Method for determining phytoplankton key monitoring sites based on t-sne and deep forest

Abstract

The application discloses a phytoplankton key monitoring station determination method based on t-SNE and deep forest, comprising the following steps: determining marine environment features and space-time features to form a candidate feature set; calculating the Pearson correlation coefficients between the features, and removing the features with high correlation to obtain a screened feature set; reducing the feature dimension of the screened feature set to 2 dimensions based on a t-distributed neighborhood embedding algorithm, and calculating the distance between the class center points in the 2 dimensions; taking the distance between the class center points in different space-time feature subsets as a prior condition, determining an optimal space-time feature subset for training a deep forest model, and training the deep forest model by using the optimal space-time feature subset; and selecting a key monitoring station according to the binary classification prediction result of the trained deep forest model. The application can ensure the determination accuracy and robustness, significantly reduce the model operation energy consumption, and realize the rapid screening and sorting of the key monitoring station.

Description

Technical Field

[0001] This invention relates to the field of marine ecological environment monitoring and intelligent information processing, and in particular to a method for determining key phytoplankton monitoring sites based on t-SNE (t-Distributed Stochastic Neighbor Embedding) and deep forests. Background Technology

[0002] Scientific justification for marine environmental monitoring station locations is a crucial step in project approval, directly impacting data representativeness, monitoring efficiency, and the return on investment. Public research indicates a trade-off between fixed and random station deployment strategies in terms of statistical performance and bias control. For example, some researchers, in comparative studies of benthic organism monitoring sampling designs, have pointed out that fixed, non-randomly selected stations have higher statistical power but are more prone to introducing systematic bias in annual variation estimation. As marine science enters the era of big data, cutting-edge technologies such as Artificial Intelligence (AI), big data, and the Internet of Things (IoT) are gradually becoming important tools for promoting marine environmental monitoring. The combination of big data and machine learning holds promise as a new paradigm for studying the evolution of complex marine phenomena. However, previous monitoring station designs largely relied on expert experience, lacking deep integration with intelligent algorithms.

[0003] Deep Forest, proposed by Zhou Zhihua et al., is a tree ensemble method. It combines the ensemble learning approach of random forests with the characteristics of deep neural networks, exhibiting significant advantages in enhancing feature diversity and improving feature classification capabilities. Its model architecture is similar to multi-layer neural networks, but each layer consists of several random forests instead of traditional neurons. The core of the Deep Forest model comprises "multi-granularity scanning" and "cascaded forests," achieving layer-by-layer processing through a cascaded structure, and the training process does not rely on backpropagation and gradient optimization. Furthermore, its multi-granularity scanning method automatically determines the model's complexity based on the data's complexity. This feature results in fewer hyperparameters and demonstrates good learning capabilities even on small datasets.

[0004] Phytoplankton forms the foundation of the marine food chain, and their spatiotemporal variations directly impact fishery resource distribution, carbon cycle processes, and the stability of marine ecosystems. For marine phytoplankton, whose density distribution is influenced by multiple sources, biomass changes exhibit strong nonlinearity, non-stationarity, and scale heterogeneity. Deep forests are susceptible to redundant and irrelevant features, resulting in slower convergence, limited parallel training efficiency, and generally slower overall training time compared to traditional models such as support vector machines and random forests. Even when good predictive performance is achieved, model validation and parameter tuning are often time-consuming, making it difficult to balance accuracy, efficiency, and engineering deployment requirements.

[0005] t-SNE is a nonlinear dimensionality reduction method that maps high-dimensional data to two- or three-dimensional space, effectively preserving local structure within the data. It utilizes the Stochastic Gradient Descent (SGD) algorithm to optimize the objective function, maintaining pairwise distances between high-dimensional inputs in a low-dimensional representation. Compared to the linear dimensionality reduction of Principal Component Analysis (PCA), t-SNE excels at handling complex nonlinear relationships between features, such as exponential relationships. Furthermore, t-SNE has the ability to control the trade-off between local and global relationships between points, thus producing more visually appealing clustering results. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention proposes a method for determining key phytoplankton monitoring sites based on t-SNE and deep forest. This method utilizes the class separability metric (inter-class centroid distance) in the two-dimensional embedding space as a priori evaluation for deep forest modeling to quickly determine the optimal spatiotemporal feature subset. This enables the invention to achieve rapid and robust determination of key monitoring sites with low computational overhead, and to achieve localized continuous updates and deployment through seasonal incremental training.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for determining key phytoplankton monitoring sites based on t-SNE and deep forest includes the following steps:

[0009] S1: Determine the marine environmental characteristics and spatiotemporal characteristics to form a candidate feature set; and by calculating the Pearson correlation coefficient between the features, remove highly correlated features to obtain the filtered feature set.

