Deep neural network-based ocean species spatial distribution prediction method
By training a deep neural network model and combining it with multi-source environmental data, the shortcomings of traditional models in predicting at fine time scales are solved, and high-timeliness and high-precision fishing ground prediction are achieved. In particular, the robustness and predictive stability of the model are improved in time periods with sparse data or many zero values.
Patent Information
- Application Number
- CN202511102828.2
- 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
Traditional species distribution models struggle to make efficient spatial distribution predictions at fine time scales, failing to meet the demands of high-efficiency operations. Furthermore, existing deep learning models are mostly geared towards classification or optimal scale selection, failing to provide precise predictions.
By training a deep neural network model using multi-source environmental variables and fishing data, and through multi-layer hidden layers and large-scale parameter learning, data features are automatically extracted for end-to-end learning. The model is then combined with multi-source environmental data to predict the spatial distribution of pelagic species.
It achieves high-precision fishing ground prediction on different time scales from 3 days to 1 year, reduces prediction error, and improves the robustness and stability of the model. In particular, it reduces the phenomenon of over-smoothing or underreporting of high-value areas in time periods with sparse data or many zero values.
Smart Images
Figure CN120952245A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of marine fishery resource prediction and ecological modeling technology, and in particular to a method for predicting the spatial distribution of oceanic species based on deep neural networks. Background Technology
[0002] The location of fishing grounds for pelagic species directly impacts yields and the effectiveness of resource management. Traditional species distribution models are mostly based on annual or monthly statistical relationships, making it difficult to predict spatial distribution at precise time scales, and therefore unable to meet the needs of high-efficiency operations.
[0003] In recent years, deep learning has become a research hotspot in the field of artificial intelligence, demonstrating significant advantages in handling complex nonlinear relationships and big data. By introducing multiple hidden layers and large-scale parameter learning, deep neural networks can automatically extract deep features from data, supporting end-to-end learning frameworks. Currently, deep learning technology has been applied in fields such as marine remote sensing and fisheries monitoring, including fish species identification, fishing vessel trajectory tracking, and fishing ground classification. Especially in fishing ground prediction, deep learning has the potential to overcome the limitations of traditional models on high-dimensional nonlinear data. However, existing deep learning-based fishing ground models primarily focus on classification or optimal scale selection, and have not yet made detailed predictions of species distribution probability values. Therefore, applying deep neural networks to spatial distribution prediction at fine temporal scales and integrating multi-source environmental data for modeling has significant innovative implications. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention utilizes multi-source environmental variables and fishing data to train a deep neural network model across multiple time windows, thereby achieving high-precision prediction of the spatial distribution of pelagic species in the study area. This overcomes the limitations of existing technologies in fine-scale time-scale prediction, complex nonlinear relationship mining, and processing of "existence-only" fishing data.
[0005] To achieve the above objectives, this invention provides a method for predicting the spatial distribution of oceanic species in the southeastern Pacific Ocean based on a deep neural network model, comprising the following steps: (1) Obtain historical fishing log data of pelagic species and environmental factor data of the target sea area, and preprocess them; (2) Aggregate the historical fishing log data and environmental factor data according to the selected time scale to obtain the catch amount per unit fishing effort (CPUE), the fishing effort of the fishing grounds in the AIS system, and the average environmental factor of each grid within the corresponding time window. (3) Generate longitude, latitude, sine time features, cosine time features, and normalized environmental factor features for each aggregated sample; (4) Build a fully connected neural network model. The number of input nodes is determined according to the feature dimension obtained in step (3). The model includes an input layer, at least two hidden layers and an output layer. The number of hidden layer nodes decreases or increases and then decreases layer by layer. The ReLU activation function is used. (5) Use the data obtained in step (3) to train the model and obtain a trained model for predicting the spatial distribution of oceanic species in the target time period.
[0006] Further environmental factor data include: sea surface temperature (SST), sea surface height (SSH), sea surface salinity (SSS), photosynthetically active radiation (PAR), multi-layer water temperature structure, ocean current velocities (u and v), primary productivity (nppv), and dissolved oxygen (O2).
