtain historical fishery production data and corresponding historical environmental factor data for specific species in the target sea area, and perform preprocessing; (2) Establish multiple datasets of fishery production data and multiple datasets of environmental factor data according to different time scales and advance times, respectively; (3) A deep learning network for fishery prediction is constructed based on the U-Net model. Its encoding part includes several downsampling layers, and a SpatialDropout2D layer is added after at least the last downsampling layer. An attention mechanism is added after at least one downsampling layer. The decoding part includes several upsampling layers corresponding to the encoding part. The input of the deep learning network is environmental factor data, and the output is the predicted spatial distribution map of the fishery. (4) Input the environmental factor data from step (2) into the deep learning network for training and evaluation, and select the best-performing time scale and advance time combination as the final model input method; (5) Obtain environmental factor data at the advance time according to the best time scale and advance time combination, input it into the trained deep learning network, and predict the current spatial distribution of the fishing ground.

[0005] Furthermore, the fisheries production data includes date, longitude, latitude, daily yield, and fishing vessels in operation; the environmental factor data includes one or more combinations of sea surface temperature (SST), sea surface height (SSH), sea surface salinity (SSS), chlorophyll content (Chla), ocean current velocity (u), and current velocity (v), with at least sea surface temperature (SST).

[0006] Furthermore, the preprocessing specifically involves: normalizing the historical environmental factor data and marking invalid values.

[0007] Furthermore, the different time scales include 3 days, 5 days, 6 days, 7 days, 10 days, 15 days, 30 days, and 90 days, and the spatial scale is uniformly set to 0.25°×0.25°. The advance timing scheme includes 2 to 15 time steps.

[0008] Furthermore, when there are multiple SpatialDropout2D layers, the layers are added sequentially from the last downsampling layer upwards. The shallower the layer, the lower the Rate, which ranges from 0.5 to 0.8, with a threshold of 0.05 for each layer.

[0009] Furthermore, when a downsampling layer contains both a SpatialDropout2D layer and an attention mechanism, the downsampling first passes through the attention mechanism and then connects to the SpatialDropout2D layer.

[0010] Furthermore, the deep learning network uses cross-entropy as the loss function during training.

[0011] Furthermore, based on the various datasets of fishery production data, the corresponding CPUE, Catch, effort, or suitable habitat index HSI area is calculated, and the spatial distribution map of the fishing grounds is obtained as the training ground's true value using any one of the following methods: quartile division, median, mean, or specific threshold.

[0012] Furthermore, the deep learning network evaluation includes accuracy, precision, recall, F1 score, ROC area, and AUC under the PR curve.

[0013] The beneficial effects of this invention are: This invention effectively captures the nonlinear spatiotemporal characteristics of environmental data using the U-Net deep learning model, improving the accuracy of fishing ground spatial distribution prediction. By introducing environmental data images from multiple consecutive time points as model input, it better captures the impact of environmental factor fluctuations on fishing ground spatial distribution. The model training and prediction methods are scientifically sound, computationally efficient, and can quickly respond to the actual needs of fisheries production, reducing production costs. This invention is applicable to different sea areas and different target fishery resources, possessing good applicability and broad promotional value. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the short-to-medium-term fishing ground spatial distribution forecasting method based on deep learning, as described in an embodiment of the present invention.

[0015] Figure 2 This is a diagram of the deep learning U-Net model architecture according to an embodiment of the present invention.

[0016] Figure 3 This is a comparison chart of the performance of the superior model in the embodiments of the present invention under different time scales and advance time steps.

[0017] Figure 4 This is a diagram showing the correspondence between the predicted central fishing grounds and the actual production data under the optimal solution of this invention (white represents the central fishing grounds, black represents non-central fishing grounds, and colored dots represent the actual yield). Detailed Implementation

[0018] 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 some embodiments of the present invention, not all embodiments. 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.

[0019] like Figure 1 As shown, this invention provides a short- to medium-term fishing ground spatial distribution forecasting method based on deep learning, comprising the following steps: S101. Obtain historical fishery production data and corresponding historical environmental factor data for specific species in the target sea area, and perform preprocessing.

[0020] This invention selected the Northwest Pacific region (longitude 145°E–165°E, latitude 36°N–48°N) as the study area, and the species used was the Northwest Pacific squid (Ommastrephes bartramii). Fisheries production data were obtained from fishing ground catch data from July to November 2020, 1998. Environmental data were selected from diurnal sea surface temperature, sea surface height (SSH), sea surface salinity (SSS), chlorophyll content (Chla), and ocean current velocities (u and v) provided by NOAA Oceanwatch within the same spatiotemporal range.

[0021] Historical environmental factor data are standardized to uniformly normalize the environmental data to the [0,1] interval. Invalid values ​​are marked (marked as -1) to construct the input data for the model.

