Deep learning-based green tide of enteromorpha prolifera marine environment assessment and prediction method

Abstract

The present application relates to the technical field of marine environment monitoring and prediction, and particularly provides a green tide of Enteromorpha prolifera marine environment assessment and prediction method based on deep learning. The method comprises collecting multi-source marine environment data and preprocessing; constructing spatial feature engineering and time feature; establishing a suitability scoring model based on the growth cycle, and combining the sea area division to mark the applicability; constructing a bidirectional spatial convolution GRU prediction model, and training and optimizing the same to predict the future marine environment, and output the prediction result. The method improves the scientificity and forward-looking nature of green tide management and prevention and control by constructing a bidirectional spatial convolution GRU prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of marine environment monitoring and prediction, and particularly relates to a green tide of Enteromorpha prolifera marine environment assessment and prediction method based on deep learning. BACKGROUND

[0002] The green tide of Enteromorpha prolifera is one of the largest and most influential algal blooms in the world. Since its first outbreak in the Yellow Sea of China in 2007, it has formed large-scale green tide disasters almost every year during late spring to summer. The special hydrographic conditions and human activities in the Yellow Sea make it a high-risk area for green tide. According to statistics, the maximum coverage area of green tide can reach tens of thousands of square kilometers, which not only seriously damages the marine ecosystem, but also causes huge economic losses to the fishing resources, coastal infrastructure and tourism industry. Therefore, effective monitoring and accurate prediction of the green tide of Enteromorpha prolifera have become a key issue in marine environmental management and disaster prevention and control.

[0003] Current monitoring and prediction of the green tide of Enteromorpha prolifera mainly rely on the following three types of technical means: (1) remote sensing monitoring technology. Remote sensing is the most widely used monitoring method at present, which mainly relies on satellite image data such as MODIS and GOCI to identify the area of Enteromorpha prolifera through vegetation index (such as NDVI and FAI). The advantage of this method is that it can provide large-scale, near-real-time green tide distribution information. However, its main limitations are that it can only be used for post-observation and cannot provide early warning; it is greatly affected by cloud cover and weather conditions, and the spatial resolution of satellite images is low, making it difficult to accurately identify details near the coast. (2) Numerical ocean model. Numerical model is a simulation method based on ocean physical processes, which usually considers the influence of dynamic factors such as sea currents and wind fields on the drift and diffusion of Enteromorpha prolifera. The advantage of this method is that it is based on physical mechanisms and has a strong theoretical basis, which can simulate the movement path of Enteromorpha prolifera. However, this method has problems such as complex calculation, parameter sensitivity and low prediction accuracy, and generally ignores the biological mechanisms and ecological processes required for the growth of Enteromorpha prolifera. (3) Statistical prediction method. Statistical model relies on historical monitoring data to predict the probability of green tide occurrence through traditional machine learning methods such as regression analysis, support vector machine (SVM) and random forest. This method is relatively simple to implement and has high computational efficiency. However, it ignores the nonlinear coupling relationship between environmental factors and the time-dependent characteristics, which weakens its prediction ability, especially in the face of rapid environmental changes and extreme events.

[0004] Through in-depth analysis of the above-mentioned technology, it can be seen that there are still many key problems in the existing Enteromorpha green tide monitoring and prediction means. First of all, there are significant deficiencies in the modeling of biological mechanisms. Most methods ignore the complete life cycle process of Enteromorpha from germination, rapid growth to decline, and cannot truly reflect its ecological dynamics, resulting in deviation of the prediction results from the actual outbreak process. Secondly, the current model lacks sufficient consideration of spatial heterogeneity. The growth of Enteromorpha is greatly affected by spatial variables such as topography, water depth, shore distance, and light intensity. If these factors are not included, it is easy to cause regional misjudgment. Because of the limited multi-factor integration capability. Temperature, nutrients, light, salinity, dissolved oxygen and other environmental factors have important effects on the growth of Enteromorpha. Existing methods often have difficulty in effectively integrating and modeling these variables, limiting the model's ability to explain and generalize. In addition, most methods have limited ability in time series modeling. Lack of effective capture of historical environmental change trends and seasonal patterns makes it difficult for the prediction model to accurately grasp the long-term change rule. More importantly, the current mainstream method has low prediction accuracy, often difficult to break through 70% accuracy, and cannot meet the precise early warning needs in practical applications. At the same time, its prediction timeliness is poor, often only providing short-term forecasts of 1-3 days, lacking the time window for advance deployment and prevention. SUMMARY

[0005] Therefore, the present application provides a deep learning-based Enteromorpha green tide marine environment assessment and prediction method to improve the scientificity and forward-looking nature of green tide management and prevention.

[0006] In a first aspect, the present application provides a deep learning-based Enteromorpha green tide marine environment assessment and prediction method, which comprises:

[0007] Step 1, collecting multi-source marine environment data and preprocessing;

[0008] Step 2, according to step 1, constructing spatial feature engineering and time feature;

[0009] Step 3, based on step 2, establishing a growth cycle-based suitability scoring model, and combining sea area division to mark applicability;

[0010] Step 4, using step 3, constructing a bidirectional spatial convolution GRU prediction model, and optimizing it to assess and predict future marine environment, outputting the prediction results.

