A method and system for identifying phytoplankton density based on ocean water temperature

Abstract

The application provides a phytoplankton density identification method and system based on ocean water temperature, the method comprising: obtaining water temperature data sets of a preset three-dimensional space of each grid region in a target ocean region within a preset time range; extracting features of each water temperature data set to obtain time sequence dynamic features, vertical stratification features, and spatial correlation features of the ocean water temperature of each grid region; inputting each time sequence dynamic feature, vertical stratification feature, and spatial correlation feature into a preset identification model to enable the identification model to perform feature fusion on each time sequence dynamic feature, vertical stratification feature, and spatial correlation feature based on an attention mechanism and a full connection network, obtain each fusion feature vector, and generate the phytoplankton density of each grid region according to each fusion feature vector, thereby improving the accuracy and efficiency of phytoplankton density identification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of phytoplankton density recognition and machine learning technology, and particularly relates to a phytoplankton density recognition method and system based on ocean water temperature. BACKGROUND

[0002] As the primary producer of marine ecosystems, the density dynamics of phytoplankton directly affect carbon cycling, red tide early warning, and biogeochemical cycling, and is a key indicator of global climate change research. Traditional phytoplankton density monitoring mainly relies on manual microscopic counting or flow cytometry, which has low efficiency, limited spatial and temporal coverage, and is difficult to meet the real-time monitoring needs of large-scale sea areas. Although rapid detection technologies such as spectral analysis and fluorescence labeling have developed in recent years, they are costly and lack adaptability to complex environments, especially in dynamic water temperature and salinity stratification, nutrient salt gradient, and other multi-coupling scenarios, making it difficult to accurately capture the nonlinear response characteristics of phytoplankton communities.

[0003] Existing water temperature data inversion methods based on remote sensing or in-situ sensors can partially correlate with phytoplankton density, but face two major technical bottlenecks: first, traditional statistical models (such as Gaussian regression and logistic regression) have limited modeling capabilities for water temperature dynamics, making it difficult to capture the complex relationship between vertical stratification parameters such as thermocline strength and mixed layer depth (MLD) and density distribution; second, the cost of fine labeling of phytoplankton density in massive monitoring data is extremely high, especially for high spatiotemporal resolution water temperature sequences. Sparse labeling leads to a significant decrease in the generalization performance of supervised learning models, making it difficult to meet the needs of cross-regional migration and long-term ecological assessment.

[0004] In recent years, deep learning technology has shown potential in underwater image classification and ecological modeling. For example, convolutional neural networks (CNN) have achieved high accuracy in phytoplankton shape recognition. In this context, the Learning from Label Proportions (LLP) algorithm uses a weakly supervised paradigm, requiring only regional density level proportion labels rather than individual sample labels, providing a new approach to reducing data labeling costs. Label proportion learning is a weakly supervised learning paradigm that trains models using class proportion information from group samples rather than individual labels, significantly reducing data labeling costs. In LLP, data is divided into multiple "bags", each providing only the global distribution of class labels without specific labels for each sample. This learning method is particularly suitable for fields such as ecological monitoring and remote sensing image analysis, which have high labeling difficulty and large data size. Therefore, it is possible to apply the label proportion learning algorithm to the phytoplankton density recognition method. SUMMARY

[0005] To solve the above technical problems, the application provides a method and system for identifying phytoplankton density based on ocean water temperature, which improves the accuracy and efficiency of phytoplankton density identification.

[0006] In a first aspect, the application provides a method for identifying phytoplankton density based on ocean water temperature, comprising:

[0007] obtaining water temperature data sets of a preset three-dimensional space in a target marine area in a preset time range;

[0008] extracting features from each water temperature data set to obtain time series dynamic features, vertical stratification features, and spatial correlation features of the ocean water temperature of each grid area;

[0009] inputting each time series dynamic feature, vertical stratification feature, and spatial correlation feature into a preset identification model to enable the identification model to perform feature fusion on each time series dynamic feature, vertical stratification feature, and spatial correlation feature based on an attention mechanism and a fully connected network, obtain each fusion feature vector, and generate the phytoplankton density of each grid area according to each fusion feature vector;

[0010] The identification model is obtained by training an initial identification model based on historical phytoplankton density data and historical water temperature data sets of a target marine area in different periods. Specifically, the historical phytoplankton density data is used as a proportion label, and the historical water temperature data sets are used as data features to construct a plurality of label proportion training data packets and perform label proportion learning training on the initial identification model to obtain the identification model. The initial identification model is obtained based on a deep learning model.

[0011] The embodiment of the present application provides a phytoplankton density identification method based on ocean water temperature, water temperature data of each position and each time point in a target ocean area is collected, a water temperature data set is constructed, time sequence dynamic features, vertical layering features and spatial correlation features are extracted from the water temperature data set, each feature is identified through an identification model, and phytoplankton density of different grid regions in the target ocean area is generated. The embodiment of the present application comprehensively captures the complex relationship between ocean water temperature and phytoplankton density by fusing time sequence dynamics, vertical layering and spatial correlation multi-dimensional features, and significantly improves the identification accuracy. The embodiment of the present application further divides the target ocean area into a plurality of grid regions in advance, and then identifies the phytoplankton density of different grid regions. This grid data processing enables the embodiment of the present application to support high-resolution fine identification, meet the real-time monitoring needs of a large range of sea areas, and better adapt to the label proportion learning algorithm. In addition, the embodiment of the present application introduces the label proportion learning algorithm based on the density characteristics of the identification task, and combines a deep learning model to construct an identification model, thereby greatly reducing the dependence on fine annotation data, reducing the data annotation cost, enhancing the generalization ability of the model in the sparse annotation scene, and improving the accuracy and efficiency of the phytoplankton density identification.

[0012] Further, the water temperature data set of each grid region in the target ocean area in the preset three-dimensional space within the preset time range comprises:

[0013] extracting a sea surface temperature data set of each of the preset three-dimensional space within the preset time range from satellite remote sensing data;

[0014] obtaining a layered water temperature data set of each of the preset three-dimensional space within the preset time range through a plurality of preset sensors;

[0015] combining each of the sea surface temperature data set and each of the layered water temperature data set to construct a water temperature data set of each of the grid region in the preset three-dimensional space within the preset time range.

[0016] The embodiment of the present application provides a water temperature data set acquisition method, which combines satellite remote sensing and a plurality of preset sensors for data acquisition, can fully collect water temperature data at different positions in each grid region, obtain multi-dimensional and multi-scale water temperature information covering the ocean surface and vertical space, improve the integrity and representativeness of the water temperature data set through complementary fusion of sea surface and layered data, lay a reliable foundation for subsequent feature extraction, and improve the accuracy of phytoplankton density identification. In addition, the embodiment of the present application simultaneously uses sensor technology and remote sensing technology to collect water temperature data, realizes simultaneous collection of multiple data sources, avoids the limitations and data errors of a single data source, and lays a reliable foundation for subsequent feature extraction.

[0017] In a possible implementation, the feature extraction on each of the water temperature data sets to obtain the time sequence dynamic feature of the ocean water temperature of each grid area comprises the following steps.

