Marine ranching environment monitoring method and system

Abstract

The invention discloses a marine ranch environment monitoring method and system, and the method comprises the steps: obtaining environment monitoring data which comprises remote sensing monitoring data, underwater unmanned plane data and sensor data; performing feature extraction on the remote sensing monitoring data based on a preset deep residual network to obtain remote sensing features; performing feature extraction on the underwater unmanned aerial vehicle data based on a preset Mask R-CNN network to obtain unmanned aerial vehicle features; performing feature extraction on the sensor data based on a preset recurrent neural network to obtain sensing features; acquiring environmental condition data, and adjusting feature weights of the remote sensing features, the unmanned aerial vehicle features and the sensing features based on a preset condition discriminant weight table, a preset reference weight and the environmental condition data; performing weighted fusion on the remote sensing features, the unmanned aerial vehicle features and the sensing features based on the feature weights to generate fusion features; and constructing an environment monitoring model based on a decision tree algorithm, inputting the fusion features into the environment monitoring model, and obtaining an environment monitoring result.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental monitoring, and particularly to a method and system for monitoring the environment of a marine ranch. Background Art

[0002] The environmental monitoring technology of marine ranches is evolving towards the development directions of intelligence, systematization, and multi-source integration. With the increasing intensity of marine resource development, the frequency and amplitude of environmental changes are constantly increasing, and the importance of environmental monitoring in ensuring the ecological security of marine ranches, improving aquaculture efficiency, and warning of disaster risks is becoming increasingly prominent. Scientific and real-time environmental monitoring not only helps to dynamically master key parameters such as sea area water quality, temperature, salinity, and dissolved oxygen, but also provides data support for ecological assessment and management decision-making.

[0003] The existing environmental monitoring methods for marine ranches mainly include remote sensing monitoring, marine buoys, sensor networks, underwater robots, and acoustic devices. Among them, remote sensing technology is suitable for large-scale sea area monitoring but has limited spatial and temporal resolutions; buoys and sensor networks have the ability to continuously collect data but have a small coverage area and high deployment and maintenance costs; although underwater robots have mobility and diverse detection functions, they are greatly limited by endurance and real-time performance. In addition, the existing monitoring methods still have deficiencies in aspects such as not considering multi-source data fusion, automatic identification and prediction of abnormal events, and stable operation under bad weather. Summary of the Invention The present invention provides a method and system for monitoring the environment of a marine ranch to improve the stability and accuracy of marine ranch environmental monitoring.

[0004] To solve the above technical problems, an embodiment of the present invention provides a method for monitoring the environment of a marine ranch, including: Obtaining environmental monitoring data, where the environmental monitoring data includes remote sensing monitoring data, underwater drone data, and sensor data; Performing feature extraction on the remote sensing monitoring data based on a preset deep residual network to obtain remote sensing features; Performing feature extraction on the underwater drone data based on a preset Mask R-CNN network to obtain drone features; Performing feature extraction on the sensor data based on a preset recurrent neural network to obtain sensing features; Obtaining environmental condition data, and adjusting the feature weights of the remote sensing features, the drone features, and the sensing features based on a preset conditional discriminant weight table, a preset reference weight, and the environmental condition data; Performing weighted fusion on the remote sensing features, the drone features, and the sensing features based on the feature weights to generate fusion features; Construct an environmental monitoring model based on the decision tree algorithm, input the fused features into the environmental monitoring model, and obtain the environmental monitoring results.

[0005] In the present invention, remote sensing data covers the macroscopic marine environment, drones capture local biological details such as algae, and sensors provide continuous time-series parameters, breaking through the limitations of a single data source, avoiding insufficient response to local or dynamic environments, and then extracting the features of each type of data based on different neural networks respectively. Thus, the feature weights are dynamically adjusted according to environmental conditions and weather conditions, avoiding inaccurate measurement of a certain type of data caused by environmental conditions and weather conditions, and enhancing the adaptability of environmental monitoring to environmental changes. Using the decision tree model for environmental monitoring improves the interpretability and decision-making efficiency of the model, and enhances the stability and accuracy of environmental monitoring in the marine ranch.

[0006] Further, the obtaining of environmental condition data, and adjusting the feature weights of the remote sensing features, the drone features, and the sensing features based on a preset condition discriminant weight table, a preset reference weight, and the environmental condition data includes: Obtain the environmental condition data, where the environmental condition data includes weather data and environmental data; Determine the importance scores of the remote sensing features, the drone features, and the sensing features based on the weather data, the environmental data, and the condition discriminant weight table; the condition discriminant weight table includes the mapping relationship between the weather data and the importance scores of each feature; Adjust the feature weights of the remote sensing features, the drone features, and the sensing features based on the importance scores and the preset reference weight.

[0007] In the present invention, by introducing environmental condition data, the scientificity and rationality of feature weight adjustment are improved. Using the condition discriminant weight table, the dynamic adjustment of feature weights is realized, and the flexibility of the system is enhanced.

