A physical law constrained marine ecological variable collaborative prediction method and a prediction system thereof

Abstract

The application discloses a physical law constrained marine ecological variable collaborative prediction method and a prediction system thereof and belongs to the technical field of intelligent data processing. Step 1: multivariate feature extraction and cross-feature fusion are performed on chlorophyll-a concentration CHL-a, dissolved oxygen concentration O2 and net primary productivity NPPV; step 2: adaptive hierarchical weighting and global coding are performed on the features fused in step 1; step 3: prediction data reconstruction and construction of a physical constraint loss are performed based on the result obtained in step 2, and seven-day collaborative prediction of the chlorophyll-a concentration CHL-a, the dissolved oxygen concentration O2 and the net primary productivity NPPV is realized. The application aims at the marine environmental change characteristics, solves the problems of insufficient spatiotemporal heterogeneity adaptation of a traditional model, rough cross-variable correlation description and inconsistency between prediction results and actual laws and the like.

Description

Technical Field

[0001] This invention belongs to the field of intelligent data processing technology, specifically relating to a collaborative prediction method and prediction system for marine ecological variables constrained by physical laws. Background Technology

[0002] Accurate monitoring and forecasting of marine ecological variables are of great significance for ecological protection, global climate research, and human activities. Among these, analyzing the changing patterns of core ecological variables such as chlorophyll-a concentration, dissolved oxygen concentration, and net primary productivity is a core issue in this field. To predict these variables, existing technologies mainly fall into two categories: traditional numerical algorithms and deep learning algorithms. Both methods have demonstrated certain advantages during their development, but they also have significant limitations. Furthermore, the insufficient functionality of related supporting systems further restricts the performance of these algorithms.

[0003] Traditional numerical algorithms are based on marine ecology principles, simulating variable changes by solving ecological dynamic equations. For example, chlorophyll-a prediction methods based on phytoplankton growth models can simulate the effects of light and nutrients on algal reproduction; dissolved oxygen estimation methods based on biogeochemical cycle models can reflect the balance between photosynthesis and respiration. These methods have clear ecophysical significance and show stability in long-term trend predictions at macroscale. However, traditional methods struggle to capture short-term high-frequency changes and small-to-medium scale processes, such as algal blooms or the formation of hypoxic zones within a 1-5 km range. Furthermore, traditional methods are often designed for single variables and cannot adapt to dynamic relationships between variables (such as the seasonal variation in the effect of chlorophyll concentration on dissolved oxygen). Additionally, most methods rely on manual data processing and lack automatic data preprocessing, resulting in data preparation times on the order of hours, making it difficult to support efficient algorithm iteration.

[0004] In recent years, deep learning algorithms have gradually become a research hotspot due to their powerful nonlinear fitting capabilities, and have spurred the development of a number of new intelligent forecasting systems. Early research focused primarily on single-variable prediction. For example, Ding et al. constructed a CNN-LSTM deep hybrid network to predict chlorophyll concentration and algal blooms using satellite data; Ma et al. designed an Oxyformer network that can accurately learn dissolved oxygen-related information and estimate global high-resolution dissolved oxygen concentrations; Wang et al. also combined CNN and LSTM, using convolutional layer spatial feature extraction and LSTM temporal information processing as the core design. In scenarios where dissolved oxygen drops sharply due to typhoons, the model accuracy was significantly improved compared to a single LSTM. Compared to traditional algorithms, these methods have achieved superior performance in single-variable prediction.

[0005] With the development of deep learning technology, some studies have extended univariate prediction to multivariate prediction. Li et al. proposed a multi-task learning framework that strengthens key feature information by combining spatiotemporal attention mechanisms; the "Wenhai" large-scale model from Laoshan Laboratory, based on the Transformer architecture, improves variable prediction accuracy by strengthening global dependency characteristics; the "Xichen" system, also based on the Transformer architecture, demonstrates high efficiency in the field of meteorological multivariate prediction by strengthening the global spatiotemporal correlation of meteorological elements. Although existing multivariate prediction research attempts to innovate in terms of feature fusion and network architecture, many problems still exist: variable fusion strategies are mostly limited to conventional tensor splicing and fixed feature weighting, lacking targeted fusion designs based on oceanographic laws; network optimization has also not fully incorporated oceanographic constraints, leading to prediction results that may violate real-world laws. These shortcomings overlap and collectively restrict the accuracy and scientific validity of collaborative oceanographic variable prediction.

[0006] Meanwhile, regarding the integrated process of variable prediction, most existing studies are equipped with algorithm-based processing systems. Existing intelligent forecasting processing systems such as "XiChen," "FUXI-Ocean," and "WenHai" have achieved efficient automated data processing and batch prediction compared to traditional methods. However, they have significant shortcomings in the interaction of related variables and the coverage of core variables: Regarding the interaction of related variables, although XiChen, Pangu, FUXI-Ocean, and WenHai claim to support multivariate prediction, they generally lack effective fusion mechanisms between variables; for related variables, these systems have not designed targeted feature interaction mechanisms, resulting in isolated variable predictions; regarding the coverage of core variables, XiChen and Pangu focus on meteorological elements, while FUXI-Ocean and WenHai only cover marine physical variables, all lacking ecological elements such as chlorophyll and dissolved oxygen, and thus unable to analyze the ecological chain relationships of related elements.