[0010] S2: Based on the t-distribution neighborhood embedding algorithm, the feature dimension in the filtered feature set is reduced to 2-dimensional, and the inter-class centroid distance is calculated in 2-dimensional;

[0011] S3: Using the distance between class centroids under different spatiotemporal feature subsets as a prior condition, determine the optimal spatiotemporal feature subset for training the deep forest model, and train the deep forest model using the optimal spatiotemporal feature subset;

[0012] S4: Select key monitoring sites based on the binary classification prediction results of the trained deep forest model.

[0013] Furthermore, the marine environmental characteristics include silicates, phosphates, nitrates, salinity, sea surface temperature, ocean currents, and dissolved oxygen.

[0014] Furthermore, the spatiotemporal features include year, month, day, longitude, and latitude.

[0015] Furthermore, the Pearson correlation coefficient of the filtered feature set in S1 is in the range of [-0.9, 0.9].

[0016] Furthermore, S3 specifically includes:

[0017] The feature set obtained in S2 is re-divided into three candidate feature subsets according to spatiotemporal features: containing spatiotemporal features, containing only time features, and containing only space features. The inter-class centroid distance of each of the three candidate feature subsets is calculated, and the candidate feature subset with the largest inter-class centroid distance is selected as the optimal spatiotemporal feature subset.

[0018] Furthermore, in S4, when resource constraints exist, for the set of monitoring stations with a predicted label of 1, the stations are sorted in descending order according to the visual probabilities generated by the Local Interpretable Model-agnostic Explanations (LIME) method, and the top K monitoring stations are selected as key monitoring stations.

[0019] Furthermore, after each phytoplankton monitoring is completed, the newly collected data from that monitoring is incorporated into the existing training set. While maintaining consistency in the sampling method, an incremental learning mechanism is used to redetermine the optimal spatiotemporal feature subset, and the deep forest model is retrained.

[0020] A device for determining key phytoplankton monitoring sites based on t-SNE and deep forest includes one or more processors for implementing a method for determining key phytoplankton monitoring sites based on t-SNE and deep forest.

[0021] An electronic device, comprising:

[0022] One or more processors;

[0023] A storage device for storing one or more programs, which, when executed by the electronic device, enable the electronic device to implement a method for determining key phytoplankton monitoring sites based on t-SNE and deep forest.

[0024] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a method for determining key phytoplankton monitoring sites based on t-SNE and deep forest.

[0025] The beneficial effects of this invention are as follows:

[0026] (1) The method of the present invention guides the S3 spatiotemporal modeling by S1 feature redundancy removal and S2 separability prior (t-SNE two-dimensional center point distance), which significantly reduces the energy consumption of model operation while ensuring the accuracy and robustness of the judgment.

[0027] (2) The present invention achieves rapid screening and sorting of key monitoring stations through deep forest binary classification, and combines localized incremental training to effectively overcome the problem of limited sampling window, reduce invalid sampling, and improve timeliness and prediction performance.

[0028] (3) This invention uses the distance between class centroids to perform prior evaluation of separability and suppress the adverse effects of redundant dimensions, thereby improving the speed, stability and reusability of key monitoring station determination while ensuring the prediction performance of deep forest, improving training efficiency, providing quantitative basis for key monitoring station decision-making, and being able to be stably migrated and deployed under cross-year conditions, with significant engineering application and promotion value. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating a method for determining key phytoplankton monitoring sites based on t-SNE and deep forests, according to one embodiment of the present invention.

[0030] Figure 2 The final feature set of one embodiment of the present invention Pearson correlation coefficient plot for any two features.

[0031] Figure 3 The final feature set provided for an embodiment of this application A 2D-reduced visualization.

[0032] Figure 4 Inter-class centroid distance diagram of different candidate subsets provided in an embodiment of this application.