[0008] Further, the adjusting the feature weights of the remote sensing features, the drone features, and the sensing features based on the importance scores and the preset reference weight includes: Obtain the reference weights of each feature, map each reference weight into a two-dimensional plane, construct a reference triangle with the reference weights as vertices, and calculate the centroid of the reference triangle; For each vertex of the reference triangle, calculate the unit vector pointing from the center to the vertex, and translate the vertex of the reference triangle along the direction of the unit vector according to the importance score to generate a deformed triangle; Calculate the centroid offset vector of the deformed triangle, and map the centroid offset vector into the direction vectors of each vertex to generate the regression factors of each feature; Incremental regression is performed on the reference weights of each feature by the regression factors based on each feature to adjust the feature weights of the remote sensing feature, the UAV feature, and the sensing feature.

[0009] The present invention intuitively reflects the balance or imbalance degree of the importance scores of the three types of features through geometric offset, and directly drives the weight adjustment based on the centroid offset, without setting complex thresholds or empirical rules.

[0010] Further, the feature extraction of the remote sensing monitoring data based on a preset deep residual network to obtain remote sensing features includes: The deep residual network includes a number of residual units, and each residual unit is connected by a skip connection; Feature extraction is performed on the remote sensing monitoring data based on each residual unit to obtain feature vectors; The remote sensing features are obtained by fusing each feature vector based on the skip connection.

[0011] Through the skip connection of the residual network in the present invention, the problem of gradient disappearance in the deep network is alleviated, and the depth and accuracy of feature extraction are improved. And by fusing multiple feature vectors, the expression ability of the remote sensing features is enhanced, and the accuracy of environmental monitoring is improved.

[0012] Further, the feature extraction of the underwater UAV data based on a preset Mask R-CNN network to obtain UAV features includes: The Mask R-CNN network includes a backbone network and a region proposal network; Feature extraction is performed on the underwater UAV data based on the backbone network to obtain a multi-scale feature map set; A candidate region set is obtained based on the region proposal network and the multi-scale feature map set; the candidate region set includes a number of candidate regions and the coordinate information of each candidate region; Feature alignment processing and feature fusion processing are performed based on the candidate region set and the multi-scale feature map set to obtain UAV features.

[0013] Through the multi-scale feature extraction and region proposal mechanism of the Mask R-CNN network in the present invention, the detection and recognition ability of underwater targets is improved. Through feature alignment and fusion processing, the accuracy and robustness of the UAV features are enhanced.

[0014] Further, the feature extraction of the sensor data based on a preset recurrent neural network to obtain sensing features includes: The recurrent neural network includes a number of LSTM units; Obtain sensor data, where the sensor data includes temperature data, salinity data, dissolved oxygen data, pH data, conductivity data, and turbidity data; Construct a first learning projection matrix based on the temperature data, salinity data, and conductivity data; Construct a second learning projection matrix based on the dissolved oxygen data and pH data; Construct a third learning projection matrix based on the turbidity data; Perform dynamic scaling on the first learning projection matrix, the second learning projection matrix, and the third learning projection matrix to achieve cross-parameter interaction and generate a fusion vector; Based on the recurrent neural network, perform feature extraction on the fusion vector to obtain sensing features.

[0015] Through the LSTM network, the present invention effectively captures the time series features in the sensor data, improving the timeliness and accuracy of feature extraction. By inputting multi-parameter sensor data, the perception ability of environmental changes is enhanced.

[0016] Further, after constructing an environmental monitoring model based on the decision tree algorithm, inputting the fusion features into the environmental monitoring model, and obtaining the environmental monitoring results, it further includes: Generate an error evaluation value and a credibility value based on the monitoring results, and generate a comprehensive evaluation score based on the error evaluation value and the credibility value; Incrementally adjust the benchmark weight according to the comprehensive evaluation score.

[0017] Through the decision tree model, the present invention improves the decision-making efficiency and interpretability of environmental monitoring. Introducing error evaluation and credibility analysis realizes self-correction of the monitoring results and enhances the adaptive ability of the system. By incrementally adjusting the benchmark weight, the stability and accuracy of the system during long-term operation are ensured.

[0018] In a second aspect, the present invention provides a marine ranch environmental monitoring system, including: a data acquisition module, a feature extraction module, a feature fusion module, and a monitoring module; The data acquisition module is used to obtain environmental monitoring data, where the environmental monitoring data includes remote sensing monitoring data, underwater drone data, and sensor data; The feature extraction module is used to perform feature extraction on the remote sensing monitoring data based on a preset deep residual network to obtain remote sensing features; perform feature extraction on the underwater drone data based on a preset Mask R-CNN network to obtain drone features; perform feature extraction on the sensor data based on a preset recurrent neural network to obtain sensing features; The feature fusion module is used to obtain environmental condition data, and adjust the feature weights of the remote sensing features, the UAV features, and the sensing features based on a preset conditional discriminant weight table, a preset reference weight, and the environmental condition data; perform weighted fusion on the remote sensing features, the UAV features, and the sensing features based on the feature weights to generate fused features; The monitoring module is used to construct an environmental monitoring model based on a decision tree algorithm, input the fused features into the environmental monitoring model, and obtain an environmental monitoring result.

