Method for constructing marine ecological environment system
Through distributed model training and online collaborative reasoning technology, the problems of blind spots in drone information acquisition technology are solved, and efficient and accurate information generation and protection of the marine ecological environment are achieved.
Patent Information
- Application Number
- CN202410934946.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-11-12
AI Technical Summary
The existing UAV information collection technology has problems such as blind spots in the marine ecological environment monitoring, excessive data transmission bandwidth and insufficient computing power, resulting in the inability to achieve global environmental protection.
Using distributed model training and online collaborative inference technology, through the codec model training of drones and buoy terminals, data transmission is compressed, intermediate features are generated, and data sharing and model training are carried out at buoy nodes to form a complete marine ecological environment. system.
It improves data transmission efficiency and computing efficiency, reduces data transmission bandwidth and computing load, realizes accurate and efficient generation of global ecological environment information, and supports effective protection of marine ecological environment.
Smart Images

Figure CN118939415A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of marine environment monitoring, and in particular relates to a method for constructing a marine ecological environment system. Background Art
[0002] At present, environmental protection has become a key issue of concern to all sectors of society. For marine ecosystems, monitoring the specific conditions of their ecosystem environment can provide a scientific basis for the protection and management of marine ecosystems. However, due to the vast scope of the sea area, it is impossible to have overall control over the marine ecological environment, which leads to a series of problems.
[0003] Due to the complexity of the marine environment, it is difficult to effectively deploy monitoring equipment. At present, researchers have turned their attention to the use of drones to collect marine ecological environment information. Drones are relatively small and flexible. They are equipped with high-definition cameras to obtain clearer and more accurate environmental information, further monitor the marine ecology and effectively protect the environment. This method can also reduce certain equipment costs and reduce the impact on the environment. However, in the real application context, due to the overly complex marine ecological environment and the large sea area, research based on drone information collection technology still has many shortcomings in this context:
[0004] First, the existing drone information collection technology mainly uses data collection with a limited field of view of a single drone. As time goes by, there will be blind spots in the field of view, and there will be a lot of missing ecological and environmental information. At the same time, only relatively sparse marine environmental information can be obtained, and global environmental protection cannot be achieved. Secondly, for the existing intelligent technology, the original environmental data is transmitted to the buoy node. This process consumes too much bandwidth and the transmission delay is too long, and the environmental information of the marine ecology cannot be grasped in real time. And as time goes by, the amount of data reaches a certain level. Due to bandwidth limitations, the transmission of raw data will bring great problems. This solution is also difficult to meet the protection of the marine ecosystem. At the same time, the model constructed by the marine ecological environment information is complex and requires large computing power, which is difficult for drones to complete independently.
[0005] Distributed model training is a distributed way of training models. It can effectively solve the problems of high load and insufficient computing power when training drones. It can compress the original data transmitted by drones into intermediate representations of features and transmit them to buoy nodes to reduce data transmission bandwidth. It uses drones and buoys to distribute data training to achieve more efficient and fast model training. Collaborative perception is a distributed intelligent collaborative technology that combines ecological and environmental data from different devices or locations while reducing data transmission bandwidth. It obtains a shared multi-perspective environmental information scene map under the buoy node. Based on the location information of multiple devices, these local ecological and environmental information can be integrated to generate a global marine ecosystem information map.
[0006] Therefore, the research and design of efficient distributed collaborative methods has become an important topic in the field of marine ecology. Through the design of efficient distributed collaborative methods, it is possible to combine multiple edge nodes for data training under the premise of limited data bandwidth, and provide a high-accuracy and high-efficiency global ecological environment information generation model for the marine ecological environment protection platform, thereby solving the problem of incomplete ecological environment information in a single edge node scenario, and achieving more effective environmental protection. Summary of the invention
[0007] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a method for constructing a marine ecological environment system, comprising the following steps:
[0008] S1. Each UAV terminal and the corresponding buoy terminal train their own codec models to obtain multiple optimal models;
[0009] S2, multiple UAV terminals respectively obtain the ecological environment information of the local area and perform preliminary processing, and then transmit it to the corresponding buoy terminal. Each buoy terminal constructs the obtained ecological environment information through the corresponding optimal model obtained in step S1 to obtain multiple local ecological environment information;
[0010] S3, uploading the multiple local ecological environment information obtained in step S2 to the cloud server, and the cloud server integrates and splices the multiple local ecological environment information into sparse global environmental information;
[0011] S4. Supplement the sparse global environmental information obtained in step S3 to generate complete environmental information, namely, the marine ecological environment system.
