Method and system for reconstructing marine three-dimensional chlorophyll-a concentration based on deep learning

Abstract

This invention provides a method and system for reconstructing three-dimensional chlorophyll a concentration in the ocean based on deep learning. The method first acquires multi-source input data from the target sea area, including bio-optical parameters, vertical physical profile data, and geographic coordinates. Then, the multi-source data is preprocessed and feature fused to form a unified-dimensional input feature vector. This vector is then input into a trained deep neural network model, whose hidden layers include a spatial attention mechanism module and a residual correction structure module, to output predicted chlorophyll a concentration values ​​at multiple discrete depth layers. Finally, a three-dimensional chlorophyll a concentration field is reconstructed based on these predicted values. This invention integrates multi-source heterogeneous data and, by introducing an attention mechanism and residual structure, significantly improves the accuracy and spatiotemporal characterization capability of three-dimensional reconstruction of complex nonlinear marine ecological processes, providing reliable technical support for marine primary productivity estimation, carbon flux research, and ecosystem dynamic monitoring.

Description

Technical Field

[0001] This invention relates to the fields of marine remote sensing and artificial intelligence technology, and in particular to a method and system for reconstructing three-dimensional chlorophyll a concentration in the ocean based on deep learning. Background Technology

[0002] Chlorophyll a (Chla) concentration is a core ecological parameter characterizing marine phytoplankton biomass and primary productivity, crucial for studying global carbon cycling, marine ecosystem dynamics, and climate change response mechanisms. Currently, large-scale, continuous Chla observations primarily rely on satellite remote sensing technology, but this technology has a fundamental limitation: it can only provide sea surface information and cannot directly obtain the vertical profile structure of the water body. However, phytoplankton distribution exhibits significant heterogeneity, especially in regions such as the thermocline and halocline, where vertical gradients change dramatically. Simply extrapolating surface remote sensing information to the entire water column would severely distort estimates of key ecological processes such as primary productivity and carbon flux.

[0003] To compensate for this deficiency, existing technologies mainly attempt the following approaches: Empirical / semi-empirical models: such as using a generalized Gaussian model combined with surface Chla and remotely sensed reflectance ( R rs Inverting vertical profiles. This type of method has greatly different applicability in different sea areas and seasons, makes it difficult to capture complex physical-biological coupling processes, and is sensitive to errors in the input data, resulting in high uncertainty in the inversion results.

[0004] Numerical Models: The three-dimensional distribution of Chla is simulated using a marine biogeochemical circulation model. These methods are limited by the accuracy of model parameterization schemes, initial fields, and boundary conditions, and are computationally expensive, making it difficult to quickly generate high-resolution products for large-scale, long-term series.

[0005] Early attempts based on machine learning: Some studies have begun to use shallow models such as artificial neural networks (ANNs) to establish statistical relationships between surface parameters and the Chla profile. However, shallow networks have limited ability to represent complex nonlinear relationships and often fail to fully consider the inherent spatial heterogeneity of ocean data and the synergistic effects between multiple data sources.

[0006] In recent years, deep learning technology, especially deep neural networks (DNNs), has demonstrated powerful advantages in fields such as image recognition and natural language processing, as it can learn extremely complex mapping relationships through multi-layer nonlinear transformations. In the field of ocean remote sensing, existing research has explored the application of DNNs for sea surface parameter inversion, but challenges remain in three-dimensional field reconstruction. Insufficient model architecture targeting: The general DNN architecture is not optimized for the uneven spatial distribution of ocean data and the variation of feature importance with region / depth, resulting in insufficient attention to key areas and limited generalization ability.

[0007] Insufficient fusion of multi-source heterogeneous data: How to efficiently fuse data from satellites (e.g.) R rs The technical challenge lies in forming a unified feature vector that strongly characterizes the vertical distribution of Chla, using multi-source, heterogeneous, and different-scale data, including PAR, field observation or reanalysis data (such as temperature, salinity, and velocity profiles), and geographical location information.

[0008] The long-term spatiotemporal analysis capability is lacking: existing methods mostly focus on single or short-term profile inversion, and lack a systematic solution for analyzing long-term (such as interdecadal) spatiotemporal change trends based on reconstructed 3D data products.

[0009] In particular, for sea areas like the Northwest Pacific (NWPO) with complex circulation systems (such as the Kuroshio and Oyashio), high productivity, and significant impact on global climate, there is an urgent need for a technical solution that can stably and accurately reconstruct its long-term three-dimensional Chla structure and deeply analyze its spatiotemporal evolution mechanism. Summary of the Invention

[0010] The present invention aims to solve at least one of the technical problems existing in the prior art, and proposes a method and system for reconstructing three-dimensional chlorophyll a concentration in the ocean based on deep learning.

[0011] In a first aspect, the present invention provides a deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in marine environments, comprising: Acquire multi-source input data for the target sea area; wherein, the multi-source input data includes bio-optical parameters, vertical physical profile data, and geographic location coordinates; The multi-source input data is preprocessed and its features are fused to form an input feature vector of a unified dimension; The input feature vector is input into a pre-trained deep neural network model to obtain the predicted chlorophyll a concentration values ​​on multiple discrete deep layers output by the deep neural network model; wherein, the hidden layer of the deep neural network model includes a spatial attention mechanism module and a residual correction structure module; Based on the predicted chlorophyll a concentration values ​​at the multiple discrete depth layers, the three-dimensional chlorophyll a concentration field of the target sea area is reconstructed.