[0033] Figure 5 A comparison chart of binary classification prediction performance of different candidate subsets provided in an embodiment of this application. Detailed Implementation

[0034] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0035] like Figure 1 As shown, one embodiment of the present invention provides a method for determining key phytoplankton monitoring sites based on t-SNE and deep forest, which includes the following steps one through four.

[0036] Step 1: Selecting Marine Environmental and Spatiotemporal Features: Based on literature and sampling schemes, determine marine environmental and spatiotemporal features to form a candidate feature set F; calculate the Pearson correlation coefficient between features and remove highly correlated features according to a threshold to obtain the filtered feature set. .

[0037] In this embodiment, the candidate feature set F consists of two parts: first, marine environmental features that significantly affect phytoplankton density in the study area, selected based on historical literature, including silicate (Si), phosphate (PO4), nitrate (NO3), salinity (SSS), sea surface temperature (SST), ocean currents (CV), and dissolved oxygen (DO); second, spatiotemporal features determined based on the sampling scheme, including year, month, day, longitude, and latitude. The specific calculation formula is as follows:

[0038]

[0039]

[0040] In the formula, E i Z represents marine environmental characteristics that significantly affect phytoplankton density in this study area, selected based on historical literature. j Spatiotemporal characteristics; This means that under sampling scheme S, if a spatiotemporal variable has at least two different values ​​in this sampling, it will be included in F; otherwise, if the variable has only one value in this sampling, it will not be included.

[0041] The formula for calculating the Pearson correlation coefficient r between features is as follows:

[0042]

[0043] In the formula, x and y are any two features. and y and y are the sample means, respectively, and n is the number of samples.

[0044] In this embodiment, highly correlated features are removed by retaining features whose Pearson correlation coefficient r falls within the range of [-0.9, 0.9]. The final feature set obtained is... The Pearson correlation coefficient plot of any two features is shown below. Figure 2 As shown, the correlation coefficients between any two features fall within the range of [-0.880588229179228, 0.64542021609752]. Therefore, all features are retained.

[0045] Step 2: Based on the t-distributed neighborhood embedding algorithm, reduce the feature dimension of the filtered feature set to 2-dimensional and calculate the inter-class centroid distance in 2-dimensional.

[0046] The t-distributed neighborhood embedding algorithm, also known as the t-SNE algorithm, is implemented in Python software in this embodiment (computer configuration: Intel(R) Core(TM) i7-14650HX CPU, NVIDIA GeForce RTX4060 Laptop GPU, 32GB RAM). The system calls data processing and machine learning modules in the Python software to perform t-SNE dimensionality reduction calculations on the standardized multidimensional feature data. Specifically, the input to the algorithm is a high-dimensional feature set. The output consists of the perplexity ρ, the number of iterations T, and the learning rate η. The dimensionality reduction result is output in two-dimensional coordinate form, such as... Figure 3 As shown, it can be used for subsequent visualization and category distribution analysis, helping researchers understand the clustering characteristics and separation of different categories of samples in the feature space.

[0047] In this embodiment, the method for calculating the distance between the center points of classes is as follows:

[0048] On the two-dimensional embedding representation matrix Z, the samples are divided into subsets of each category based on the label vector y. The geometric centroid of each subset is calculated, and the Euclidean distance between the geometric centroids of the two categories is calculated. The specific calculation formula is as follows:

[0049]

[0050]

[0051]

[0052]

[0053] In the formula, n is the number of samples; z represents the sample label vector; i This represents the coordinate vector of the i-th sample in the two-dimensional embedding space. Z k Let represent the set of vectors of the k-th class of samples in the two-dimensional embedding space; k is the class index, and since it is a binary classification, k can take the values ​​0 and 1, where 0 represents a low-risk region and 1 represents a high-risk region; c k This represents the geometric centroid of the k-th class sample in two-dimensional space; Represents set Z k The cardinality is the number of samples in the k-th class; d is the Euclidean distance between the geometric centroids of the high-risk region and the low-risk region. This represents the Euclidean second norm.

[0054] Step 3: Using the distance between class centroids under different spatiotemporal feature subsets as prior conditions, determine the optimal spatiotemporal feature subset for training the deep forest model, and train the deep forest model using the optimal spatiotemporal feature subset.

