Fishing situation prediction method and system based on multi-source heterogeneous data

Abstract

The application relates to the technical field of data processing, and discloses a fishing situation prediction method and system based on multi-source heterogeneous data. The method comprises the following steps: acquiring heterogeneous data such as sea surface temperature, salinity, fishing boat trajectory and sea current, and converting the data into a space-time data matrix; extracting environmental parameters to calculate a water body layering index and a nutrient salt enrichment index, and combining the indexes to form a fishery feature vector; inputting the four-layer BP neural network to train and establish a prediction model; calculating a space-time distance to distribute weights and generate a correlation prediction result; adding real-time data to form a dynamic training set, detecting environmental changes, updating model parameters and outputting a prediction result. The application solves the problem that multi-source heterogeneous marine data cannot be effectively fused and intelligently processed, and improves the accuracy and real-time adaptability of fishing situation prediction.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for predicting fishery conditions based on multi-source heterogeneous data. Background Technology

[0002] Traditional fishery forecasting methods primarily rely on fishermen's experience and single-sensor monitoring. They predict fishing ground location and catch volume by analyzing historical catch records, seasonal patterns of marine environmental changes, and simple statistical analyses. Existing technologies include independent data collection methods such as sea surface temperature monitoring based on satellite remote sensing, water quality parameter measurements based on ocean buoys, and catch statistics based on fishing vessel reports. These methods provide, to some extent, a reference for fishery production.

[0003] However, existing technologies have significant shortcomings: First, the data sources are singular and isolated, and there is a lack of effective data fusion mechanisms between various monitoring systems, making it difficult to reflect the diverse correlations and complexity of the marine environment; second, the heterogeneous data processing capabilities are insufficient, and it is difficult to uniformly process and comprehensively analyze marine environmental data, meteorological data, fishing vessel trajectory data, etc., in different formats and at different spatiotemporal scales, resulting in poor generalization of prediction models; third, there is a lack of intelligent feature extraction and mapping learning capabilities, making it impossible to effectively capture the nonlinear and complex relationship between environmental factors and fishing conditions.

[0004] Because fishery forecasting not only needs to consider the environmental conditions of the current location, but also needs to analyze the influence patterns of the surrounding area and historical time, existing technologies cannot establish an effective spatiotemporal weight allocation mechanism; at the same time, static forecasting models cannot adapt to the dynamic changes in the marine environment. When the marine ecosystem undergoes sudden changes, traditional models lack online learning and adaptive adjustment capabilities, making it difficult to maintain forecast accuracy and timeliness. Summary of the Invention

[0005] This application provides a method and system for fishery condition prediction based on multi-source heterogeneous data, which solves the problem of ineffective fusion and intelligent processing of multi-source heterogeneous marine data, and improves the accuracy and real-time adaptability of fishery condition prediction.

[0006] Firstly, this application provides a method for predicting fishery conditions based on multi-source heterogeneous data, the method comprising:

[0007] Step S101: Acquire sea surface temperature data, salinity data, fishing vessel trajectory data, and ocean current data, and convert the heterogeneous data into a spatiotemporal data matrix;

[0008] Step S102: Extract water temperature, salinity, and chlorophyll concentration from the spatiotemporal data matrix, calculate the water stratification index and nutrient enrichment index, and combine them with seasonal and location information to obtain a fishery feature vector.

[0009] Step S103: Input the fishery feature vector into a four-layer BP neural network for training, learn the mapping relationship between environmental factors and catch, and establish a fishery situation prediction model.

[0010] Step S104: Calculate the spatiotemporal distance between the predicted location and the observation point, assign influence weights to each observation data, and generate spatiotemporal correlation prediction results;

[0011] Step S105: Add real-time data to the spatiotemporal correlation prediction results to obtain a dynamic training set. When environmental pattern changes are detected, update the model parameters and output the fishing situation prediction results.

[0012] Optionally, step S101 includes:

[0013] Sea surface temperature data is collected by satellite remote sensing system, and the temperature values ​​are digitally encoded with high precision and short time intervals to obtain a sea surface temperature time series dataset.

[0014] Salinity and chlorophyll concentration data were collected by an ocean buoy network, and the vertical profile was sampled in layers at equal depth intervals to obtain a vertical dataset of the marine environment.

[0015] Fishing vessel trajectory data is collected by the fishing vessel positioning system, and latitude and longitude coordinates and catch records are associated and matched according to timestamps to obtain a fishery operation dataset.

[0016] The sea surface temperature time series dataset, marine environment vertical dataset, and fishery operation dataset are subjected to coordinate system transformation and time standardization to obtain a spatiotemporal data matrix.

[0017] Optionally, step S102 includes:

[0018] Water temperature, salinity, and chlorophyll concentration parameters are extracted from the spatiotemporal data matrix. The difference between water temperature and water depth data is calculated to obtain the water stratification index.

[0019] The nutrient enrichment index is obtained by multiplying the chlorophyll concentration and dissolved oxygen content and then dividing by the product of water temperature and salinity.

[0020] Seasonal information is encoded using periodic functions, and monthly and tidal information are converted into combinations of sine and cosine functions to obtain time feature codes. The water stratification index, nutrient enrichment index, time feature codes, and location information are then concatenated into vectors to obtain fishery feature vectors.

[0021] Optionally, step S103 includes:

[0022] The fishery feature vector is input into the input layer of a four-layer BP neural network. The input layer contains 58 neuron nodes that receive feature data, and the output signal of the input layer is obtained.

[0023] The input layer output signal is passed to the first hidden layer, which contains 128 neurons. The input signal is processed by the ReLU activation function to obtain the first layer feature map.

[0024] The first layer feature map is passed to the second hidden layer, which contains 64 neurons. The Tanh activation function is used to perform high-order feature combination processing to obtain the second layer feature map.

[0025] The number of neurons in the output layer is set according to the type of prediction task. One neuron is set for the catch prediction task, ten neurons are set for the fish species prediction task, and eight neurons are set for the fishing area prediction task, thus obtaining the output layer structure configuration.

[0026] The second layer feature map is linearly combined with the output layer weight matrix, and then processed by the Sigmoid or Softmax activation function to obtain the predicted output value.

[0027] The error of the loss function is calculated based on the predicted output value and the actual catch data. The weight parameters of each layer are updated through the backpropagation algorithm. The training is repeated until the loss function converges, and the fishery prediction model is obtained.

[0028] Optionally, step S104 includes:

[0029] The spatial distance matrix is ​​obtained by calculating the Euclidean distance between the predicted location of the fishery situation prediction model and the latitude and longitude coordinates of the historical observation points.

[0030] The time difference between the predicted location and the historical observation point is processed by an exponential decay function to obtain the time correlation matrix.

[0031] A weighted combination operation is performed based on the spatial distance matrix and the temporal correlation matrix, with the spatial weight coefficient set to 0.6 and the temporal weight coefficient set to 0.4, to obtain the spatiotemporal correlation matrix. An influence weight is assigned to each historical observation data according to the spatiotemporal correlation matrix, and the weight values ​​are weighted and summed with the corresponding catch data to obtain the spatiotemporal correlation prediction result.

[0032] Optionally, step S105 includes:

[0033] The newly collected real-time sea surface temperature data, salinity data, fishing vessel trajectory data, and ocean current data are merged with the spatiotemporal correlation prediction results in chronological order to obtain a dynamic training set.

[0034] The difference between the current sample mean and the historical sample mean in the dynamic training set is calculated and divided by the standard deviation. When the difference exceeds a set threshold, the model update mechanism is triggered to obtain the update trigger signal.

[0035] Based on the update trigger signal, the dynamic training set is input into the fishery condition prediction model, and the weight parameters of each layer of the neural network are adjusted by the gradient descent algorithm to obtain the updated model parameters.

[0036] The environmental data of the target sea area is input into the updated fishery forecasting model for calculation and processing, and the fishery forecasting results, including the catch value and confidence interval range, are output.