[0019] Further, the fusion module is used to obtain environmental condition data, and adjust the feature weights of the remote sensing features, the UAV features, and the sensing features based on a preset conditional discriminant weight table, a preset reference weight, and the environmental condition data, including: Obtain the environmental condition data, where the environmental condition data includes weather data and environmental data; Determine the importance scores of the remote sensing features, the UAV features, and the sensing features based on the weather data, the environmental data, and the conditional discriminant weight table; the conditional discriminant weight table includes the mapping relationship between the weather data and the importance scores of each feature; Adjust the feature weights of the remote sensing features, the UAV features, and the sensing features based on the importance scores and the preset reference weight.

[0020] Further, the fusion module is used to adjust the feature weights of the remote sensing features, the UAV features, and the sensing features based on the importance scores and the preset reference weight, including: Obtain the reference weight of each feature, map each reference weight into a two-dimensional plane, construct a reference triangle with the reference weight as the vertex, and calculate the centroid of the reference triangle; For each vertex of the reference triangle, calculate the unit vector pointing from the center to the vertex, and translate the vertex of the reference triangle along the direction of the unit vector according to the importance score to generate a deformed triangle; Calculate the centroid offset vector of the deformed triangle, and map the centroid offset vector into the direction vectors of each vertex to generate the regression factor of each feature; Perform incremental regression on the reference weights of each feature based on the regression factors of each feature to adjust the feature weights of the remote sensing features, the UAV features, and the sensing features. Description of the Drawings

[0021] Figure 1 It is a schematic flowchart of a method for monitoring the environment of a marine ranch provided by an embodiment of the present invention. Detailed Embodiments

[0022] The specific embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. The following embodiments are used to illustrate the present invention, but are not used to limit the scope of the present invention.

[0023] The terms "first" and "second" etc. in the description, claims and drawings of this application are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally further includes steps or units not listed, or optionally further includes other steps or units inherent to these processes, methods, products or devices.

[0024] Referring to "embodiments" herein means that the specific features, structures or characteristics described in conjunction with the embodiments can be included in at least one embodiment of this application. The phrase appears in various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described herein can be combined with other embodiments.

[0025] Embodiment 1 See Figure 1 , Figure 1 which is a schematic flowchart of a method for monitoring the marine ranch environment provided by an embodiment of the present invention. An embodiment of the present invention provides a method for monitoring the marine ranch environment, including steps 101 to 107, specifically as follows: Step 101: Obtain environmental monitoring data, where the environmental monitoring data includes remote sensing monitoring data, underwater drone data and sensor data; In this embodiment, three types of data sources are used to collect marine ranch environment information respectively.

[0026] In this embodiment, remote sensing monitoring data, based on satellite or space-based remote sensing, can obtain parameters such as sea surface temperature (SeaSurface Temperature), sea color (such as chlorophyll a concentration), sea surface salinity, and sea surface wind speed. Among them, the chlorophyll a concentration can reflect the biomass of phytoplankton and the water bloom situation. The remote sensing data has a wide coverage range, which is conducive to the dynamic monitoring of the macro environment.

[0027] In this embodiment, macro picture data of the marine ranch is obtained through remote sensing monitoring data, so as to monitor the environment of the marine ranch from a macro perspective.

[0028] In this embodiment, an underwater unmanned aerial vehicle (AUV / ROV) is equipped with sensors such as a high-definition camera and a sonar, and can acquire images and videos of the seabed environment. Through autonomous cruising or a preset route, the unmanned aerial vehicle can observe underwater targets such as fish distribution, corals / seaweeds, and artificial structures (such as aquaculture facilities), and provide high-resolution visual feature information. Continuous operation of the unmanned aerial vehicle can monitor ecological dynamics in real time and obtain underwater unmanned aerial vehicle data, where the underwater unmanned aerial vehicle data includes continuously generated picture data or video data.

[0029] In this embodiment, multi-parameter environmental sensors are deployed on aquaculture cages or buoys to collect physical and chemical parameters of water quality such as water temperature, salinity (conductivity), dissolved oxygen, pH, conductivity, and turbidity in real time. These sensors can continuously monitor fine-grained time-series data, supplement local environmental information of remote sensing and image data, and thus generate sensor data.

[0030] Step 102: Extract features from the remote sensing monitoring data based on a preset deep residual network to obtain remote sensing features; In this embodiment, The extracting features from the remote sensing monitoring data based on a preset deep residual network to obtain remote sensing features includes: The deep residual network includes a number of residual units, and each residual unit is connected through a skip connection; Extract feature vectors from the remote sensing monitoring data based on each residual unit; Evaluate the similarity between pairwise feature vectors, and construct a feature vector topology graph based on the similarity; Determine the skip paths of each feature based on the feature vector topology graph and the maximum weight path algorithm, and fuse each feature vector based on the skip paths and skip connections to obtain remote sensing features.