[0022] S102. Establish multiple datasets of fishery production data and multiple datasets of environmental factor data according to different time scales and advance times, respectively, based on the historical fishery production data and historical environmental factor data.

[0023] This invention employs various time scales, including 3 days, 5 days, 6 days, 7 days, 10 days, 15 days, 30 days, and 90 days, with a uniform spatial scale of 0.25° × 0.25°. The advance timing schemes include 2, 3, 4, ..., 15 time steps. Alternatively, a combination of the aforementioned time scales and advance timings can be selected for verification.

[0024] First, historical fishery production data and historical environmental factor data are interpolated and resampled according to the selected time scale and spatial scale, respectively, to ensure that the environmental data and fishery data have the same spatiotemporal resolution.

[0025] Then, based on the combination of time scale and advance time, rasterization and unified processing are performed to establish multiple datasets of fishery production data and multiple datasets of environmental factor data.

[0026] S103. A deep learning network for fishery prediction is constructed based on the U-Net model. Its encoding part includes several downsampling layers, and a SpatialDropout2D layer is added after at least the last downsampling layer. An attention mechanism is added after at least one downsampling layer. The decoding part includes several upsampling layers corresponding to the encoding part. The input of the deep learning network is environmental factor data, and the output is the predicted spatial distribution map of the fishery.

[0027] This invention constructs a deep learning network architecture for fishery prediction based on the U-Net model. The number of upsampling and downsampling layers is 2 to 5. At least one SpatialDropout2D layer is added after the last downsampling layer. When multiple SpatialDropout2D layers exist, they are added sequentially from the last downsampling layer upwards. The shallower the layer, the lower the Rate, ranging from 0.5 to 0.8, with a threshold of 0.05 for each layer. Simultaneously, an attention mechanism is added after at least one downsampling layer. When a downsampling layer contains both a SpatialDropout2D layer and an attention mechanism, the downsampling layer first passes through the attention mechanism before connecting to the SpatialDropout2D layer.

[0028] The deep learning network takes environmental factor data as input and outputs a predicted spatial distribution map of fishing grounds.

[0029] like Figure 2 As shown, this paper introduces a deep learning network using a 3-layer downsampling encoding part and a 3-layer upsampling decoding part as an example. The downsampling layer consists of two convolutional layers, an activation function layer, and one max pooling layer. The convolutional kernel size in the convolutional layer is 3×3, with a step size of 1. The number of convolutional kernels in the first convolution is adjusted to 16. After the first convolution, the number of feature channels of the sample changes from 1 to 16. The activation function is the ReLU function, which mainly aims to add nonlinearity to the output of the convolutional layer and enhance the learning of nonlinear features. The max pooling layer has a filter size of 2×2 and a step size of 2. After pooling, the sample size is reduced to 1 / 2×1 / 2, while the number of feature channels remains unchanged. The main purpose is to compress features, simplify network complexity, reduce parameters and computational cost, increase the receptive field of convolution, achieve scale transformation invariance of detected features, and improve the network's robustness to noise and clutter. To address the overfitting issue, SpatialDropout2D layers were added to the last convolutional layer after the second and third downsampling operations. An attention mechanism layer was also added between the last convolutional layer after the second downsampling operation and the SpatialDropout2D layer. In the latter part, symmetrical upsampling was performed three times. Because the model is dealing with a binary classification problem, the activation function was replaced with a sigmoid function.

[0030] S104. Input the various datasets of environmental factor data from step S102 into the deep learning network for training and evaluation, and select the combination of time scale and advance time with the best performance as the final model input method.

[0031] To reduce the amount of validation required, this embodiment of the invention uses 4 time scales and 7 advance times to design 28 prediction schemes for model training and evaluation, as shown in Table 1.

[0032] Table 1 Data from 1998 to 2019 was used as the model training and validation set, with 80% of the data used for training and 20% for validation and hyperparameter tuning. Data from 2020 was used as the test set for final model performance evaluation. Cross-entropy was used as the loss function for model training, optimized using the Adam optimizer with an initial learning rate of 0.001 and a training epoch count of 300. Early stopping was used to monitor the validation set loss and prevent overfitting.

[0033] The corresponding CPUE, Catch, effort, or HSI area is calculated based on various datasets of fishery production data. The spatial distribution map of the fishing grounds is obtained as the training ground using any one of the following methods: quartile division, median, mean, or specific threshold.