[0037] Optionally, the step of inputting the dynamic training set into the fishery condition prediction model based on the update trigger signal, and adjusting the weight parameters of each layer of the neural network using the gradient descent algorithm to obtain the updated model parameters includes:

[0038] The fishery feature vectors in the dynamic training set are paired and combined with the corresponding real catch data to construct training sample pairs and obtain online learning data pairs.

[0039] The online learning data is used to perform forward propagation calculations on the input four-layer BP neural network. The difference between the network output value and the actual catch is calculated to obtain the current error value.

[0040] Based on the current error value, the gradient information of the weights of each layer is calculated using the backpropagation algorithm. The gradient value is multiplied by the learning rate of 0.001 to perform weight update calculation and obtain the weight increment.

[0041] The weight increments are added to the original weight parameters of each layer of the BP neural network, and elastic weights are added to reinforce the regularization term to prevent the forgetting of historical knowledge, thus obtaining the updated model parameters.

[0042] Secondly, this application provides a fishery condition prediction system based on multi-source heterogeneous data, the fishery condition prediction system based on multi-source heterogeneous data comprising:

[0043] The data acquisition module is used to acquire sea surface temperature data, salinity data, fishing vessel trajectory data and ocean current data, and convert the above heterogeneous data into a spatiotemporal data matrix;

[0044] The feature extraction module is used to extract water temperature, salinity, and chlorophyll concentration from the spatiotemporal data matrix, calculate the water stratification index and nutrient enrichment index, and combine them with seasonal and location information to obtain fishery feature vectors.

[0045] The model training module is used to input fishery feature vectors into a four-layer BP neural network for training, learn the mapping relationship between environmental factors and catch, and establish a fishery condition prediction model.

[0046] The correlation prediction module is used to calculate the spatiotemporal distance between the predicted location and the observation point, assign influence weights to each observation data, and generate spatiotemporal correlation prediction results.

[0047] The dynamic update module is used to add real-time data to the spatiotemporal correlation prediction results to form a dynamic training set, update the model parameters when environmental pattern changes are detected, and output the fishing situation prediction results.

[0048] Thirdly, a fishery condition prediction device based on multi-source heterogeneous data is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the fishery condition prediction device based on multi-source heterogeneous data to execute the aforementioned fishery condition prediction method based on multi-source heterogeneous data.

[0049] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described method for predicting fishing conditions based on multi-source heterogeneous data.

[0050] The technical solution provided in this application solves the fundamental problems of isolated and inconsistent data formats in traditional fisheries forecasting by acquiring sea surface temperature, salinity, fishing vessel trajectory, and ocean current data and converting heterogeneous data into a spatiotemporal data matrix. This enables collaborative analysis of marine environmental information from different monitoring systems within a unified spatiotemporal coordinate system. The feature engineering method extracts water temperature, salinity, and chlorophyll concentration from the spatiotemporal data matrix and calculates water stratification and nutrient enrichment indices, transforming raw environmental parameters into composite indicators with fisheries ecological significance. This effectively captures key ecological factors influencing fish distribution and aggregation, providing stronger predictive targeting compared to traditional single-parameter analysis. The technical feature vector of fisheries is input into a four-layer BP neural network for training and learning the mapping relationship between environmental factors and catch. Through multi-layer nonlinear transformation, an intelligent correlation model between complex environmental conditions and fisheries status is established, overcoming the limitations of traditional empirical prediction and linear statistical methods. This approach can handle multi-factor interactions and nonlinear response relationships in marine ecosystems.

[0051] A technical solution was developed to calculate the spatiotemporal distance between the predicted location and the observation point, and to assign influence weights to each observation data point to generate spatiotemporally correlated prediction results. This established a model for spatial continuity and temporal continuity in fishery condition prediction, overcoming the shortcomings of traditional methods that neglect the influence of neighboring areas and historical periods. This makes the prediction results more consistent with the spatiotemporal evolution of the marine environment and fish behavior. An adaptive mechanism was implemented, which incorporates real-time data into the spatiotemporally correlated prediction results to obtain a dynamic training set and updates model parameters when environmental pattern changes are detected. This endows the fishery condition prediction system with the ability to continuously learn and dynamically optimize. When the marine environment experiences seasonal changes, climate anomalies, or ecological disturbances, the system can automatically adjust its prediction strategy to maintain the stability and timeliness of prediction accuracy. Particularly in the specific application field of fishery condition prediction, the multi-level feature abstraction capability of the four-layer BP neural network matches the hierarchical structure of the marine ecosystem. The network's input layer, hidden layer, and output layer correspond to the hierarchical processing of environmental data collection, ecological relationship modeling, and fishery condition prediction, respectively. This ensures that the algorithm architecture is highly compatible with the inherent logic of the application scenario, thereby achieving accurate, timely, and reliable fishery condition prediction results in complex and ever-changing marine environments. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of a fishery situation prediction method based on multi-source heterogeneous data according to this application;

[0054] Figure 2 This is a diagram illustrating a process for fishery situation prediction based on multi-source heterogeneous data, as described in this application.

[0055] Figure 3 This is a schematic diagram of the structure of a fishery situation prediction system based on multi-source heterogeneous data according to this application;

[0056] Figure 4 This is a schematic diagram of the structure of a fishery condition prediction device based on multi-source heterogeneous data according to this application. Detailed Implementation

[0057] This application provides a method and system for fishery situation prediction based on multi-source heterogeneous data. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the fishing situation prediction method based on multi-source heterogeneous data in this application includes:

[0059] Step S101: Acquire sea surface temperature data, salinity data, fishing vessel trajectory data, and ocean current data, and convert the heterogeneous data into a spatiotemporal data matrix;

[0060] Step S102: Extract water temperature, salinity, and chlorophyll concentration from the spatiotemporal data matrix, calculate the water stratification index and nutrient enrichment index, and combine them with seasonal and location information to obtain the fishery feature vector.

[0061] Step S103: Input the fishery feature vector into a four-layer BP neural network for training, learn the mapping relationship between environmental factors and catch, and establish a fishery situation prediction model.

[0062] Step S104: Calculate the spatiotemporal distance between the predicted location and the observation point, assign influence weights to each observation data, and generate spatiotemporal correlation prediction results;

[0063] Step S105: Add real-time data to the spatiotemporal correlation prediction results to obtain a dynamic training set. When environmental pattern changes are detected, update the model parameters and output the fishing situation prediction results.

[0064] It is understood that the executing entity of this application can be a fishery situation prediction system based on multi-source heterogeneous data, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0065] Specifically, sea surface temperature data is collected through satellite remote sensing systems, and the temperature values ​​are digitized and encoded with high precision and short time intervals to obtain a sea surface temperature time-series dataset. Salinity and chlorophyll concentration data are collected through an ocean buoy network, and vertical profiles are sampled at equal depth intervals to obtain a marine environmental vertical dataset. Fishing vessel trajectory data is collected through a fishing vessel positioning system, and latitude and longitude coordinates and catch records are correlated and matched according to timestamps to obtain a fisheries operation dataset. Ocean current data is also collected. Coordinate system transformation and time standardization are performed on the sea surface temperature time-series dataset, marine environmental vertical dataset, fisheries operation dataset, and ocean current data to unify the spatial reference and temporal granularity, resulting in a spatiotemporal data matrix. This process solves the problems of single and isolated data sources and inconsistent formats in traditional fisheries forecasting, enabling collaborative analysis of marine data from different monitoring systems within a unified spatiotemporal framework.

[0066] Then, water temperature, salinity, and chlorophyll concentration parameters are extracted from the spatiotemporal data matrix. The difference between water temperature and depth data is used to calculate the water stratification index. The nutrient enrichment index is obtained by multiplying chlorophyll concentration and dissolved oxygen content and then dividing by the product of water temperature and salinity. Seasonal information is encoded using periodic functions, and monthly and tidal information are converted into sine and cosine functions to obtain temporal feature codes. Location information is extracted from the spatiotemporal data matrix. The water stratification index, nutrient enrichment index, temporal feature codes, and location information are vector-concatenated to obtain a fisheries feature vector. This vector transforms the original environmental parameters into composite indicators meaningful for fisheries, addressing the lack of intelligent feature extraction capabilities and the inability to capture the nonlinear relationship between environmental factors and fisheries conditions.