[0031] In this embodiment, by obtaining the original remote sensing monitoring image data as the remote sensing monitoring data, and inputting the original remote sensing monitoring image data into a preset deep residual network, where the network includes multiple residual units (ResidualBlock), and each residual unit internally learns an identity mapping and a residual mapping through a skip connection to extract spatial features from low level to high level respectively.

[0032] In this embodiment, after passing through the residual units of each layer of the deep residual network in sequence, the output of the last layer of the network is the remote sensing features that fuse multi-scale information.

[0033] In this embodiment, in the initial stage, the deep residual network focuses on extracting low-level and texture-level features. Usually, the basic residual block (Basic Block) is adopted. This block consists of two 3×3 convolutional layers and a skip connection, and the number of input and output channels remains the same, which helps to stabilize the learning of shallow features.

[0034] In this embodiment, as the network deepens and enters the stage with higher requirements for feature expression ability and parameter efficiency, the bottleneck residual block (Bottleneck Block) is used instead. Each bottleneck block consists of a 1×1 dimensionality reduction convolution, a 3×3 main convolution, and a 1×1 dimensionality increase convolution, which can control the number of parameters while ensuring the trainability of the deep network.

[0035] In this embodiment, through the output of the original remote sensing monitoring image data; in the residual unit feature extraction, I is sequentially input into the deep residual network. The deep residual network includes a number of residual blocks (Residual Block), and each residual block adds the input to the output of the consecutive convolutional layers within the unit to learn the residual mapping, so as to extract multi-level feature vectors { }.

[0036] In this embodiment, for any two feature vectors, the similarity is calculated using cosine similarity to obtain the similarity matrix S:

[0037] In this embodiment, an undirected graph is constructed, and the node set corresponds to the feature vectors; where the edges and weights in the undirected graph indicate that if (a preset threshold), then an edge is connected between , and the weight takes the similarity value.

[0038] In this embodiment, the "input feature" node and the "output feature" node are defined, and they are unconditionally connected to and respectively.

[0039] In this embodiment, in the graph , using the maximum weight path algorithm (such as the modified Dijkstra or maximum flow path), the strongest connected path P∗ from to is found; In this embodiment, the optimal skip path list is obtained.

[0040] In this embodiment, based on the optimal jump path The list sequentially takes the corresponding vectors , and for each segment of the jump, a preset fusion function (such as weighted sum or small MLP) is used for fusion to obtain the fusion result h. The fusion result h is added to the original input or the features of the previous stage to achieve the effect of "skip connection" and obtain the final remote sensing features:

[0041] In this embodiment, the similarity map can reflect the correlation degree of the features of each residual layer, enabling the jump path to focus on the most relevant levels, improving the efficiency and accuracy of feature fusion. At the same time, the path search based on the graph structure can skip redundant or noisy layers and only fuse highly relevant features, enhancing the model's ability to capture details in complex scenarios.

[0042] Step 103: Based on a preset Mask R-CNN network, extract features from the underwater drone data to obtain drone features; In this embodiment, the process of extracting features from the underwater drone data based on a preset Mask R-CNN network to obtain drone features includes: The Mask R-CNN network includes a backbone network and a region proposal network; Based on the backbone network, extract features from the underwater drone data to obtain a multi-scale feature map set; Based on the region proposal network and the multi-scale feature map set, obtain a candidate region set; the candidate region set includes several candidate regions and the coordinate information of each candidate region; Based on the candidate region set and the multi-scale feature map set, perform feature alignment processing and feature fusion processing to obtain drone features.

[0043] In this embodiment, the original image or video frame data collected by the underwater drone is obtained.

[0044] In this embodiment, the image frame is input into a preset Mask R-CNN backbone network (backbone), and the backbone network includes an FPN (Feature Pyramid Network) for multi-scale feature extraction, generating a feature map set {P2, P3, P4, P5} at different scales.

[0045] In this embodiment, on each feature map, the Region Proposal Network (RPN) generates candidate target regions (RoI proposals); and performs RoIAlign operations on the candidate regions respectively, maps them to feature blocks of a unified size, and inputs them into the classification and bounding box regression branches and the mask branch respectively.

[0046] In this embodiment, the classification and bounding box regression branch performs classification scoring and bounding box coordinate regression on each RoI feature block, and outputs the category probability and bounding box adjustment amount of each target. In this embodiment, the mask branch performs a number of convolution and transposed convolution operations on each RoI feature block to generate a binary mask. In this embodiment, the intermediate fully-connected layer feature vector output by the classification branch, the features output by the bounding box branch, and the convolution features output by the mask branch are concatenated or weighted and fused according to a preset rule to obtain the final drone features.

[0047] Step 104: Extract features from the sensor data based on a preset recurrent neural network to obtain sensing features. In this embodiment, the extracting features from the sensor data based on a preset recurrent neural network to obtain sensing features includes: The recurrent neural network includes several layers of LSTM units. Obtain sensor data, where the sensor data includes temperature data, salinity data, dissolved oxygen data, pH data, conductivity data, and turbidity data. Construct a first learning projection matrix based on the temperature data, salinity data, and conductivity data. Construct a second learning projection matrix based on the dissolved oxygen data and pH data. Construct a third learning projection matrix based on the turbidity data. Perform dynamic scaling on the first learning projection matrix, the second learning projection matrix, and the third learning projection matrix to achieve cross-parameter interaction, and generate a fusion vector. Extract features from the fusion vector based on the recurrent neural network to obtain sensing features.