[0034] Deep learning network evaluation includes accuracy, precision, recall, F1 score, ROC area, and AUC under the PR curve. This invention provides a comprehensive evaluation using overall accuracy (OA), precision, recall, and F1 score, calculated as follows: ×100% Precision : Recall : = in, N TP (TP stands for true positive) indicates the number of true central fishing grounds that are predicted to be central fishing grounds, i.e., the number of true samples; N TN (TN stands for true negative) represents the number of true non-central fishing grounds that are predicted as non-central fishing grounds, i.e., the number of true negative samples; N FP (FP stands for false positive) indicates the number of true non-central fishing grounds that are predicted as central fishing grounds, i.e., the number of false positive samples; N FN(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. The results show that the scheme with a 15-day timescale and an advance prediction of 4 moments (Scheme 12) performs best, achieving an overall accuracy of 87.98% on the test set and an F1 score of 0.8833 (e.g., ...). Figure 3 (As shown).

[0035] The optimal forecasting scheme (15-day scale, 4 time advance) was applied to the spatial forecasting of the central squid fishing grounds in 2020, and the results are shown in Table 2.

[0036] Table 2 As can be seen from the table above, the prediction accuracy rate remained stable above 80% for each period, with the highest accuracy rate in the first half of July (OA at 93.43% and F1 score at 0.9545). Overall, the prediction results effectively support the formulation of plans for deep-sea squid fishing operations.

[0037] S105. Obtain environmental factor data at the advance time according to the best time scale and advance time combination, input it into the trained deep learning network, and predict the current spatial distribution of the fishing grounds.

[0038] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A short- to medium-term fishing ground spatial distribution forecasting method based on deep learning, characterized in that, Includes the following steps: (1) Obtain historical fishery production data and corresponding historical environmental factor data for specific species in the target sea area, and perform preprocessing; (2) Establish multiple datasets of fishery production data and multiple datasets of environmental factor data according to different time scales and advance times, respectively; (3) A deep learning network for fishery prediction is constructed based on the U-Net model. Its encoding part includes several downsampling layers, and a SpatialDropout2D layer is added after at least the last downsampling layer. An attention mechanism is added after at least one downsampling layer. The decoding part includes several upsampling layers corresponding to the encoding part. The input of the deep learning network is environmental factor data, and the output is the predicted spatial distribution map of the fishery. (4) Input the environmental factor data from step (2) into the deep learning network for training and evaluation, and select the best-performing time scale and advance time combination as the final model input method; (5) Obtain environmental factor data at the advance time according to the best time scale and advance time combination, input it into the trained deep learning network, and predict the current spatial distribution of the fishing ground.

2. The method for short- to medium-term spatial distribution forecasting of fishing grounds based on deep learning according to claim 1, characterized in that: The fisheries production data includes date, longitude, latitude, daily yield, and fishing vessels in operation; the environmental factor data includes one or more combinations of sea surface temperature (SST), sea surface height (SSH), sea surface salinity (SSS), chlorophyll (Chla), ocean current velocity (u), and (v), with at least sea surface temperature (SST).

3. The method for short- to medium-term spatial distribution forecasting of fishing grounds based on deep learning according to claim 1, characterized in that, The preprocessing specifically involves normalizing the historical environmental factor data and marking invalid values.

4. The method for short- to medium-term spatial distribution forecasting of fishing grounds based on deep learning according to claim 1, characterized in that: The different time scales include 3 days, 5 days, 6 days, 7 days, 10 days, 15 days, 30 days, and 90 days, and the spatial scale is uniformly set to 0.25°×0.25°. The advance timing scheme includes 2 to 15 time steps.

5. The method for short- to medium-term spatial distribution forecasting of fishing grounds based on deep learning according to claim 1, characterized in that: When there are multiple SpatialDropout2D layers, the layers are added sequentially from the last downsampling layer upwards. The shallower the layer, the lower the Rate. The Rate ranges from 0.5 to 0.8, and is increased with a threshold of 0.

05.

6. The method for short- to medium-term spatial distribution forecasting of fishing grounds based on deep learning according to claim 1, characterized in that: When a downsampling layer contains both a SpatialDropout2D layer and an attention mechanism, the downsampling first passes through the attention mechanism and then connects to the SpatialDropout2D layer.

7. The method for short- to medium-term spatial distribution forecasting of fishing grounds based on deep learning according to claim 1, characterized in that: The deep learning network uses cross-entropy as the loss function during training.

8. The method for short- to medium-term spatial distribution forecasting of fishing grounds based on deep learning according to claim 1, characterized in that: The corresponding CPUE, Catch, effort, or HSI area is calculated based on various datasets of fishery production data. The spatial distribution map of the fishing grounds is obtained as the training ground using any one of the following methods: quartile division, median, mean, or specific threshold.

9. The method for short- to medium-term spatial distribution forecasting of fishing grounds based on deep learning according to claim 1, characterized in that: The evaluation of the deep learning network includes accuracy, precision, recall, F1 score, ROC area, and AUC under the PR curve.

Images