[0007] Compared to traditional analysis and numerical simulation, deep learning methods have significant advantages in feature extraction and nonlinear modeling. However, in predicting marine ecological variables, they suffer from weak integration with physical laws and limited generalization performance. Furthermore, existing research often focuses on single-variable prediction, neglecting the intrinsic relationships between marine elements. More importantly, how to achieve collaborative prediction of variables while ensuring physical consistency remains a critical technological gap that urgently needs to be addressed.

[0008] Early deep learning methods often focused on single-variable predictions. For example, Ding et al.'s CNN-LSTM network only targeted chlorophyll concentration and did not consider the limiting effect of dissolved oxygen on phytoplankton growth, resulting in a sharp drop in accuracy in eutrophic waters. Ma et al.'s Oxyformer network, while efficient in predicting dissolved oxygen, did not correlate with variables such as net primary productivity, failing to support overall ecosystem analysis. Wang et al.'s CNN-LSTM model for predicting dissolved oxygen significantly improved accuracy in typhoon scenarios in the Northwest Pacific, but its regional adaptability was weak, with a significant decrease in performance in areas such as Haikou.

[0009] For multivariate prediction, existing methods lack specificity in feature fusion and network optimization. While Li et al.'s multi-task learning framework enhances feature information through spatiotemporal attention, feature fusion still relies on weighted operations, making it difficult to adaptively model the dynamic relationships between variables. Although the main models of prediction systems like "XiChen" and "WenHai" employ Transformer architecture to strengthen global dependencies, their feature fusion strategies all use concatenation, failing to flexibly capture nonlinear interactions between variables, leading to decreased prediction accuracy in complex regions. Furthermore, these methods have significant shortcomings in ensuring the physical consistency of multivariate prediction: multivariate predictions lack effective collaboration, with each variable prediction process relatively independent, potentially causing predictions of related variables to violate physical laws; and the system models do not embed relevant physical constraints in network optimization, failing to enforce physical consistency in the prediction results.

[0010] At the forecasting process design level, the functional limitations of the aforementioned systems further amplify the algorithmic deficiencies: In terms of variable coverage, these mainstream systems are all based on single-type elements. Among them, XiChen and PanGu focus on meteorological elements and lack ocean-related variables, while FUXI-Ocean, WenHai, and other systems only cover ocean physical elements and lack ocean ecological elements, thus failing to support ocean ecological analysis. From the perspective of overall functionality, XiChen and PanGu focus on global meteorological circulation simulation, while FUXI-Ocean, WenHai, and other systems focus on ocean physical element forecasting, all of which lack professional ocean ecological data processing and ecosystem collaborative prediction analysis. Summary of the Invention

[0011] This invention provides a method and system for collaborative prediction of marine ecological variables constrained by physical laws. It addresses the problems of insufficient spatiotemporal heterogeneity adaptation, coarse characterization of cross-variable correlations, and inconsistencies between prediction results and actual laws in traditional models, which are designed to address the characteristics of marine environmental changes.

[0012] This invention is achieved through the following technical solution: A method for collaborative prediction of marine ecological variables constrained by physical laws, the method comprising the following steps: Step 1: Combine chlorophyll-a concentration (CHL-a) and dissolved oxygen concentration. The net primary productivity (NPPV) is subjected to multivariate feature extraction and cross-feature fusion through first-level feature extraction, first-level fusion of chlorophyll-a concentration (CHL-a) and NPPV based on ConvLSTM, second-level feature extraction, second-level fusion of chlorophyll-a concentration (CHL-a) and NPPV based on spatiotemporal attention, third-level feature extraction, and third-level fusion of chlorophyll-a concentration (CHL-a) and NPPV based on spatiotemporal attention. Step 2: Perform adaptive hierarchical weighting and global encoding on the features fused in Step 1; Step 3: Based on the results obtained in Step 2, reconstruct the predicted data and construct the physical constraint loss to realize the chlorophyll-a concentration (CHL-a) and dissolved oxygen concentration. 7-day co-forecast of net primary productivity (NPPV).

[0013] Furthermore, before performing the first-level feature extraction, data input and channel expansion are first performed. Specifically, the original data of the three variables are first decoupled into average data AVG and deviation data Δ, and the 21-day deviation data is then processed. Channel dimension expansion is performed to transform the original shape [B, 21, 8, 8] into [B, 1, 21, 8, 8], unifying the data format to support parallel extraction of multivariate features.