[0012] Furthermore, the bio-optical parameters include at least remote sensing reflectance. R rs Satellite inversion of surface chlorophyll a concentration (Chla) and photosynthetically active radiation (PAR); The vertical physical profile data includes at least seawater temperature, seawater salinity, eastward current velocity (ESWV), and northward current velocity (NSWV). The geographic coordinates include at least latitude and longitude.

[0013] Furthermore, the plurality of discrete depth layers cover a water depth range of 0-200 meters, including at least 30 non-uniformly distributed depth layers.

[0014] Furthermore, the deep neural network model is obtained through training in the following manner: Construct a training dataset, which includes samples of fused multi-source data from historical periods. The input features of each sample are the multi-source input data, and the output labels are the measured or reanalyzed chlorophyll a concentrations at the corresponding locations and times in multiple depth layers. Construct an initial deep neural network whose hidden layers contain a spatial attention mechanism module and a residual correction structure module; The initial deep neural network is trained using the training dataset. The root mean square error loss function between the predicted and true values ​​is minimized by using the Adam optimizer combined with the gradient descent algorithm, and dropout regularization is used to prevent overfitting until the model converges.

[0015] Furthermore, the hidden layers of the deep neural network model are set to 4 layers, with the number of neurons being 2048, 512, 128 and 32 respectively.

[0016] Furthermore, the method also includes: Based on the reconstructed three-dimensional chlorophyll a concentration field, the variation trend of chlorophyll a concentration at different depths in a specified sea area within a specific time period is calculated. Perform a significance test on the calculated trend of change; Output and visualize the spatial distribution trend after significance testing.

[0017] Furthermore, the method also includes: Obtain at least one independent validation dataset, which contains measured chlorophyll a profile data of a portion of the target sea area at certain locations and / or times; The multi-source input data corresponding to the validation dataset is input into the trained deep neural network model to obtain the predicted chlorophyll a profile; The predicted chlorophyll a profile was compared with the measured chlorophyll a profile in the validation dataset, and the root mean square error (RMSD), mean absolute relative error (MARD), bias, and coefficient of determination (R²) were calculated. 2 Metrics are used to evaluate the accuracy and robustness of the model.

[0018] Secondly, the present invention provides a deep learning-based system for reconstructing three-dimensional chlorophyll a concentration in marine environments, comprising: The data acquisition module is used to acquire multi-source input data of the target sea area; wherein, the multi-source input data includes bio-optical parameters, vertical physical profile data, and geographic location coordinates; The data processing module is used to preprocess and fuse the multi-source input data to form an input feature vector of a unified dimension. The model reconstruction module internally stores a pre-trained deep neural network model, which is used to receive the input feature vector output by the data processing module and output the predicted values ​​of chlorophyll a concentration on multiple discrete deep layers; wherein, the hidden layer of the deep neural network model includes a spatial attention mechanism module and a residual correction structure module. The three-dimensional field construction module is used to reconstruct the three-dimensional chlorophyll a concentration field of the target sea area based on the predicted chlorophyll a concentration values ​​on the multiple discrete depth layers.

[0019] Thirdly, the present invention provides an electronic device, comprising: One or more processors; A storage unit is used to store one or more programs that, when executed by one or more processors, enable the one or more processors to implement the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean as described above.

[0020] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, enables the implementation of the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean as described above.

[0021] Compared with the prior art, the present invention has the following beneficial effects: By fusing multi-source heterogeneous data (bio-optics, physical environment, and geographic location) and utilizing deep neural networks incorporating spatial attention mechanisms and residual structures, high-precision vertical chlorophyll a profile of water bodies can be effectively inferred from sea surface and near-surface information, significantly improving the reconstruction accuracy and spatial resolution of the three-dimensional chlorophyll a field. Deep neural networks can effectively learn the complex nonlinear relationship between the vertical distribution of chlorophyll a and multi-source driving factors, overcoming the limitations of traditional empirical models' poor adaptability to different sea areas and seasons. The spatial attention mechanism enables the model to adaptively focus on key feature regions that contribute more to predictions, enhancing the model's adaptability to the spatial heterogeneity of different sea areas. The residual structure ensures effective training of the deep network, improving the model's stability and generalization performance. Based on the reconstructed long-term series three-dimensional data, the long-term variation trend, seasonal characteristics, and relationship with environmental factors of chlorophyll a can be systematically analyzed, providing a powerful data tool for marine ecological research and climate change assessment. Attached Figure Description

[0022] Figure 1 This is a flowchart of a deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean according to an embodiment of the present invention; Figure 2 A schematic diagram of the deep neural network model structure constructed according to an embodiment of the present invention; Figure 3 Figures showing the comparison and verification results between reanalysis data and independent field observation data (JMA and BGC-Argo); (a) Japan Meteorological Agency (JMA) observation data; (b) BGC-Argo observation data; Figure 4 A scatter plot comparing the model's predictions at a depth of 5 meters in each month of 2021 with the reanalysis data; Figure 5 This is a summary graph of the importance of model input features evaluated using the SHAP method in this embodiment of the invention; Figure 6 The result is a graph showing the addition of random noise to the satellite input data to test the robustness of the model. Figure 7 Spatial distribution comparison and error distribution map of the three-dimensional chlorophyll a field reconstructed for the model and the reanalysis data at multiple representative depths in July 2021; Figure 8 A scatter plot showing the comparison of model predictions and reanalysis data at multiple typical depths, along with a schematic diagram of 3D structure reconstruction. Figure 9 The graph shows the variation of the overall accuracy evaluation index of the model with depth across 31 discrete depth layers from 0 to 200 meters. Figure 10This is a graph showing the time series variation of chlorophyll a at different depths in a specific oligotrophic sea area (12–21°N, 144–162°E) based on long-term data reconstructed from a model. Figure 11 This is a spatial distribution map showing the long-term trend of chlorophyll a concentration at different depths from 2000 to 2023, calculated based on model-reconstructed data. Figure 12 This is a schematic diagram of the structure of a deep learning-based marine three-dimensional chlorophyll a concentration reconstruction system according to another embodiment of the present invention. Detailed Implementation

[0023] Hereinafter, exemplary embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present invention, and not all embodiments of the present invention. It should be understood that the present invention is not limited to the exemplary embodiments described herein.