[0018] The data statistics are performed on each of the water temperature data sets by using a time sliding window with a preset size, and the water temperature mean value, the standard deviation, and the gradient change rate of each of the grid areas in each time period are calculated.

[0019] The time sequence decomposition is performed on each of the water temperature data sets based on a time sequence decomposition algorithm, and the long-term trend feature, the seasonal cycle feature, and the residual term of the water temperature change of each of the grid areas are obtained.

[0020] The time sequence dynamic feature of the ocean water temperature of each of the grid areas is constructed by combining the water temperature mean value, the standard deviation, the gradient change rate, the long-term trend feature, the seasonal cycle feature, and the residual term.

[0021] The embodiment of the present application provides a time sequence dynamic feature extraction method, which calculates the water temperature mean value, the standard deviation, and the gradient change rate by using a data statistics method based on a time sliding window, effectively captures the short-term fluctuation feature of the ocean water temperature, performs time sequence decomposition on the water temperature data set by using a time sequence decomposition algorithm, obtains the long-term trend feature, the seasonal cycle feature, and the residual term, enhances the analysis capability of the model on the dynamic change of the water temperature, and improves the time sequence adaptability of the density prediction. Finally, the short-term feature and the long-term feature are combined to construct the time sequence dynamic feature of the ocean water temperature of each of the grid areas, so that the model can accurately capture the change rule of the water temperature with time, and then the change rule of the water temperature is mapped to the change rule of the phytoplankton density, and the accuracy of the phytoplankton density identification is improved.

[0022] In a possible implementation, the feature extraction on each of the water temperature data sets to obtain the vertical stratification feature of the ocean water temperature of each of the grid areas comprises the following steps.

[0023] The mixed layer depth and the thermocline depth of each of the grid areas are determined according to each of the water temperature data sets.

[0024] The integral temperature gradient of each of the grid areas in the corresponding range is obtained by performing gradient integral calculation on the water temperature data between the mixed layer depth and the thermocline depth according to each of the water temperature data sets.

[0025] The vertical stratification feature of the ocean water temperature of each of the grid areas is constructed by combining the mixed layer depth, the thermocline depth, and the integral temperature gradient.

[0026] The embodiment of the application provides a vertical stratification feature extraction method, which quantifies the physical characteristics of vertical water temperature stratification by mixing the calculation of layer depth, thermocline depth and integral temperature gradient, so that the model can accurately capture the change rule of water temperature with space, and then accurately represent the regulation of thermocline strength on the vertical distribution of phytoplankton, solve the problem of insufficient modeling of stratification parameters in traditional models, and improve the accuracy of phytoplankton density identification.

[0027] In a possible implementation manner, the feature extraction on each water temperature data set to obtain the spatial correlation features of marine water temperature of each grid area comprises:

[0028] According to each water temperature data set, a four-dimensional continuous water temperature field containing longitude, latitude, depth and time is generated;

[0029] According to each water temperature data set, the longitudinal water temperature gradient, the latitudinal water temperature gradient, the vertical water temperature gradient and the time water temperature gradient of each sampling point of the four-dimensional continuous water temperature field are calculated respectively;

[0030] According to each longitudinal water temperature gradient, latitudinal water temperature gradient, vertical water temperature gradient and time water temperature gradient, the target marine area is modeled by a graph convolution network to obtain the spatial correlation features of marine water temperature of each grid area.

[0031] The embodiment of the application provides a spatial correlation feature extraction method, which considers the flow characteristics of the ocean. The water temperature change characteristics of the same grid area are affected by the time and space of the region, as well as the surrounding grid areas. Therefore, the embodiment of the application builds a four-dimensional continuous water temperature field and models by combining a graph convolution network, effectively captures the spatial correlation of longitude, latitude, depth and time dimensions, reveals the local and global patterns of water temperature change, enhances the adaptability of the model to complex marine spatial structures, and further improves the accuracy of phytoplankton density identification.

[0032] Further, the identification model performs feature fusion on each time series dynamic feature, vertical stratification feature and spatial correlation feature based on an attention mechanism and a fully connected network, obtains each fusion feature vector, and generates the phytoplankton density of each grid area according to each fusion feature vector, comprising:

[0033] Each time series dynamic feature and each vertical stratification feature are respectively feature-encoded by a preset bidirectional gated recurrent unit to obtain each corresponding time series feature vector and vertical feature vector;

[0034] Each spatial correlation feature is respectively multi-scale pooled by a preset pooling layer to obtain each corresponding spatial feature vector;

[0035] For each of the grid regions, input the time sequence feature vector, the vertical feature vector and the spatial feature vector of the grid region into a preset full connection layer, so that the full connection layer generates an attention weight corresponding to each feature vector based on an attention mechanism;

[0036] According to each of the attention weights, perform feature fusion on the time sequence feature vector, the vertical feature vector and the spatial feature vector of each of the grid regions through a full convolution network and a LeakyReLU activation function, to generate a fusion feature vector corresponding to each grid region;

[0037] According to each of the fusion feature vectors, generate the phytoplankton density of each of the grid regions through a Softmax function and a label proportion learning parameter.

[0038] The embodiment of the present application provides a phytoplankton density identification method based on an identification model, which converts input features into corresponding feature vectors through feature encoding and multi-scale pooling, then performs feature weighted fusion on each feature vector based on an attention mechanism, and finally generates the phytoplankton density of each grid region according to each fusion feature vector. The embodiment of the present application dynamically allocates the weight of different features through the attention mechanism, focuses on key information, suppresses noise interference, optimizes the feature fusion effect, combines the full convolution network with the LeakyReLU activation function, improves the nonlinear expression ability of the model, enhances the modeling ability of the nonlinear response characteristics of the phytoplankton density, and improves the accuracy and efficiency of the phytoplankton density identification.

[0039] In a possible implementation manner, the initial identification model is trained according to the historical phytoplankton density data and the historical water temperature data set of the target marine region in different periods to obtain the identification model, including:

[0040] The initial identification model is obtained based on deep learning model construction, and the initial identification model includes an input layer, a bidirectional gated recurrent unit, a pooling layer, a full connection layer, a convolution layer and an output layer;

[0041] According to the grid region division of the target marine region, the historical phytoplankton density data and the historical water temperature data set, divide a plurality of training phytoplankton density data and a plurality of training water temperature data set corresponding to each grid region;

[0042] Feature extraction is performed on each of the training water temperature data sets to obtain a corresponding training water temperature feature set;

[0043] According to each of the training phytoplankton density data and each of the training water temperature feature set, a plurality of label proportion training data packets are created for each grid area, wherein any of the training phytoplankton density data is taken as a proportion label in the corresponding label proportion training data packet, and any of the training water temperature feature set is taken as a data feature in the corresponding label proportion training data packet;

[0044] According to a preset hybrid loss function and each of the label proportion training data packet, the initial recognition model is trained to determine deep learning parameters and label proportion learning parameters, and then the recognition model is obtained by updating, wherein the hybrid loss function includes a deep learning loss function and a label proportion learning loss function, and the label proportion learning loss function is constructed based on minimizing the Wasserstein distance between a prediction result and a proportion label.