[0011] Optionally, the step 1 comprises:

[0012] The multi-source marine environment data includes seawater surface temperature data, marine optical data, nutrient salt concentration data, dissolved oxygen concentration, seawater salinity data, and the distribution area of Enteromorpha algae bloom on the sea; wherein the marine optical data includes photosynthetically active radiation and water transparency;

[0013] The preprocessing includes outlier detection and processing, adopting rule to remove extreme values; missing value filling, adopting bidirectional linear interpolation method; data normalization, normalizing all numerical variables to [0, 1]; time alignment, unified to daily or monthly scale; spatial interpolation, adopting bilinear interpolation method, mapping data to regular grid.

[0014] Optionally, the step 2 comprises:

[0015] Firstly, multi-dimensional spatial geographical features are extracted, including: geographical coordinates (latitude , longitude ); distance from the center of the Yellow Sea ; the shortest distance from the coastline ; season encoding, winter = 0, spring = 1, summer = 2, autumn = 3;

[0016] Supposing the latitude of the position point is , and the longitude is , the spatial features include:

[0017] a. Geographical coordinates: (latitude Latitude= , longitude Longitude= );

[0018] b. Distance from the center of the Yellow Sea: the center coordinates of the Yellow Sea are set as , the Haversine function is used to calculate the spherical distance between two points, and the expression is:

[0019] ;

[0020] wherein, R is the radius of the earth;

[0021] c. The shortest distance from the coastline , the expression is:

[0022] ;

[0023] wherein, ; C is the point set on the coastline; is the spherical distance; the minimum value is the nearest coastal point;

[0024] Time features: season encoding features are used as input variables of the time dimension; season encoding is winter = 0, spring = 1, summer = 2, autumn = 3; the time features discretize the four seasons of a year into four categories of values, so as to identify the regulation of different seasons on the growth of Enteromorpha by the model, and the mathematical expression form is:

[0025] Let the current month be m, then the season code is determined by the following function:

[0026] .

[0027] Optionally, the step 3 comprises:

[0028] The suitability score model is divided into 12 growth stages in units of months, and combined with the basic suitability score and the temperature response coefficient, a suitability score mechanism reflecting the whole year growth process is constructed;

[0029] Define the suitability function , wherein m is the month, taking values from 1 to 12; T is the average sea surface temperature in the month; L is the average light intensity in the month; N is the average nitrate concentration in the month; P is the average phosphate concentration in the month;

[0030] The monthly setting parameters include the basic suitability score , temperature factor , light factor , nutrient salt factor ; if the environmental mean vector contains temperature, light, nitrogen, and phosphorus, then the temperature vector ; the light vector ; the nitrogen salt vector ; the phosphorus salt vector ;

[0031] The suitability function is:

[0032] ;

[0033] The suitability function output value is used for suitability labeling, when is greater than or equal to the threshold , it is labeled as suitable, otherwise it is labeled as unsuitable.

[0034] Optionally, the step 4 comprises:

[0035] First, a feature fusion layer is constructed to integrate heterogeneous input, the feature fusion layer includes environmental features, spatial features, and suitability scores based on suitability score model, the above features are vectorized and dimensioned through multilayer perceptron MLP to ensure the integration of heterogeneous information, i.e. feature fusion;

[0036] Subsequently, a multi-scale convolution layer is embedded to capture spatial dependencies of different scales, convolution kernels of sizes 3x3, 5x5 and 7x7 are applied in parallel to extract local fine-grained changes and spatial trends, and their outputs are connected to form a multi-scale representation;

[0037] The core component of the bidirectional spatial convolution GRU prediction model is a bidirectional spatial convolution GRU network, in which a time series is processed by a forward sequence and a reverse sequence at the same time, wherein the forward sequence encodes historical dependency relationships, and the reverse sequence integrates future information; the outputs of the two sequence directions are spliced to capture complete time dependency relationships; a convolution operation is used to replace the fully connected layer in the bidirectional spatial convolution GRU network to retain spatial structure information during gating and state updating;

[0038] In order to further enhance the spatial differentiation, a spatial attention mechanism is introduced; the attention weight is applied to the output of the bidirectional spatial convolution GRU network, and a spatial matrix is generated;

[0039] Finally, the output layer designs a double-function head, which is a segmentation head and a prediction head; the segmentation head uses a convolution layer to generate a spatial distribution map for visualizing the suitability distribution; the prediction head integrates a convolution layer, a ReLU layer and a fully connected layer to generate a time series prediction.

[0040] Optionally, the core update of the basic unit of the bidirectional spatial convolution GRU network is expressed as:

[0041] ;

[0042] Among them, represents an update gate, represents a reset gate, represents a candidate hidden state, represents a final hidden state, represents the input feature vector of time t , represents the hidden state of the previous moment, and W represent trainable weight parameters, U represents a bias term, b represents element-wise multiplication, represents an activation function Sigmoid; In order to further introduce spatial structure information, the bidirectional spatial convolution GRU network is improved in two aspects:

[0043] a. Multi-scale spatial convolution embedding, different scales of spatial perception information are extracted from the input feature

[0044] by multi-scale convolution, and its expression is:

[0045] ;

[0046] Among them, represents a kernel size of ​two-dimensional convolution operation of the input data, K representing the scale number;

[0047] b, a spatial attention gating mechanism, introducing position attention gate to adjust the contribution of the spatial region to the updated state, and its expression is:

[0048] ;

[0049] ;

[0050] the final output representing the memory state of the current time step fusion spatial attention.