[0024] As indicated in this invention and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0025] While this invention makes various references to certain modules in systems according to embodiments of the invention, any number of different modules can be used and run on user terminals and / or servers. The modules are merely illustrative, and different aspects of the systems and methods may use different modules.

[0026] This invention uses flowcharts to illustrate the operations performed by the system according to embodiments of the invention. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously, as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0027] Existing marine three-dimensional chlorophyll a (Chla) reconstruction techniques, in highly nonlinear and heterogeneous marine environments, mostly rely on traditional empirical models or shallow neural networks. These techniques are generally based on linear or simplified nonlinear assumptions, making it difficult to fully characterize the complex and spatially heterogeneous coupling relationship between Chla vertical distribution and multiple driving factors (such as sea surface optical signals, water body physical profiles, and geographical location). This assumption has significant shortcomings in real marine systems because it ignores the different dominant physical-biological processes in different sea areas (such as the subtropical circulation zone and the Kuroshio-Oyashio confluence zone), leading to significant spatial variations and depth dependence in the relationship between Chla profile morphology and surface signals. Therefore, any technical approach that attempts to use a global, fixed model relationship to invert Chla concentrations at different depths across the entire sea area will inevitably result in model mismatches at different regions and depths, limiting reconstruction accuracy.

[0028] Therefore, the technical solution of this invention proposes a deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean. This method aims to overcome the accuracy limitations of traditional models due to their neglect of spatial heterogeneity and complex depth dependencies. It achieves high-precision three-dimensional reconstruction by constructing a deep learning model that can adaptively focus on key regions and possesses deep, stable learning capabilities. Specifically, this method first simultaneously acquires multi-source input data, including bio-optical parameters from satellites (such as multi-band remote sensing reflectance). R rs The system first obtains surface Chla (chlorophyll a), photosynthetically active radiation (PAR), vertical physical profiles (such as temperature, salinity, and current velocity) from field observations or reanalysis data, and geographic coordinates. Then, it preprocesses and fuses the features of the multi-source data to form a feature vector with a unified dimension. This feature vector is then input into a deep neural network model, which innovatively incorporates a spatial attention mechanism module and a residual correction structure module in its hidden layers. The spatial attention mechanism enables the model to automatically learn the importance weights of different geographic regions and features based on the input, thus effectively capturing spatial heterogeneity; the residual structure ensures the stability of the deep network training, enabling the model to learn the complex mapping from sea surface information to Chla concentrations at various depth layers. Finally, the model outputs predicted Chla concentration values ​​at multiple discrete depth layers, thereby reconstructing a three-dimensional chlorophyll a concentration field and solving the mismatch problem of traditional models in complex, nonlinear, and spatially heterogeneous marine environments.

[0029] Figure 1 This is a flowchart illustrating a deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in marine environments according to an embodiment of the present invention. Figure 1As shown, the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean according to an embodiment of the present invention includes the following steps: S100, acquiring multi-source input data of the target sea area, wherein the multi-source input data includes bio-optical parameters, vertical physical profile data, and geographic coordinates; S200, preprocessing and feature fusion of the multi-source input data to form an input feature vector of uniform dimension; S300, inputting the input feature vector into a pre-trained deep neural network model to obtain chlorophyll a concentration prediction values ​​at multiple discrete depth layers output by the deep neural network model, wherein the hidden layer of the deep neural network model includes a spatial attention mechanism module and a residual correction structure module; S400, reconstructing the three-dimensional chlorophyll a concentration field of the target sea area based on the chlorophyll a concentration prediction values ​​at the multiple discrete depth layers.

[0030] Specifically, in step S100, multi-source input data for the target sea area is acquired. It is understandable that the vertical distribution of marine chlorophyll a (Chla) is affected by sea surface optical properties (such as remote sensing reflectance). R rs The Chla layer exhibits complex and nonlinear synergistic regulation by multiple factors, including photosynthetically active radiation (PAR), water physical states (such as temperature and salinity profiles), and geographical location (latitude and longitude). Traditional single-source or simple combination data are insufficient to fully characterize this multidimensional coupling relationship, resulting in inadequate predictive power for deep Chla layers. Therefore, the technical solution of this invention acquires multi-source input data from the target sea area to provide subsequent deep learning models with comprehensive information sources from different observation platforms, reflecting different physical / biological processes. This establishes an information foundation for the model, enabling it to not only observe optical signals from the sea surface but also understand the underwater physical environment and spatial location, thus making it more likely to learn the complex patterns of inferring the entire Chla layer structure from surface information.