[0007] Furthermore, the preprocessing includes: (1.1) Resample all data to a uniform grid resolution of 0.25°×0.25°; (1.2) Delete or correct outliers in environmental factor data and erroneous records in historical fishing log data; (1.3) For cases where only recorded data exists, pseudo-missing samples are generated. The pseudo-missing samples are generated by randomly sampling unsampled grid areas and marking CPUE as zero, or by using a classifier to identify blank areas.
[0008] Furthermore, the time scale includes 3 days, 6 days, 10 days, 15 days, 30 days, 90 days, and one year, and the time scale selected in step (2) includes at least 3 days, 6 days, 10 days, 15 days, and 30 days.
[0009] Furthermore, the sinusoidal time characteristics and cosine time characteristics are respectively: Time_sin=sin(2πt / P) Time_cos=cos(2πt / P) Where: t is the cycle number of the aggregated sample in a year, and P is the number of time windows in a year.
[0010] Furthermore, the fully connected neural network model has 2 to 5 hidden layers.
[0011] Furthermore, the fully connected neural network model is trained using a GPU-accelerated deep learning framework, employing the Adam optimization algorithm and an early stopping strategy, with mean squared error as the loss function.
[0012] Furthermore, the training process of the fully connected neural network model is set with a learning rate of 0.001 and a batch size of 128, 256, or 512, and early stopping is triggered when the verification loss does not decrease for 10 consecutive iterations.
[0013] Furthermore, the fully connected neural network model is evaluated using mean squared error (MSE), mean absolute error (MAE), and area under the precision-recall curve (PR-AUC).
[0014] Furthermore, the fully connected neural network model is used to generate a predicted heatmap of CPUE for each grid in the research sea area, and outputs the latitude and longitude coordinates and abundance level of high CPUE hotspot areas.
[0015] The beneficial effects of this invention are: (1) It supports fishery forecasting analysis at different time scales from 3 days to 1 year, filling the gap that traditional models are mostly limited to monthly scales, and achieving more timely fishery forecasts.
[0016] (2) The proposed DNN model can automatically extract complex nonlinear relationship features and has lower prediction error compared with traditional generalized additive model (GAM), gradient boosting decision tree (XGBoost) and shallow neural network (ANN).
[0017] (3) Through careful design of input variables (introducing latitude, longitude and periodic time features) and cleaning of training data (removing unsampled noise), the model has better robustness. Especially in fine-grained time periods with sparse data and a large number of zero values, the model can still maintain good prediction stability and reduce the phenomenon of over-smoothing or underreporting of high-value areas. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the method for predicting the spatial distribution of oceanic species in the southeastern Pacific Ocean based on a deep neural network model, according to an embodiment of the present invention.
[0019] Figure 2 This is a schematic diagram of the deep neural network model structure described in an embodiment of the present invention.
[0020] Figure 3 This is a comparison chart of model performance at different time scales according to an embodiment of the present invention.
[0021] Figure 4 This is a schematic diagram comparing the predicted spatial distribution of stem squid using embodiments of the present invention with actual observations (where the dark areas represent stem squid fishing grounds with high CPUE). Detailed Implementation
[0022] The present invention will be further explained and described below with reference to the accompanying drawings and embodiments.
[0023] like Figure 1 As shown, this invention provides a method for predicting the spatial distribution of pelagic species based on deep neural networks, comprising the following steps: S101. Obtain historical fishing log data of oceanic species and environmental factor data of the target sea area, and preprocess them.
[0024] This invention uses the stem squid (Dosidicus gigas) as an example. The study area was selected as 95–75°W and 20–8°S, and divided into 0.25×0.25° grid cells, resulting in 3840 grids. Fishing logs and related environmental data were collected daily from 2012 to 2021.
[0025] Historical fishing log data includes date, geographical longitude and latitude, daily catch, and number of fishing vessels in operation.