[0067] Next, the fishery feature vector is input into the input layer of a four-layer BP neural network. The 58 neurons in the input layer receive the feature data and produce the input layer output signal. This output signal is then passed to the first hidden layer with 128 neurons, where a non-linear transformation using the ReLU activation function yields the first layer feature map. It is then passed to the second hidden layer with 64 neurons, where a higher-order feature combination using the Tanh activation function yields the second layer feature map. The number of neurons in the output layer is adjusted according to the prediction task type. The second layer feature map is linearly combined with the output layer weight matrix and processed using either the Sigmoid or Softmax activation function to obtain the predicted output value. The loss function error is calculated based on the predicted output value and actual catch data. The weight parameters of each layer are updated using the backpropagation algorithm. This process is repeated until the loss function converges, establishing a fishery prediction model. This model overcomes the limitations of traditional linear statistical methods and solves the problem of handling the complex non-linear relationship between environmental factors and fishery conditions.

[0068] Furthermore, the spatial distance matrix is ​​obtained by calculating the Euclidean distance between the predicted location of the fishery forecasting model and the latitude and longitude coordinates of historical observation points. The temporal correlation matrix is ​​obtained by applying an exponential decay function to the time difference between the predicted location and the historical observation points. Based on the spatial distance matrix and the temporal correlation matrix, a weighted combination operation is performed with a spatial weight coefficient of 0.6 and a temporal weight coefficient of 0.4 to obtain the spatiotemporal correlation matrix. An influence weight is assigned to each historical observation data point according to the spatiotemporal correlation matrix. The weight values ​​are then weighted and summed with the corresponding catch data to generate the spatiotemporal correlation prediction result. This process establishes a spatiotemporal weight allocation mechanism to address the problem of neglecting the influence of surrounding areas and historical time.

[0069] Finally, the newly collected real-time sea surface temperature data, salinity data, fishing vessel trajectory data, and ocean current data are merged with the spatiotemporal correlation prediction results in chronological order to obtain a dynamic training set. The difference between the current sample mean and the historical sample mean in the dynamic training set is calculated and divided by the standard deviation. When the difference exceeds a set threshold, a model update mechanism is triggered, generating an update trigger signal. Based on the update trigger signal, the dynamic training set is input into the fishery prediction model, and the weight parameters of each layer of the neural network are adjusted using the gradient descent algorithm to obtain the updated model parameters. Environmental data from the target sea area are input into the updated fishery prediction model for calculation and processing, outputting fishery prediction results including catch values ​​and confidence intervals. This dynamic update mechanism addresses the problem that static models cannot adapt to dynamic changes in the marine environment and lack adaptive adjustment capabilities.

[0070] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0071] Sea surface temperature data is collected by satellite remote sensing system, and the temperature values ​​are digitally encoded with high precision and short time intervals to obtain a sea surface temperature time series dataset.

[0072] Salinity and chlorophyll concentration data were collected by an ocean buoy network, and the vertical profile was sampled in layers at equal depth intervals to obtain a vertical dataset of the marine environment.

[0073] Fishing vessel trajectory data is collected by the fishing vessel positioning system, and latitude and longitude coordinates and catch records are associated and matched according to timestamps to obtain a fishery operation dataset.

[0074] Coordinate system transformation and time standardization were performed on the sea surface temperature time series dataset, marine environment vertical dataset, and fishery operation dataset to obtain a spatiotemporal data matrix.

[0075] Specifically, sea surface temperature data was collected using a satellite remote sensing system. The temperature values ​​were digitized with an accuracy of 0.1℃ and hourly time intervals to obtain a sea surface temperature time-series dataset, which includes a triplet of latitude and longitude coordinates, a timestamp, and a temperature value. Salinity and chlorophyll concentration data were collected using an ocean buoy network. A CTD profiler was used for stratified sampling at 5-meter depth intervals. Each sampling point recorded four-dimensional parameters: depth, salinity, chlorophyll concentration, and dissolved oxygen content, resulting in a marine environmental vertical dataset. Fishing vessel trajectory data was collected using a fishing vessel positioning system. This system, based on AIS signals and catch logs, records latitude and longitude coordinates every minute. The latitude and longitude coordinates are then correlated and matched with catch data according to timestamps, and fish species information is also incorporated to obtain a fisheries operation dataset, which includes location, time, catch amount, and fish species information.

[0076] Furthermore, coordinate system transformation and time standardization were performed on the aforementioned sea surface temperature (SST) time-series dataset, marine environment vertical dataset, and fisheries operation dataset. For coordinate system transformation, the SST time-series dataset adopted Mercator projection to transform the coordinate system, and the resolution of the SST raster was set to 1km × 1km after transformation. The marine environment vertical dataset converted discrete profile data into continuous raster data through linear interpolation, and a bilinear interpolation relationship was established between the spatial positions of this continuous raster data and the SST raster. The fisheries operation dataset adopted the UTC time standard to unify the time base, and the nearest neighbor method was used to match fishing vessel trajectory points to environmental data raster cells. For time standardization, a sliding time window mechanism was used to unify data from different collection frequencies to an hourly time granularity. The SST data retained its original time series, the marine buoy data used cubic spline interpolation to complete the time intervals, and the fisheries operation data was aggregated into hourly statistics according to the operation period.

[0077] After the coordinate system transformation and time standardization processes described above, a spatiotemporal data matrix is ​​obtained. This matrix contains a three-dimensional tensor of environmental parameters (longitude × latitude × time) and a two-dimensional table of fishery indicators (time × catch attribute). Each spatiotemporal unit in the matrix contains 12-dimensional environmental features and 4-dimensional fishery labels. The high spatiotemporal resolution of the sea surface temperature time series dataset compensates for the sparse sampling points of the marine environment vertical dataset. The vertical profile data provided by the marine environment vertical dataset corrects the limitation that the sea surface temperature time series dataset can only reflect the sea surface state. The fishery operation dataset provides data support for subsequent analysis of the correlation between environmental factors and catch, thereby solving the problem of incompatibility between data formats from traditional fishery forecasting systems such as satellite remote sensing, buoy monitoring, and fishing vessel reports. This allows marine information from different sources and with different levels of precision to be correlated and analyzed within a unified spatiotemporal framework, while overcoming the limitations of a single data source in terms of spatiotemporal coverage and parameter dimensions.

[0078] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0079] Water temperature, salinity, and chlorophyll concentration parameters are extracted from the spatiotemporal data matrix. The difference between water temperature and water depth data is calculated to obtain the water stratification index.

[0080] The nutrient enrichment index is obtained by multiplying the chlorophyll concentration and dissolved oxygen content and then dividing by the product of water temperature and salinity.

[0081] Seasonal information is encoded using periodic functions, and monthly and tidal information are converted into combinations of sine and cosine functions to obtain time feature codes.

[0082] The water stratification index, nutrient enrichment index, time feature codes, and location information are concatenated into a vector to obtain the fishery feature vector.

[0083] Specifically, water temperature, salinity, and chlorophyll concentration parameters are extracted from the spatiotemporal data matrix. The water temperature parameter corresponds to the temperature dimension of the three-dimensional tensor of environmental parameters in the spatiotemporal data matrix. A depth-weighted average algorithm is used for extraction, dividing the 0-200 meter water layer into 20 layers at 10-meter intervals. The water temperature data of each layer is multiplied by the corresponding depth weight coefficient and then summed. Simultaneously, water depth data is extracted from the marine environmental vertical data associated with the spatiotemporal data matrix. The difference between the extracted water temperature and the corresponding water depth data is calculated, and the temperature difference between the surface water temperature and the 50-meter water temperature reflects the thermocline strength, resulting in a water stratification index. This index corresponds to the temperature difference in the vertical direction of the water body and is related to the vertical distribution of fish. A strong thermocline can be correlated with changes in the aggregation of fish in the middle and upper layers.