[0045] The embodiments of the present application provide a training method of a recognition model. According to the characteristics of the phytoplankton density recognition task, the water temperature of each position point cannot reflect the phytoplankton density of the entire grid area, and the phytoplankton density of the entire grid area is equivalent to the proportion label of the bag in the label proportion learning. Therefore, the recognition task can be regarded as a label proportion learning task, and then the initial recognition model is trained based on the label proportion learning algorithm. According to the characteristics of the weak supervision paradigm of the label proportion learning algorithm, only the regional density proportion label is required to train the model without labeling the label of each data, which significantly reduces the labeling cost and solves the problem of sparse labeling. In addition, the Wasserstein distance loss is introduced in the hybrid loss function in the embodiments of the present application. The Wasserstein distance measures the distribution difference through the optimal transport theory. Even if there is no overlap between the support sets of the real label distribution and the prediction distribution, effective gradient signals can still be provided. In the process of the model based on LLP, when the class distribution difference of multiple label bags is large, the Wasserstein distance can avoid the training stagnation caused by distribution separation, and improve the robustness of the model.

[0046] In a second aspect, the embodiments of the present application provide a phytoplankton density recognition system based on marine water temperature, which includes an acquisition module, a feature extraction module, and a recognition module.

[0047] The acquisition module is configured to acquire a water temperature data set of each grid area in a target marine area in a preset time range.

[0048] The feature extraction module is configured to perform feature extraction on each of the water temperature data sets to obtain time sequence dynamic features, vertical layering features, and spatial correlation features of the marine water temperature of each grid area.

[0049] The identification module is configured to input each of the time-series dynamic features, vertical layering features and spatial correlation features into a preset identification model, so that the identification model performs feature fusion on each of the time-series dynamic features, vertical layering features and spatial correlation features based on an attention mechanism and a full connection network, obtains each fusion feature vector, and generates phytoplankton density of each of the grid regions according to each of the fusion feature vectors.

[0050] The identification model is obtained by training an initial identification model according to historical phytoplankton density data and historical water temperature data sets of a target marine region in different periods. Specifically, the historical phytoplankton density data is taken as a proportion label, and the historical water temperature data sets are taken as data features. A plurality of label proportion training data packets are constructed, and the initial identification model is trained by label proportion learning to obtain the identification model. The initial identification model is obtained based on a deep learning model.

[0051] Further, the identification model performs feature fusion on each of the time-series dynamic features, vertical layering features and spatial correlation features based on an attention mechanism and a full connection network, obtains each fusion feature vector, and generates phytoplankton density of each of the grid regions according to each of the fusion feature vectors, including:

[0052] Each of the time-series dynamic features and each of the vertical layering features is feature-encoded by a preset bidirectional gated recurrent unit to obtain each corresponding time-series feature vector and vertical feature vector.

[0053] Each of the spatial correlation features is multi-scale pooled by a preset pooling layer to obtain each corresponding spatial feature vector.

[0054] For each of the grid regions, the time-series feature vector, the vertical feature vector and the spatial feature vector of the grid region are input into a preset full connection layer, so that the full connection layer generates an attention weight corresponding to each feature vector based on an attention mechanism.

[0055] According to each of the attention weights, the time-series feature vector, the vertical feature vector and the spatial feature vector of each of the grid regions are feature-fused by a full convolutional network and a LeakyReLU activation function to generate a fusion feature vector corresponding to each grid region.

[0056] According to each of the fusion feature vectors, the phytoplankton density of each of the grid regions is generated by a Softmax function and a label proportion learning parameter.

[0057] In a possible implementation manner, the phytoplankton density identification system further includes a model training module, configured to train an initial identification model according to a plurality of historical phytoplankton density data and a plurality of historical water temperature data sets of a target marine region in different periods to obtain the identification model, including a model construction unit, a training data set division unit, a feature extraction unit, a training data construction unit, and a training unit;

[0058] The model construction unit is configured to obtain the initial identification model based on a deep learning model construction, and the initial identification model includes an input layer, a bidirectional gated recurrent unit, a pooling layer, a full connection layer, a convolution layer, and an output layer.

[0059] The training data set division unit is configured to divide, according to the grid region division of the target marine region, the plurality of historical phytoplankton density data and the plurality of historical water temperature data sets, a plurality of training phytoplankton density data and a plurality of training water temperature data sets corresponding to each grid region.

[0060] The feature extraction unit is configured to perform feature extraction on each of the training water temperature data sets to obtain a corresponding training water temperature feature set.

[0061] The training data construction unit is configured to create, according to each of the training phytoplankton density data and each of the training water temperature feature set, a plurality of label proportion training data packets for each grid region, wherein any training phytoplankton density data is taken as a proportion label in a corresponding label proportion training data packet, and any training water temperature feature set is taken as a data feature in the corresponding label proportion training data packet.

[0062] The training unit is configured to train the initial identification model according to a preset hybrid loss function and each of the label proportion training data packets to determine deep learning parameters and label proportion learning parameters, and then update the identification model, wherein the hybrid loss function includes a deep learning loss function and a label proportion learning loss function, and the label proportion learning loss function is constructed based on minimizing the Wasserstein distance between a prediction result and a proportion label. BRIEF DESCRIPTION OF DRAWINGS

[0063] Figure 1 A flowchart of a phytoplankton density identification method based on marine water temperature provided by an embodiment of the present application;

[0064] Figure 2 A structure diagram of an identification model in a phytoplankton density identification method based on marine water temperature provided by an embodiment of the present application;

[0065] Figure 3A structural schematic diagram of a phytoplankton density identification system based on ocean water temperature is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0067] It should be noted that the step numbers in the text are only for the convenience of explaining the specific embodiments, and do not serve as the function of limiting the execution sequence of the steps. In the description of the present application, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include one or more of the features.

[0068] Embodiment one:

[0069] As shown in Figure 1 Embodiment one provides a method for identifying the density of phytoplankton based on ocean water temperature, comprising steps S1-S3:

[0070] Step S1, obtaining water temperature data sets of a preset three-dimensional space of each grid region in a target ocean region within a preset time range;

[0071] Step S2, extracting features from each of the water temperature data sets to obtain time sequence dynamic features, vertical layering features, and spatial correlation features of the ocean water temperature of each grid region;

[0072] Step S3, inputting each of the time sequence dynamic features, vertical layering features, and spatial correlation features into a preset identification model, so that the identification model performs feature fusion on each of the time sequence dynamic features, vertical layering features, and spatial correlation features based on an attention mechanism and a full connection network, obtains each fusion feature vector, and generates the density of phytoplankton of each of the grid regions according to each of the fusion feature vectors;

[0073] The identification model is obtained by training an initial identification model according to historical phytoplankton density data and historical water temperature data sets of a target marine area at different periods. Specifically, the historical phytoplankton density data is taken as a proportion label, and the historical water temperature data sets are taken as data features, a plurality of label proportion training data packets are constructed, and the initial identification model is trained by label proportion learning to obtain the identification model. The initial identification model is obtained based on a deep learning model.