[0051] In a second aspect, an embodiment of the present application provides a computer readable storage medium, the computer readable storage medium comprising a stored program, wherein the program controls the device where the computer readable storage medium is located to execute the deep learning-based Enteromorpha green tide marine environment assessment and prediction method in the first aspect or any possible implementation manner of the first aspect when the program is running.

[0052] In a third aspect, an embodiment of the present application provides an electronic device, comprising: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and the one or more computer programs comprise instructions that, when executed by the device, cause the device to execute the deep learning-based Enteromorpha green tide marine environment assessment and prediction method in the first aspect or any possible implementation manner of the first aspect.

[0053] The technical scheme provided by the present application comprises collecting multi-source marine environment data and preprocessing; constructing spatial feature engineering and time feature; establishing a suitability scoring model based on the growth cycle, and combining the sea area division to mark the applicability; constructing a bidirectional spatial convolution GRU prediction model, and optimizing the model to predict the future marine environment, and output the prediction result. The method improves the scientificity and forward-looking of green tide management and prevention by constructing a bidirectional spatial convolution GRU prediction model. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0055] Figure 1 The schematic diagram of the deep learning-based Enteromorpha green tide marine environment assessment and prediction method provided by the embodiments of the present application;

[0056] Figure 2a A schematic diagram of a geographical coordinate feature provided for an embodiment of the present application;

[0057] Figure 2b A schematic diagram of a distance from the center of the Yellow Sea feature provided for an embodiment of the present application;

[0058] Figure 3 A curve of annual growth cycle adaptability of Enteromorpha provided for an embodiment of the present application;

[0059] Figure 4 A time distribution diagram of annual growth stages of Enteromorpha provided for an embodiment of the present application;

[0060] Figure 5a A change trend of temperature provided for an embodiment of the present application;

[0061] Figure 5b A change trend of dissolved oxygen provided for an embodiment of the present application;

[0062] Figure 5c A change trend of nutrient salt provided for an embodiment of the present application;

[0063] Figure 5d A change trend of phosphorus provided for an embodiment of the present application;

[0064] Figure 5e A change trend of nitrogen provided for an embodiment of the present application;

[0065] Figure 6 An architecture diagram of a bidirectional spatial convolution GRU prediction model provided for an embodiment of the present application;

[0066] Figure 7 A schematic diagram of a spatial convolution GRU network basic unit provided for an embodiment of the present application;

[0067] Figure 8 A schematic diagram of Enteromorpha suitability spatial gridding provided for an embodiment of the present application;

[0068] Figure 9 A schematic diagram of an Enteromorpha growth suitability evaluation result provided for an embodiment of the present application;

[0069] Figure 10 A schematic diagram of an electronic device provided for an embodiment of the present application. DETAILED DESCRIPTION

[0070] In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0071] It should be clear that the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0072] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "said" and "the" used in the embodiments of the present application are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0073] It should be understood that the term "and / or" used herein is only to describe the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B, which can represent the three cases of A alone, A and B together, and B alone. In addition, the character " / " in this paper generally represents that the front and rear associated objects are a "or" relationship.

[0074] Depending on the context, the word "if" as used herein can be interpreted as "when" or "upon" or "in response to a determination" or "in response to a detection". Similarly, depending on the context, the phrase "if determined" or "if detecting (a stated condition or event)" can be interpreted as "when determined" or "in response to a determination" or "when detecting (a stated condition or event)" or "in response to a detection (a stated condition or event)".

[0075] The present application provides a deep learning-based Enteromorpha green tide marine environment assessment and prediction method, as shown in Figure 1 The method comprises the following steps:

[0076] Step 1, collecting multi-source marine environment data and preprocessing.

[0077] In the embodiments of the present application, step 1 comprises:

[0078] The multi-source marine environment data comprises seawater surface temperature data, marine optical data, nutrient salt concentration data (such as nitrate and phosphate), dissolved oxygen concentration, seawater salinity data, and the distribution area of Enteromorpha algae bloom on the sea; wherein the marine optical data comprises photosynthetically active radiation and water transparency. The present application provides a deep learning-based Enteromorpha green tide marine environment assessment and prediction method, as shown in Figure 1 The method comprises the following steps:

[0076] Step 1, collecting multi-source marine environment data and preprocessing.

[0077] In the embodiments of the present application, step 1 comprises:

[0078] The multi-source marine environment data comprises seawater surface temperature data, marine optical data, nutrient salt concentration data (such as nitrate and phosphate), dissolved oxygen concentration, seawater salinity data, and the distribution area of Enteromorpha algae bloom on the sea; wherein the marine optical data comprises photosynthetically active radiation and water transparency.

[0079] Preprocessing includes outlier detection and handling, employing... The rules include: removing extreme values; filling missing values ​​using bilinear interpolation; normalizing data by normalizing all numerical variables to [0,1]; aligning time to a daily or monthly scale; and spatial interpolation using bilinear interpolation to map the data to a regular grid.