[0048] In this embodiment, obtain the original sensor time series data sequence D = {d1, d2, …, dT} as the sensor data, where each includes parameter vectors such as temperature, salinity, dissolved oxygen, pH, conductivity, and turbidity collected at time t. In this embodiment, input D into a normalization module, perform standardization processing on each parameter dimension, and output the standardized data sequence .

[0049] In this embodiment, a first learning projection matrix is constructed based on the temperature data, salinity data, and conductivity data : A second learning projection matrix is constructed based on the dissolved oxygen data and pH data ; A third learning projection matrix is constructed based on the turbidity data .

[0050] In this embodiment, a learnable scaling factor is assigned to each projection :

[0051] where is the activation function

[0052] In this embodiment, dynamic scaling is performed on the first learning projection matrix, the second learning projection matrix, and the third learning projection matrix to achieve cross-parameter interaction, and a fusion vector is generated :

[0053] In this embodiment, the importance of various environmental parameters is adaptively adjusted by the scaling ratio, which helps to eliminate the data scale difference and enhance the coupling modeling ability

[0054] In this embodiment, the fusion vector sequence at each moment is fed into a deep recurrent network, and a multi-layer LSTM unit (e.g., 2 - 3 layers) is used to model the fusion vector sequence to extract long-term dependence features, and finally a sensor time series feature vector is obtained

[0055] In this embodiment, the standardized time series data is gradually input into a preset recurrent neural network (RNN) unit set in chronological order. The RNN includes several layers of LSTM At the t-th moment, the hidden state at the previous moment and the current input are fed into the RNN unit together. According to the recurrent unit calculation function, the hidden state is updated, and the current hidden state is output In this embodiment, after processing the end of the sequence , the hidden state at the last moment and (optionally) the set of intermediate hidden state vectors at each time step are extracted In this embodiment, the hidden state and / or the vector obtained by pooling or attention weighting on is input into the fully connected layer. After linear transformation and activation function processing, a sensing feature with a fixed dimension is output​

[0056] Step 105: Obtain environmental condition data, and adjust the feature weights of the remote sensing features, the UAV features, and the sensing features based on a preset conditional discriminant weight table, a preset reference weight, and the environmental condition data. In this embodiment, the obtaining of the environmental condition data and the adjustment of the feature weights of the remote sensing features, the UAV features, and the sensing features based on a preset conditional discriminant weight table, a preset reference weight, and the environmental condition data include: Obtain the environmental condition data, where the environmental condition data includes weather data and environmental data; Determine the importance scores of the remote sensing features, the UAV features, and the sensing features based on the weather data, the environmental data, and the conditional discriminant weight table; the conditional discriminant weight table includes the mapping relationship between the weather data and the importance scores of each feature; Adjust the feature weights of the remote sensing features, the UAV features, and the sensing features based on the importance scores and the preset reference weight.

[0057] In this embodiment, a conditional discriminant weight table is established. The conditional discriminant weight table represents the mapping relationship between the weather data and the importance scores of each feature. The rows of the conditional discriminant weight table represent different environmental scenarios or weather conditions (such as "strong wind", "low visibility", "pollution warning", etc.), and the columns correspond to different feature types (remote sensing features, underwater UAV features, sensor features). The adjustment coefficients or target values of the feature weights in this environment are filled in the table.

[0058] In this embodiment, the conditional discriminant weight table can define the reference weight values of each type of feature and the increase and decrease coefficients in different scenarios. For example, in a conventional scenario, the remote sensing feature can be set to 0.4, the underwater UAV feature to 0.4, and the sensor feature to 0.2. When encountering "strong wind", the rule table can specify that the weight of the UAV is reduced by 0.2, the weight of the remote sensing is increased by 0.1, and the weight of the sensor is increased by 0.1.

[0059] In this embodiment, if the current environment simultaneously meets multiple conditions (such as strong wind and pollution warning), the rule table can set compound entries or apply them in order of priority. For example, first apply the "strong wind" rule with the highest priority, and then consider the secondary adjustment brought by the "pollution warning".

[0060] Please refer to Table 1. Table 1 is a schematic conditional discriminant weight table provided by an embodiment of the present invention.

[0061] Table 1

[0062] In this embodiment, the importance score ranges from 0.0 to 1.0. The larger the value, the more credible the data source is under this condition. During actual fusion, the system can first obtain the current weather and environmental information, then look up the corresponding scores in a table, and use these scores for calculation as the weight coefficients for feature fusion (such as weighted average) to adjust the contribution degrees of each modality data. For example, in rainy or cloudy weather, it indicates that the credibility of remote sensing data decreases (the score takes a smaller value), the UAV data still has a certain reliability, and the sensor data has a higher credibility; the system can then reduce the influence of remote sensing features according to the weight ratio in the table and enhance the weight of sensor features, thereby improving the robustness of the fusion result. Through this conditional discriminant weight adjustment strategy, it can dynamically adapt to the changing marine environment and improve the overall monitoring accuracy.