[0074] The embodiment of the present application provides a phytoplankton density identification method based on marine water temperature. The water temperature data set is constructed by collecting water temperature data of each position and each time point in the target marine area, and then the time sequence dynamic feature, the vertical layering feature and the spatial correlation feature are extracted from the water temperature data set and identified by the identification model to generate the phytoplankton density of different grid regions in the target marine area. The embodiment of the present application comprehensively captures the complex relationship between marine water temperature and phytoplankton density by fusing the multi-dimensional features of time sequence dynamics, vertical layering and spatial correlation, and significantly improves the identification accuracy. The target marine area is divided into a plurality of grid regions in advance in the embodiment of the present application, and then the phytoplankton density of different grid regions is identified. This grid data processing enables the embodiment of the present application to support high-resolution fine identification, meet the real-time monitoring needs of a large range of sea areas, and better adapt to the label proportion learning algorithm. In addition, the label proportion learning algorithm is introduced based on the density characteristics of the identification task, and the identification model is constructed by combining the deep learning model, which greatly reduces the dependence on fine annotation data, reduces the data annotation cost, enhances the generalization ability of the model in the sparse annotation scene, and improves the accuracy and efficiency of the phytoplankton density identification.

[0075] Further, in step S1, the water temperature data set of each grid region in the target marine area in a preset three-dimensional space within a preset time range is obtained, including:

[0076] extracting a sea surface temperature data set of each of the preset three-dimensional spaces within the preset time range from satellite remote sensing data;

[0077] obtaining a layered water temperature data set of each of the preset three-dimensional spaces within the preset time range by a plurality of preset sensors;

[0078] combining each of the sea surface temperature data sets and each of the layered water temperature data sets to construct the water temperature data set of each of the grid regions in the preset three-dimensional space within the preset time range.

[0079] The embodiment of the present application provides a water temperature data set acquisition method, which combines satellite remote sensing and a plurality of preset sensors to perform data acquisition, can sufficiently acquire water temperature data of different positions in each grid area, obtains multi-dimensional and multi-scale water temperature information covering the ocean surface layer and vertical space, improves the integrity and representativeness of the water temperature data set through complementary fusion of the sea surface and the layered data, and lays a reliable foundation for subsequent feature extraction and improves the accuracy of phytoplankton density identification. In addition, the sensor technology and the remote sensing technology are used simultaneously to acquire the water temperature data in the embodiment of the present application, the simultaneous acquisition of a plurality of data sources is realized, the limitation of a single data source and data error are avoided, a reliable foundation is laid for subsequent feature extraction.

[0080] In a preferred embodiment, before the density identification is performed, the target marine area is divided into a plurality of three-dimensional grids according to longitude, latitude and depth, and the size of each grid can be set according to the monitoring resolution requirement. After the grid division, a unique identifier is assigned to each grid, and then a plurality of grid areas are obtained. It should be noted that each grid contains the sea surface. In the water temperature data acquisition process, the sea surface temperature (SST) data set of the target area is acquired from the MODIS or VIIRS satellite sensor, the time range is the past 1 year, the time resolution is daily, and then the sea surface temperature data set of each grid is obtained. Through the deployment of Argo buoy sensors at different depths in the target marine area, the vertical layered water temperature data in the grid area is collected, such as 0-200 meters, one Argo buoy sensor is deployed every 10 meters, the time range is synchronized with the satellite data, and then the layered water temperature data set of each grid is obtained. The two-dimensional sea surface temperature data of satellite remote sensing and the three-dimensional layered water temperature data of the buoy sensor are aligned, missing values are filled through an interpolation algorithm, and a time and space continuous water temperature data set of each grid is generated in the form of a four-dimensional tensor (longitude, latitude, depth, time).

[0081] In a possible implementation manner, in step S2, the feature extraction is performed on each water temperature data set to obtain the time sequence dynamic feature of the marine water temperature of each grid area, including:

[0082] The data statistics are performed on each water temperature data set through a preset size time sliding window, the water temperature mean value, the standard deviation and the gradient change rate of each grid area in each time period are calculated;

[0083] The time sequence decomposition is performed on each water temperature data set based on a time sequence decomposition algorithm, and the long-term trend feature, the seasonal cycle feature and the residual term of the water temperature change of each grid area are obtained;

[0084] The time sequence dynamic feature of the marine water temperature of each grid area is constructed in combination with the water temperature mean value, the standard deviation, the gradient change rate, the long-term trend feature, the seasonal cycle feature and the residual term.

[0085] The embodiment of the present application provides a kind of extraction method of time sequence dynamic characteristics, the mean, standard deviation and gradient change rate of water temperature are calculated by the data statistical method based on time sliding window, effectively capture the short-term fluctuation characteristics of ocean water temperature;The time series decomposition algorithm is used to the time sequence decomposition of water temperature data set, obtains long-term trend characteristics, seasonal periodic characteristics and residual term, enhances the analytical ability of model to water temperature dynamic change, improves the time sequence adaptability of density prediction.Finally, short-term characteristics and long-term characteristics are combined, the time sequence dynamic characteristics of the ocean water temperature of each grid area are obtained, so that the model can accurately capture the change rule of water temperature with time, and then the change rule of water temperature is mapped to the change rule of phytoplankton density, improve the accuracy of phytoplankton density identification.

[0086] In a preferred embodiment, when the time sequence dynamic characteristics of ocean water temperature are extracted, the original water temperature data set is smoothed by Savitzky-Golay filter, the long-term trend is retained while the short-time disturbance is inhibited.Then set 30 days of time window, and the sliding step is 7 days.The mean, standard deviation and gradient change rate in window are calculated for each grid water temperature data set.In the process of time sequence decomposition, Fourier series constraint is introduced in seasonal component, to solve the defect that traditional STL is sensitive to non-integer period.For example, for the composite mode of annual cycle (365 days) superimposed on monthly cycle (30 days), different scale period items are separated by frequency domain filtering.In addition, autocorrelation test (ACF / PACF) is carried out on the residual term after decomposition, to identify the mutation event that is not captured by the model, and it is used as abnormal event feature vector.

[0087] In a possible implementation manner, in step S2, the feature extraction is carried out on each water temperature data set to obtain the vertical stratification characteristics of ocean water temperature of each grid area, including:

[0088] The mixed layer depth and thermocline depth of each grid area are calculated and determined according to each water temperature data set respectively;

[0089] According to each water temperature data set, the water temperature data between the mixed layer depth and the thermocline depth is gradient integrated to calculate the integrated temperature gradient of each grid area in the corresponding range;

[0090] The vertical stratification characteristics of ocean water temperature of each grid area are constructed by combining the mixed layer depth, thermocline depth and integrated temperature gradient.

[0091] The embodiment of the application provides a vertical stratification feature extraction method, which quantifies the physical characteristics of vertical water temperature stratification by calculating the mixed layer depth, the thermocline depth and the integral temperature gradient, so that the model can accurately capture the variation law of water temperature with space, and then accurately represent the regulation of the thermocline strength on the vertical distribution of phytoplankton, solve the problem of insufficient modeling of stratification parameters in the traditional model, and improve the accuracy of phytoplankton density identification.