[0026] Environmental data include a variety of marine environmental factors closely related to the habitat of oceanic species, including sea surface temperature (SST), sea surface height (SSH), sea surface salinity (SSS), photosynthetically active radiation (PAR), ocean current velocities (u and v), primary productivity (nppv), multi-layer water temperature structure (5m, 10m, 15m, 20m, 50m, 100m), and dissolved oxygen (O2).
[0027] Environmental data covering the same timeframe as historical fishing log data (2012–2021) were sampled daily and obtained through remote sensing observations or ocean reanalysis data. Specifically, the SST data had a spatial resolution of approximately 0.05°, the SSH and SSS approximately 0.25°, and the PAR approximately 4 km.
[0028] Preprocessing includes: (1) All data were uniformly resampled to a grid resolution of 0.25°×0.25° by bilinear interpolation.
[0029] (2) Delete or correct outliers in environmental factor data and erroneous records in historical fishing log data.
[0030] (3) For cases where only recorded data exists, pseudo-missing samples are generated. The pseudo-missing samples are generated by randomly sampling unsampled grid areas and marking CPUE as zero, or by using a classifier to identify blank areas.
[0031] S102. Aggregate the historical fishing log data and environmental factor data according to the selected time scale to obtain the catch amount per unit fishing effort (CPUE), the fishing effort of the fishing grounds and the average value of environmental factors of each grid within the corresponding time window.
[0032] This invention defines multiple time scale windows for different forecasting needs, including 3 days, 6 days (week), 10 days (ten-day period), 15 days (half-month), 30 days (month), 90 days (quarter), and one year. The time scales selected in this invention include at least 3 days, 6 days, 10 days, 15 days, and 30 days.
[0033] For each time scale, the original daily data is segmented and aggregated according to the period, and the average environmental variable value of each grid in each time window is calculated, as well as the corresponding catch-per-unit effort (CPUE) index and the fishing effort of the AIS system.
[0034] The formula for calculating CPUE is: CPUE = Catch / Effort Where: Catch is the cumulative catch (tons) during the time period, and Effort is the number of days the fishing vessel operated.
[0035] Through the above calculations, the CPUE value matrix of the study sea area under each time window is obtained, consisting of 48×80 grids (0.25° resolution covering the specified latitude and longitude range). As the time window shortens, the number of periods increases, resulting in more data entries; for example, at a 3-day scale, approximately 1200 time periods are divided from 2012 to 2021, with approximately 4.608×10⁶ data entries, while at a 30-day (monthly) scale, there are approximately 120 time periods, with approximately 4.608×10⁵ data entries.
[0036] S103. Generate longitude, latitude, sine time features, cosine time features, and normalized environmental factor features for each aggregated sample.
[0037] The sine time characteristics and cosine time characteristics are as follows: Time_sin=sin(2πt / P) Time_cos=cos(2πt / P) Where: t is the cycle number of the aggregated sample in a year, and P is the number of time windows in a year.
[0038] All environmental variable data are normalized (mapped to the 0-1 range), and outliers or missing values are replaced with -1 to improve the convergence efficiency and robustness of model training.
[0039] S104. Construct a fully connected neural network model. The number of input nodes is determined according to the feature dimensions obtained in step S103. The model includes an input layer, at least two hidden layers, and an output layer. The number of hidden layer nodes decreases layer by layer or increases first and then decreases. The ReLU activation function is used.
[0040] A fully connected neural network model was constructed to fit the relationship between input features and CPUE output, which was used to generate predicted heatmaps of CPUE for each grid in the study area, and output the latitude and longitude coordinates and abundance levels of high CPUE hotspot areas.
[0041] The fully connected neural network model used in this embodiment of the invention is a deep fully connected feedforward neural network (i.e., a DNN model). The model structure includes: an input layer, corresponding to the feature vectors constructed above; several hidden layers, adopting a fully connected architecture; and an output layer.
[0042] The number of input nodes in the input layer is determined based on the feature dimensions obtained in step S103, and includes all or some of the feature dimensions.
[0043] The hidden layers consist of 2 to 5 hidden layers, with the number of nodes decreasing layer by layer or increasing first and then decreasing. The number of hidden layer nodes can be 8, 16, 32, 64, 128, 256, or 512. All neurons in the hidden layers use the Rectified Linear Activation Function (ReLU) to ensure non-negative outputs while mitigating the vanishing gradient problem.