[0063] Step 106: Perform weighted fusion on the remote sensing feature, the UAV feature, and the sensing feature based on the feature weights to generate a fusion feature; In this embodiment, the adjusting the feature weights of the remote sensing feature, the UAV feature, and the sensing feature based on the importance score and a preset reference weight includes: Obtain the reference weights of each feature, map each reference weight into a two-dimensional plane, construct a reference triangle with the reference weights as vertices, and calculate the centroid of the reference triangle; For each vertex of the reference triangle, calculate the unit vector pointing from the centroid to the vertex, and translate the vertex of the reference triangle along the direction of the unit vector according to the importance score to generate a deformed triangle; Calculate the centroid offset vector of the deformed triangle, and map the centroid offset vector into the direction vectors of each vertex to generate the regression factor of each feature; Perform incremental regression on the reference weights of each feature based on the regression factor of each feature to adjust the feature weights of the remote sensing feature, the UAV feature, and the sensing feature.

[0064] In this embodiment, the feature weights are updated according to the importance score queried from the rule table.

[0065] In this embodiment, first initialize the triangle base point, construct a reference triangle on the two-dimensional plane with three preset reference weights as vertices, and calculate the centroid of the reference triangle. For each vertex of the reference triangle, calculate its outer normal direction, that is, the unit vector pointing from the centroid to the vertex, and perform translation along the direction of the unit vector according to the importance score. Specifically: (1) Among them, is the vertex of the reference triangle , as the importance score of the feature is a unit vector are the vertices of the deformed triangle represents remote sensing represents an unmanned aerial vehicle represents a sensor

[0066] In this embodiment, three new vertices form a deformed triangle to calculate the centroid of the deformed triangle, and calculate the centroid offset vector based on this centroid. Specifically: (2) (3) where represents the centroid of the deformed triangle respectively represent the three vertices of the deformed triangle represents the centroid of the reference triangle represents the centroid offset

[0067] In this embodiment, project onto the direction vector of each vertex , and obtain the regression factor of each feature: (4) where is the regression factor, which represents the component of the "imbalance" degree of the current triangle in this direction

[0068] In this embodiment, perform incremental regression on the reference weights based on this regression factor: (5) where are the regression weights of each feature are the reference weights of each feature is the regression coefficient

[0069] In this embodiment, trim and normalize the regression weights of each feature to obtain the final feature weights of each feature

[0070] In this embodiment, the balance or imbalance degree of the importance scores of the three types of features is intuitively reflected through geometric offset, and the weight adjustment is directly driven based on the centroid offset, without setting complex thresholds or empirical rules

[0071] Step 107: Construct an environmental monitoring model based on the decision tree algorithm, input the fusion features into the environmental monitoring model, and obtain the environmental monitoring results

[0072] In this embodiment, after weighted fusion of remote sensing features, underwater drone features, and sensing features, a fused feature vector is obtained as the input of the environmental monitoring model.

[0073] In this embodiment, a decision tree model is trained using historical environmental monitoring data. The CART (Classification and Regression Tree) algorithm is selected. This algorithm uses the Gini index as the splitting criterion and is applicable to classification and regression tasks.

[0074] In this embodiment, the environmental monitoring results include water quality grades (such as Class I to Class V), predicted pollutant concentration values (such as ammonia nitrogen, phosphorus, heavy metals, etc.), early warnings for abnormal events (such as red tides, hypoxia, etc.), and ecological risk level assessments.

[0075] In this embodiment, after constructing the environmental monitoring model based on the decision tree algorithm, inputting the fused features into the environmental monitoring model, and obtaining the environmental monitoring results, the following steps are further included: Generating an error evaluation value and a credibility value based on the monitoring results, and generating a comprehensive evaluation score based on the error evaluation value and the credibility value; Incrementally adjusting the benchmark weights according to the comprehensive evaluation score.

[0076] In this embodiment, by comparing the model prediction results with actual observation data, the error evaluation value e is calculated. The error evaluation value e includes the mean absolute error (MAE) and the root mean square error (RMSE). Combining the stability performance of the environmental detection model in multiple batch evaluations, the result reliability score r, such as classification accuracy or regression value, is calculated to quantify the credibility of this detection result.

[0077] In this embodiment, the comprehensive evaluation score is specifically: (3) Where and are preset weight coefficients used to balance the influence of error and reliability on the feedback.

[0078] In this embodiment, incrementally adjusting the benchmark weights according to the comprehensive evaluation score includes: (4) Where is the benchmark weight, where is the feature type (remote sensing, drone, sensing) sensitivity coefficient, which can be calibrated offline in advance.