[0092] In a preferred embodiment, in the feature extraction of the vertical stratification feature of the ocean water temperature, the mixed layer depth of each grid area is determined according to each water temperature data set by using the density change threshold method; meanwhile, the water temperature gradient distribution of each grid area is calculated, the depth layer of the maximum water temperature gradient in each grid area is located by using the curvature analysis method, and then the thermocline depth of each grid area is determined. Finally, the vertical water temperature gradient between the mixed layer depth and the thermocline depth in each grid area is normalized, and the index can represent the hindering strength of the thermocline on the vertical migration of phytoplankton.

[0093] In a possible implementation manner, in step S2, the feature extraction is performed on each water temperature data set to obtain the spatial correlation features of the ocean water temperature of each grid area, including:

[0094] According to each water temperature data set, a four-dimensional continuous water temperature field containing longitude, latitude, depth and time is generated;

[0095] According to each water temperature data set, the meridional water temperature gradient, the latitudinal water temperature gradient, the vertical water temperature gradient and the time water temperature gradient of each sampling point of the four-dimensional continuous water temperature field are calculated respectively;

[0096] According to each meridional water temperature gradient, latitudinal water temperature gradient, vertical water temperature gradient and time water temperature gradient, the target marine area is modeled by using a graph convolution network to obtain the spatial correlation features of the ocean water temperature of each grid area.

[0097] The embodiment of the application provides a spatial correlation feature extraction method, considering the flow characteristics of the ocean, the water temperature variation characteristics of the same grid area are affected by the time and space of the region, and are also affected by the surrounding grid areas, therefore, the embodiment of the application builds a whole four-dimensional continuous water temperature field and models by combining a graph convolution network, effectively captures the spatial correlation of longitude, latitude, depth and time dimensions, reveals the local and global patterns of water temperature variation, enhances the adaptability of the model to the complex marine spatial structure, and then improves the accuracy of phytoplankton density identification.

[0098] In a preferred embodiment, when performing feature extraction on the spatial correlation characteristics of the ocean water temperature, the meridional, latitudinal, vertical and time gradients of the four-dimensional continuous water temperature field are calculated respectively according to each of the water temperature data sets using a Sobel operator to form a 32-dimensional gradient tensor. Then, through covariance matrix analysis, the principal components (PCA) of the gradient direction are quantified to extract the dominant transmission direction characteristics and obtain the meridional water temperature gradient, latitudinal water temperature gradient, vertical water temperature gradient and time water temperature gradient of each sampling point. Finally, taking the grid as the node and the gradient as the edge weight, a graph structure is constructed, the neighborhood information is aggregated through a graph convolution network (GCN), and the spatial correlation feature vector of each grid is output.

[0099] Further, in step S3, the identification model performs feature fusion on each of the time series dynamic features, vertical layering features and spatial correlation features based on an attention mechanism and a fully connected network, obtains each fusion feature vector, and generates the phytoplankton density of each grid area according to each of the fusion feature vectors, including:

[0100] Each of the time series dynamic features and each of the vertical layering features is feature-encoded by a preset bidirectional gated recurrent unit to obtain each corresponding time series feature vector and vertical feature vector;

[0101] Each of the spatial correlation features is multi-scale pooled by a preset pooling layer to obtain each corresponding spatial feature vector;

[0102] For each of the grid areas, the time series feature vector, vertical feature vector and spatial feature vector of the grid area are input into a preset fully connected layer, so that the fully connected layer generates an attention weight corresponding to each feature vector based on an attention mechanism;

[0103] According to each of the attention weights, the time series feature vector, vertical feature vector and spatial feature vector of each of the grid areas are feature-fused by a fully convolutional network and a LeakyReLU activation function to generate a fusion feature vector corresponding to each grid area;

[0104] According to each of the fusion feature vectors, the phytoplankton density of each of the grid areas is generated by a Softmax function and a label proportion learning parameter.

[0105] The embodiment of the present application provides a phytoplankton density identification method based on an identification model. The input features are converted into corresponding feature vectors through feature coding and multi-scale pooling. Then, the feature vectors are weighted and fused based on an attention mechanism. Finally, the phytoplankton density of each grid region is generated according to the fused feature vectors. The embodiment of the present application dynamically allocates the weights of different features through the attention mechanism, focuses on key information, suppresses noise interference, optimizes the feature fusion effect, combines the fully convolutional network with the LeakyReLU activation function, improves the nonlinear expression ability of the model, enhances the modeling ability of the nonlinear response characteristics of the phytoplankton density, and improves the accuracy and efficiency of the phytoplankton density identification.

[0106] In a possible implementation manner, the initial identification model is trained according to the historical phytoplankton density data and the historical water temperature data sets of the target marine region in different periods to obtain the identification model, and the method comprises the following steps.

[0107] The initial identification model is obtained based on deep learning model construction, and the initial identification model comprises an input layer, a bidirectional gated recurrent unit, a pooling layer, a fully connected layer, a convolutional layer, and an output layer.

[0108] According to the grid region division of the target marine region, the historical phytoplankton density data and the historical water temperature data sets, the training phytoplankton density data and the training water temperature data sets corresponding to each grid region are divided.

[0109] The feature extraction is performed on each training water temperature data set to obtain a corresponding training water temperature feature set.

[0110] According to each training phytoplankton density data and each training water temperature feature set, a plurality of label proportion training data packets are created for each grid region, wherein any training phytoplankton density data is used as a proportion label in the corresponding label proportion training data packet, and any training water temperature feature set is used as a data feature in the corresponding label proportion training data packet.

[0111] The initial identification model is trained according to a preset hybrid loss function and each label proportion training data packet to determine deep learning parameters and label proportion learning parameters, and then the identification model is obtained by updating, wherein the hybrid loss function comprises a deep learning loss function and a label proportion learning loss function, and the label proportion learning loss function is constructed based on the minimization of the Wasserstein distance between the prediction result and the proportion label.

[0112] The embodiment of the application provides a training method of a recognition model. According to the characteristics of the phytoplankton density recognition task, the water temperature of each position point cannot reflect the phytoplankton density of the entire grid area, and the phytoplankton density of the entire grid area is equivalent to the proportion of the bag in the label proportion learning. Therefore, the recognition task can be regarded as a label proportion learning task, and then the initial recognition model is trained based on a label proportion learning algorithm. After introducing the label proportion learning algorithm, according to the characteristics of the weak supervision paradigm of the label proportion learning algorithm, only the proportion of the regional density label is required to train the model without labeling the label of each data, which significantly reduces the labeling cost and solves the problem of sparse labeling. In addition, the Wasserstein distance loss is introduced in the mixed loss function. The Wasserstein distance measures the distribution difference through the optimal transport theory. Even if there is no overlap between the support set of the real label distribution and the predicted distribution, effective gradient signals can still be provided. In the process of the model based on LLP, when the class distribution difference of multiple label bags is large, the Wasserstein distance can avoid the training stagnation caused by distribution separation, and improve the robustness of the model.