[0012] Preferably, step S1 comprises the following steps:
[0013] S11, each UAV terminal collects its own position and ecological environment information within the field of view, and uses its own encoder to encode and compress the data to obtain multiple intermediate features, and sends them to the corresponding buoy terminal respectively;
[0014] S12. Each buoy terminal uses its own decoder to decode the received intermediate features. Based on the decoded information, each buoy terminal calculates the training loss of its own codec model and feeds it back to the corresponding UAV terminal to update the parameters of the codec model. Repeat the training multiple times to obtain multiple optimal models.
[0015] Preferably, step S1 comprises the following steps:
[0016] S111, each UAV terminal collects its own position and ecological environment information within the field of view, and uses its own encoder to encode and compress the data to obtain multiple intermediate features, and sends them to the corresponding buoy terminal respectively;
[0017] S122. Each buoy terminal receives the sent intermediate features respectively, and shares the received intermediate features with multiple adjacent buoy terminals. Then each buoy terminal uses its own decoder to decode the multiple intermediate features obtained. Each buoy terminal calculates the training loss of its own codec model based on the decoded information, and feeds back to the corresponding UAV terminal to update the parameters of the codec model. Repeat the training multiple times to obtain multiple optimal models.
[0018] Preferably, in step S12 and step S122, the calculation formula of the training loss is:
[0019]
[0020] Where i is the drone terminal index, k is the voxel index, L is the total number of voxels, c is the number of categories, and c=2, m i,k,c represents the kth voxel of the i-th drone terminal and is a one-hot encoded vector, represents the prediction of the kth voxel of the ith UAV terminal.
[0021] Preferably, in step S12 and step S122, the optimal model is obtained by the following method:
[0022] The voxel-level IOU values between the voxel-level labels predicted by each buoy terminal and the corresponding true labels of the UAV terminal are compared. When the voxel-level IOU value is greater than 0.5 and the average predicted accuracy is the highest, the optimal model is obtained.
[0023] Preferably, step S2 comprises the following steps:
[0024] S21, multiple UAV terminals respectively obtain ecological environment information of local areas, and use their own encoders to encode and compress data respectively, obtain multiple intermediate features, and send them to corresponding buoy terminals respectively;
[0025] S22. Each buoy terminal uses its own decoder and the corresponding optimal model obtained in step S1 to predict and decode the received intermediate features to obtain multiple local ecological environment information.
[0026] Preferably, the construction method provided by the present invention further includes step S5, and step S5 includes the following steps:
[0027] The cloud server distributes the marine ecological environment system obtained in step S4 to multiple buoy terminals for storage according to access requirements, and updates it locally and provides data support for multiple drone terminals.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] (1) The marine ecological environment system construction method provided by the present invention can utilize the information encoding shared by each edge end to splice into a complete marine ecosystem through distributed model training, online collaborative reasoning, and global ecological environment construction and joint design, thereby improving the execution efficiency of the task;
[0030] (2) The present invention adopts the design of distributed model training, whereby local ecological environment data can be generated by multiple drone nodes without sharing the original data, thereby reducing data transmission and computing load. The codec model is distributed on multiple different drones and buoys, and the training process is completed by the drone and buoy equipment. The training loss generated by the buoy equipment is used to update the parameters of the network. By collecting the intermediate representation of the data, the training efficiency is improved, and model training in a distributed environment is supported. Through the above technical means, a balance between bandwidth and performance can be achieved, and data can be shared efficiently. The original data is retained on the local device of the encoder, and only the intermediate features are shared. At the same time, the computing load is dispersed, and the computing efficiency and responsiveness of model training are improved.