[0084] The salinity parameter corresponds to the parameter dimension retained after processing the marine environmental vertical dataset in the spatiotemporal data matrix. The halocline intensity index is calculated by the salinity difference between the surface and bottom layers. The chlorophyll concentration parameter also corresponds to the parameter dimension after processing the marine environmental vertical dataset, taking the integral value from the surface to the euphotic layer. Dissolved oxygen content parameters are extracted from the spatiotemporal data matrix, and the vertical distribution curve of dissolved oxygen content is fitted using the least squares method to obtain characteristic parameters. The extracted chlorophyll concentration and dissolved oxygen content parameters are multiplied; this product represents the primary productivity level. The result is then divided by the product of water temperature and salinity in the same spatiotemporal unit to eliminate environmental interference, yielding a nutrient enrichment index. This index reflects the material basis required for fishery formation and quantifies the impact of marine nutrient levels on fish aggregation.

[0085] Next, seasonal information is extracted from the time dimension of the spatiotemporal data matrix. The seasonal information includes monthly and tidal information. When encoding the seasonal information using periodic functions, Fourier series expansion is used to convert the monthly information into a pair of sine and cosine functions with an annual cycle (such as sin(2π×month / 12) and cos(2π×month / 12)). The tidal information is then used to construct a phase code with a synodic month as the cycle, forming a time feature code containing 6-dimensional periodic parameters. During the encoding process, normalized angular frequency is used to ensure the coordination between the annual and monthly cycles.

[0086] Then, location information is extracted from the spatial dimension of the spatiotemporal data matrix. The location information is the latitude and longitude coordinates of the corresponding unit. After extraction, Mercator projection coordinates and geostrophic parameters are combined for processing. Longitude is converted into a cosine function to take into account the Earth's rotation effect. Coriolis force correction factor is added to the latitude parameter, and the location information is converted into a relative value in the 0-1 interval.

[0087] Furthermore, the water stratification index and nutrient enrichment index were Z-score standardized, while the temporal feature encoding retained the original trigonometric function values. Then, the water stratification index, nutrient enrichment index, temporal feature encoding, and location information were concatenated according to vector dimensions. Specifically, the product of chlorophyll concentration and dissolved oxygen content was logarithmically transformed to avoid numerical overflow, and the vertical gradient of water temperature was calculated using the central difference method to ensure numerical stability. This resulted in a 58-dimensional fisheries feature vector, where the first 12 dimensions represent environmental parameters, the middle 6 dimensions represent temporal encoding, and the last 40 dimensions represent spatial location and derived features.

[0088] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0089] The fishery feature vector is input into the input layer of a four-layer BP neural network. The input layer contains 58 neuron nodes that receive feature data, and the output signal of the input layer is obtained.

[0090] The input layer output signal is passed to the first hidden layer, which contains 128 neurons. The input signal is processed by the ReLU activation function to obtain the first layer feature map.

[0091] The first layer feature map is passed to the second hidden layer, which contains 64 neurons. The Tanh activation function is used to perform high-order feature combination processing to obtain the second layer feature map.

[0092] The number of neurons in the output layer is set according to the type of prediction task. One neuron is set for the catch prediction task, ten neurons are set for the fish species prediction task, and eight neurons are set for the fishing area prediction task, thus obtaining the output layer structure configuration.

[0093] The second-layer feature map is linearly combined with the output layer weight matrix, and then processed by the Sigmoid or Softmax activation function to obtain the predicted output value.

[0094] The error of the loss function is calculated based on the predicted output value and the actual catch data. The weight parameters of each layer are updated through the backpropagation algorithm. The training is repeated until the loss function converges, and the fishery prediction model is obtained.

[0095] Specifically, after standardizing and preprocessing the fishery feature vector, it is input into the input layer of a four-layer backpropagation (BP) neural network. The input layer contains 58 neurons, the number of which strictly corresponds to the 58 dimensions of the fishery feature vector. Each neuron receives data from one feature dimension of the fishery feature vector. The input layer performs linear weighting on the received feature data to obtain the input layer output signal, which serves as the input to the first hidden layer. The linear weighting process is implemented using a weight matrix and a bias vector. W The first dimension is set to 58×128, and the bias vector is... b With dimension 128, the linear transformation result satisfies Z 1 =X × W 1+ b 1 ( X (This is the standardized fishery feature vector).

[0096] Secondly, the input layer output signal is passed to the first hidden layer, which contains 128 neurons, and the result of the linear transformation of the input is... Z 1. Nonlinear transformation is performed using the ReLU activation function, which is defined as max(0, x ), that is, when Z When the value in step 1 is positive, the value is directly output; when it is negative, 0 is output. This process preserves the positive features of the input signal and suppresses the negative signals, enhancing the sparse representation capability of the network and resulting in a 128-dimensional first-layer feature map. H 1.

[0097] Then, the first layer of features is mapped. H 1 is passed to the second hidden layer, which contains 64 neurons, and the first layer's feature mapping... H 1. First, use the weight matrix W 2 (dimension 128×64) and bias vector b A linear transformation is performed on 2 (dimension 64) to obtain Z 2= H 1× W 2+ b 2. Then, apply the Tanh activation function to... Z 2. Perform high-order feature combination processing; the Tanh activation function calculation formula is as follows: ,Will Z The numerical mapping to the interval of -1 to 1 captures higher-order feature interactions to reflect the complex nonlinear relationship between environmental factors and fishing conditions. Furthermore, its symmetry facilitates gradient propagation and alleviates the gradient vanishing problem in deep networks, resulting in a 64-dimensional second-layer feature mapping. H 2.

[0098] Furthermore, the number of neurons in the output layer is set according to the prediction task type: 1 neuron for catch prediction, 10 neurons for fish species prediction, and 8 neurons for fishing area prediction, resulting in the output layer structure configuration. The output layer calculation relies on the second layer feature mapping. H 2. Weight matrix W 3 and bias vector b 3. Achieve, satisfying Y= f ( H 2× W 3+ b 3) f (The activation function) is the weight matrix in the fish catch prediction task. W The 3-dimensional matrix is ​​64×1. A Sigmoid activation function is used to map the linear combination results to the [0,1] interval to represent the standardized catch probability. The weight matrix is ​​used in both the fish species prediction task and the fishing area prediction task. W The three dimensions are 64×10 and 64×8 respectively. The linear combination result is converted into a probability distribution using the Softmax activation function.

[0099] Finally, the loss function error is calculated based on the predicted output value Y and the actual catch data. The actual catch data comes from the catch amount, fish species, and fishing area information recorded in the fishery operation dataset. The catch amount prediction task uses mean squared error loss, while the fish species prediction task and the fishing area prediction task use cross-entropy loss. The weight parameters of each layer of the neural network are updated through the backpropagation algorithm. When calculating the output layer error, it is obtained by element-wise multiplication of the gradient of the loss function with respect to the output layer activation value and the derivative of the output layer activation function. When calculating the error of the first hidden layer, it is obtained by element-wise multiplication of the transpose of the next layer weight matrix with the error of the next layer and the derivative of the first hidden layer activation function. The weight update adopts the stochastic gradient descent method with momentum. The update amount is composed of the product of the learning rate, the current layer error, the outer product of the previous layer activation value, and the product of the momentum coefficient and the weight update amount of the previous round. The learning rate is set to 0.001, and the momentum coefficient is set to 0.9. During training, mini-batch gradient descent is implemented, selecting 70% of the samples in each batch for training. The validation set loss is calculated after each training round. Multiple measures are employed to prevent overfitting: weights are initialized using a He normal distribution, bias is initialized to 0, L2 weight decay (decay coefficient set to 0.01) and Dropout (dropout ratio set to 0.2) are introduced, and the training data is shuffled before training to ensure the independence of each batch. An early stopping mechanism is enabled; training terminates if the validation set loss does not decrease after five consecutive rounds. This training process is repeated until the loss function converges, resulting in the fishing situation prediction model.

[0100] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0101] The spatial distance matrix is ​​obtained by calculating the Euclidean distance between the predicted location of the fishery condition prediction model and the latitude and longitude coordinates of the historical observation points.