[0113] In a preferred embodiment, as shown in Figure 2 The input layer is used for receiving the external input of the time sequence dynamic feature, the vertical layering feature and the spatial correlation feature. The encoding layer is used for processing the time sequence dynamic feature and the vertical layering feature through a bidirectional gated recurrent unit (BiGRU) to output a 128-dimensional time sequence feature vector and a vertical feature vector. The pooling layer is used for extracting the multi-resolution information of the spatial correlation feature through a multi-scale pooling operation (maximum pooling, average pooling) to output a 128-dimensional spatial feature vector. The fully connected layer is used for calculating the attention weight of each feature vector, and the attention weight and each feature vector are input into the convolution layer. The convolution layer is used for weighting and fusing each feature vector according to each attention weight through a fully convolutional network (FCN) and a LeakyReLU activation function. The output layer is used for outputting the phytoplankton proportion of each grid area through a Softmax function combined with a label proportion parameter, and converting the phytoplankton proportion into the corresponding phytoplankton density. When the initial recognition model is trained, the time sequence / vertical / spatial water temperature features of each grid and the corresponding phytoplankton proportion are constructed into a label proportion training data bag, wherein the phytoplankton proportion can be obtained by normalizing calculation according to the real phytoplankton density of the area, and the label proportion training data bag is input into the initial recognition model to make the initial recognition model output the corresponding prediction result, and the difference between the prediction result and the bag label is calculated based on the cross-entropy loss and the Wasserstein distance, and then the model parameters are updated. Based on the above training process, the Adam optimizer is used, the learning rate is set to 0.001, the batch size is 32, and the iteration is trained until the loss converges, and the final recognition model is obtained.

[0114] Further, the embodiment of the present application can also map the phytoplankton density to the corresponding phytoplankton density level according to the preset mapping relationship table, input the phytoplankton density level of each grid area to the geographic information system (GIS), and generate a phytoplankton distribution heat map of the target marine area, so as to realize the visual display of the density recognition result. Since the embodiment of the present application can recognize the phytoplankton density in real time, the phytoplankton distribution heat map at different time points can be continuously generated over time, so as to reveal the change rule of marine phytoplankton and provide data support for subsequent marine research work.

[0115] Embodiment two:

[0116] As shown in Figure 3 , embodiment two provides a phytoplankton density recognition system based on marine water temperature, which comprises an acquisition module 10, a feature extraction module 20, and a recognition module 30.

[0117] The acquisition module 10 is configured to acquire water temperature data sets of a preset three-dimensional space of each grid area in a target marine area within a preset time range.

[0118] The feature extraction module 20 is configured to extract features from each water temperature data set to obtain time sequence dynamic features, vertical layering features, and spatial correlation features of the marine water temperature of each grid area.

[0119] The recognition module 30 is configured to input each time sequence dynamic feature, vertical layering feature, and spatial correlation feature into a preset recognition model, so that the recognition model performs feature fusion on each time sequence dynamic feature, vertical layering feature, and spatial correlation feature based on an attention mechanism and a full connection network, obtains each fusion feature vector, and generates the phytoplankton density of each grid area according to each fusion feature vector.

[0120] The recognition model is obtained by training an initial recognition model according to a plurality of historical phytoplankton density data and a plurality of historical water temperature data sets of the target marine area at different periods. Specifically, the historical phytoplankton density data is taken as a proportion label, the historical water temperature data set is taken as a data feature, a plurality of label proportion training data packets are constructed, and the initial recognition model is trained by label proportion learning to obtain the recognition model. The initial recognition model is obtained based on a deep learning model.

[0121] Further, the acquisition module 10 acquires water temperature data sets of a preset three-dimensional space of each grid area in a target marine area within a preset time range, comprising:

[0122] extracting a sea surface temperature dataset of each of the preset three-dimensional spaces within a preset time range from satellite remote sensing data;

[0123] obtaining a stratified water temperature dataset of each of the preset three-dimensional spaces within a preset time range through a preset number of sensors;

[0124] combining each of the sea surface temperature datasets and each of the stratified water temperature datasets to construct a water temperature dataset of each of the grid regions within a preset time range in a preset three-dimensional space.

[0125] In a possible implementation manner, the feature extraction module 20 performs feature extraction on each of the water temperature datasets to obtain a time sequence dynamic feature of the ocean water temperature of each grid region, including:

[0126] performing data statistics on each of the water temperature datasets through a preset size of a time sliding window to calculate a water temperature mean value, a standard deviation, and a gradient change rate of each of the grid regions within each time period;

[0127] performing time sequence decomposition on each of the water temperature datasets based on a time sequence decomposition algorithm to obtain a long-term trend feature, a seasonal cycle feature, and a residual term of water temperature change of each of the grid regions;

[0128] combining each of the water temperature mean value, the standard deviation, the gradient change rate, the long-term trend feature, the seasonal cycle feature, and the residual term to construct the time sequence dynamic feature of the ocean water temperature of each grid region.

[0129] In a possible implementation manner, the feature extraction module 20 performs feature extraction on each of the water temperature datasets to obtain a vertical stratification feature of the ocean water temperature of each grid region, including:

[0130] determining a mixed layer depth and a thermocline depth of each of the grid regions according to each of the water temperature datasets;

[0131] performing gradient integral calculation on water temperature data between the mixed layer depth and the thermocline depth according to each of the water temperature datasets to obtain an integral temperature gradient of each grid region within a corresponding range;

[0132] combining each of the mixed layer depth, the thermocline depth, and the integral temperature gradient to construct the vertical stratification feature of the ocean water temperature of each grid region.

[0133] In a possible implementation manner, the feature extraction module 20 performs feature extraction on each of the water temperature datasets to obtain a spatial correlation feature of the ocean water temperature of each grid region, including:

[0134] generate a four-dimensional continuous water temperature field containing longitude, latitude, depth, and time according to each water temperature data set;

[0135] calculate the longitudinal water temperature gradient, the latitudinal water temperature gradient, the vertical water temperature gradient, and the time water temperature gradient of each sampling point of the four-dimensional continuous water temperature field respectively according to each water temperature data set;

[0136] model the target marine area by a graph convolutional network according to each longitudinal water temperature gradient, latitudinal water temperature gradient, vertical water temperature gradient, and time water temperature gradient, and obtain the spatial correlation features of marine water temperature of each grid area.

[0137] Further, the recognition model performs feature fusion on each of the time-series dynamic features, vertical layering features, and spatial correlation features based on an attention mechanism and a fully connected network, obtains each fusion feature vector, and generates the phytoplankton density of each grid area according to each fusion feature vector, including:

[0138] encode each of the time-series dynamic features and each of the vertical layering features by a preset bidirectional gated recurrent unit to obtain each corresponding time-series feature vector and vertical feature vector;

[0139] perform multi-scale pooling on each of the spatial correlation features by a preset pooling layer to obtain each corresponding spatial feature vector;

[0140] For each grid area, input the time-series feature vector, vertical feature vector, and spatial feature vector of the grid area into a preset fully connected layer to enable the fully connected layer to generate an attention weight corresponding to each feature vector based on an attention mechanism;

[0141] fuse the time-series feature vector, vertical feature vector, and spatial feature vector of each grid area by a fully convolutional network and a LeakyReLU activation function according to each attention weight to generate a fusion feature vector corresponding to each grid area;

[0142] generate the phytoplankton density of each grid area by a Softmax function and a label proportion learning parameter according to each fusion feature vector.