[0031] More specifically, in a specific example of the present invention, the acquisition of the multi-source input data includes: acquiring bio-optical parameters by downloading monthly average L3 data from publicly available satellite fusion products (such as GlobColour), including remote sensing reflectance in the 412, 443, 490, 555, and 670 nm bands. R rsThe data included photosynthetically active radiation (PAR) and surface Chla concentration retrieved using the OCx algorithm, with a spatial resolution of approximately 9 km. Vertical physical profile data was acquired by downloading monthly average data from global ocean reanalysis datasets (such as the Copernicus Marine Service's GLORYS12V1 product), including profiles of sea surface temperature, salinity, eastward current velocity (ESWV), and northward current velocity (NSWV) at 31 standard depths within the 0-200 meter range, with a spatial resolution of approximately 0.25° × 0.25°. Geographic coordinates were directly mapped to the latitude and longitude of each data grid point. These data underwent temporal matching (aligned to the 15th of each month) and spatial interpolation (interpolating the reanalysis data to a grid matching the satellite data), and were standardized (subtracting the mean and dividing by the standard deviation) to form a spatiotemporally consistent sample data pool. Validation data came from an independent Japan Meteorological Agency (JMA) field profile dataset, with sampling locations widely distributed across the Northwest Pacific, containing 893 profiles, used to independently evaluate model performance.

[0032] Specifically, in step S200, the multi-source input data is preprocessed and its features are fused to form an input feature vector of uniform dimension. It is understood that the original multi-source input data obtained in the previous step comes from different databases and has different dimensions (e.g., ...). R rs The data consists of independent datasets with different spatial resolutions and physical meanings, representing a two-dimensional sea surface field and a three-dimensional temperature profile. These datasets cannot be directly used as model input and require structuring and fusion. Deep learning models require a fixed-dimensional, numerical input vector to perform forward propagation. Therefore, in the technical solution of this invention, the multi-source input data is further preprocessed and feature-fused to form a unified-dimensional input feature vector. This transforms and concatenates these heterogeneous, multi-dimensional data into a one-dimensional feature vector corresponding to a single sample point through spatiotemporal matching, standardization, feature extraction, and dimensionality reduction. This vector incorporates sea surface optical signals (…). R rs The system includes wavelength bands, PAR (particulate matter ratio), surface Chla (water surface area), and water physical parameters at specific depths (such as temperature and salinity values ​​in multiple layers from 0 to 200 m), as well as latitude and longitude information. This provides a structured, information-rich, and uniformly formatted input for subsequent deep neural networks, enabling them to simultaneously consider the combined effects of surface light field, water physical structure, and geographical location on the vertical distribution of Chla, thus laying a data foundation for high-precision modeling.

[0033] Specifically, in step S300, the input feature vector is input to a pre-trained deep neural network model to obtain predicted chlorophyll a concentration values ​​at multiple discrete depth layers output by the deep neural network model. It is understood that although the input feature vector formed in the preceding steps incorporates multi-source information, its mapping relationship with the target (Chla concentration at each depth layer within 0-200 meters) is extremely complex, exhibiting high nonlinearity and spatial heterogeneity. This relationship cannot be accurately expressed by simple linear regression or shallow networks. Traditional empirical models or shallow ANNs, due to their limited capacity and lack of specialized handling for spatial heterogeneity, struggle to learn the changes in dominant factors and coupling strengths in different sea areas. Therefore, in the technical solution of this invention, the input feature vector is further input to a pre-trained deep neural network model to obtain predicted chlorophyll a concentration values ​​at multiple discrete depth layers. Specifically, as... Figure 2 As shown, a deep neural network model is used, whose hidden layers include a spatial attention mechanism module and a residual correction structure module. This model inputs fused feature vectors into the network, performs multiple nonlinear transformations, and ultimately outputs Chla concentration predictions for 31 discrete depth layers (e.g., 1, 2, 3, ..., 200 meters) simultaneously. This allows the powerful function approximation capability of deep networks to characterize complex nonlinear relationships; simultaneously, the spatial attention mechanism enables the model to adaptively "focus" on regional features more important to the current prediction point (e.g., temperature in subtropical regions, salinity in nearshore regions, etc.). R rs (443)) thus effectively handles spatial heterogeneity; the residual structure ensures gradient stability during the deep network training process, avoiding gradient vanishing / exploding, enabling the model to effectively converge and learn complex relationships from the surface to the deep layers. Ultimately, high-precision, integrated prediction of the vertical Chla structure of water bodies is achieved.

[0034] To gain a deeper understanding of the model's internal decision-making mechanism and verify whether the learned feature-target relationship aligns with physical cognition, this invention employs the SHAP (Shapley Additive exPlanations) method to perform interpretability analysis on the trained model. For example... Figure 5 As shown, this analysis summarizes the average absolute contribution (SHAP) of each input feature to the model's predicted Chla concentration. The results indicate that latitude, sea surface temperature, and remotely sensed reflectance in key bands (such as...) all contribute to the model's prediction of Chla concentration. R rs(443) was identified as the feature that has the most significant impact on the model output. This finding is highly consistent with the classic understanding in oceanography that Chla distribution is dominated by latitude (sunlight, climate zone), water temperature (affecting phytoplankton growth rate), and sea surface optical properties (directly reflecting surface phytoplankton abundance). This indirectly proves that the deep neural network model of this invention is not a "black box" and that its learning process has clear physical rationality, and can effectively capture the key factors driving the vertical distribution of Chla in the ocean.