[0102] The time difference between the predicted location and the historical observation point is processed by an exponential decay function to obtain the time correlation matrix;

[0103] A weighted combination operation is performed based on the spatial distance matrix and the temporal correlation matrix, with the spatial weight coefficient set to 0.6 and the temporal weight coefficient set to 0.4, to obtain the spatiotemporal correlation matrix.

[0104] Based on the spatiotemporal correlation matrix, an influence weight is assigned to each historical observation data. The weight values ​​are then weighted and summed with the corresponding catch data to obtain the spatiotemporal correlation prediction results.

[0105] Specifically, the Euclidean distance is calculated between the predicted location of the fishery condition prediction model and the latitude and longitude coordinates of historical observation points. The predicted location is the latitude and longitude coordinates of the target sea area where the fishery condition is to be predicted, while the latitude and longitude coordinates of historical observation points are derived from the latitude and longitude information of fishing vessel operation locations and ocean buoys and satellite remote sensing monitoring points recorded in the fishery operation dataset. Before the calculation, the latitude and longitude coordinates of both are converted to a Cartesian coordinate system using Mercator projection. x , y Coordinate transformation uses standard formulas. x = R × λ , y = R ×ln[tan(π / 4+ φ / 2)]( R For the Earth's radius, λ Longitude φ (Latitude), the conversion process uses the PROJ library to ensure projection accuracy, and the converted coordinate values ​​are in meters to eliminate the influence of the Earth's curvature on distance calculation. The Euclidean distance formula is then substituted into the plane coordinate system. ,in( ) represents the coordinates of the predicted point, ( The coordinates of the historical observation points are used as a reference. The distance between each predicted location and all historical observation points is calculated, and the results are sorted according to the number of predicted points. n Compared with historical observation points m The corresponding relationships are arranged to form n×m 3D spatial distance matrix D .

[0106] Next, the time difference between the predicted location and the historical observation point is processed using an exponential decay function. The timestamp of the predicted location is the time node of the public opinion to be predicted, and the timestamp of the historical observation point is the time of historical observation data collection (such as fishing boat operation time, buoy sampling time). The time difference Δ between the two is calculated using the Unix timestamp as the benchmark. t Time processing uses the chrono library to manage timestamps, and Δ t Substitute into the exponential decay function Attenuation constant τ Set to 2,592,000 seconds corresponding to 30 days, the calculation of the exponential function is accelerated by using a lookup table, and the function output value varies with Δ. t The value decreases as the number of recent observations increases, meaning the output value is larger. The processing result generates a spatial distance matrix. D Same-dimensional time correlation matrix T The matrix elements take values ​​in the range (0,1).

[0107] Furthermore, based on the spatial distance matrix D Time correlation matrix TThe spatiotemporal correlation matrix is ​​obtained by performing weighted combination operations. First, the spatial distance matrix is ​​processed. D Perform normalization by dividing each element by the maximum value of the matrix. Time correlation matrix T Keeping the original values ​​unchanged, according to the formula The spatiotemporal correlation matrix was calculated, with spatial weight coefficients of 0.6 and temporal weight coefficients of 0.4 determined through grid search. This matrix performed optimally on the validation set. D norm This is used to convert the "smaller the spatial distance, the stronger the correlation" property into a "larger the numerical value, the stronger the correlation" property, in relation to the time correlation matrix. T The numerical logic remains consistent. C The spatiotemporal correlation matrix has an element value range of [0,1]. The larger the value, the stronger the correlation between the predicted location and the historical observation point. The weighted combination mechanism coordinates the relative importance of spatiotemporal factors.

[0108] Finally, based on the spatiotemporal correlation matrix C Assign influence weights to each historical observation data point and apply them to the matrix. C Each row undergoes a softmax transformation to ensure a reasonable weight distribution. Weight allocation employs multi-threaded parallel processing, with each prediction point assigned an independent thread to perform the softmax operation. Memory management utilizes data block loading, and the large matrix is ​​divided into sub-matrices for batch processing. Historical observation data includes catch data corresponding to each historical observation point (derived from a fisheries operation dataset). The influence weight value of each historical observation data point is multiplied by its corresponding catch data to obtain a weighted catch data. All weighted catch data are then summed to obtain the spatiotemporal correlation prediction result for that prediction location. If multiple prediction locations exist, the result for each location is calculated separately using the above method to form the overall spatiotemporal correlation prediction result. This process strengthens the influence of nearby spatial and recent historical data through weight allocation, addressing the problem that traditional models' "static predictions cannot adapt to the spatiotemporal evolution of the marine environment," making the prediction results more closely reflect the spatiotemporal distribution characteristics of fish in the sea.

[0109] In one specific embodiment, the process of executing step S105 may specifically include the following steps:

[0110] The newly collected real-time sea surface temperature data, salinity data, fishing vessel trajectory data, and ocean current data are merged with the spatiotemporal correlation prediction results in chronological order to obtain a dynamic training set.

[0111] The difference between the current sample mean and the historical sample mean in the dynamic training set is calculated and divided by the standard deviation. When the difference exceeds the set threshold, the model update mechanism is triggered to obtain the update trigger signal.

[0112] Based on the update trigger signal, the dynamic training set is input into the fishery condition prediction model, and the weight parameters of each layer of the neural network are adjusted through the gradient descent algorithm to obtain the updated model parameters.

[0113] The environmental data of the target sea area is input into the updated fishery forecasting model for calculation and processing, and the fishery forecasting results, including the catch value and confidence interval range, are output.

[0114] Specifically, the newly collected real-time sea surface temperature data, salinity data, fishing vessel trajectory data, and ocean current data are merged with the spatiotemporal correlation prediction results in chronological order. Real-time sea surface temperature data is collected via satellite remote sensing, real-time salinity data via an ocean buoy network, real-time fishing vessel trajectory data via a fishing vessel positioning system, and real-time ocean current data via ocean observation stations or ocean current sensors mounted on buoys. The collection frequency of these four types of real-time data is consistent with the corresponding data in step S101 (e.g., real-time sea surface temperature data is collected once per hour, and real-time fishing vessel trajectory data is collected once per minute). The spatiotemporal correlation prediction results are the result data generated in step S104, containing the correlation values ​​of catch yield at each predicted location. During merging, timestamps are used as the basis, and real-time data with the same timestamp are correlated with the spatiotemporal correlation prediction results. Real-time data with inconsistent timestamps are supplemented to the corresponding time nodes through linear interpolation, forming a dynamic training set. This training set includes both the latest environmental and operational data and integrates historical prediction results, addressing the problem that static prediction models cannot adapt to dynamic changes in the marine environment and lack online learning data, thus providing data support for dynamic model updates.

[0115] Then, the difference between the current sample mean and the historical sample mean in the dynamic training set is calculated and divided by the standard deviation. The current sample refers to the samples within the most recent time window in the dynamic training set (e.g., samples from the last 7 days), and the historical samples refer to all samples in the dynamic training set before the current sample. First, the current sample mean and the historical sample mean of the correlation between environmental parameters (sea surface temperature, salinity, ocean current) and catch values ​​in the dynamic training set are calculated separately. Then, the difference between the two is calculated (taking the absolute value), and the standard deviation of the historical samples is also calculated. The difference is divided by the standard deviation to obtain the coefficient of variation. A threshold is set (this threshold is determined through historical data validation, such as setting the coefficient of variation threshold for environmental parameters to 2 and the coefficient of variation threshold for catch correlation values ​​to 1.5). When the coefficient of variation of any parameter exceeds the corresponding set threshold, it is determined that the environmental pattern has changed, triggering the model update mechanism and obtaining an update trigger signal. This mechanism identifies environmental changes by quantifying the degree of data variation, solving the problem of being unable to detect sudden changes in the marine environment in a timely manner and the model update lagging behind, ensuring that the model adjusts in a timely manner when the environment is abnormal.