[0080] Step 2: Based on Step 1, construct spatial feature engineering and temporal features.

[0081] In this embodiment of the invention, step 2 includes:

[0082] First, extract multidimensional spatial geographic features, including: geographic coordinates (latitude) ,longitude ),like Figure 2a As shown; distance from the center of the Yellow Sea ,like Figure 2b As shown; the shortest distance to the coastline Seasonal codes: Winter = 0, Spring = 1, Summer = 2, Autumn = 3;

[0083] Let the latitude of the location point be... Longitude is The spatial features include:

[0084] a. Geographic coordinates: (Latitude= Longitude = );

[0085] b. Based on the distance from the center of the Yellow Sea: The coordinates of the center of the Yellow Sea are set as follows: The spherical distance between two points is calculated using the Haversine function, which is expressed as follows:

[0086] ;

[0087] Where R is the Earth's radius (approximately 6371 km).

[0088] c. Shortest distance from the coastline Its expression is:

[0089] ;

[0090] in, C represents the set of points on the coastline. The spherical distance; minimum value The nearest coastline;

[0091] A time feature is used to reflect the influence of seasonal changes in the growth process of Enteromorpha prolifera, and a season coding feature is used as an input variable in the time dimension, as shown in Table 1, season coding is winter = 0, spring = 1, summer = 2, and autumn = 3; the time feature discretizes the four seasons of a year into four types of values, so that the model can identify the regulation of different seasons on the growth of Enteromorpha prolifera, and the mathematical expression form is:

[0092] Suppose the month of the current time is m, then the season coding is determined by the following function:

[0093] .

[0094] After the feature extraction is completed, further feature evaluation is performed, and the feature evaluation includes:

[0095] a. Calculate the correlation coefficient between each feature and the Enteromorpha prolifera suitability score result;

[0096] b. Use principal component analysis (PCA) or mutual information (MI) method to evaluate the explanatory degree of the feature to the target variable;

[0097] c. According to the feature importance ranking, filter the features with low contribution degree, and ensure that the input features can retain the main environmental driving factors and reduce redundancy and noise.

[0098] Table 1 Season coding

[0099] .

[0100] In the embodiment of the present application, as shown in Figure 1 , the time series is combined after the time series information of various environmental elements is input; the evaluation is an evaluation on whether Enteromorpha prolifera exists.

[0101] Step 3, based on step 2, an suitability score model based on the growth cycle is established, and the applicability is marked in combination with the sea area division.

[0102] In the embodiment of the present application, the applicability is marked in combination with the sea area division, and the environmental conditions of the study sea area are divided into two categories: suitable and unsuitable.

[0103] In the embodiment of the present application, step 3 includes:

[0104] In order to effectively improve the biological accuracy of the Enteromorpha prolifera growth prediction, an suitability score model based on the growth cycle is established. The suitability score model is divided into 12 growth stages in units of months, and combined with the basic suitability score and the temperature response coefficient, an suitability score mechanism reflecting the whole year growth process is constructed;

[0105] The growth of Enteromorpha has significant seasonal characteristics, and it experiences a complete cycle of dormancy, germination, growth, flourishing, and decline in a year under different environmental conditions (temperature, light, nutrient salt, etc.). Therefore, the present application proposes a segmented function divided by months, and gives different basic growth suitability and basic potential in different growth periods.

[0106] Definition of suitability function , wherein m is the month, taking values from 1 to 12; T is the average sea surface temperature in the month; L is the average light intensity in the month; N is the average nitrate concentration in the month; P is the average phosphate concentration in the month;

[0107] The monthly set parameters include the basic suitability score (Base Suitability) , the temperature factor (Temperature Factor) , the light factor (Light Factor) , and the nutrition factor (Nutrition Factor) ; if the environmental mean vector contains temperature, light, nitrogen, and phosphorus, then the temperature vector ; the light vector ; the nitrogen salt vector ; and the phosphorus salt vector .

[0108] Suitability function is:

[0109] .

[0110] The output value of the suitability function is used for suitability labeling, and when is greater than or equal to the threshold value , it is labeled as suitable, otherwise it is labeled as unsuitable.

[0111] The values of the above parameters are set according to the measured data, and the specific values are shown in Table 2.

[0112] Table 2 Values of different factors in different periods

[0113] .

[0114] In order to improve the timing modeling ability of Enteromorpha growth suitability, the present application constructs a continuous suitability score model covering 12 months based on the life cycle rule of Enteromorpha, such as Figure 3 and Figure 4As shown, the peak period is in June, and the suitability score is 0.74; the dormant period is in December, and the suitability score is 0.02. The model takes biological characteristics as the core, dynamically integrates the temperature factor to adjust the monthly score, thereby enhancing the sensitivity to seasonal changes and environmental responses. At the same time, the method has good scalability and can be used in combination with time series models to realize continuous prediction and spatiotemporal modeling of the growth process of Enteromorpha, further improving the ecological consistency and accuracy of the prediction.

[0115] The method of the present application has significant advantages in ecological modeling, such as Figures 5a to 5e As shown, on the one hand, by introducing biological driving mechanisms, the model can more accurately reflect the change law of the growth intensity of Enteromorpha in different time periods throughout the year, improving the ecological rationality of the prediction results; on the other hand, the method has good adaptability and can be flexibly embedded in a multi-factor modeling framework to build an ecological-environmental coupled prediction system with environmental variables, providing more scientific decision support for green tide prevention and ecological early warning.