[0014] Furthermore, the first-level feature extraction in step 1 specifically involves extracting the first-level features of the expanded three types of variable data through three-dimensional convolution and ReLU activation, expanding the number of channels from 1 to 3, forming a feature body with the shape [B, 3, 20, 7, 7], and using 2×2×2 convolution kernels to capture the temporal changes of adjacent days and the spatial correlation of the 2×2 grid; The first-level fusion of chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV) based on ConvLSTM in step 1 specifically involves normalizing the first-level features of CHL-a and NPPV to eliminate dimensional differences, then inputting the features into ConvLSTM for fusion through feature splicing, using a gating mechanism to screen key temporal information and filter short-term noise, and finally obtaining weighted fusion features through residual connections.

[0015] Furthermore, the second-level feature extraction in step 1 specifically involves first balancing spatial size and computational efficiency through three-dimensional adaptive pooling to obtain a feature body of [B, 3, 20, 6, 6], and then expanding the number of channels to 6 through convolution to output a feature body of shape [B, 6, 19, 5, 5]. The second-level fusion of chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV) based on spatiotemporal attention in step 1 specifically involves applying a joint spatiotemporal attention mechanism to the second-level features in order to enhance the feature representation of key regions. This mechanism includes channel attention for screening key ecological channels and spatial attention for increasing the weight of sensitive areas. The final output is the fused feature obtained through residual connections.

[0016] Furthermore, the third-level feature extraction in step 1 specifically involves further extracting the second-level features of the three variables, obtaining a feature body of [B, 3, 19, 4, 4] through three-dimensional adaptive pooling, and then expanding the number of channels to 9 through convolution to obtain a feature body of shape [B, 9, 18, 3, 3]. The third-level fusion of chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV) based on spatiotemporal attention in step 1 specifically involves applying spatiotemporal attention to fuse the third-level features again, capturing the key information change process through the attention mechanism.

[0017] Furthermore, step 2 includes the following steps: Step 2.1: Extract higher-order features from the features fused in Step 1; Step 2.2: Perform adaptive weighted hierarchical fusion on the features extracted in Step 2.1; Step 2.3: Perform Transformer global encoding on the features fused in Step 2.2.

[0018] Furthermore, step 2.1 specifically involves... The first, second, and third level features, and the first, second, and third level features after CHL-a and NPPV fusion, are uniformly mapped to the 24-channel space through the corresponding high-order encoders to form hierarchical standard features to support cross-level fusion.

[0019] Furthermore, step 2.2 specifically involves assigning learnable weight coefficients to the higher-order features of different variables at each level, and then summing them by weight to obtain the fusion features of the three variables. Each weight coefficient is dynamically updated during the training process. Furthermore, step 2.3 specifically involves inputting the fused features of the three variables into the corresponding Transformer, and capturing global spatiotemporal correlations through multi-head attention.

[0020] Furthermore, the specific steps of the prediction data reconstruction in step 3 involve generating a 7-day prediction bias from the Transformer output corresponding to each variable through a fully connected layer, then expanding it to a spatial grid and comparing it with the regional mean data of the corresponding variables. Superimpose and mask the land to eliminate invalid values. Finally, the prediction results for the three variables were obtained. ; The physical constraint loss in step 3 specifically involves designing a dual loss function based on the relationship between chlorophyll concentration and primary productivity empirical equations to optimize the network, thereby balancing numerical accuracy and ecological rationality.

[0021] A collaborative prediction system for marine ecological variables constrained by physical laws, the system employing the aforementioned collaborative prediction method for marine ecological variables constrained by physical laws, the system comprising: Data download module: It acquires multi-source ocean data through the Copernicus Marine service, allowing users to select data, choose specific variables, and freely set the spatiotemporal range for downloading; it adopts a multi-threaded design to ensure that the download process does not block the main interface; Monitoring and Analysis Module: Equipped with historical and real-time monitoring functions; the real-time monitoring automatically updates the inversion data every 2 hours, highlighting the latest monitoring points; it provides a dual view mode, which can display spatiotemporal data visualization: a geographic projection display based on Cartopy, allowing you to view the spatiotemporal changes of data through a time slider; and it also supports single-point time series analysis: displaying the changing trends of marine environmental and ecological variables at a specific geographical location. Data processing module: It can add pre-loaded variable data and build a complete data preprocessing workflow; from data standardization, geographic correction, gridding, regional segmentation to feature extraction, each step can be monitored in real time through logs, realizing a fully automated data processing workflow at the minute level; Model Prediction Module: The backend loads a pre-trained model architecture: a marine ecological variable collaborative prediction network based on physical law constraints, which can achieve full-area collaborative prediction of three core variables; users can visualize the 7-day prediction results of the three variables by adjusting the interactive time axis; at the same time, this module integrates marine knowledge expert analysis function based on the DeepSeek large model for the prediction results, answering user questions based on the prediction results and analysis word prompts, helping users to scientifically understand the prediction results and related ecological phenomena and significance; Help Center: Provides complete system documentation to help users quickly get started with the system and expand more business functions.

[0022] The beneficial effects of this invention are: This invention significantly improves prediction accuracy and physical consistency. Attached Figure Description

[0023] Figure 1 This is a schematic diagram for verifying the effectiveness of the physical information loss constraint of the present invention.

[0024] Figure 2 This is a flowchart of the method of the present invention.