[0031] (3) Through the design of online collaborative reasoning, the present invention can combine ecological environment data from different devices or locations while reducing data transmission bandwidth, and obtain multiple shared local environmental information under the buoy node;
[0032] (4) The present invention can splice and supplement multiple local environmental information through the design of global ecological environment construction, and finally form a marine ecological environment system. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] Figure 1A diagram showing the architecture of the global marine ecological environment scenario provided by the present invention;
[0034] Figure 2 Shows a flow chart of Embodiment 1 of the present invention;
[0035] Figure 3 It shows a signaling diagram of Embodiment 1 of the present invention;
[0036] Figure 4 Shows a flow chart of Embodiment 2 of the present invention;
[0037] Figure 5 It shows a signaling diagram of Embodiment 2 of the present invention;
[0038] Figure 6 Shows a flow chart of Embodiment 3 of the present invention;
[0039] Figure 7 It shows a signaling diagram of Embodiment 3 of the present invention;
[0040] Figure 8 Shows a flow chart of Embodiment 5 of the present invention;
[0041] Fig. 9 It shows the signaling diagram of embodiment 5 of the present invention. DETAILED DESCRIPTION
[0042] The following is a description of the embodiments of the present invention. Figures 1 to 9 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0043] like Figure 1 As shown, the core technical ideas of the present invention mainly include the following aspects:
[0044] (1) Distributed model training
[0045] The present invention introduces the idea of distributed model training and proposes an efficient training method for UAV terminals and buoy terminals. In the method proposed by the present invention, local ecological environment data are generated by multiple UAV terminals, and the codec model is distributed on multiple different UAV and buoy terminals. The training process is completed by the UAV terminal and the buoy terminal, and the training loss generated by the buoy terminal is used to update the parameters of the network. This distributed training method can reduce data transmission and computing load by collecting intermediate representations of data without sharing the original data, improve the efficiency of training, and support model training in a distributed environment.
[0046] (2) Online Collaborative Reasoning
[0047] The present invention generates a multi-perspective ecological environment map based on the idea of collaborative perception. During the reasoning process, each buoy terminal is equipped with the same codec model. Each training can synchronously generate multi-perspective local environmental information by sharing the environmental information encoding of multiple adjacent edge servers.
[0048] (3) Construction of the global ecological environment
[0049] This invention introduces a point cloud scene reconstruction method, which is based on the splicing and complementation of multiple local environmental information to eventually form a global marine ecological environment. First, the buoy terminal collects data to obtain local local environmental information. Then, these local environmental information are uploaded to the cloud server for information splicing to form sparse global ecological environment information. Finally, by complementing the environmental information, the sparse global ecological environment information is converted into dense global ecological environment information, and distributed to the buoy terminal for storage. At the same time, the present invention provides two access strategies: one is the access to the global ecological environment, and the other is on-demand access.
[0050] Example 1
[0051] Consider the following scenario: when training the codec model, the buoy terminal does not receive data from other buoy terminals, but only uses its own UAV terminal data to train the codec model. Therefore, in this scenario, the UAV terminal first sends a data transmission request to the corresponding buoy terminal based on the collected environmental information, and then the buoy terminal trains the codec model shared with the UAV based on the data features. After training the model once, the buoy terminal feeds back the training loss to the UAV terminal, thus completing a data training. Multiple UAV terminals and the corresponding buoy terminals can use their own data for training respectively, without the need to share data with other buoy terminals, and can be trained in a distributed and asynchronous manner.
[0052] In this embodiment, when training the codec model, it is assumed that the buoy terminal does not receive data from other drone terminals, and only uses its own drone terminal data to train the codec model. Based on the above scenario, an embodiment of the present invention provides a training method for a drone terminal and a buoy terminal, such as Figure 2 and Figure 3 As shown, the following steps are included:
[0053] Step 1: Encoding phase of codec model training
[0054] Each UAV terminal collects its own location and ecological environment information within its field of view, and uses its own encoder to encode data features and location information. At the same time, it uses spatial sampling and timestamps to further compress the encoded features to obtain multiple intermediate features, which are sent to the corresponding buoy terminals.