[0055] Specifically, the feature set obtained from S2 is re-divided into three candidate feature subsets according to spatiotemporal features: those containing spatiotemporal features (year, month, day, longitude, latitude), those containing only time features (year, month, day), and those containing only spatial features (longitude, latitude). The inter-class centroid distance of each of the three candidate feature subsets is calculated, and the candidate feature subset with the largest inter-class centroid distance is selected as the optimal spatiotemporal feature subset.

[0056] like Figure 4 As shown, in this embodiment, the feature subset containing only spatial features has the largest inter-class centroid distance, which is 10.2734. The average intra-class distance for class 0 is 27.9518, and for class 1 it is 29.3352. The feature subset containing spatiotemporal features has an inter-class centroid distance of 5.3943, with an average intra-class distance of 28.0122 for class 0 and 31.7359 for class 1. The feature subset containing only temporal features has the smallest inter-class centroid distance, which is 2.8494. The average intra-class distance for class 0 is 28.9446, and for class 1 it is 31.7181. Therefore, the feature subset corresponding to only spatial features (with the largest inter-class centroid distance) is used to train the deep forest model.

[0057] In this embodiment, the input to the deep forest model is marine phytoplankton monitoring data from 1999 to 2019 (1: high-risk area, 0: low-risk area; the training set is marine phytoplankton monitoring data from 1999 to 2017, and the test set is marine phytoplankton monitoring data from 2018 and 2019) and corresponding different feature subsets. The model prediction result output is 0 or 1.

[0058] In this embodiment, the phytoplankton data collection method for inputting the deep forest model is as follows: all phytoplankton samples were from scientific cruise monitoring, and their processing and counting were performed in accordance with the "Marine Monitoring Specifications". Qualitative samples were collected using a shallow-water Type III plankton net, vertically trawled from 2m below the seabed to the sea surface, and fixed with 5% formaldehyde. Quantitative samples were collected using a GO-FLOW water sampler to collect surface water, and fixed with 6-8 mL of Lugol's solution per liter of water sample.

[0059] In this embodiment, the training and optimization of the deep forest model are implemented using Python software (computer configuration: i5-12400H CPU, 16G memory), and the key hyperparameters of the model are jointly optimized using a grid search method. Specifically, the system pre-defines a hyperparameter grid space containing multiple candidate values, the hyperparameters being as follows: 'max_layers': [5, 10, 15]; 'n_estimators': [2, 5, 10, 15]; 'n_trees': [50, 100, 150, 200, 250, 300, 350, 400, 450, 500]; 'min_samples_split': [2, 5, 10, 15]; 'min_samples_leaf': [1, 2, 5, 10, 15]. Subsequently, the system uses the GridSearchCV grid search method to traverse and search the above hyperparameter combinations. During the hyperparameter search process, we employ a five-fold cross-validation strategy to partition the training data into five non-overlapping subsets. In each training round, four subsets are selected for model training, and the remaining subset is used for performance validation. This validation process is repeated for all subsets. For each set of hyperparameters, the system calculates a model performance evaluation index based on the results of the five validations, and uses this index to assess the model's stability and generalization ability under that parameter combination.

[0060] After evaluating all hyperparameter combinations, the system automatically selects the optimal hyperparameter configuration based on preset performance comparison criteria to construct the final deep forest model. By designing a training strategy combining grid search and cross-validation, the system effectively reduces the model's dependence on the results of a single data partitioning, improves the objectivity and robustness of model parameter selection, and thus enhances the classification performance and stability of the deep forest model in real-world applications. The final results are as follows:

[0061] 1) Under the candidate subset training scheme with only spatial features, the best prediction performance is achieved, with a model prediction performance (F1 score) of 70.04% (parameters: {'max_layers': 5, 'min_samples_leaf': 2, 'min_samples_split': 15, 'n_estimators': 5, 'n_trees': 450}). Figure 5 As shown;

[0062] 2) Under the training scheme of candidate subsets of spatiotemporal features, the model achieves suboptimal prediction performance, with a prediction performance (F1 score) of 66.67% (parameters: {'max_layers': 5, 'min_samples_leaf': 1, 'min_samples_split': 10, 'n_estimators': 2, 'n_trees': 200}).

[0063] 3) Under the training scheme using only temporal features as candidate subsets, the worst prediction performance was achieved, with a model prediction performance (F1 score) of 61.79% (parameters: {'max_layers': 10, 'min_samples_leaf': 2, 'min_samples_split': 10, 'n_estimators': 2, 'n_trees': 150}). Therefore, when constructing high-risk phytoplankton areas based on the deep forest model, only spatial features need to be selected.