[0025] Figure 3 This is a flowchart of the multivariate feature extraction and cross-feature fusion process of the present invention.

[0026] Figure 4 This is a flowchart of the adaptive hierarchical weighting and global encoding process of the present invention.

[0027] Figure 5 This is a flowchart of the physical information loss constraint of the present invention.

[0028] Figure 6 Table 1 shows the experimental results comparing the present invention with existing methods.

[0029] Figure 7 This is a schematic diagram of the overall framework of the system of the present invention.

[0030] Figure 8 This is the automated data processing page of the system of this invention.

[0031] Figure 9 This is the prediction result visualization and large model analysis page of the system of this invention.

[0032] Figure 10 The present invention is CHL-a, Visualization diagrams of the NPPV prediction results for days 1, 3, 5, and 7, where (a) is the visualization of the chlorophyll-a concentration prediction results for days 1, 3, 5, and 7, (b) is the visualization of the dissolved oxygen concentration prediction results for days 1, 3, 5, and 7, and (d) is the visualization of the net primary productivity prediction results for days 1, 3, 5, and 7. Detailed Implementation

[0033] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of this application with unnecessary detail.

[0034] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0035] It should also be understood that the terminology used in this application specification is for the purpose of describing particular embodiments only and is not intended to limit the application. As used in this application specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0036] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0037] Many specific details are set forth in the following description in order to provide a full understanding of this application. However, this application may also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0038] Implementation Method 1 To capture multi-level spatiotemporal features from local details to global trends, strengthen the cross-variable ecological association between chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV), and improve the model's ability to discriminate ecological anomalies in complex marine environments, inspired by the successful application of 3D convolution in spatiotemporal feature extraction, ConvLSTM in spatiotemporal association modeling, and attention mechanisms in key regions, we achieve multivariable feature mining and cross-variable dynamic modeling through multi-level feature extraction and cross-feature fusion.

[0039] This embodiment provides a method for collaborative prediction of marine ecological variables constrained by physical laws, the method comprising the following steps: Step 1: Combine chlorophyll-a concentration (CHL-a) and dissolved oxygen concentration. The net primary productivity (NPPV) is subjected to multivariate feature extraction and cross-feature fusion through first-level feature extraction, first-level fusion of chlorophyll-a concentration (CHL-a) and NPPV based on ConvLSTM, second-level feature extraction, second-level fusion of chlorophyll-a concentration (CHL-a) and NPPV based on spatiotemporal attention, third-level feature extraction, and third-level fusion of chlorophyll-a concentration (CHL-a) and NPPV based on spatiotemporal attention. The multivariate feature extraction and cross-feature fusion of this invention: First, the input data is processed to extract hierarchical features of different variables and model the ecological relationship between CHL-a and NPPV. While fully mining multi-level information, the intrinsic connection of key elements is established, laying the foundation for subsequent fusion and prediction.

[0040] Step 2: Perform adaptive hierarchical weighting and global encoding on the features fused in Step 1; The adaptive hierarchical weighting and global encoding of this invention integrates multi-scale information and strengthens global spatiotemporal dependence through adaptive multi-level feature fusion and Transformer global encoding, breaking down the barriers between features of different scales, enabling the model to better grasp the global spatiotemporal change patterns, and improving the model's adaptability to complex marine environmental changes.

[0041] Step 3: Based on the results obtained in Step 2, reconstruct the predicted data and construct the physical constraint loss to realize the chlorophyll-a concentration (CHL-a) and dissolved oxygen concentration. 7-day co-forecast of net primary productivity (NPPV).

[0042] The physical information loss constraint of this invention converts the prediction deviation into the actual prediction result by averaging it, and adds ecological variable proportional loss to constrain the physical laws, so that the prediction result is consistent with the marine ecological laws and improves the physical consistency of the prediction result.

[0043] Furthermore, the process is as follows: Figure 3 As shown, before performing the first-level feature extraction, data input and channel expansion are performed. Specifically, the original data of the three variables are first decoupled into average data (AVG) and bias data (Δ). Simultaneously, to ensure the input data conforms to the "batch-channel-time-height-width" format of 3D convolution, the 21-day bias data... ( i ∈{ΔCHL-a、Δ The channel dimension is expanded using ΔNPPV, transforming the original shape [B, 21, 8, 8] into [B, 1, 21, 8, 8], unifying the data format to support parallel extraction of multivariate features. The operation is as shown in equation (1):

[0044] Furthermore, the first-level feature extraction in step 1 specifically involves extracting the first-level features of the three types of variable data after expansion by using first-level three-dimensional convolution and ReLU activation to capture local detailed features of the nearshore sea area. The number of channels is expanded from 1 to 3, forming a feature body with the shape [B, 3, 20, 7, 7]. A 2×2×2 convolution kernel is used to capture the temporal changes of adjacent days and the spatial correlation of the 2×2 grid, as shown in formula (2).