[0055] Step 2: Decoding phase of codec model training
[0056] Each buoy terminal uses its own decoder to decode the received intermediate features, and performs self-supervised training on its own codec model based on the decoded information. It calculates the training loss and feeds it back to the corresponding UAV terminal to update the parameters of the codec model. The training is repeated multiple times to obtain multiple optimal models and save them.
[0057] In this embodiment, the calculation formula of the training loss is:
[0058]
[0059] Where i is the drone terminal index, k is the voxel index, L is the total number of voxels, c is the number of categories, and c=2, m i,k,c represents the kth voxel of the i-th drone terminal and is a one-hot encoded vector, represents the prediction of the kth voxel of the ith UAV terminal.
[0060] In this embodiment, the optimal model is obtained by the following method:
[0061] The voxel-level IOU values between the voxel-level labels predicted by each buoy terminal and the corresponding true labels of the UAV terminal are compared. When the voxel-level IOU value is greater than 0.5 and the average predicted accuracy is the highest, the optimal model is obtained.
[0062] Example 2
[0063] Consider the following scenario: when training the codec model, the buoy terminal shares the data received from the drone terminal with other buoy terminals, requiring the data of multiple buoy terminals to be synchronized, and using the data of multiple drone terminals to train the codec model. Therefore, in this scenario, the drone terminal first sends a data transmission request to the buoy terminal based on the collected ecological environment information, and then the buoy terminal shares the received data with the adjacent buoy terminal, saves the training data synchronization, and the multiple buoy terminals sharing the data synchronously train the codec model. After training the model once, each buoy terminal feeds back the training loss to the drone terminal, thereby completing a model training and updating the model parameters.
[0064] In this embodiment, when the codec model is trained, the buoy terminal shares the data received from the drone terminal with other buoy terminals, requiring the data of multiple buoy terminals to be synchronized, and the codec model is trained synchronously using the data of multiple drone terminals. Based on the above scenario, an embodiment of the present invention provides a training method for a drone terminal and a buoy terminal, such as Figure 4 and Figure 5 As shown, the following steps are included:
[0065] Step 1: Encoding phase of codec model training
[0066] Each UAV terminal collects its own location and ecological environment information within its field of view, and uses its own encoder to encode and compress the data to obtain multiple intermediate features, which are then sent to the corresponding buoy terminal.
[0067] Step 2: Decoding phase of codec model training
[0068] Each buoy terminal receives the sent intermediate features respectively, and shares the received intermediate features with multiple adjacent buoy terminals. Then each buoy terminal uses its own decoder to decode the multiple intermediate features obtained, and generates a local environment information map based on the decoded information. The training loss of each codec model is calculated and fed back to the corresponding UAV terminal to update the parameters of the codec model. The training is repeated multiple times to obtain multiple optimal models and save them.
[0069] In this embodiment, the calculation formula of the training loss is:
[0070]
[0071] Where i is the drone terminal index, k is the voxel index, L is the total number of voxels, c is the number of categories, and c=2, m i,k,c represents the kth voxel of the i-th drone terminal and is a one-hot encoded vector, represents the prediction of the kth voxel of the ith UAV terminal.
[0072] In this embodiment, the optimal model is obtained by the following method:
[0073] The voxel-level IOU values between the voxel-level labels predicted by each buoy terminal and the corresponding true labels of the UAV terminal are compared. When the voxel-level IOU value is greater than 0.5 and the average predicted accuracy is the highest, the optimal model is obtained.