[0116] Furthermore, based on the update trigger signal, the dynamic training set is input into the fishery prediction model. The environmental parameters (sea surface temperature, salinity, ocean current) and fishing vessel trajectory data in the dynamic training set are converted into fishery feature vectors according to the processing method in step S102, and paired with the corresponding spatiotemporal correlation prediction results (as label data) to form training sample pairs. The weight parameters of each layer of the four-layer BP neural network are adjusted by the gradient descent algorithm. During the adjustment, the training sample pairs are first forward propagated into the neural network to obtain the predicted values. The loss function error between the predicted values ​​and the label data is calculated. Then, the gradient of the weights of each layer (such as the weight gradient between the input layer and the first hidden layer, the first hidden layer and the second hidden layer, and the second hidden layer and the output layer) is calculated by the backpropagation algorithm. The weight parameters are gradually adjusted according to the gradient direction. At the same time, elastic weights are added to consolidate the regularization term to prevent the forgetting of historical knowledge, resulting in updated model parameters. This parameter adjustment process enables the model to adapt to the new environmental mode, solving the problem that traditional model parameters are fixed and cannot be adaptively adjusted.

[0117] Finally, the environmental data of the target sea area is input into the updated fisheries prediction model for calculation and processing. The environmental data of the target sea area includes real-time sea surface temperature data, salinity data, and ocean current data (collected using the same method as real-time data). This environmental data is first converted into fisheries feature vectors according to the feature extraction process in step S102, and then input into the updated fisheries prediction model. The model calculates and outputs predicted values ​​through forward propagation. Simultaneously, based on the error distribution of the loss function during model training, the Bootstrap method is used to calculate the confidence interval range of the predicted values, outputting fisheries prediction results that include the catch amount and the confidence interval range. This result provides both a specific reference for the catch amount and reflects the reliability of the prediction, addressing the problem of insufficient uncertainty assessment and practicality in prediction results, and providing a more comprehensive basis for fisheries production decisions.

[0118] In one specific embodiment, a dynamic training set is input into the fishery condition prediction model based on an update trigger signal. The weight parameters of each layer of the neural network are adjusted using a gradient descent algorithm to obtain the updated model parameters, including:

[0119] The fishery feature vectors in the dynamic training set are paired with the corresponding real catch data to construct training sample pairs and obtain online learning data pairs.

[0120] The online learning data is used to perform forward propagation calculations on the input four-layer BP neural network. The difference between the network output value and the actual catch is calculated to obtain the current error value.

[0121] Based on the current error value, the backpropagation algorithm is used to calculate the gradient information of the weights of each layer. The gradient value is multiplied by the learning rate of 0.001 to calculate the weight update and obtain the weight increment.

[0122] The weight increments are added to the original weight parameters of each layer of the BP neural network, and elastic weights are added to reinforce the regularization term to prevent the forgetting of historical knowledge, resulting in updated model parameters.

[0123] Specifically, the fishery feature vectors in the dynamic training set are paired and combined with the corresponding real catch data. The fishery feature vectors in the dynamic training set are obtained through the processing flow of step S102—extracting parameters such as water temperature, salinity, and chlorophyll concentration from the real-time sea surface temperature data, salinity data, fishing vessel trajectory data, and ocean current data contained in the dynamic training set, calculating the water stratification index and nutrient enrichment index, and then combining them with seasonal information and location information to form a vector containing 58 features (the first 12 dimensions are environmental parameters, the middle 6 dimensions are time codes, and the last 40 dimensions are spatial location and derived features). The actual catch data comes from the catch logs associated with the fishing vessel trajectory data collected by the fishing vessel positioning system. It includes actual operation records such as catch amount, fish species information, and fishing area. When pairing and combining data, the timestamp and location information are used as the basis to match the fishery feature vectors under the same spatiotemporal unit with the actual catch data one by one to construct training sample pairs. Each sample pair contains a mapping relationship of "input feature - real label". This results in online learning data pairs. These data pairs establish the correlation between dynamic environmental features and actual fishing results, solving the problem that static models lack online learning data support and cannot adapt to new environmental data. This provides a training basis for adjusting model parameters.

[0124] Then, the online learning data is input into a four-layer BP neural network for forward propagation. The fishery feature vectors from the online learning data are input into the input layer of the neural network (58 neurons, corresponding one-to-one with the feature vector dimension), linearly passed to the first hidden layer (128 neurons, ReLU activation function), and a non-linear transformation is performed to obtain the first layer feature map. The first layer feature map is passed to the second hidden layer (64 neurons, Tanh activation function), and high-order feature combination is performed to obtain the second layer feature map. The second layer feature map is linearly combined with the output layer weight matrix, and then processed by a Sigmoid or Softmax activation function to obtain the network output value (the output value for the catch prediction task, and the output probability distribution for the fish species or fishing area prediction task). The difference between the network output value and the actual catch data in the online learning data is calculated. The absolute difference between the two values ​​is used for the catch prediction task, while the cross-entropy difference of the probability distribution is used for the fish species or fishing area prediction task. The current error value is obtained. This error value quantifies the deviation between the current prediction result of the model and the actual fishing situation, solves the problem of not being able to evaluate the model prediction accuracy in real time and the difficulty in locating the direction of parameter adjustment, and provides an error basis for backpropagation.

[0125] Furthermore, based on the current error value, the backpropagation algorithm is used to calculate the gradient information of the weights of each layer. Starting from the output layer, the derivative of the output layer activation function is calculated based on the current error value, and the gradient of the output layer weights is obtained by combining the linear combination result of the output layer. The gradient of the output layer weights is backpropagated to the second hidden layer, and the gradient of the second hidden layer weights is calculated by combining the derivative of the second hidden layer Tanh activation function with the feature map of the first layer. This backpropagation continues to the first hidden layer, and the gradient of the first hidden layer weights is calculated by combining the derivative of the first hidden layer ReLU activation function with the input layer output signal. At the same time, the gradients of the bias parameters of each layer are calculated to form gradient information. The gradient value of each layer weight is multiplied by the learning rate of 0.001. The learning rate is determined through historical training verification to ensure the stability of parameter adjustment and avoid slow convergence. This yields the weight increment, which reflects the magnitude and direction of the weight adjustment required for each layer, solving the problem of lack of quantitative basis for parameter updates and low adjustment efficiency.

[0126] Finally, the weight increments are added to the original weight parameters of each layer of the BP neural network, namely the original weight matrices of the input layer and the first hidden layer, the first hidden layer and the second hidden layer, and the second hidden layer and the output layer. These are then added to the corresponding layer's weight increments to obtain the initially updated weight parameters. Simultaneously, an elastic weight consolidation regularization term is added. This regularization term constrains weights with large update magnitudes by preserving the value characteristics of important weights from historical training, preventing the forgetting of historical environmental factors and catch correlation knowledge due to new data training. For example, for weight parameters reflecting the correlation between the thermocline layer and catch, the regularization term reduces their update magnitude to maintain the original correlation characteristics. Through the above weight increment superposition and regularization processing, updated model parameters are obtained. These parameters adapt to the new environmental features in the dynamic training set while retaining historical learning results, solving the problem of model updates easily losing historical knowledge and long-term prediction accuracy decay.

[0127] The above describes the fishery situation prediction method based on multi-source heterogeneous data in the embodiments of this application. The following describes the fishery situation prediction processing procedure 200 based on multi-source heterogeneous data in the embodiments of this application. Please refer to [link / reference]. Figure 2 The processing revolves around a closed-loop operation for fishery condition prediction based on multi-source heterogeneous data. From the acquisition and conversion of multi-source heterogeneous data to the output of fishery condition prediction results, each step is connected according to a fixed logic. The specific process is as follows:

[0128] The data acquisition and processing process 201 involved collecting data through satellite remote sensing, ocean buoys, fishing vessel positioning systems, and ocean current monitoring equipment. The satellite remote sensing system encoded sea surface temperature data with an accuracy of 0.1℃ and hourly intervals, forming a sea surface temperature time-series dataset containing latitude, longitude, timestamps, and temperature values. Ocean buoys used a CTD profiler to collect salinity and chlorophyll concentration data at 5-meter intervals, recording four-dimensional parameters to form a vertical dataset of the marine environment. The fishing vessel positioning system recorded latitude and longitude every minute and correlated it with catch volume, forming a fisheries operation dataset containing location, time, catch volume, and fish species. Ocean current monitoring equipment collected ocean current data. Subsequently, coordinate system transformation and time standardization were performed on the four types of data. Sea surface temperature data was converted to Mercator projection and set to a 1km×1km grid. Buoy data was interpolated to a continuous grid. Fishing vessel data was standardized to UTC time and matched to the grid, with time uniformized to an hourly granularity, forming a spatiotemporal data matrix containing a three-dimensional tensor of environmental parameters and a two-dimensional table of fisheries indicators.