[0064] In this embodiment, the F1 score is calculated as follows: First, three types of quantities are counted: the number of samples that are actually positive and predicted as positive, the number of samples that are actually negative but predicted as positive, and the number of samples that are actually positive but predicted as negative. Then, two intermediate quantities are calculated: one represents "how many of the samples predicted as positive are actually positive", which is calculated by dividing the number of samples that are actually positive and predicted as positive by the number of samples that are predicted as positive; the other represents "how many of the samples that are actually positive were found", which is calculated by dividing the number of samples that are actually positive and predicted as positive by the number of samples that are actually positive. Finally, these two intermediate quantities are combined: they are multiplied together, then multiplied by two, and the result is divided by the sum of the two intermediate quantities. The result is the F1 score for the positive samples.

[0065] In this embodiment, to improve the applicability and stability of the method under different data distribution conditions, the present invention does not limit itself to a single optimal solution during the hyperparameter optimization process, but retains ten sets of hyperparameter combinations with the best performance indicators as a candidate parameter set. This candidate parameter set is used for adaptive selection or rapid switching based on environmental noise levels, sample characteristics, and computational resource constraints during actual deployment. This avoids performance fluctuations caused by overfitting a single parameter configuration to specific data, improves the algorithm's robustness, engineering feasibility, and application scope, and reduces the time cost of repeated on-site parameter tuning. The top 10 optimal parameter combinations are as follows:

[0066] The top 10 optimal parameter combinations for training candidate subsets based solely on spatial features are as follows:

[0067] Parameters: {'max_layers': 5, 'min_samples_leaf': 2, 'min_samples_split':15, 'n_estimators': 5, 'n_trees': 450};

[0068] Parameters: {'max_layers': 10, 'min_samples_leaf': 2, 'min_samples_split':15, 'n_estimators': 5, 'n_trees': 450};

[0069] Parameters: {'max_layers': 15, 'min_samples_leaf': 2, 'min_samples_split':15, 'n_estimators': 5, 'n_trees': 450};

[0070] Parameters: {'max_layers': 5, 'min_samples_leaf': 2, 'min_samples_split':15, 'n_estimators': 10, 'n_trees': 300};

[0071] Parameters: {'max_layers': 10, 'min_samples_leaf': 2, 'min_samples_split':15, 'n_estimators': 10, 'n_trees': 300};

[0072] Parameters: {'max_layers': 15, 'min_samples_leaf': 2, 'min_samples_split':15, 'n_estimators': 10, 'n_trees': 300};

[0073] Parameters: {'max_layers': 5, 'min_samples_leaf': 2, 'min_samples_split':15, 'n_estimators': 10, 'n_trees': 150};

[0074] Parameters: {'max_layers': 10, 'min_samples_leaf': 2, 'min_samples_split':15, 'n_estimators': 10, 'n_trees': 150};

[0075] Parameters: {'max_layers': 15, 'min_samples_leaf': 2, 'min_samples_split':15, 'n_estimators': 10, 'n_trees': 150};

[0076] Parameters: {'max_layers': 5, 'min_samples_leaf': 1, 'min_samples_split':15, 'n_estimators': 10, 'n_trees': 100}.

[0077] The top 10 optimal parameter combinations for training candidate subsets of spatiotemporal features are as follows:

[0078] Parameters: {'max_layers': 5, 'min_samples_leaf': 1, 'min_samples_split':10, 'n_estimators': 2, 'n_trees': 200};

[0079] Parameters: {'max_layers': 10, 'min_samples_leaf': 1, 'min_samples_split':10, 'n_estimators': 2, 'n_trees': 200};

[0080] Parameters: {'max_layers': 15,'min_samples_leaf': 1,'min_samples_split': 10, 'n_estimators': 2, 'n_trees': 200};

[0081] Parameters: {'max_layers': 5,'min_samples_leaf': 2,'min_samples_split': 2, 'n_estimators': 2, 'n_trees': 250};

[0082] Parameters: {'max_layers': 5,'min_samples_leaf': 2,'min_samples_split': 2, 'n_estimators': 5, 'n_trees': 300};

[0083] Parameters: {'max_layers': 10,'min_samples_leaf': 2,'min_samples_split': 2, 'n_estimators': 2, 'n_trees': 250};