[0044] The output layer consists of a single neuron that directly outputs the predicted CPUE value.
[0045] The number of layers and neurons per layer can be adjusted according to the actual data scale and complexity; this invention is not limited to a specific structure. When the input dimension changes, the size of the hidden layers can be adjusted accordingly to maintain the consistency of the model's fitting ability. Compared with traditional shallow neural networks (ANNs) with only one hidden layer, the DNN model of this invention can extract more abstract feature patterns by increasing network depth, thereby enhancing the model's ability to represent complex nonlinear relationships and its generalization performance.
[0046] like Figure 2 As shown, this embodiment of the invention preferably uses three hidden layers, with each layer of the fully connected neural network model having 8, 32, 16, 8, and 1 nodes. The total number of parameters in the model is approximately (8×32 + 32×16 + 16×8 + 8×1) ≈ 1040. In contrast, a traditional single-hidden-layer ANN model requires more than 100 neurons in its hidden layer to achieve a similar parameter scale. Therefore, this embodiment of the DNN improves feature extraction capabilities through its deep structure while maintaining a comparable number of parameters.
[0047] S105. Use the data obtained in step S103 to train the model and obtain a trained model, which is used to predict the spatial distribution of oceanic species in the target time period.
[0048] Model training is performed on a GPU-accelerated environment (such as an NVIDIA RTX 2080Ti or equivalent computing device) using a deep learning framework (such as TensorFlow 2.4.1). Adaptive optimization algorithms (such as Adam or SGD) are employed during training to minimize the error between predicted and true values. The loss function can use regression metrics such as mean squared error (MSE) or mean absolute error (MAE). Considering the sparse distribution and high number of zero values in CPUE data, this embodiment of the invention may also introduce the area under the precision-recall curve (PR curve AUC) as an auxiliary metric during model evaluation to better measure the model's ability to capture rare events (high CPUE regions). The fully connected neural network model training process is set with a learning rate of 0.001, batch sizes of 128, 256, or 512, and early stopping is triggered when the validation loss does not decrease for 10 consecutive iterations.
[0049] The deep neural network model is trained and optimized using historical data. Preferably, data from 2012 to 2020 is used as the training and validation sets, while data from 2021 is reserved as an independent test set to evaluate the model's generalization performance. Before training, the constructed sample dataset is randomly divided into training and validation sets, for example, using a 4:1 cross-validation tuning ratio. The 2021 samples are used as the test set to evaluate the model's performance. Taking a 30-day scale dataset as an example, the number of training samples is approximately 120 segments × 3840 grids ≈ 460,000 records. In contrast, the number of training samples on a 3-day scale can be as high as approximately 4.6 million records. To address the data imbalance problem, samples with catches and blank samples are sampled at different ratios during training; simultaneously, an early stopping strategy and regularization techniques are introduced to prevent overfitting. Training stops when the validation set loss no longer decreases, and the trained model parameters for each scale are finally saved.
[0050] During the testing phase, data from various scales in 2021 were input into the corresponding models to predict the spatial distribution of *Schizothorax chinensis* in the study area. Taking the 30-day scale model as an example, the test results were compared with actual observations (e.g., ...). Figure 3(As shown). The results show that the DNN model basically captures the outline of the high CPUE area and its monthly variation trend. For example, August to November is the high-yield season, and the high CPUE fishing grounds predicted by the model are mainly located in the waters off Peru, which highly matches the actual high-catch areas. In months with low catches (such as February), the range of the predicted fishing grounds output by the model is correspondingly smaller. In some months, the high-value centers predicted by the model have a shift of about 1-2 grids relative to the actual measurements, but the overall trend is correct. Quantitative evaluation results of each model show that at the 30-day scale, the DNN model has a mean squared error (MSE) of ≈16 and a mean absolute error (MAE) of ≈2.7, both of which are the best among the four models; at the 3-day scale, the error is improved (MSE about 30, MAE ≈3.9), but the DNN still maintains the lowest error. The AUC index of the Precision-Recall curve also shows that the DNN model has the best overall performance in terms of prediction accuracy and recall for high CPUE grids across all time scales. These results verify the effectiveness and superiority of the method of this invention.