[0129] The feature extraction process 202 extracts water temperature, salinity, and chlorophyll concentration parameters from the spatiotemporal data matrix. Water temperature is processed using depth-weighted averaging, salinity is calculated based on the difference between surface and bottom layers, and chlorophyll concentration is calculated using integral values. Dissolved oxygen content is also extracted and its curve is fitted. The water stratification index (difference between water temperature and depth) and the nutrient enrichment index (product of chlorophyll and dissolved oxygen divided by the product of water temperature and salinity) are calculated. Seasonal information is encoded as a 6-dimensional temporal feature, and location information is processed into 0-1 interval values. After standardization, the parameters are concatenated to form a 58-dimensional fisheries feature vector.

[0130] In the model training process 203, the standardized feature vectors are input into a four-layer backpropagation (BP) neural network: the input layer (58 neurons) receives data, which is then linearly weighted and passed to the first hidden layer (128 neurons, ReLU activation), and then to the second hidden layer (64 neurons, Tanh activation). The output layer has 1, 10, or 8 neurons depending on the task, and the predicted value is obtained by combining the activation function. The error with the actual catch data is calculated, the weights are updated through backpropagation, and mini-batch training is performed with various anti-overfitting measures until the loss converges to obtain the model.

[0131] In the correlation prediction process 204, the distance between the predicted location and the historical observation point coordinates is converted to a Cartesian coordinate system to obtain a spatial matrix. The time difference is obtained by exponential decay to obtain a time matrix. The two matrices are weighted and combined to form a spatiotemporal matrix. After weighting, the spatiotemporal correlation prediction result is obtained by weighted summation. GPU acceleration and anomaly handling are enabled in the process.

[0132] The dynamic update and output process 205 merges real-time data and prediction results to form a dynamic training set, calculates the coefficient of variation, and updates model parameters if the coefficient exceeds a threshold; otherwise, the original model is used. Target marine environmental data is converted into feature vectors and input into the model, outputting results including catch amount and confidence intervals. Quantization accelerates the inference phase, and deployment is optimized to support high concurrency.

[0133] The above describes the fishery condition prediction method based on multi-source heterogeneous data in the embodiments of this application. The following describes the fishery condition prediction system based on multi-source heterogeneous data in the embodiments of this application. Please refer to [link / reference]. Figure 3 One embodiment of the fishery situation prediction system based on multi-source heterogeneous data in this application includes:

[0134] The data acquisition module 301 is used to acquire sea surface temperature data, salinity data, fishing vessel trajectory data and ocean current data, and convert the above heterogeneous data into a spatiotemporal data matrix.

[0135] The feature extraction module 302 is used to extract water temperature, salinity, and chlorophyll concentration from the spatiotemporal data matrix, calculate the water stratification index and nutrient enrichment index, and combine them with seasonal and location information to obtain a fishery feature vector.

[0136] The model training module 303 is used to input fishery feature vectors into a four-layer BP neural network for training, learn the mapping relationship between environmental factors and catch, and establish a fishery condition prediction model.

[0137] The correlation prediction module 304 is used to calculate the spatiotemporal distance between the predicted location and the observation point, assign influence weights to each observation data, and generate spatiotemporal correlation prediction results.

[0138] The dynamic update module 305 is used to add real-time data to the spatiotemporal correlation prediction results to form a dynamic training set, update the model parameters when environmental pattern changes are detected, and output the fishing situation prediction results.

[0139] Through the collaborative efforts of the aforementioned components, the system constructs a complete technical chain for multi-source heterogeneous marine data, from acquisition and processing to intelligent prediction. This enables a transformation in fisheries forecasting from traditional experience-driven to data-driven approaches, effectively addressing the core issues of data isolation, difficulty in handling heterogeneous data, and static models. The data acquisition module 301 first acquires sea surface temperature data, salinity and chlorophyll concentration data, fishing vessel trajectory and catch data, and ocean current data through satellite remote sensing systems, marine buoy networks, fishing vessel positioning systems, and ocean current detection equipment. Then, through coordinate system transformation (such as Mercator projection to unify the spatial reference) and time standardization (sliding time windows to unify to hourly granularity), the heterogeneous data with different formats and spatiotemporal scales are integrated into a spatiotemporal data matrix. This matrix uses a three-dimensional framework of "longitude-latitude-time" to link environmental parameters and fisheries operation data, providing structurally unified basic data for subsequent modules. The feature extraction module 302 first extracts key environmental parameters such as water temperature, salinity, and chlorophyll concentration from the spatiotemporal data matrix. Then, it calculates the water stratification index through difference calculation and the nutrient enrichment index through multi-parameter operations. Simultaneously, it performs periodic function encoding on seasonal information and extracts latitude and longitude coordinates as location information. These parameters are then concatenated into a 58-dimensional fisheries feature vector. The model training module 303 uses the fisheries feature vector as the core input and imports it into a four-layer BP neural network. The input layer has 58 neurons corresponding to the feature dimensions that receive data. This data undergoes nonlinear transformation and high-order feature combination through the first hidden layer (128 neurons, ReLU activation) and the second hidden layer (64 neurons, Tanh activation). The number of neurons in the output layer is configured according to different prediction tasks such as catch volume, fish species, and fishing area. The predicted output value is obtained by combining a Sigmoid or Softmax activation function. The error is calculated using a loss function and the weights are updated via backpropagation. Training is repeated until the loss converges, establishing a fisheries prediction model. The correlation prediction module 304 calculates the Euclidean distance between the predicted location output by the model and the latitude and longitude of historical observation points to generate a spatial distance matrix. It then performs exponential decay processing on the time difference to generate a temporal correlation matrix, and finally weights these matrices with a weight ratio of 0.6:0.4 to obtain a spatiotemporal correlation matrix. Based on this, it assigns weights to historical observation data and sums them with the catch amount to generate a spatiotemporal correlation prediction result. The dynamic update module 305 merges newly collected real-time sea surface temperature, salinity, fishing vessel trajectory, and ocean current data with the spatiotemporal correlation prediction result output by the correlation prediction module 304, forming a dynamic training set by timestamp. By calculating the mean difference and coefficient of variation between the current sample and historical samples, it determines whether to trigger a model update. When the coefficient of variation exceeds a threshold, the dynamic training set is input into the fishery prediction model. Gradient descent is used to adjust the weight parameters, and a regularization term is added to prevent the forgetting of historical knowledge. The updated model performs calculations on the target sea area environmental data and outputs prediction results containing catch amounts and confidence intervals. This enables the system to adaptively respond to dynamic changes in the marine environment, providing real-time and accurate decision support for fishery production.

[0140] above Figure 3 The fishery condition prediction system based on multi-source heterogeneous data in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The fishery condition prediction device based on multi-source heterogeneous data in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0141] Reference Figure 4 This invention also provides a fishery condition prediction device 400 based on multi-source heterogeneous data. This device can be a server, and its internal structure can be as follows: Figure 4 As shown. The fishery condition prediction device 400 based on multi-source heterogeneous data includes a processor 402, a memory 403, a display screen 404, an input device 405, a network interface 406, and a database 407 connected via a system bus 401. The processor 402, designed as a computer, provides computing and control capabilities. The memory 403 of the fishery condition prediction device includes a non-volatile storage medium 4031 and internal memory 4032. The non-volatile storage medium 4031 stores the operating system and computer programs. The internal memory 4032 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database 407 of the fishery condition prediction device based on multi-source heterogeneous data stores the data corresponding to this embodiment. The network interface 406 of the fishery condition prediction device based on multi-source heterogeneous data is used for communication with external terminals via a network connection. The computer program, executed by the processor, can implement the above-described method.