[0143] In one possible implementation, the phytoplankton density recognition system further includes a model training module configured to train an initial recognition model according to several historical phytoplankton density data and several historical water temperature data sets of the target marine area in different periods to obtain the recognition model, including a model construction unit, a training data set division unit, a feature extraction unit, a training data construction unit, and a training unit.

[0144] The model construction unit is configured to construct the initial identification model based on a deep learning model, the initial identification model comprising an input layer, a bidirectional gated recurrent unit, a pooling layer, a full connection layer, a convolution layer, and an output layer.

[0145] The training data set division unit is configured to divide, according to the grid division of the target marine area, the historical phytoplankton density data and the historical water temperature data set, a plurality of training phytoplankton density data and a plurality of training water temperature data set corresponding to each grid area.

[0146] The feature extraction unit is configured to extract features from each of the training water temperature data sets to obtain a corresponding training water temperature feature set.

[0147] The training data construction unit is configured to create a plurality of label ratio training data packets for each grid area according to each of the training phytoplankton density data and each of the training water temperature feature set, wherein any of the training phytoplankton density data is used as a proportion label in the corresponding label ratio training data packet, and any of the training water temperature feature set is used as a data feature in the corresponding label ratio training data packet.

[0148] The training unit is configured to train the initial identification model according to a preset hybrid loss function and each of the label ratio training data packet to determine deep learning parameters and label ratio learning parameters, and then update the identification model, wherein the hybrid loss function comprises a deep learning loss function and a label ratio learning loss function, and the label ratio learning loss function is constructed based on minimizing the Wasserstein distance between the prediction result and the proportion label.

[0149] The embodiment of the present application provides a phytoplankton density identification system based on ocean water temperature, acquires water temperature data of each position and each time point in a target ocean area, constructs a water temperature data set, extracts time sequence dynamic features, vertical layering features and spatial correlation features from the water temperature data set, identifies each feature through an identification model, and generates phytoplankton density of different grid areas in the target ocean area. The embodiment of the present application comprehensively captures the complex relationship between ocean water temperature and phytoplankton density by fusing multi-dimensional features of time sequence dynamics, vertical layering and spatial correlation, and significantly improves the identification accuracy. The embodiment of the present application also divides the target ocean area into multiple grid areas in advance, and then identifies the phytoplankton density of different grid areas. This grid data processing enables the embodiment of the present application to support high-resolution fine identification, meet the real-time monitoring needs of a large range of sea areas, and better adapt to the label proportion learning algorithm. In addition, the embodiment of the present application introduces the label proportion learning algorithm based on the density characteristics of the identification task, and combines the identification model constructed through the deep learning model, greatly reduces the dependence on fine annotation data, reduces the data annotation cost, enhances the generalization ability of the model in the sparse annotation scene, and improves the accuracy and efficiency of the phytoplankton density identification.

[0150] The more detailed working principle and step flow of the embodiment can be but not limited to refer to the related description of embodiment one.

[0151] The above specific embodiments further illustrate the purpose, technical solutions and advantages of the present application. It should be understood that the above description is only for specific embodiments of the present application and is not intended to limit the protection scope of the present application. It is particularly pointed out that any modification, equivalent replacement, improvement, etc. made by those skilled in the art within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for identifying phytoplankton density based on ocean water temperature, characterized by, The method comprises the following steps: obtaining a preset three-dimensional space of each grid region in a target marine area within a preset time range; extracting features from each of the water temperature data sets to obtain time series dynamic features, vertical stratification features, and spatial correlation features of the ocean water temperature of each grid region; inputting each of the time series dynamic features, vertical stratification features, and spatial correlation features into a preset identification model to enable the identification model to perform feature fusion based on an attention mechanism and a fully connected network on each of the time series dynamic features, vertical stratification features, and spatial correlation features, obtain each fusion feature vector, and generate phytoplankton density for each of the grid regions according to each of the fusion feature vectors; wherein the identification model is obtained by training an initial identification model based on historical phytoplankton density data and historical water temperature data sets of the target marine area at different periods, specifically: the historical phytoplankton density data is used as a proportion label, and the historical water temperature data sets are used as data features to construct a plurality of label proportion training data packages and perform label proportion learning training on the initial identification model to obtain the identification model, and the initial identification model is obtained based on a deep learning model; the feature extraction from each of the water temperature data sets to obtain the time series dynamic features of the ocean water temperature of each grid region comprises: performing data statistics on each of the water temperature data sets through a preset size of a time sliding window to calculate the water temperature mean, standard deviation, and gradient change rate of each of the grid regions within each time period; performing time series decomposition on each of the water temperature data sets based on a time series decomposition algorithm to obtain long-term trend features, seasonal cycle features, and residual terms of the water temperature change of each of the grid regions; combining each of the water temperature mean, standard deviation, gradient change rate, long-term trend features, seasonal cycle features, and residual terms to obtain the time series dynamic features of the ocean water temperature of each grid region; the feature extraction from each of the water temperature data sets to obtain the vertical stratification features of the ocean water temperature of each grid region comprises: determining the mixed layer depth and thermocline depth of each of the grid regions according to each of the water temperature data sets; performing gradient integral calculation on the water temperature data between the mixed layer depth and the thermocline depth according to each of the water temperature data sets to obtain the integral temperature gradient of each grid region within the corresponding range; combining each of the mixed layer depth, thermocline depth, and integral temperature gradient to obtain the vertical stratification features of the ocean water temperature of each grid region; the feature extraction from each of the water temperature data sets to obtain the spatial correlation features of the ocean water temperature of each grid region comprises: generating a four-dimensional continuous water temperature field containing longitude, latitude, depth, and time according to each of the water temperature data sets; calculating the meridional water temperature gradient, latitudinal water temperature gradient, vertical water temperature gradient, and time water temperature gradient of each sampling point of the four-dimensional continuous water temperature field according to each of the water temperature data sets; According to the meridional water temperature gradient, the latitudinal water temperature gradient, the vertical water temperature gradient, and the time water temperature gradient, the target marine area is modeled by a graph convolution network to obtain spatial correlation characteristics of marine water temperature of each grid area.

2. The method of claim 1, wherein the ocean water temperature is determined by a satellite. The water temperature data set of each grid area in the target marine area in a preset three-dimensional space within a preset time range is obtained, including: The sea surface temperature data set of each of the preset three-dimensional space within the preset time range is extracted from satellite remote sensing data; The layered water temperature data set of each of the preset three-dimensional space within the preset time range is obtained by a plurality of preset sensors; The water temperature data set of each of the grid area in the preset three-dimensional space within the preset time range is constructed by combining each of the sea surface temperature data set and each of the layered water temperature data set.