[0116] The suitability score model proposed in the present application can not only be used independently for Enteromorpha ecological suitability evaluation, but also can be used as prior input for deep learning models. By introducing the suitability score into the subsequent bidirectional spatial convolution GRU network, the model can provide time-seasonal constraints consistent with biological mechanisms, avoiding the ecological unreasonable prediction results caused by relying only on data fitting.

[0117] Step 4, using step 3, a bidirectional spatial convolution GRU prediction model is constructed and optimized to evaluate and predict future marine environments, and the prediction results are output.

[0118] In the embodiment of the present application, step 4 includes:

[0119] The present application designs a bidirectional spatial convolution GRU (BSC-GRU) prediction model, and the structure is as shown in Figure 6 The structure not only integrates multi-source environmental characteristics and spatial characteristics, but also embeds the biological suitability score obtained above as prior feature input. Through multi-scale convolution and spatial attention mechanism, BSC-GRU can learn complex spatiotemporal dependencies while maintaining the response consistency of the prediction results to the life cycle of Enteromorpha.

[0120] The BSC-GRU architecture of the present application integrates multiple functional modules to enhance spatio-temporal modeling capabilities. First, a feature fusion layer is constructed to integrate heterogeneous inputs, including environmental features (such as sea surface temperature, nitrate, phosphate, salinity, dissolved oxygen, and light), spatial features (such as latitude, longitude, and distance to coast), and suitability scores based on suitability scoring models. These features are vectorized and dimensioned using a multi-layer perceptron (MLP) to ensure the integration of heterogeneous information, i.e., feature fusion.

[0121] Subsequently, a multi-scale convolution layer is embedded to capture spatial dependencies at different scales, and convolution kernels of sizes 3x3, 5x5, and 7x7 are applied in parallel to extract local fine-grained changes and spatial trends. The outputs of these are concatenated to form a multi-scale representation. This design allows the model to consider both local dynamics at small scales and background processes at large scales simultaneously, thereby improving its adaptability to complex marine environments.

[0122] The core component of the bidirectional spatial convolution GRU prediction model is the bidirectional spatial convolution GRU network, in which the time series is processed by both forward and reverse sequences. The forward sequence encodes historical dependencies, while the reverse sequence integrates future information. The outputs of the two sequence directions are concatenated to capture the complete temporal dependencies, Figure 6 The input sequence includes a time series and an evaluation of the presence of Enteromorpha; convolution operations are used instead of fully connected layers in the bidirectional spatial convolution GRU network to preserve spatial structure information during gating and state updating.

[0123] To further enhance spatial differentiation, a spatial attention mechanism is introduced. Attention weights are applied to the output of the bidirectional spatial convolution GRU network, and a spatial matrix is generated. This matrix adaptively emphasizes key areas, such as high-risk areas near the coast and areas with environmental mutations. By redistributing attention between spatial domains, this mechanism improves the ecological relevance and robustness of the predictions. As shown in Figure 6As shown, the input is a C×H×W original feature map, where C is the number of channels (depth of the original feature map), H is the height (number of rows of the original feature map), and W is the width (number of columns of the original feature map). Feature mapping represents operations (such as convolution, pooling, or linear transformation) on the input original feature map to obtain intermediate features required for spatial attention calculation. Weight calculation represents calculating attention weights in the spatial dimension, that is, calculating a weight for each spatial location (h, w) to represent the importance of that location in the overall features. The attention weights are represented by grids representing different spatial locations, and different colors representing different weights; darker colored grids represent areas with higher importance and greater weights, while lighter colored areas represent areas with lower importance and smaller weights. Weighted fusion represents performing element-wise weighted operations on the input original feature map and the calculated attention weights, and then fusing them to obtain a new output feature map C×H×W. The output image size specification C×H×W means that the output feature map has the same size as the original input feature map, but after the spatial attention mechanism, the feature value at each position has been adjusted according to the attention weight.

[0124] Finally, the output layer features a dual-function head: a segmentation head and a prediction head. The segmentation head uses convolutional layers to generate a spatial distribution map for visualizing suitability distribution. The prediction head integrates convolutional layers, ReLU layers, and fully connected layers to generate time-series predictions and classifies the future marine environment as suitable or unsuitable. These output designs collectively provide spatially defined and temporally continuous predictions, offering comprehensive support for ecological monitoring and management. Figure 6 As shown, the segmentation head uses a 3×3 convolutional layer with a kernel size of 3 to extract spatial features from the input features, capturing local spatial information. The ReLU activation function provides a non-linear mapping, setting negative values ​​to 0 and retaining positive values ​​to increase the model's expressive power. A 1×1 convolutional layer with a kernel size of 1 performs feature mapping, mapping the convolution result to the category dimension required for the segmentation task. The prediction head uses adaptive pooling to unify input feature maps of different sizes into a fixed-size output, ensuring consistent input dimensions for subsequent fully connected layers. The flattening layer flattens multi-dimensional feature maps into one-dimensional vectors for easier input to the fully connected layer. The linear layer (appearing for the first time) is the fully connected layer, mapping the input features to a new feature space. The ReLU activation function provides a non-linear mapping, setting negative values ​​to 0 and retaining positive values ​​to increase the model's expressive power. The linear layer (appearing for the second time) outputs the prediction result, specifically the time series regression value. The prediction result is judged as appropriate or inappropriate.