[0045] in, These are the first-level features of the three variables. This is a 3D convolution operation. The kernel size is [size]. Step size, For fill size, For activation functions; The specific steps of step 1, which involves the first-level fusion of chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV) based on ConvLSTM, are as follows: In order to capture the long-term temporal correlation of "chlorophyll-a-net primary productivity", the first-level features of CHL-a and NPPV are first normalized to eliminate the difference in dimensions, and then the features are spliced ​​and input into ConvLSTM for fusion. The gating mechanism filters key temporal information and filters short-term noise. Finally, the weighted fusion features are obtained through residual connection, as shown in formula (3):

[0046] in, The first-level normalized feature of ΔCHL-a In scientific notation, it means "multiplied by ten". This represents the first-level normalized feature of ΔNPPV. The feature is the result of concatenating two normalized features. Output the hidden state at time t. The output of the cell state at time t. For ConvLSTM cells, The splicing and fusion features at time t are... This represents the first-level feature after ΔCHL-a fusion. To select the first three channels, This represents the first-level feature after ΔNPPV fusion; This process significantly improved the temporal correlation between CHL-a and NPPV, enhancing the model's ability to express the characteristics of long-term ecological events.

[0047] Furthermore, the second-level feature extraction in step 1 specifically involves further extracting the first-level features of the three variables to capture more feature information. First, a feature body of [B, 3, 20, 6, 6] is obtained by balancing spatial size and computational efficiency through three-dimensional adaptive pooling. Then, the number of channels is expanded to 6 through convolution, and a feature body of shape [B, 6, 19, 5, 5] is output, as shown in formula (4).

[0048] in, For adaptive 3D average pooling, This is a second-level feature; The second-level fusion of chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV) based on spatiotemporal attention in step 1 is as follows: to enhance the feature expression of key regions, a joint spatiotemporal attention mechanism is applied to the second-level features, wherein channel attention filters key ecological channels, spatial attention increases the weight of sensitive areas, and the final output is the fused feature obtained through residual connection, as shown in formula (5):

[0049] in, This is the second-level normalized feature of ΔCHL-a. This is the second-level normalized feature of ΔNPPV. For channel attention mechanism, For spatial attention mechanisms, The output after passing through the spatiotemporal attention module. This represents the second-level feature after ΔCHL-a fusion. This represents the second-level feature after ΔNPPV fusion; This operation amplifies subtle ecological changes and differences in various feature information by dynamically enhancing the intensity of feature responses, thereby significantly improving the ability to identify ecological anomalies in complex marine environments.

[0050] Furthermore, the third-level feature extraction in step 1 specifically involves further extracting the second-level features of the three variables, obtaining a feature body of [B, 3, 19, 4, 4] through three-dimensional adaptive pooling, and then expanding the number of channels to 9 through convolution to obtain a feature body of shape [B, 9, 18, 3, 3], as shown in formula (6):

[0051] in, These are the third-level features of the three variables; The third-level fusion of chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV) based on spatiotemporal attention in step 1 is as follows: To further enhance the variable correlation characteristics, the third-level features of spatiotemporal attention are fused again, and the key information change process is captured through the attention mechanism, as shown in formula (7):

[0052] in, This is the third-level normalized feature of ΔCHL-a. This is the third-level normalized feature of ΔNPPV. The feature is the result of concatenating two normalized features. This represents the third-level feature after ΔCHL-a fusion. This represents the third-level feature after ΔNPPV fusion. To retrieve the data from the first 9 channels.

[0053] Furthermore, the process is as follows: Figure 4 As shown, step 2 includes the following steps: Step 2.1: Extract higher-order features from the features fused in Step 1; Step 2.2: Perform adaptive weighted hierarchical fusion on the features extracted in Step 2.1; Step 2.3: Perform Transformer global encoding on the features fused in Step 2.2.

[0054] Furthermore, step 2.1 specifically involves, in order to eliminate the dimensional differences between features at different levels of the same variable, [the following is implied:] ... The first, second, and third level features, and the first, second, and third level features after CHL-a and NPPV fusion, are uniformly mapped to the 24-channel space through the corresponding high-order encoders to form hierarchical standard features to support cross-level fusion, as shown in formula (8):

[0055] in, The high-dimensional features are obtained by uniformly encoding the first-level fusion features of ΔCHL-a. This is the higher-order encoder corresponding to the first-level fusion feature of ΔCHL-a. The high-dimensional features are obtained by uniformly encoding the second-level fusion features of ΔCHL-a. This is the higher-order encoder corresponding to the second-level fusion feature of ΔCHL-a. This refers to the first-level fusion feature of ΔCHL-a obtained previously. This refers to the second-level fusion feature of ΔCHL-a obtained previously. The high-dimensional features are obtained by uniformly encoding the third-level fusion features of ΔCHL-a. This refers to the previously obtained ΔCHL-a third-level fusion feature. for The high-dimensional features are obtained after the first-level features are uniformly encoded. for The higher-order encoder corresponds to the first-level features. For what was obtained before First-level features, For what was obtained before Second-level features, for The high-dimensional features are obtained after the second-level features are uniformly encoded. for The higher-order encoder corresponds to the second-level features.