[0051] 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 method for predicting the spatial distribution of pelagic species based on deep neural networks, characterized in that, Includes the following steps: (1) Obtain historical fishing log data of pelagic species and environmental factor data of the target sea area, and preprocess them; (2) Aggregate the historical fishing log data and environmental factor data according to the selected time scale to obtain the catch amount per unit fishing effort (CPUE), the fishing effort of the fishing grounds in the AIS system, and the average environmental factor of each grid within the corresponding time window. (3) Generate longitude, latitude, sine time features, cosine time features, and normalized environmental factor features for each aggregated sample; (4) Build a fully connected neural network model. The number of input nodes is determined according to the feature dimension obtained in step (3). The model includes an input layer, at least two hidden layers and an output layer. The number of hidden layer nodes decreases or increases and then decreases layer by layer. The ReLU activation function is used. (5) Use the data obtained in step (3) to train the model and obtain a trained model for predicting the spatial distribution of oceanic species in the target time period.
2. The spatial distribution prediction method for pelagic species based on deep neural networks according to claim 1, characterized in that, Environmental factor data include: sea surface temperature (SST), sea surface height (SSH), sea surface salinity (SSS), photosynthetically active radiation (PAR), multi-layer water temperature structure, ocean current velocities (u and v), primary productivity (nppv), and dissolved oxygen (O2).
3. The method for predicting the spatial distribution of pelagic species based on deep neural networks according to claim 1, characterized in that, The preprocessing includes: (1.1) Resample all data to a uniform grid resolution of 0.25°×0.25°; (1.2) Delete or correct outliers in environmental factor data and erroneous records in historical fishing log data; (1.3) For cases where only recorded data exists, pseudo-missing samples are generated. The pseudo-missing samples are generated by randomly sampling unsampled grid areas and marking CPUE as zero, or by using a classifier to identify blank areas.
4. The method for predicting the spatial distribution of pelagic species based on deep neural networks according to claim 1, characterized in that: The time scales include 3 days, 6 days, 10 days, 15 days, 30 days, 90 days, and one year. The time scale selected in step (2) includes at least 3 days, 6 days, 10 days, 15 days, and 30 days.
5. The method for predicting the spatial distribution of pelagic species based on deep neural networks according to claim 1, characterized in that, The sinusoidal time characteristics and cosine time characteristics are respectively: Time_sin=sin(2πt / P) Time_cos=cos(2πt / P) Where: t is the cycle number of the aggregated sample in a year, and P is the number of time windows in a year.
6. The method for predicting the spatial distribution of pelagic species based on deep neural networks according to claim 1, characterized in that: The fully connected neural network model has 2 to 5 hidden layers.
7. The method for predicting the spatial distribution of pelagic species based on deep neural networks according to claim 1, characterized in that: The fully connected neural network model is trained using a GPU-accelerated deep learning framework, employing the Adam optimization algorithm and an early stopping strategy, with mean squared error as the loss function.
8. The method for predicting the spatial distribution of pelagic species based on deep neural networks according to claim 6, characterized in that: The training process of the fully connected neural network model is set with a learning rate of 0.001 and a batch size of 128, 256 or 512, and early stopping is triggered when the verification loss does not decrease for 10 consecutive iterations.
9. The method for predicting the spatial distribution of pelagic species based on deep neural networks according to claim 1, characterized in that: The fully connected neural network model was evaluated using mean squared error (MSE), mean absolute error (MAE), and area under the precision-recall curve (PR-AUC).
10. The method for predicting the spatial distribution of pelagic species based on deep neural networks according to claim 1, characterized in that: The fully connected neural network model is used to generate a predicted heatmap of CPUE for each grid in the study area, and outputs the latitude and longitude coordinates and abundance levels of high CPUE hotspot areas.