[0125] In embodiments of the present invention, such as Figure 7The core update of the bidirectional spatial convolutional GRU network basic unit is shown, and its expression is:

[0126] ;

[0127] wherein, represents the update gate, represents the reset gate, represents the candidate hidden state, represents the final hidden state, represents the input feature vector of time t (including rasterized environmental elements such as SST, salinity, etc.), represents the hidden state of the previous time, W and U represent the trainable weight parameters, b represents the bias term, represents element-wise multiplication, represents the activation function Sigmoid;

[0128] In order to further introduce the spatial structure information, the bidirectional spatial convolutional GRU network is improved in two aspects:

[0129] a. Multi-scale spatial convolutional embedding, by multi-scale convolution on the input feature extract different scales of spatial perception information, and its expression is:

[0130] ;

[0131] wherein, represents a two-dimensional convolution operation with a kernel size of , and K represents the number of scales;

[0132] b. Spatial attention gating mechanism (SAG), a position attention gate is introduced to adjust the contribution of spatial regions to the update state, and its expression is:

[0133] ;

[0134] ;

[0135] The final output represents the memory state of the current time step fused with spatial attention.

[0136] The improved structure significantly enhances the modeling capability of the model for regional heterogeneity and temporal variability. In actual marine multivariate prediction tasks, compared with the traditional GRU, the BSC-GRU can significantly improve the root mean square error (RMSE) and RMSE and MAE in the above table are significantly improved, especially in complex sea areas (such as the Yellow Sea nearshore).

[0137] In the model training and optimization process, first, the mean squared error (MSE) is selected as the loss function to measure the difference between the predicted value and the true observed value. The MSE squares the prediction bias, which can give higher penalty weight to larger errors, so as to make the model converge to the true value faster in the training process, and has good performance in numerical stability and controllability of optimization.

[0138] In the parameter update aspect, the Adam adaptive optimization algorithm is preferably used, which combines the advantages of momentum gradient descent and RMSProp. Through the dynamic estimation of the first and second moments, the adaptive learning rate adjustment of the parameters is realized, which can effectively improve the convergence speed and reduce the oscillation phenomenon in the non-convex optimization space. In order to further improve the stability and generalization performance of the training, the learning rate scheduling strategy is introduced, which dynamically adjusts the learning rate according to the performance of the validation set. When the validation set error does not improve within the preset number of rounds, the learning rate is automatically reduced to promote the model to perform more detailed parameter search in the later stage. At the same time, the gradient clipping mechanism is introduced in the back propagation process to constrain the gradient norm and prevent numerical instability caused by gradient explosion.

[0139] In addition, the early stopping mechanism is also introduced, which automatically terminates the training and preserves the best model parameters when the validation set performance does not improve significantly within a certain number of continuous training cycles, which not only avoids the risk of overfitting of the model in the later training stage, but also significantly reduces the waste of computing resources. Through the synergistic effect of the above loss function design, optimization algorithm selection, learning rate scheduling, gradient clipping and early stopping mechanism, the model constructed by the present application has significant advantages in convergence speed, prediction accuracy and generalization ability, which can meet the high precision and high stability requirements of complex spatio-temporal prediction tasks.

[0140] In the grid prediction stage, the biological suitability factor and the deep learning prediction result jointly act, the former is the prior adjustment, and the latter is the spatio-temporal dynamic learning result, and after the fusion of the two, the ecological rationality and accuracy of the prediction are further improved. First, the spatial gridding method is used to systematically divide the target sea area to realize high-precision and full-coverage spatial prediction. Specifically, the target sea area is regularly gridded according to the latitude and longitude coordinate system, such as Figure 8As shown, the spatial resolution of each grid cell can be flexibly set according to application requirements, for example, it can be 0.01°x0.01° or higher accuracy, to balance the calculation cost and prediction accuracy.

[0141] In the prediction process, first, the corresponding spatial features (including geographic location, sea depth, distance from the shore, lighting conditions, etc.) and time features (including season encoding, historical environmental factor trends, etc.) of each grid cell are extracted, and these features are input into the trained bidirectional spatial convolution GRU prediction model to obtain the future time step prediction value of the corresponding grid. This method can ensure the continuity and consistency of the prediction results in space, while taking into account the influence of time series trends and local environmental differences.

[0142] In order to ensure that the prediction has the ability of long time span, the present application supports the introduction of multi-step rolling prediction mechanism in spatial grid prediction. That is, after completing the first prediction step, the prediction results are used as one of the inputs for the next step, and the iteration is run to the required time length, thereby realizing the whole cycle prediction from short term (1-3 days) to medium and long term (several weeks or even months). In this process, the model can capture the dynamic processes of spatial migration, local outbreak and decline of green tide at the same time, and meet the needs of marine ecological monitoring, emergency response and resource scheduling and other practical application requirements.

[0143] In addition, by griding the prediction results, high-resolution spatial distribution maps can be generated to provide intuitive and operable spatial information support for decision-making departments. This method significantly enhances the applicability and visualization value of the model in actual marine management while maintaining the prediction accuracy.