[0035] More specifically, in a specific example of the present invention, the deep neural network model is pre-trained as follows: A training dataset is constructed, with sample inputs being the uniform-dimensional input feature vectors formed in step S200, and sample labels being high-quality multi-layer Chla concentration data (e.g., from reanalysis data products or BGC-Argo measured profiles) corresponding to the location and time. An initial deep neural network is constructed, embedding the spatial attention mechanism module and the residual correction structure module in its hidden layers. The hidden layers are set to 4 layers, with the number of neurons being 2048, 512, 128, and 32, respectively. The output layer has 31 neurons, corresponding to the preset 31 depth layers. The initial deep neural network model is trained using the training dataset, minimizing the root mean square error loss function between the predicted and true values ​​using the Adam optimizer combined with the gradient descent algorithm, and employing dropout regularization (dropout rate = 0.1) to prevent overfitting, until the model's performance on the validation set converges. The spatial attention mechanism module learns a weight matrix to weight the importance of different geographical regions or feature combinations; the residual correction structure module introduces skip connections to directly pass the output of a certain layer to deeper layers to alleviate the gradient problem.

[0036] Specifically, in step S400, the three-dimensional chlorophyll a concentration field of the target sea area is reconstructed based on the predicted chlorophyll a concentration values ​​at the multiple discrete depth layers. It is understood that the model output in the previous step is the predicted Chla concentration values ​​for all spatial grid points within the study area at multiple preset discrete depth layers. These predicted values ​​are independent, point-like data and have not yet been integrated into a continuous three-dimensional spatial data field. To intuitively display and analyze the three-dimensional distribution structure of Chla across the entire water body, it is necessary to systematically organize and reconstruct the prediction results of these discrete points according to their spatial location (longitude, latitude) and depth information. Therefore, in the technical solution of this invention, the three-dimensional chlorophyll a concentration field of the target sea area is further reconstructed based on the predicted chlorophyll a concentration values ​​at the multiple discrete depth layers, thereby integrating and gridding the predicted values ​​at each depth layer output by the model for all grid points according to three-dimensional spatial coordinates (longitude, latitude, depth). Specifically, for each spatial location (a latitude and longitude grid point) in the target sea area, there are model-predicted Chla concentration values ​​for 31 depth layers, forming a vertical profile. By arranging the vertical profile data of all spatial grid points according to rules, a three-dimensional data cube (i.e., a three-dimensional field) covering the entire study area and containing multiple depth layers can be constructed. In this way, the model's point-like prediction results can be transformed into a standardized three-dimensional spatial data product that can be used for visualization, quantitative analysis, and further scientific research (such as calculating the water column integral Chla, analyzing vertical structural characteristics, etc.), thereby ultimately achieving a complete reconstruction from multi-source surface information to a three-dimensional marine ecological parameter field.

[0037] More specifically, in a concrete example of the present invention, the specific processing procedure of the reconstruction process is as follows: For the entire target sea area (e.g., the Northwest Pacific region: 120°E-180°E, 0°N-50°N), at each sampling time (e.g., July 2021), the processing unit receives 31-dimensional Chla prediction vectors from all grid points (e.g., approximately tens of thousands of grid points) output by the model reconstruction module. Each vector represents the Chla concentration profile from the surface to 200 meters at that point. Subsequently, the processing unit performs spatial interpolation (e.g., bilinear interpolation) and sorts all data points according to a preset three-dimensional grid coordinate system (longitude grid, latitude grid, depth layer sequence), filling them into a three-dimensional array (or a NetCDF / HDF5 format file). The first dimension of this array is longitude, the second dimension is latitude, and the third dimension is depth. Through this operation, a spatially continuous three-dimensional chlorophyll a concentration field data product is finally generated. This product can be used to draw horizontal distribution maps, vertical profile maps, animation displays, and subsequent spatiotemporal trend analysis, etc. For example, it can be observed that at a depth of 5 meters, Chla exhibits low values ​​in the subtropical circulation zone and high values ​​in the Kuroshio-Oyashio confluence zone; while at a depth of 200 meters, the overall Chla concentration is low and the distribution tends to be uniform.

[0038] To visually demonstrate the spatial distribution characteristics and accuracy of the reconstruction results, such as Figure 7 As shown, the horizontal distribution of the reconstructed 3D Chla field and the reanalysis data at several representative depths (5 m, 54 m, 108 m, and 200 m) in July 2021 was compared, and the spatial error distribution was analyzed. The results show that the model reconstruction results and the reanalysis data are highly consistent in terms of horizontal spatial pattern, and the error is at a low level in most sea areas, verifying the model's ability to capture the spatial distribution characteristics of Chla. Furthermore, Figure 8 (a)-(d) show scatter plots comparing model predictions and reanalysis data at the four typical depths mentioned above. The data points are closely distributed along the 1:1 line and remain stable across different concentration ranges. To more precisely evaluate the ability to reconstruct vertical structures, Figure 8 (e) and (f) show the vertical distribution of Chla concentration along specific longitude and latitude sections, respectively. The model reconstruction results are compared with the reanalysis data, and the two are in good agreement in terms of vertical gradient and core features. Figure 8 (g) vividly illustrates the reconstructed three-dimensional structure of Chla, clearly showing the vertical gradient of concentration decay with increasing depth and the complex spatial heterogeneity. The final data product can be used for various analyses. For example, it can be observed that at a depth of 5 meters, Chla exhibits low values ​​in the subtropical circulation zone and high values ​​in the Kuroshio-Oyashio confluence zone; while at a depth of 200 meters, the overall Chla concentration is low and the distribution tends to be uniform.