[0084] Parameters: {'max_layers': 10,'min_samples_leaf': 2,'min_samples_split': 2, 'n_estimators': 5, 'n_trees': 300};

[0085] Parameters: {'max_layers': 15,'min_samples_leaf': 2,'min_samples_split': 2, 'n_estimators': 2, 'n_trees': 250};

[0086] Parameters: {'max_layers': 15,'min_samples_leaf': 2,'min_samples_split': 2, 'n_estimators': 5, 'n_trees': 300};

[0087] Parameters: {'max_layers': 5, 'min_samples_leaf': 2, 'min_samples_split':10, 'n_estimators': 2, 'n_trees': 50}.

[0088] The top 10 optimal parameter combinations for training candidate subsets based solely on temporal features are as follows:

[0089] Parameters: {'max_layers': 10, 'min_samples_leaf': 2, 'min_samples_split':10, 'n_estimators': 2, 'n_trees': 150};

[0090] Parameters: {'max_layers': 15, 'min_samples_leaf': 2, 'min_samples_split':10, 'n_estimators': 2, 'n_trees': 150};

[0091] Parameters: {'max_layers': 10, 'min_samples_leaf': 15, 'min_samples_split':2, 'n_estimators': 2, 'n_trees': 400};

[0092] Parameters: {'max_layers': 10, 'min_samples_leaf': 15, 'min_samples_split':5, 'n_estimators': 2, 'n_trees': 400};

[0093] Parameters: {'max_layers': 10, 'min_samples_leaf': 15, 'min_samples_split':10, 'n_estimators': 2, 'n_trees': 400};

[0094] Parameters: {'max_layers': 10, 'min_samples_leaf': 15, 'min_samples_split':15, 'n_estimators': 2, 'n_trees': 400};

[0095] Parameters: {'max_layers': 15, 'min_samples_leaf': 15, 'min_samples_split':2, 'n_estimators': 2, 'n_trees': 400};

[0096] Parameters: {'max_layers': 15, 'min_samples_leaf': 15, 'min_samples_split':5, 'n_estimators': 2, 'n_trees': 400};

[0097] Parameters: {'max_layers': 15, 'min_samples_leaf': 15, 'min_samples_split':10, 'n_estimators': 2, 'n_trees': 400};

[0098] Parameters: {'max_layers': 15, 'min_samples_leaf': 15, 'min_samples_split':15, 'n_estimators': 2, 'n_trees': 400}.

[0099] Step 5: Select key monitoring sites based on the binary classification prediction results of the trained deep forest model.

[0100] Under ideal conditions where resources are not a constraint, the determination of key monitoring sites can be made directly based on the prediction results of each monitoring site from the trained deep forest model: all monitoring sites whose model output indicates a high-risk area and whose prediction label is "1" are uniformly included in the key monitoring scope. At this time, the set of key monitoring sites is consistent with the set of sites "predicted as 1", which can maximize the coverage of potential high-risk areas identified by the model. This is suitable for scenarios with sufficient budget, personnel, and equipment, thereby achieving comprehensive tracking and continuous monitoring of high-risk sites.

[0101] When resource constraints exist (e.g., limited monitoring budget, number of available inspection personnel, and number of available monitoring devices), to ensure that limited resources are prioritized for sites with higher risks and greater urgency, further refined screening and ranking are needed within the set of sites with a "predicted label of 1". Specifically, the process is as follows: First, sites with a predicted label of "1" are used as the candidate set, ensuring that all selected sites fall within the high-risk range determined by the model. Then, the Locally Interpretable Method (LIME) is introduced to locally interpret the model output for each candidate site, generating interpretability results related to that site and obtaining a visualized probability (used to characterize the confidence level or risk intensity of the site being classified as high-risk). Based on this, the candidate sites are ranked from highest to lowest according to the visualized probabilities generated by LIME, forming a priority sequence. Finally, the top K monitoring sites from the ranking results are selected as the set of key monitoring sites. This strategy enables the prioritization of monitoring sites most likely to be at risk under resource constraints, improving the efficiency of key monitoring resources and the targeting of risk management. It also helps to form traceable and explainable criteria for site selection, facilitating subsequent operation and maintenance decisions, task assignment, and verification of monitoring results.