[0144] The visualization and analysis of the results of the present application present the prediction results of the model in an intuitive and easy-to-understand form, facilitating researchers, management departments and related industries to quickly obtain key information and assist decision-making. The system supports multiple visualization output methods, and combines with statistical analysis module to interpret the results and issue risk warnings, such as Figure 9 as shown.

[0145] The constructed bidirectional spatial convolution GRU (Bi-Spatial ConvGRU, BSC-GRU) combines multi-scale spatial convolution and spatial attention gating mechanism, realizes the fusion modeling of environmental features, spatial features and biological suitability factors, and realizes high-precision spatiotemporal prediction of the growth process of Enteromorpha. The method overcomes the shortcomings of traditional models in spatial dependence, time coupling, scale perception and regional difference modeling, and introduces biological prior constraints to ensure that the prediction results have ecological rationality, accurately reflect the biological characteristics and seasonal variation of Enteromorpha, effectively integrate multi-source marine environmental data and spatial geographic information, realize continuous prediction from short-term to long-term, and provide visual prediction results and risk assessment to assist decision-making.

[0146] Compared with the prior art, the present application has the following beneficial effects: the bidirectional time series modeling improves the ability to capture the dynamic of the life cycle of Enteromorpha; the multi-scale convolution structure is introduced to effectively consider local small-scale changes and large-scale trends; the spatial attention mechanism is used to highlight the key area prediction effect and improve the model generalization ability; the biological suitability model is combined to ensure that the prediction results have ecological rationality; and the continuous prediction from short-term to long-term is supported to provide scientific basis for green tide prevention and control.

[0147] In the technical scheme provided by the present application, the method comprises collecting multi-source marine environmental data and preprocessing; constructing spatial feature engineering and time feature; establishing a suitability scoring model based on the growth cycle and combining sea area division for applicability annotation; constructing a bidirectional spatial convolution GRU prediction model and training and optimizing it to evaluate and predict the future marine environment and output the prediction results. The method improves the scientificity and forward-looking nature of green tide management and prevention by constructing a bidirectional spatial convolution GRU prediction model.

[0148] The various steps of the embodiment of the present application can be executed by an electronic device. The electronic device includes but is not limited to a tablet computer, a portable PC, a desktop computer, etc.

[0149] The embodiment of the present application provides a computer readable storage medium, which comprises a stored program, wherein when the program runs, the electronic device where the computer readable storage medium is located executes the embodiment of the above-mentioned green tide marine environment evaluation and prediction method based on deep learning.

[0150] Figure 10 A schematic diagram of an electronic device provided by the embodiment of the present application is shown in Figure 10As shown, the electronic device 21 comprises a processor 211, a memory 212, and a computer program 213 stored in the memory 212 and executable on the processor 211, which, when executed by the processor 211, implements the deep learning-based Enteromorpha prolifera green tide marine environment assessment prediction method in the embodiments. For the sake of brevity and conciseness, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be described herein.

[0151] The electronic device 21 comprises, but is not limited to, the processor 211 and the memory 212. Those skilled in the art can understand that the electronic device 21 can further comprise other components, such as an input / output device, a network access device, a bus, etc. Figure 10 The electronic device 21 is only an example and does not constitute a limitation on the electronic device 21, and can comprise more or fewer components than those shown in the figure, or combine certain components or different components, for example, the electronic device can further comprise an input / output device, a network access device, a bus, etc.

[0152] The processor 211 can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor.

[0153] The memory 212 can be an internal storage unit of the electronic device 21, such as a hard disk or a memory of the electronic device 21. The memory 212 can also be an external storage device of the electronic device 21, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, etc. equipped on the electronic device 21. Further, the memory 212 can comprise both the internal storage unit and the external storage device of the electronic device 21. The memory 212 is used to store computer programs and other programs and data required by the network device. The memory 212 can also be used to temporarily store data that has been output or will be output.

[0154] Those skilled in the art can clearly understand that, for the sake of brevity and conciseness, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiments, which will not be described herein.

[0155] The above merely provides the preferred embodiment of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of protection of the present application.

Claims

1. A deep learning-based method for assessing and predicting marine environmental conditions related to *Ulva prolifera* green tides, characterized in that... The method includes: Step 1: Collect multi-source marine environmental data and perform preprocessing; Step 2: Based on Step 1, construct spatial feature engineering and temporal features; Step 3: Based on Step 2, establish a suitability scoring model based on the growth cycle, and combine it with sea area division to perform suitability labeling; Step 4: Using the method from Step 3, construct a bidirectional spatial convolutional GRU prediction model, train and optimize it to assess and predict the future marine environment, and output the prediction results. Step 4 includes: First, a feature fusion layer is constructed to integrate heterogeneous inputs. The feature fusion layer includes environmental features, spatial features, and suitability scores based on a suitability scoring model. The above features are vectorized and dimension-unified through a multilayer perceptron (MLP) to ensure the integration of heterogeneous information, i.e., feature fusion. Subsequently, by embedding multi-scale convolutional layers to capture spatial dependencies at different scales, convolutional kernels of sizes 3×3, 5×5 and 7×7 are applied in parallel to extract local fine-grained changes and spatial trends, and their outputs are concatenated to form a multi-scale representation. The core component of the bidirectional spatial convolutional GRU prediction model is the bidirectional spatial convolutional GRU network. In the bidirectional spatial convolutional GRU network, the time series is processed simultaneously by forward and backward sequences. The forward sequence encodes historical dependencies, while the backward sequence integrates future information. The outputs of the two sequences are concatenated to capture the complete temporal dependencies. Convolutional operations are used to replace the fully connected layers in the bidirectional spatial convolutional GRU network to preserve spatial structure information during gating and state updates. To further enhance spatial differentiation, a spatial attention mechanism is introduced; attention weights are applied to the output of the bidirectional spatial convolutional GRU network to generate a spatial matrix. Finally, the output layer is designed with a dual-function head: a segmentation head and a prediction head. The segmentation head uses convolutional layers to generate a spatial distribution map for visualizing the suitability distribution. The prediction head integrates convolutional layers, ReLU layers, and fully connected layers to generate time series predictions.