[0079] In this embodiment, the macro ocean environment is covered by remote sensing data, the UAV captures local biological details such as algae, the sensor provides continuous time-series parameters, breaking through the limitation of a single data source, avoiding insufficient response to local or dynamic environments, and then extracting the features of each type of data based on different neural networks respectively. Thus, the feature weights are dynamically adjusted according to environmental conditions and weather conditions, avoiding inaccurate measurement of a certain type of data caused by environmental conditions and weather conditions, and enhancing the adaptability of environmental monitoring to environmental changes. Using a decision tree model for environmental monitoring improves the interpretability and decision-making efficiency of the model, and enhances the stability and accuracy of the environmental monitoring of the marine ranch.

[0080] In an embodiment of the present invention, a terminal device is further provided, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the above-mentioned method for environmental monitoring of the marine ranch is implemented.

[0081] In an embodiment of the present invention, a computer-readable storage medium is further provided. The computer-readable storage medium includes a stored computer program, wherein when the computer program runs, it controls the device where the computer-readable storage medium is located to execute the above-mentioned method for environmental monitoring of the marine ranch.

[0082] Exemplarily, the computer program can be divided into one or more modules. One or more modules are stored in the memory and executed by the processor to complete the present invention. One or more modules can be a series of computer program instruction segments capable of performing specific functions, and these instruction segments are used to describe the execution process of the computer program in the terminal device.

[0083] The terminal device can be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The terminal device may include, but is not limited to, a processor, a memory, and a display. Those skilled in the art can understand that the above components are only examples of the terminal device and do not constitute a limitation to the terminal device. It may include more or fewer components than those described above, or combine some components, or different components. For example, the terminal device may further include input / output devices, network access devices, a bus, etc.

[0084] The so-called processor may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc. The processor is the control center of the terminal device and connects all parts of the entire terminal device through various interfaces and lines.

[0085] The memory can be used to store computer programs and / or modules. By running or executing the computer programs and / or modules stored in the memory, and by calling the data stored in the memory, the processor realizes various functions of the terminal device. The memory mainly includes a program storage area and a data storage area. Among them, the program storage area can store an operating system, application programs required for at least one function (such as a sound playback function, a text conversion function, etc.); the data storage area can store data created according to the use of the mobile phone (such as audio data, text message data, etc.). In addition, the memory may include high-speed random access memory and may also include non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0086] Among them, when the module for marine ranch environmental monitoring is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0087] The specific embodiments described above further elaborate on the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only for the specific embodiments of the present invention and is not used to limit the protection scope of the present invention. In particular, it is pointed out that for those skilled in the art, any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An environmental monitoring method for a marine ranch, characterized in that, Including: Obtain environmental monitoring data, where the environmental monitoring data includes remote sensing monitoring data, underwater drone data, and sensor data; Extract features from the remote sensing monitoring data based on a preset deep residual network to obtain remote sensing features; Extract features from the underwater drone data based on a preset Mask R-CNN network to obtain drone features; Extract features from the sensor data based on a preset recurrent neural network to obtain sensing features; Obtain environmental condition data, and adjust the feature weights of the remote sensing features, the drone features, and the sensing features based on a preset conditional discriminant weight table, a preset reference weight, and the environmental condition data; Perform weighted fusion on the remote sensing features, the drone features, and the sensing features based on the feature weights to generate fused features; Construct an environmental monitoring model based on a decision tree algorithm, input the fused features into the environmental monitoring model, and obtain an environmental monitoring result.

2. The method for monitoring the marine ranch environment according to claim 1, characterized in that, The obtaining of the environmental condition data, and adjusting the feature weights of the remote sensing features, the drone features, and the sensing features based on a preset conditional discriminant weight table, a preset reference weight, and the environmental condition data includes: Obtain the environmental condition data, where the environmental condition data includes weather data and environmental data; Determine the importance scores of the remote sensing features, the drone features, and the sensing features based on the weather data, the environmental data, and the conditional discriminant weight table; the conditional discriminant weight table includes the mapping relationship between the weather data and the importance scores of each feature; Adjust the feature weights of the remote sensing features, the drone features, and the sensing features based on the importance scores and the preset reference weight.

3. The method for monitoring the ocean ranch environment according to claim 2, characterized in that, The adjusting the feature weights of the remote sensing features, the drone features, and the sensing features based on the importance scores and the preset reference weight includes: Obtain the reference weights of each feature, map each reference weight to a two-dimensional plane, construct a reference triangle with the reference weights as vertices, and calculate the centroid of the reference triangle; For each vertex of the reference triangle, calculate the unit vector pointing from the center to the vertex, and translate the vertex of the reference triangle along the direction of the unit vector according to the importance score to generate a deformed triangle; Calculate the centroid offset vector of the deformed triangle, and map the centroid offset vector into the direction vectors of each vertex to generate the regression factors of each feature; Perform incremental regression on the reference weights of each feature based on the regression factors of each feature to adjust the feature weights of the remote sensing features, the drone features, and the sensing features.