[0056] Furthermore, step 2.2 specifically involves assigning learnable weight coefficients to the higher-order features of different variables at each level to adaptively balance their contributions, and then summing the weighted values ​​to obtain the fused features of the three variables. Each weight coefficient is dynamically updated during the training process, as shown in equation (9).

[0057] in, The normalized learnable parameter set corresponds to the three variables. These are the weight parameters for the first-level high-order features. These are the weight parameters for the second-level higher-order features. These are the weight parameters for the third-level higher-order features. The multi-level fusion features correspond to the three variables. for , These are the first-level high-order coding features corresponding to the three variables. for , These are the second-level higher-order coding features corresponding to the three variables. for , These are the third-level higher-order coding features corresponding to the three variables; The above process can dynamically adjust the contribution ratio of different levels of features in the same variable (such as deeper features usually having higher weights), breaking through the limitations of traditional fixed weights.

[0058] Furthermore, step 2.3 specifically involves, in order to model global cross-regional correlations, inputting the fusion features of the three variables into the corresponding Transformers, and capturing global spatiotemporal correlations through multi-head attention, as shown in equation (10):

[0059] in, This refers to the multi-level fusion features after compressing the dimensions of the three variables. For dimensional compression operations, The multi-level fusion features correspond to the three variables. These are the global enhancement features obtained by applying the Transformer to the three variables. This is a neural network architecture based on a self-attention mechanism.

[0060] Furthermore, the specific method for reconstructing the predicted data in step 3 is as follows: to convert the biased predicted data into actual predicted data, the Transformer output corresponding to each variable is passed through a fully connected layer to generate a 7-day prediction bias, which is then extended to a spatial grid and compared with the regional mean data of the corresponding variables. Superimpose and mask the land to eliminate invalid values. Finally, the prediction results for the three variables were obtained. (j∈{CHL-a、 , NPPV}, where the order of variables in i and j corresponds one-to-one), the formula is as shown in equation (11):

[0061] in, The prediction results are obtained after spatial expansion. The prediction results are obtained after passing through the fully connected layer. For land masking, () represents a space expansion operation; The above process ultimately yielded predicted results for the three variables, with units consistent with the measured data, achieving a 100% interference elimination rate in the land area. Furthermore, the process is as follows: Figure 5 As shown, the physical constraint loss in step 3 is specifically designed to ensure that the variable prediction results conform to the physical laws of marine ecology. Based on the empirical equation of chlorophyll concentration-primary productivity, a double loss function is designed to optimize the network, balancing numerical accuracy and ecological rationality. Specifically, a 7-day prediction bias is generated through a fully connected layer, expanded to a spatial grid, and superimposed with the regional mean. Invalid values ​​are eliminated using a land mask, ultimately yielding prediction results consistent with the measured data units. The land area interference elimination rate reaches 100%, ensuring the results have practical physical meaning. The double loss function balances numerical accuracy and ecological rationality, where the numerical accuracy loss (… Minimize the error between the predicted value and the true value, and reduce the ecological rationality ratio loss. The NPPV / CHL-a ratio is constrained to be within a reasonable range; the total loss ( ) Optimize to ensure that the prediction meets both the numerical accuracy requirements and follows the laws of marine ecology, as shown in formula (12):

[0062] in, For the total amount of data, This is the seven-day forecast value for NPPV. This is the seven-day predicted value for CHL-a. This represents the seven-day true value of NPPV. The seven-day true value of CHL-a. For the ecological rationality ratio loss, This is the sum of the numerical precision losses for the three variables; Numerical precision loss ( MAE Minimize the error between the predicted value and the true value, and reduce the ecological rationality ratio loss. The NPPV / CHL-a ratio is constrained to be within a reasonable range.

[0063] To verify the effectiveness of the proposed method, it was compared with 10 novel spatiotemporal prediction methods on global ocean areas using relevant evaluation metrics, including 3D CNN-BiGRU, 3D CNN-BiLSTM, 3D CNN-GRU, 3D CNN-LSTM, 3DCNN, CoTCN, TemproNet, BTVFormer, EarthFormer, and LSTM-Transformer. The evaluation metrics used included mean absolute error (MAE), root mean square error (RMSE), mean relative error (MRE), normalized weighted composite score (NWCS), and model size (MB). The principle of NWCS is to normalize all method values ​​for each error metric to (0, 1), and then take the arithmetic mean of different normalized error metrics for the same method to obtain the score. A lower score indicates better overall model performance.

[0064] The specific comparison results are shown in Table 1. This method was applied to chlorophyll-a concentration (CHL-a) and dissolved oxygen concentration (CHL-a). In the 7-day co-forecasting mission of net primary productivity (NPPV), it significantly outperforms existing methods.

[0065] like Figure 6 Table 1 shows the experimental results compared with existing methods. Note: The top three results are represented in red, blue and green respectively in the table.

[0066] As shown in Table 1, the proposed method ranks among the top three in most indicators, effectively validating its superior performance. Compared with other methods, this method is the best in chlorophyll-a concentration (MAE and RMSE) and third in MRE; it is the best in all three dissolved oxygen concentration indicators; it ranks third in net primary productivity (MAE and RMSE); and it has the smallest NWCS. Furthermore, this method also has the smallest model size.