[0039] Specifically, in a specific example of the present invention, the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean further includes: step S500: performing long-term spatiotemporal variation characteristic analysis based on the reconstructed three-dimensional chlorophyll a concentration field.

[0040] More specifically, step S500 includes: calculating the variation trend of chlorophyll a concentration at different depths in a specified sea area within a specific time period based on the reconstructed three-dimensional chlorophyll a concentration field; performing a significance test on the calculated variation trend; and outputting and visually displaying a variation trend map with spatial distribution characteristics after the significance test.

[0041] Understandably, since the preceding steps have already generated a long-term, high-spatial-resolution three-dimensional Chla concentration field data product, this dataset can not only be used for single-time-point and single-profile analysis, but also contains rich information on long-term evolution and spatial distribution patterns. If only static or isolated analyses are performed, the data value cannot be fully explored to reveal the spatiotemporal dynamics of the regional marine ecosystem and its correlation with environmental factors. Therefore, in the technical solution of this invention, based on the reconstructed three-dimensional chlorophyll a concentration field, the changing trend of chlorophyll a concentration in a specified sea area at different depths over a specific time period is calculated, and its significance is tested and visualized, thereby achieving a quantitative characterization and spatial pattern mining of the long-term evolution process of the marine ecosystem.

[0042] More specifically, in a specific example of the present invention, the long-term variation characteristic analysis is performed according to the following steps: Trend Calculation: First, extract the monthly mean Chla concentration time series data for all grid points at the target depth (e.g., 5 meters) in a specified study area (e.g., a rectangular sea area within a specific latitude and longitude range, or a polygonal region) from January 2000 to December 2023. Then, apply a linear regression model to the time series data for each grid point to calculate the rate of change (slope) of Chla concentration over time at each spatial location, in mg / m³. -3 yr -1 This slope represents the long-term linear trend of change at that location during the analysis period.

[0043] Significance Test: To avoid interference from natural fluctuations in trend judgment, a statistical significance test is performed on the linear trend of each grid point calculated above. For example, a non-parametric Mann-Kendall trend test or a t-test of the linear regression results can be used, with the significance level set to p < 0.05. Only trends that pass the significance test are considered to have statistical significance.

[0044] Visualization: Finally, the trend results that passed the significance test are combined with their spatial location and output and visualized in the form of a geographic raster image. For example, ... Figure 11 As shown in the figure, this figure illustrates the long-term trends (mg / m³) of Chla concentration at depths of 5 meters, 54 meters, and 108 meters between 2000 and 2023, calculated based on model reconstruction data. -3 yr -1 The spatial distribution of Chla concentration is shown in the figure. Warm colors (such as red) indicate a significant upward trend, cool colors (such as blue) indicate a significant downward trend, and white areas indicate no significant trend. This figure visually reveals the vertical differences and spatial heterogeneity of Chla concentration changes. For example, there is a significant increasing trend of Chla in the surface layer (5 meters) of tropical seas (about 10°N and south), while the surface layer of the subtropical circulation region (about 20°–30°N) shows a significant decreasing trend.

[0045] To analyze the details of changes in specific typical sea areas in greater depth, key regions can be selected for time series analysis. For example... Figure 10 As shown, taking a typical oligotrophic sea area in the Northwest Pacific (12–21°N, 144–162°E, the yellow rectangular area in subfigure a) as an example, Figure 10 Images (b)-(e) present the monthly average time series of Chla concentrations in this region from 2000 to 2023 at depths of 5m, 54m, 108m, and 200m, respectively, clearly depicting the seasonal fluctuations and long-term evolution of Chla concentrations at different depths. This analytical step transforms the reconstructed static three-dimensional field into a dynamic scientific understanding with spatiotemporal evolution patterns, providing crucial evidence for studying marine ecological responses under the background of climate change.

[0046] Specifically, in a specific example of the present invention, the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean further includes: step S600: evaluating the model accuracy and robustness based on an independent validation dataset.

[0047] More specifically, step S600 includes: acquiring at least one independent validation dataset, the validation dataset containing measured chlorophyll a profile data of some locations and / or times in the target sea area; inputting the multi-source input data corresponding to the validation dataset into the trained deep neural network model to obtain a predicted chlorophyll a profile; comparing the predicted chlorophyll a profile with the measured chlorophyll a profile in the validation dataset, and calculating the root mean square error RMSD, mean absolute relative error MARD, bias, and coefficient of determination R. 2 Metrics are used to evaluate the accuracy and robustness of the model.

[0048] Understandably, although trained deep neural network models perform well on the training set and some reanalysis validation data, their true generalization ability and reliability must be rigorously verified on real-world test data completely independent of the training process. Existing model validation techniques often limit themselves to comparisons with reanalysis data from the same source, lacking systematic, multi-indicator quantitative comparisons with field-measured profile data, thus failing to comprehensively evaluate the model's accuracy and robustness in practical applications. Therefore, in the technical solution of this invention, at least one independent validation dataset is further obtained, and this dataset is used to evaluate the model's accuracy and robustness, thereby establishing an objective and rigorous third-party evaluation mechanism.

[0049] More specifically, in a specific example of the present invention, the verification process is as follows: Obtaining an independent validation dataset: First, obtain a third-party measured dataset completely independent of the model training data source. In this example, shipborne field observation Chla profile data from 2005 to 2022 provided by the Japan Meteorological Agency (JMA), and use data not used in model training (e.g., reanalysis data for specific years, such as 2021-2023 data) as an independent validation set. The vertical observation depth of the JMA dataset is the standard layer (0, 10, 20, 30, 50, 75, 100, 150, 200 meters). Spatiotemporally match these measured data with the model input data according to time and geographical location (longitude, latitude).