4. The marine ranch environment monitoring method according to claim 1, characterized in that The extracting features from the remote sensing monitoring data based on a preset deep residual network to obtain remote sensing features includes: The deep residual network includes a number of residual units, and each residual unit is connected by a skip connection; Extract feature vectors from the remote sensing monitoring data based on each residual unit; Evaluate the similarity between pairwise feature vectors, and construct a feature vector topology graph based on the similarity; Determine the jump paths of each feature vector based on the feature vector topology graph and the maximum weight path algorithm, and fuse each feature vector based on the jump paths and jump connections to obtain remote sensing features.

5. The method for monitoring the marine ranch environment according to claim 1, characterized in that, The feature extraction of the underwater drone data based on the preset MaskR-CNN network to obtain drone features includes: The Mask R-CNN network includes a backbone network and a region proposal network; Based on the backbone network, perform feature extraction on the underwater drone data to obtain a multi-scale feature map set; Based on the region proposal network and the multi-scale feature map set, obtain a candidate region set; the candidate region set includes several candidate regions and the coordinate information of each candidate region; Based on the coordinate information of each candidate region and the multi-scale feature map set, map each candidate region to a feature block of a fixed size to obtain a feature block set; Fuse the feature block set to obtain drone features.

6. The marine ranch environment monitoring method according to claim 1, wherein, The feature extraction of the sensor data based on the preset recurrent neural network to obtain sensing features includes: The recurrent neural network includes several layers of LSTM units; Obtain sensor data, which includes temperature data, salinity data, dissolved oxygen data, pH data, conductivity data, and turbidity data; Based on the temperature data, salinity data, and conductivity data, construct a first learning projection matrix; Based on the dissolved oxygen data and pH data, construct a second learning projection matrix; Based on the turbidity data, construct a third learning projection matrix; Perform dynamic scaling on the first learning projection matrix, the second learning projection matrix, and the third learning projection matrix to achieve cross-parameter interaction and generate a fusion vector; Based on the recurrent neural network, perform feature extraction on the fusion vector to obtain sensing features.

7. A method for monitoring the marine ranch environment according to any one of claims 1 to 6, characterized in that, After constructing an environmental monitoring model based on the decision tree algorithm, inputting the fusion features into the environmental monitoring model, and obtaining an environmental monitoring result, it further includes: Generate an error evaluation value and a credibility value based on the monitoring result, and generate a comprehensive evaluation score based on the error evaluation value and the credibility value; Incrementally adjust the benchmark weight according to the comprehensive evaluation score.

8. An environmental monitoring system for a marine ranch, characterized in that, It includes: A data acquisition module, a feature extraction module, a feature fusion module, and a monitoring module; The data acquisition module is used to obtain environmental monitoring data, and the environmental monitoring data includes remote sensing monitoring data, underwater drone data, and sensor data; The feature extraction module is used to perform feature extraction on the remote sensing monitoring data based on the preset deep residual network to obtain remote sensing features; Perform feature extraction on the underwater drone data based on the preset Mask R-CNN network to obtain drone features; perform feature extraction on the sensor data based on the preset recurrent neural network to obtain sensing features; The feature fusion module is used to obtain environmental condition data, and adjust the feature weights of the remote sensing features, the drone features, and the sensing features based on a preset conditional discriminant weight table, a preset benchmark weight, and the environmental condition data; Perform weighted fusion on the remote sensing features, the UAV features, and the sensing features based on the feature weights to generate fused features; The monitoring module is configured to construct an environmental monitoring model based on a decision tree algorithm, input the fused features into the environmental monitoring model, and obtain an environmental monitoring result.

9. The marine ranch environment monitoring system according to claim 8, wherein The fusion module is configured to obtain environmental condition data, and adjust the feature weights of the remote sensing features, the UAV features, and the sensing features based on a preset condition discriminant weight table, a preset reference weight, and the environmental condition data, including: Obtain the environmental condition data, where the environmental condition data includes weather data and environmental data; Determine the importance scores of the remote sensing features, the UAV features, and the sensing features based on the weather data, the environmental data, and the condition discriminant weight table; the condition discriminant weight table includes the mapping relationship between the weather data and the importance scores of each feature; Adjust the feature weights of the remote sensing features, the UAV features, and the sensing features based on the importance scores and the preset reference weight.

10. The marine ranch environment monitoring system according to claim 9, characterized in that, The fusion module is configured to adjust the feature weights of the remote sensing features, the UAV features, and the sensing features based on the importance scores and the preset reference weight, including: Obtain the reference weights of each feature, map each reference weight to a two-dimensional plane, construct a reference triangle with the reference weights as vertices, and calculate the centroid of the reference triangle; For each vertex of the reference triangle, calculate the unit vector pointing from the centroid to the vertex, and translate the vertex of the reference triangle along the direction of the unit vector according to the importance score to generate a deformed triangle; Calculate the centroid offset vector of the deformed triangle, and map the centroid offset vector into the direction vectors of each vertex to generate the regression factors of each feature; Perform incremental regression on the reference weights of each feature based on the regression factors of each feature to adjust the feature weights of the remote sensing features, the UAV features, and the sensing features.

Images