2. The method according to claim 1, characterized in that, Step 1 includes: Multi-source marine environmental data includes sea surface temperature data, marine optical data, nutrient concentration data, dissolved oxygen concentration, seawater salinity data, and the distribution area of ​​Ulva lactuca blooms in the ocean; among which, marine optical data includes photosynthetically active radiation and water transparency. Preprocessing includes outlier detection and handling, employing... The rules include: removing extreme values; filling missing values ​​using bilinear interpolation; normalizing data by normalizing all numerical variables to [0,1]; aligning time to a daily or monthly scale; and spatial interpolation using bilinear interpolation to map the data to a regular grid.

3. The method according to claim 2, characterized in that, Step 2 includes: First, extract multi-dimensional spatial geographic features, including: geographic coordinates (latitude) ,longitude ); Distance from the center of the Yellow Sea The shortest distance from the coastline Seasonal codes: Winter = 0, Spring = 1, Summer = 2, Autumn = 3; Let the latitude of the location point be... Longitude is The spatial features include: a. Geographic coordinates: (Latitude= Longitude = ); b. Based on the distance from the center of the Yellow Sea: The coordinates of the center of the Yellow Sea are set as follows: The spherical distance between two points is calculated using the Haversine semi-versine function, and its expression is: ； Where R is the Earth's radius; c. Shortest distance from the coastline Its expression is: ； in, C represents the set of points on the coastline. The spherical distance; minimum value The nearest coastline; The temporal feature uses seasonal coding features as the input variable for the time dimension; the seasonal coding is winter=0, spring=1, summer=2, autumn=3; the temporal feature discretizes the four seasons into four numerical categories so that the model can identify the regulation of Ulva prolifera growth by different seasons. Its mathematical expression is as follows: Let m be the month of the current time, then the season code is... Determined by the following function: 。 4. The method according to claim 3, characterized in that, Step 3 includes: The suitability scoring model is divided into 12 growth stages based on months, and combines the basic suitability score with the temperature response coefficient to construct a suitability scoring mechanism that reflects the growth process throughout the year. Define suitability function Where m represents the month, ranging from 1 to 12; T represents the average sea surface temperature for the month; L represents the average solar radiation intensity for the month; N represents the average nitrate concentration for the month; and P represents the average phosphate concentration for the month. Monthly parameters include basic suitability score. Temperature factor Light factor Nutrient factors If the environmental mean vector includes temperature, light, nitrogen, and phosphorus, then the temperature vector... Lighting vector Nitrogen salt vector ; Phosphate vector ; Suitability Function for: ； The suitability function output value is used for suitability labeling, when Greater than or equal to the threshold If the condition is met, mark it as appropriate; otherwise, mark it as inappropriate.

5. The method according to claim 1, characterized in that, The core update of the basic unit of the bidirectional spatial convolutional GRU network is expressed as follows: ； in, Indicates an update to the door. This indicates that the door is being reset. This represents a candidate hidden state. Indicates the final hidden state. Indicates time t The input feature vector, This represents the hidden state at the previous time step. W and U This represents the weight parameters that can be trained. b Indicates the bias term. This represents element-wise multiplication. This represents the activation function Sigmoid; To further incorporate spatial structure information, the bidirectional spatial convolutional GRU network is improved in two aspects: a) Multi-scale spatial convolutional embedding: This method embeds input features using multi-scale convolution. The expression for extracting spatial perception information at different scales is as follows: ； in, Indicates the kernel size as Two-dimensional convolution operation, K Indicates the scale number; b. Spatial attention gating mechanism, introducing positional attention gating. To adjust the contribution of the spatial region to the updated state, its expression is: ； ； Final output This indicates the memory state of the fused spatial attention at the current time step.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device on which the computer-readable storage medium is located to perform the deep learning-based marine environmental assessment and prediction method for Ulva prolifera green tides as described in any one of claims 1 to 5.

7. An electronic device, characterized in that, include: One or more processors; Memory; And one or more computer programs, wherein the one or more computer programs are stored in the memory, the one or more computer programs including instructions that, when executed by the device, cause the device to perform the deep learning-based marine environmental assessment and prediction method for Ulva prolifera green tides as described in any one of claims 1 to 5.

Images