[0067] In addition, to verify the effectiveness of the physical information loss constraint, a loss ablation experiment was conducted, and the specific results are as follows: Figure 1As shown in the figure. In the figure, "origin" and "new" represent the model without and after using the physical information loss constraint, respectively. The horizontal axis represents the number of training iterations, and the vertical axis represents the scaling factor loss. ), This reflects the level of physical consistency; the closer to 0, the better the consistency. The trend of the curves in the graph clearly shows that after applying the physical information loss constraint, the proportional gain loss gradually approaches 0 as the model optimizes, indicating that the prediction results better conform to ecological constraints and the physical consistency of the model's predictions is better. In contrast, the proportional gain loss of the model without the physical information loss constraint oscillates continuously and may even exhibit abnormal losses.

[0068] Implementation Method 2 This embodiment provides a collaborative prediction system for marine ecological variables constrained by physical laws. The system uses a collaborative prediction method for marine ecological variables constrained by physical laws as described in Embodiment 1. The system adopts a hierarchical modular architecture, builds a GUI framework based on PyQt5, and integrates the entire workflow from data acquisition to advanced analysis. The overall system framework is as follows: Figure 7 As shown, it includes: Data Download Module: This module acquires multi-source ocean data through the Copernicus Marine service, allowing users to select data, choose specific variables, and freely set the spatiotemporal range for downloading; a multi-threaded design ensures that the download process does not block the main interface; Monitoring and Analysis Module: This module is equipped with historical and real-time monitoring functions. Real-time monitoring automatically updates the inversion data every 2 hours, highlighting the latest monitoring points. It provides a dual-view mode, which can display spatiotemporal data visualization: a geographic projection display based on Cartopy, allowing you to view the spatiotemporal changes of data through a time slider; and also supports single-point time series analysis: displaying the changing trends of marine environmental and ecological variables at a specific geographical location. Data processing module: This module can add pre-loaded variable data and build a complete data preprocessing workflow; from data standardization, geographic correction, gridding, regional segmentation to feature extraction, each step can be monitored in real time through logs, realizing a fully automated data processing workflow at the minute level; Model Prediction Module: This module loads a pre-trained model architecture on the backend: a marine ecological variable collaborative prediction network based on physical law constraints, which can achieve full-region collaborative prediction of three core variables. Users can visualize the 7-day prediction results of the three variables by adjusting the interactive time axis. At the same time, this module integrates marine knowledge expert analysis function based on the DeepSeek large model for the prediction results. It can answer user questions based on the prediction results and analysis word prompts, helping users to scientifically understand the prediction results and related ecological phenomena and significance. Help Center: This module provides complete system documentation, including interactive user guides, data source details, technical parameters, and access to technical support, helping users quickly get started with the system and expand more business functions.

[0069] Figure 8 This displays the system's automated data processing page. Figure 9 The system's prediction results visualization and large model analysis page are displayed. Figure 10 The system was demonstrated in the Bohai Sea region for the observation of CHL-a, Visualization of NPPV prediction results on days 1, 3, 5, and 7 (Test period: February 22, 2025 - February 28, 2025).

Claims

1. A method for collaborative prediction of marine ecological variables constrained by physical laws, characterized in that, The method includes the following steps: Step 1: Multivariate feature extraction and cross-feature fusion are performed on chlorophyll-a concentration (CHL-a), dissolved oxygen concentration (O2), and net primary productivity (NPPV) through first-level feature extraction, first-level fusion of chlorophyll-a concentration (CHL-a) and NPPV based on ConvLSTM, second-level feature extraction, second-level fusion of chlorophyll-a concentration (CHL-a) and NPPV based on spatiotemporal attention, third-level feature extraction, and third-level fusion of chlorophyll-a concentration (CHL-a) and NPPV based on spatiotemporal attention. Step 2: Perform adaptive hierarchical weighting and global encoding on the features fused in Step 1; Step 3: Based on the results obtained in Step 2, the predicted data is reconstructed and the physical constraint loss is constructed to achieve a 7-day co-forecast of chlorophyll-a concentration (CHL-a), dissolved oxygen concentration (O2), and net primary productivity (NPPV). The first-level fusion of chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV) based on ConvLSTM in step 1 specifically involves normalizing the first-level features of CHL-a and NPPV to eliminate dimensional differences, then inputting the features into ConvLSTM for fusion through feature splicing, using a gating mechanism to filter key temporal information and short-term noise, and finally obtaining weighted fusion features through residual connections. The second-level fusion of chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV) based on spatiotemporal attention in step 1 is as follows: In order to enhance the feature expression of key regions, a joint spatiotemporal attention mechanism is applied to the second-level features, wherein channel attention filters key ecological channels, spatial attention increases the weight of sensitive areas, and finally outputs the fused features obtained through residual connections. The third-level fusion of chlorophyll-a concentration (CHL-a) and net primary productivity (NPPV) based on spatiotemporal attention in step 1 specifically involves applying spatiotemporal attention to fuse the third-level features again, and capturing the key information change process through the attention mechanism. The specific steps of step 3, the reconstruction of the predicted data, involve generating a 7-day prediction bias from the Transformer output corresponding to each variable through a fully connected layer, then expanding it to a spatial grid and comparing it with the regional mean data of the corresponding variables. Superimpose and mask the land to eliminate invalid values. Finally, the prediction results for the three variables were obtained. ; The physical constraint loss in step 3 specifically involves designing a dual loss function based on the relationship between chlorophyll concentration and primary productivity empirical equation variables to optimize the network, thereby balancing numerical accuracy and ecological rationality.