3. The method of claim 1, wherein the ocean water temperature is determined by a method comprising: obtaining a satellite image of the ocean water; and determining the ocean water temperature from the satellite image. The recognition model is based on an attention mechanism and a fully connected network to fuse features of each of the time series dynamic features, the vertical layered features, and the spatial correlation features, to obtain each of the fusion feature vectors, and to generate the phytoplankton density of each of the grid areas according to each of the fusion feature vectors, including: Each of the time series dynamic features and each of the vertical layered features is respectively feature-encoded by a preset bidirectional gated recurrent unit to obtain each of the corresponding time series feature vectors and vertical feature vectors; Each of the spatial correlation features is respectively multi-scale pooled by a preset pooling layer to obtain each of the corresponding spatial feature vectors; For each of the grid areas, the time series feature vector, the vertical feature vector, and the spatial feature vector of the grid area are input into a preset fully connected layer, so that the fully connected layer generates an attention weight corresponding to each of the feature vectors based on an attention mechanism; According to each of the attention weights, the time series feature vector, the vertical feature vector, and the spatial feature vector of each of the grid areas are fused by a fully convolutional network and a LeakyReLU activation function to generate a fusion feature vector corresponding to each of the grid areas; According to each of the fusion feature vectors, the phytoplankton density of each of the grid areas is generated by a Softmax function and a label proportion learning parameter.

4. A method of identifying phytoplankton density based on ocean water temperature according to any one of claims 1 to 3, wherein, The initial recognition model is trained according to a plurality of historical phytoplankton density data and a plurality of historical water temperature data sets of the target marine area in different periods to obtain the recognition model, including: The initial recognition model is constructed based on a deep learning model, including an input layer, a bidirectional gated recurrent unit, a pooling layer, a fully connected layer, a convolutional layer, and an output layer; According to the grid area division of the target marine area, the plurality of historical phytoplankton density data, and the plurality of historical water temperature data sets, a plurality of training phytoplankton density data and a plurality of training water temperature data sets corresponding to each grid area are divided; Each of the training water temperature data sets is feature-extracted to obtain each of the corresponding training water temperature feature sets; According to each of the training phytoplankton density data and each of the training water temperature feature set, a plurality of label proportion training data packets are created for each grid area, wherein any of the training phytoplankton density data is taken as a proportion label in the corresponding label proportion training data packet, and any of the training water temperature feature set is taken as data features in the corresponding label proportion training data packet; According to a preset hybrid loss function and each of the label proportion training data packet, the initial recognition model is trained to determine deep learning parameters and label proportion learning parameters, and then the recognition model is updated, wherein the hybrid loss function includes a deep learning loss function and a label proportion learning loss function, and the label proportion learning loss function is constructed based on minimizing the Wasserstein distance between the prediction result and the proportion label.

5. A marine water temperature-based phytoplankton density identification system for implementing the marine water temperature-based phytoplankton density identification method according to claim 1, characterized by The method comprises an acquisition module, a feature extraction module, and a recognition module. The acquisition module is configured to acquire water temperature data sets of a preset three-dimensional space of each grid area in a target marine region within a preset time range. The feature extraction module is configured to perform feature extraction on each of the water temperature data sets to obtain time series dynamic features, vertical layering features, and spatial correlation features of marine water temperature of each grid area. The recognition module is configured to input each of the time series dynamic features, vertical layering features, and spatial correlation features into a preset recognition model, so that the recognition model performs feature fusion on each of the time series dynamic features, vertical layering features, and spatial correlation features based on an attention mechanism and a fully connected network, obtains each fusion feature vector, and generates phytoplankton density of each grid area according to each of the fusion feature vectors. The recognition model is obtained by training an initial recognition model based on a plurality of historical phytoplankton density data and a plurality of historical water temperature data sets of the target marine region in different periods, specifically: the historical phytoplankton density data are taken as proportion labels, the historical water temperature data sets are taken as data features, a plurality of label proportion training data packets are constructed, and the initial recognition model is trained for label proportion learning to obtain the recognition model. The initial recognition model is constructed based on a deep learning model.

6. A phytoplankton density identification system based on ocean water temperature as claimed in claim 5, wherein, The recognition model performs feature fusion on each of the time series dynamic features, vertical layering features, and spatial correlation features based on an attention mechanism and a fully connected network, obtains each fusion feature vector, and generates phytoplankton density of each grid area according to each of the fusion feature vectors, including: Each of the time series dynamic features and each of the vertical layering features is feature-encoded by a preset bidirectional gated recurrent unit to obtain each corresponding time series feature vector and vertical feature vector; Each of the spatial correlation features is multi-scale pooled by a preset pooling layer to obtain each corresponding spatial feature vector; For each grid area, the time series feature vector, vertical feature vector, and spatial feature vector of the grid area are input into a preset fully connected layer to enable the fully connected layer to generate attention weights corresponding to each feature vector based on an attention mechanism. According to each of the attention weights, the time sequence feature vector, the vertical feature vector and the spatial feature vector of each of the grid regions are fused by a full convolution network and a LeakyReLU activation function to generate a fused feature vector corresponding to each of the grid regions; According to each of the fused feature vectors, the phytoplankton density of each of the grid regions is generated by a Softmax function and a label proportion learning parameter.

7. A phytoplankton identification system based on ocean water temperature as claimed in claim 5 or 6, wherein, The phytoplankton density recognition system further comprises a model training module configured to train an initial recognition model according to a plurality of historical phytoplankton density data and a plurality of historical water temperature data sets of a target marine region in different periods to obtain the recognition model, including a model construction unit, a training data set division unit, a feature extraction unit, a training data construction unit and a training unit. The model construction unit is configured to obtain the initial recognition model based on a deep learning model construction, and the initial recognition model comprises an input layer, a bidirectional gated recurrent unit, a pooling layer, a full connection layer, a convolution layer and an output layer. The training data set division unit is configured to divide a plurality of training phytoplankton density data and a plurality of training water temperature data sets corresponding to each grid region according to the grid region division of the target marine region, the plurality of historical phytoplankton density data and the plurality of historical water temperature data sets. The feature extraction unit is configured to extract features from each of the training water temperature data sets to obtain a corresponding training water temperature feature set. The training data construction unit is configured to create a plurality of label proportion training data packets for each grid region according to each of the training phytoplankton density data and each of the training water temperature feature sets, wherein any of the training phytoplankton density data is used as a proportion label in the corresponding label proportion training data packet, and any of the training water temperature feature sets is used as a data feature in the corresponding label proportion training data packet. The training unit is configured to train the initial recognition model according to a preset hybrid loss function and each of the label proportion training data packets to determine a deep learning parameter and a label proportion learning parameter, and then update the recognition model, wherein the hybrid loss function comprises a deep learning loss function and a label proportion learning loss function, and the label proportion learning loss function is constructed based on minimizing the Wasserstein distance between the prediction result and the proportion label.

Images