[0050] Model Prediction and Comparison: Subsequently, the multi-source input data (including...) corresponding to each sample point in the above-matched validation dataset will be... R rs (Inputs such as temperature and salinity) are fed into the trained deep neural network model to obtain the Chla concentration profile predicted by the model at the corresponding location, time, and depth layer. The model-predicted profile is then compared point-by-point and layer-by-layer with the measured profile in the validation dataset.

[0051] Accuracy metrics calculation: Finally, based on paired predicted and measured values, a set of widely accepted, multi-faceted accuracy evaluation metrics are calculated, including: Root mean square error (RMSD): Reflects the overall average deviation between predicted and measured values. For example... Figure 3 As shown, the RMSD of the model evaluated on the JMA dataset is approximately 0.49 mg m. -3 This indicates that there is an average difference of approximately 0.49 mg m between the predicted and measured values. -3 The deviation demonstrates the model's good generalization ability in real marine environments.

[0052] Mean Absolute Relative Error (MARD) and Bias: These reflect the average relative magnitude of the prediction error with respect to the measured value. They are dimensionless and facilitate comparison of accuracy across different concentration ranges. Bias reflects the average direction and magnitude of the systematic deviation of the predicted value from the measured value. For example, ... Figure 9 As shown, the model's MARD is below 0.2 in most shallow layers (such as below 20 meters) across 31 depth layers in the 0-200 meter water column range, meaning that the average relative error is better than 20%. Meanwhile, the model's Bias is close to 0 in all depth layers, indicating that the model's predictions do not have a significant tendency to overestimate or underestimate, and the prediction results are unbiased.

[0053] Coefficient of determination (R) 2 (This reflects the strength of the linear relationship between the model's predicted values ​​and the measured values, and is used to evaluate the model's goodness of fit. For example, ...) Figure 4 As shown, the model's predictions at a depth of 5 meters are highly consistent with the reanalysis data for each month of 2021, with the monthly scatter points closely distributed along a 1:1 line. R 2 Generally higher than 0.78 (e.g., January R) 2 The value can reach 0.81, indicating that the model can explain most of the changes in the observed data and is stable in different seasons.

[0054] Overall accuracy profile: Figure 9 The exhibition comprehensively showcased RMSD, MARD, Bias, and R. 2 The profile varies with depth. It can be seen that RMSD has a peak at the subsurface layer (approximately 27 meters deep), at about 0.13 mg m. -3 However, it remains at a low level; while R 2 The values ​​remain high within the true light layer. This systematically quantifies the model's performance at different depths, confirming that its accuracy meets the requirements for scientific applications.

[0055] To test the model's robustness to uncertainties in the input data, key input variables (such as satellite remote sensing reflectance) were examined. R rs (443) Artificial random noise of varying amplitudes was applied. For example... Figure 6 As shown, when R rs When the deviation of (443) reaches 10% (this range is close to the actual error level of satellite products), the average change of the Chla concentration prediction value output by the model is controlled within 5%, which shows that the model of the present invention has good anti-interference ability against input data error and strong robustness.

[0056] Through the above systematic evaluation, the model's performance on real-world data can be quantified, its accuracy can be confirmed to meet the requirements of scientific applications, and the robustness of the model to unknown data can be evaluated, thus providing solid empirical evidence for the promotion and application of the model in a wider range of time and space.

[0057] In summary, the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean according to embodiments of the present invention acquires and fuses multi-source heterogeneous input data, and utilizes a deep neural network integrating spatial attention mechanisms and residual structures for modeling. This method effectively learns and infers high-precision vertical chlorophyll a profile from sea surface and near-surface information, ultimately reconstructing a continuous three-dimensional concentration field. This method overcomes the limitations of traditional models in complex, nonlinear, and spatially heterogeneous marine environments, providing a powerful technical means for refined research and long-term change monitoring of marine ecosystems.

[0058] Based on the same inventive concept, such as Figure 12 As shown, this invention provides a deep learning-based system for reconstructing three-dimensional chlorophyll a concentration in the ocean, comprising: The data acquisition module 100 is used to acquire multi-source input data of the target sea area; wherein, the multi-source input data includes bio-optical parameters, vertical physical profile data, and geographic location coordinates; The data processing module 200 is used to preprocess and fuse the multi-source input data to form a unified dimension input feature vector. The model reconstruction module 300 internally stores a pre-trained deep neural network model, which is used to receive the input feature vector output by the data processing module and output the predicted values ​​of chlorophyll a concentration on multiple discrete deep layers; wherein, the hidden layer of the deep neural network model includes a spatial attention mechanism module and a residual correction structure module. The three-dimensional field construction module 400 is used to reconstruct the three-dimensional chlorophyll a concentration field of the target sea area based on the predicted values ​​of chlorophyll a concentration on the multiple discrete depth layers.

[0059] The deep learning-based marine three-dimensional chlorophyll a concentration reconstruction system according to embodiments of the present invention acquires and fuses multi-source heterogeneous input data, and utilizes a deep neural network integrating spatial attention mechanisms and residual structures for modeling. This system effectively learns and infers high-precision vertical chlorophyll a profile from sea surface and near-surface information, ultimately reconstructing a continuous three-dimensional concentration field. This method overcomes the limitations of traditional models in complex, nonlinear, and spatially heterogeneous marine environments, providing a powerful technical means for refined research and long-term change monitoring of marine ecosystems.