2. The method according to claim 1, characterized in that, Before performing the first-level feature extraction, data input and channel expansion are performed, specifically as follows: First, the raw data of the three variables are decoupled into mean data AVG and deviation data Δ. Then, the 21-day deviation data is analyzed. ( i Channel dimension expansion is performed on ∈{ΔCHL-a, ΔO2, ΔNPPV}, transforming the original shape [B, 21, 8, 8] into [B, 1, 21, 8, 8], where B represents the batch, and the data format is unified to support parallel extraction of multivariate features.

3. The method according to claim 1, characterized in that, The first-level feature extraction in step 1 specifically involves extracting the first-level features of the three types of variable data after expansion by using three-dimensional convolution and ReLU activation, expanding the number of channels from 1 to 3, and using 2×2×2 convolution kernels to capture the temporal changes of adjacent days and the spatial correlation of the 2×2 grid.

4. The method according to claim 3, characterized in that, The second-level feature extraction in step 1 specifically involves first balancing spatial size and computational efficiency through three-dimensional adaptive pooling to obtain a feature volume of [B, 3, 20, 6, 6]. Then, the number of channels is expanded to 6 through convolution, outputting a feature volume of shape [B, 6, 19, 5, 5], where B represents the batch size.

5. The method according to claim 4, characterized in that, The third-level feature extraction in step 1 specifically involves further extracting the second-level features of the three variables, obtaining a feature body of [B, 3, 19, 4, 4] through three-dimensional adaptive pooling, and then expanding the number of channels to 9 through convolution to obtain a feature body of shape [B, 9, 18, 3, 3], where B represents the batch.

6. The method according to claim 1, characterized in that, Step 2 includes the following steps: Step 2.1: Extract higher-order features from the features fused in Step 1; Step 2.2: Perform adaptive weighted hierarchical fusion on the features extracted in Step 2.1; Step 2.3: Perform Transformer global encoding on the features fused in Step 2.

2.

7. The method according to claim 6, characterized in that, Specifically, step 2.1 involves mapping the first, second, and third level features of O2, and the first, second, and third level features after the fusion of CHL-a and NPPV, to a 24-channel space through the corresponding high-order encoders to form hierarchical standard features to support cross-level fusion.

8. The method according to claim 7, characterized in that, Specifically, step 2.2 involves assigning learnable weight coefficients to the higher-order features of different variables at each level, and then summing them by weight to obtain the fusion features of the three variables. Each weight coefficient is dynamically updated during the training process. Specifically, step 2.3 involves inputting the fused features of the three variables into the corresponding Transformer and capturing global spatiotemporal correlations through multi-head attention.

9. A collaborative prediction system for marine ecological variables constrained by physical laws, characterized in that, The system uses a collaborative prediction method for marine ecological variables constrained by physical laws as described in any one of claims 1-8, and the system comprises: Data download module: It acquires multi-source ocean data through the Copernicus Marine service, allowing users to select data, choose specific variables, and freely set the spatiotemporal range for downloading; it adopts a multi-threaded design to ensure that the download process does not block the main interface; Monitoring and Analysis Module: Equipped with historical and real-time monitoring functions; the real-time monitoring automatically updates the inversion data every 2 hours, highlighting the latest monitoring points; it provides a dual view mode, which can display spatiotemporal data visualization: a geographic projection display based on Cartopy, allowing you to view the spatiotemporal changes of data through a time slider; and it also supports single-point time series analysis: displaying the changing trends of marine environmental and ecological variables at a specific geographical location. Data processing module: It can add pre-loaded variable data and build a complete data preprocessing workflow; from data standardization, geographic correction, gridding, regional segmentation to feature extraction, each step can be monitored in real time through logs, realizing a fully automated data processing workflow at the minute level; Model Prediction Module: The backend loads a pre-trained model architecture: a marine ecological variable collaborative prediction network based on physical law constraints, which can achieve full-area collaborative prediction of three core variables; users can visualize the 7-day prediction results of the three variables by adjusting the interactive time axis; at the same time, this module integrates marine knowledge expert analysis function based on the DeepSeek large model for the prediction results, which can answer user questions based on the prediction results and analysis word prompts, helping users to scientifically understand the prediction results and related ecological phenomena and significance; Help Center: Provides complete system documentation to help users quickly get started with the system and expand more business functions.

Images