[0142] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the fishing situation prediction device based on multi-source heterogeneous data to which the present invention is applied.

[0143] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the fishery situation prediction method based on multi-source heterogeneous data.

[0144] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0145] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a fishery situation prediction device (which may be a personal computer, server, or network device, etc.) based on multi-source heterogeneous data to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0146] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting fishery conditions based on multi-source heterogeneous data, characterized in that, The method includes: Step S101: Acquire sea surface temperature data, salinity data, fishing vessel trajectory data, and ocean current data, and convert the heterogeneous data into a spatiotemporal data matrix; Step S102: Extract water temperature, salinity, and chlorophyll concentration from the spatiotemporal data matrix, calculate the water stratification index and nutrient enrichment index, and combine them with seasonal and location information to obtain a fishery feature vector. Step S103: Input the fishery feature vector into a four-layer BP neural network for training, learn the mapping relationship between environmental factors and catch, and establish a fishery situation prediction model. Step S104: Calculate the spatiotemporal distance between the predicted location and the observation point, assign influence weights to each observation data, and generate spatiotemporal correlation prediction results; Step S105: Add real-time data to the spatiotemporal correlation prediction results to obtain a dynamic training set. When environmental pattern changes are detected, update the model parameters and output the fishing situation prediction results. Step S104 includes: calculating the Euclidean distance between the predicted location of the fishery prediction model and the latitude and longitude coordinates of historical observation points to obtain a spatial distance matrix; processing the time difference between the predicted location and the historical observation points using an exponential decay function to obtain a temporal correlation matrix; performing a weighted combination operation based on the spatial distance matrix and the temporal correlation matrix, setting the spatial weight coefficient to 0.6 and the time weight coefficient to 0.4 to obtain a spatiotemporal correlation matrix; assigning influence weights to each historical observation data according to the spatiotemporal correlation matrix, and performing a weighted summation of the weight values ​​and the corresponding catch data to obtain a spatiotemporal correlation prediction result.

2. The method for predicting fishery conditions based on multi-source heterogeneous data according to claim 1, characterized in that, Step S101 includes: Sea surface temperature data is collected by satellite remote sensing system, and the temperature values ​​are digitally encoded with high precision and short time intervals to obtain a sea surface temperature time series dataset. Salinity and chlorophyll concentration data were collected by an ocean buoy network, and the vertical profile was sampled in layers at equal depth intervals to obtain a vertical dataset of the marine environment. Fishing vessel trajectory data is collected by the fishing vessel positioning system, and latitude and longitude coordinates and catch records are associated and matched according to timestamps to obtain a fishery operation dataset. The sea surface temperature time series dataset, marine environment vertical dataset, and fishery operation dataset are subjected to coordinate system transformation and time standardization to obtain a spatiotemporal data matrix.

3. The method for predicting fishery conditions based on multi-source heterogeneous data according to claim 1, characterized in that, Step S102 includes: Water temperature, salinity, and chlorophyll concentration parameters are extracted from the spatiotemporal data matrix. The difference between water temperature and water depth data is calculated to obtain the water stratification index. The nutrient enrichment index is obtained by multiplying the chlorophyll concentration and dissolved oxygen content and then dividing by the product of water temperature and salinity. Seasonal information is encoded using periodic functions, and monthly and tidal information are converted into combinations of sine and cosine functions to obtain time feature codes. The water stratification index, nutrient enrichment index, time feature encoding, and location information are concatenated into a vector to obtain the fishery feature vector.

4. The method for predicting fishery conditions based on multi-source heterogeneous data according to claim 1, characterized in that, Step S103 includes: The fishery feature vector is input into the input layer of a four-layer BP neural network. The input layer contains 58 neuron nodes that receive feature data, and the output signal of the input layer is obtained. The input layer output signal is passed to the first hidden layer, which contains 128 neurons. The input signal is processed by the ReLU activation function to obtain the first layer feature map. The first layer feature map is passed to the second hidden layer, which contains 64 neurons. The Tanh activation function is used to perform high-order feature combination processing to obtain the second layer feature map. The number of neurons in the output layer is set according to the type of prediction task: 1 neuron for the catch prediction task, 10 neurons for the fish species prediction task, and 8 neurons for the fishing area prediction task, thus obtaining the output layer structure configuration. The second layer feature map is linearly combined with the output layer weight matrix, and then processed by the Sigmoid or Softmax activation function to obtain the predicted output value. The error of the loss function is calculated based on the predicted output value and the actual catch data. The weight parameters of each layer are updated through the backpropagation algorithm. The training is repeated until the loss function converges, and the fishery prediction model is obtained.

5. The method for predicting fishery conditions based on multi-source heterogeneous data according to claim 1, characterized in that, Step S105 includes: The newly collected real-time sea surface temperature data, salinity data, fishing vessel trajectory data, and ocean current data are merged with the spatiotemporal correlation prediction results in chronological order to obtain a dynamic training set. The difference between the current sample mean and the historical sample mean in the dynamic training set is calculated and divided by the standard deviation. When the difference exceeds a set threshold, the model update mechanism is triggered to obtain the update trigger signal. Based on the update trigger signal, the dynamic training set is input into the fishery condition prediction model, and the weight parameters of each layer of the neural network are adjusted by the gradient descent algorithm to obtain the updated model parameters. The environmental data of the target sea area is input into the updated fishery forecasting model for calculation and processing, and the fishery forecasting results, including the catch value and confidence interval range, are output.

6. The method for predicting fishery conditions based on multi-source heterogeneous data according to claim 5, characterized in that, The process involves inputting the dynamic training set into the fishery condition prediction model based on the update trigger signal, adjusting the weight parameters of each layer of the neural network using the gradient descent algorithm, and obtaining the updated model parameters, including: The fishery feature vectors in the dynamic training set are paired and combined with the corresponding real catch data to construct training sample pairs and obtain online learning data pairs. The online learning data is used to perform forward propagation calculations on the input four-layer BP neural network. The difference between the network output value and the actual catch is calculated to obtain the current error value. Based on the current error value, the gradient information of the weights of each layer is calculated using the backpropagation algorithm. The gradient value is multiplied by the learning rate of 0.001 to perform weight update calculation and obtain the weight increment. The weight increments are added to the original weight parameters of each layer of the BP neural network, and elastic weights are added to reinforce the regularization term to prevent the forgetting of historical knowledge, thus obtaining the updated model parameters.

7. A fishery situation prediction system based on multi-source heterogeneous data, characterized in that, For implementing the fishery condition prediction method based on multi-source heterogeneous data as described in any one of claims 1-6, the fishery condition prediction system based on multi-source heterogeneous data comprises: The data acquisition module is used to acquire sea surface temperature data, salinity data, fishing vessel trajectory data and ocean current data, and convert the above heterogeneous data into a spatiotemporal data matrix; The feature extraction module is used to extract water temperature, salinity, and chlorophyll concentration from the spatiotemporal data matrix, calculate the water stratification index and nutrient enrichment index, and combine them with seasonal and location information to obtain fishery feature vectors. The model training module is used to input fishery feature vectors into a four-layer BP neural network for training, learn the mapping relationship between environmental factors and catch, and establish a fishery condition prediction model. The correlation prediction module is used to calculate the spatiotemporal distance between the predicted location and the observation point, assign influence weights to each observation data, and generate spatiotemporal correlation prediction results. The dynamic update module is used to add real-time data to the spatiotemporal correlation prediction results to form a dynamic training set, update the model parameters when environmental pattern changes are detected, and output the fishing situation prediction results.

8. A fishery condition prediction device based on multi-source heterogeneous data, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the fishing situation prediction method based on multi-source heterogeneous data as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the fishing situation prediction method based on multi-source heterogeneous data as described in any one of claims 1 to 6.

Images