[0060] Based on the same inventive concept, the present invention provides an electronic device, comprising: One or more processors; A storage unit is used to store one or more programs that, when executed by one or more processors, enable the one or more processors to implement the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean as described above.

[0061] Based on the same inventive concept, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can realize the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean as described above.

[0062] The computer-readable medium may be included in the apparatus, device, or system of this invention / utility model, or it may exist independently.

[0063] The computer-readable storage medium may be any tangible medium that contains or stores a program, and may be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples include, but are not limited to, electrical connections having one or more wires, portable computer disks, hard disks, optical fibers, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0064] The computer-readable storage medium may also include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code, specific examples of which include, but are not limited to, electromagnetic signals, optical signals, or any suitable combination thereof. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in marine environments, characterized in that, include: Acquire multi-source input data for the target sea area; wherein, the multi-source input data includes bio-optical parameters, vertical physical profile data, and geographic location coordinates; The multi-source input data is preprocessed and its features are fused to form an input feature vector of a unified dimension; The input feature vector is input into a pre-trained deep neural network model to obtain the predicted chlorophyll a concentration values ​​on multiple discrete deep layers output by the deep neural network model; wherein, the hidden layer of the deep neural network model includes a spatial attention mechanism module and a residual correction structure module; Based on the predicted chlorophyll a concentration values ​​at the multiple discrete depth layers, the three-dimensional chlorophyll a concentration field of the target sea area is reconstructed.

2. The method for reconstructing three-dimensional chlorophyll a concentration in the ocean based on deep learning according to claim 1, characterized in that, The bio-optical parameters include at least remote sensing reflectance. R rs Satellite inversion of surface chlorophyll a concentration (Chla) and photosynthetically active radiation (PAR); The vertical physical profile data includes at least seawater temperature, seawater salinity, eastward current velocity (ESWV), and northward current velocity (NSWV). The geographic coordinates include at least latitude and longitude.

3. The method for reconstructing three-dimensional chlorophyll a concentration in the ocean based on deep learning according to claim 1, characterized in that, The multiple discrete depth layers cover a water depth range of 0-200 meters, including at least 30 non-uniformly distributed depth layers.

4. The method for reconstructing three-dimensional chlorophyll a concentration in the ocean based on deep learning according to claim 1, characterized in that, The deep neural network model is obtained through training in the following manner: Construct a training dataset, which includes samples of fused multi-source data from historical periods. The input features of each sample are the multi-source input data, and the output labels are the measured or reanalyzed chlorophyll a concentrations at the corresponding locations and times in multiple depth layers. Construct an initial deep neural network whose hidden layers contain a spatial attention mechanism module and a residual correction structure module; The initial deep neural network is trained using the training dataset. The root mean square error loss function between the predicted and true values ​​is minimized by using the Adam optimizer combined with the gradient descent algorithm, and dropout regularization is used to prevent overfitting until the model converges.

5. The method for reconstructing three-dimensional chlorophyll a concentration in the ocean based on deep learning according to claim 1, characterized in that, The deep neural network model has 4 hidden layers with 2048, 512, 128 and 32 neurons respectively.

6. The method for reconstructing three-dimensional chlorophyll a concentration in the ocean based on deep learning according to claim 1, characterized in that, The method further includes: Based on the reconstructed three-dimensional chlorophyll a concentration field, the variation trend of chlorophyll a concentration at different depths in a specified sea area within a specific time period is calculated. Perform a significance test on the calculated trend of change; Output and visualize the spatial distribution trend after significance testing.

7. The method for reconstructing three-dimensional chlorophyll a concentration in the ocean based on deep learning according to claim 1, characterized in that, The method further includes: Obtain at least one independent validation dataset, which contains measured chlorophyll a profile data of a portion of the target sea area at certain locations and / or times; The multi-source input data corresponding to the validation dataset is input into the trained deep neural network model to obtain the predicted chlorophyll a profile; The predicted chlorophyll a profile was compared with the measured chlorophyll a profile in the validation dataset, and the root mean square error (RMSD), mean absolute relative error (MARD), bias, and coefficient of determination (R²) were calculated. 2 Metrics are used to evaluate the accuracy and robustness of the model.

8. A deep learning-based system for reconstructing three-dimensional chlorophyll a concentration in marine environments, characterized in that, include: The data acquisition module is used to acquire multi-source input data of the target sea area; wherein, the multi-source input data includes bio-optical parameters, vertical physical profile data, and geographic location coordinates; The data processing module is used to preprocess and fuse the multi-source input data to form an input feature vector of a unified dimension. The model reconstruction module internally stores a pre-trained deep neural network model, which is used to receive the input feature vector output by the data processing module and output the predicted values ​​of chlorophyll a concentration on multiple discrete deep layers; wherein, the hidden layer of the deep neural network model includes a spatial attention mechanism module and a residual correction structure module. The three-dimensional field construction module is used to reconstruct the three-dimensional chlorophyll a concentration field of the target sea area based on the predicted chlorophyll a concentration values ​​on the multiple discrete depth layers.

9. An electronic device, characterized in that, include: One or more processors; A storage unit for storing one or more programs that, when executed by one or more processors, enable the one or more processors to implement the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it can realize the deep learning-based method for reconstructing three-dimensional chlorophyll a concentration in the ocean as described in any one of claims 1 to 7.

Images