[0102] Furthermore, to achieve continuous iterative updates and online applications in response to regional environmental changes, after each phytoplankton monitoring session, the newly collected data is incorporated into the existing training set. While maintaining consistency in sampling methods, an incremental learning mechanism is used to re-determine the optimal spatiotemporal feature subset, and the deep forest model is retrained. After the update is completed, localized deployment and version management are implemented in the target area.

[0103] Corresponding to the aforementioned embodiments of the method for determining key phytoplankton monitoring sites based on t-SNE and deep forest, the present invention also provides embodiments of a device for determining key phytoplankton monitoring sites based on t-SNE and deep forest.

[0104] The phytoplankton key monitoring site determination device based on t-SNE and deep forest in this invention includes one or more processors for implementing the phytoplankton key monitoring site determination method based on t-SNE and deep forest in the above embodiments.

[0105] The embodiments of the phytoplankton key monitoring site determination device based on t-SNE and deep forest of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiments can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device reading the corresponding computer program instructions from non-volatile memory into memory and executing them. From a hardware perspective, in addition to the processor, memory, network interface, and non-volatile memory, the data processing device in the embodiments typically includes other hardware depending on its actual functions, which will not be elaborated further.

[0106] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0107] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0108] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the method for determining key phytoplankton monitoring sites based on t-SNE and deep forests as described in the above embodiments.

[0109] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0110] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A method for determining key phytoplankton monitoring sites based on t-SNE and deep forest, characterized in that, Includes the following steps: S1: Determine the marine environmental characteristics and spatiotemporal characteristics to form a candidate feature set; and by calculating the Pearson correlation coefficient between the features, remove highly correlated features to obtain the filtered feature set. S2: Based on the t-distribution neighborhood embedding algorithm, the feature dimension in the filtered feature set is reduced to 2-dimensional, and the inter-class centroid distance is calculated in 2-dimensional; S3: Using the inter-class centroid distance under different spatiotemporal feature subsets as prior conditions, determine the optimal spatiotemporal feature subset for training the deep forest model, and train the deep forest model using the optimal spatiotemporal feature subset; S3 specifically includes: The feature set obtained in S2 is re-divided into three candidate feature subsets according to spatiotemporal features: containing spatiotemporal features, containing only time features, and containing only space features. The inter-class centroid distance of each of the three candidate feature subsets is calculated, and the candidate feature subset with the largest inter-class centroid distance is selected as the optimal spatiotemporal feature subset. S4: Select key monitoring sites based on the binary classification prediction results of the trained deep forest model.

2. The t-SNE and deep forest based phytoplankton hotspot monitoring station determination method according to claim 1, characterized in that, The marine environmental characteristics include silicates, phosphates, nitrates, salinity, sea surface temperature, ocean currents, and dissolved oxygen.

3. The t-SNE and deep forest based phytoplankton watch site determination method of claim 1, wherein, The spatiotemporal features include year, month, day, longitude, and latitude.

4. The t-SNE and deep forest based phytoplankton watch site determination method of claim 1, wherein, The Pearson correlation coefficient of the feature set filtered in S1 is in the range of [-0.9, 0.9].

5. The t-SNE and deep forest based phytoplankton watch site determination method of claim 1, wherein, In step S4, when resource constraints exist, for the set of monitoring stations with a predicted label of 1, the stations are sorted in descending order of the visual probabilities generated by the Locally Interpretable Method (LIME) model, and the top K monitoring stations are selected as key monitoring stations.

6. The t-SNE and deep forest based phytoplankton focus monitoring station site determination method of claim 1, wherein, After each phytoplankton monitoring session, the newly collected data from that monitoring session is incorporated into the existing training set. While maintaining consistency in the sampling method, an incremental learning mechanism is used to re-determine the optimal spatiotemporal feature subset, and the deep forest model is retrained.

7. A device for determining a phytoplankton focus monitoring site based on t-SNE and deep forest, characterized by, It includes one or more processors for implementing the method for determining key phytoplankton monitoring sites based on t-SNE and deep forest as described in any one of claims 1 to 6.

8. An electronic device, comprising: include: One or more processors; A storage device for storing one or more programs, which, when executed by the electronic device, cause the electronic device to implement the method for determining key phytoplankton monitoring sites based on t-SNE and deep forest as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a program that, when executed by a processor, implements the method for determining key phytoplankton monitoring sites based on t-SNE and deep forests as described in any one of claims 1 to 6.

Images