[0074] Example 3
[0075] Considering the following scenario, the drone terminal (distributed device) can only obtain ecological environment information of a local area, but its data processing capability cannot meet the needs of complete reconstruction of the ecosystem environment. Therefore, in this case, the embodiment of the present invention provides a method for constructing a marine ecological environment system, which is implemented based on the training method of the drone terminal and the buoy terminal provided in Example 1, such as Figure 6 and Figure 7 As shown, the specific steps include:
[0076] S1. Each UAV terminal and the corresponding buoy terminal (edge server) train the codec model on each buoy terminal to obtain multiple optimal models, which specifically includes the following steps:
[0077] S11, each UAV terminal collects its own position and ecological environment information within the field of view, and uses its own encoder to encode and compress the data to obtain multiple intermediate features, and sends them to the corresponding buoy terminal respectively;
[0078] S12, each buoy terminal decodes the received intermediate features using its own decoder. Based on the decoded information, each buoy terminal calculates the training loss of its own codec model and feeds it back to the corresponding UAV terminal to update the parameters of the codec model. The training is repeated multiple times to obtain multiple optimal models.
[0079] In this embodiment, the calculation formula of the training loss is:
[0080]
[0081] Where i is the drone terminal index, k is the voxel index, L is the total number of voxels, c is the number of categories, and c=2, m i,k,c represents the kth voxel of the i-th drone terminal and is a one-hot encoded vector, represents the prediction of the kth voxel of the ith UAV terminal.
[0082] In this embodiment, the optimal model is obtained by the following method:
[0083] The voxel-level IOU values between the voxel-level labels predicted by each buoy terminal and the corresponding true labels of the UAV terminal are compared. When the voxel-level IOU value is greater than 0.5 and the average predicted accuracy is the highest, the optimal model is obtained.
[0084] S2, multiple drone terminals respectively obtain the ecological environment information of the local area and perform preliminary processing, and then transmit it to the corresponding buoy terminal. Each buoy terminal constructs the obtained ecological environment information through the corresponding optimal model obtained in step S1 to obtain multiple local ecological environment information, which specifically includes the following steps:
[0085] S21, multiple UAV terminals respectively obtain ecological environment information of local areas, and use their own encoders to encode and compress data respectively, obtain multiple intermediate features, and send them to corresponding buoy terminals respectively;
[0086] S22, each buoy terminal uses its own decoder and the corresponding optimal model obtained in step S1 to predict and decode the received intermediate features, and save them in the form of point clouds to obtain multiple local ecological environment information;
[0087] S3. Due to the influence of interference and other factors, the local ecological environment information may not be complete, so the multiple local ecological environment information obtained in step S2 is uploaded to the cloud server, which is pre-processed and saved by the cloud server. Since the local ecological environment information contains all the information covered by the edge node, the cloud server integrates and splices the multiple local ecological environment information, specifically mapping the local space to the global space through the alignment and splicing of the point cloud data, forming a sparse global environment information with low confidence;
[0088] S4. Select an appropriate interpolation algorithm based on the characteristics of the point cloud and scene requirements, and use the information of the surrounding known point clouds to supplement the sparse global environmental information obtained in step S3 through interpolation methods to fill in the missing or increase the point cloud density, generate complete environmental information with high confidence, and obtain the marine ecological environment system; and store it in the cloud server for reference when making command decisions.
[0089] In an embodiment of the present invention, the global ecological environment information may become sparse due to incomplete local information and loss during the splicing process, but it can be converted into complete global ecological environment information through a supplementary model and constructed into a global ecological environment scene graph to provide a global reference for command decisions.
[0090] Example 4
[0091] Considering the following scenarios, the drone terminal (distributed device) can only obtain ecological environment information of a local area, but its data processing capability cannot meet the needs of complete reconstruction of the ecosystem environment. Therefore, in this case, the embodiment of the present invention provides a method for constructing a marine ecological environment system, which is implemented based on the training method of the drone terminal and the buoy terminal provided in Example 1, and specifically includes the following steps:
[0092] S1. Each UAV terminal and the corresponding buoy terminal (edge server) train their own codec models to obtain multiple optimal models, which specifically includes the following steps:
[0093] S111, each UAV terminal collects its own position and ecological environment information within the field of view, and uses its own encoder to encode and compress the data to obtain multiple intermediate features, and sends them to the corresponding buoy terminal respectively;
[0094] S122. Each buoy terminal receives the sent intermediate features respectively, and shares the received intermediate features with multiple adjacent buoy terminals. Then each buoy terminal uses its own decoder to decode the multiple intermediate features obtained. Each buoy terminal calculates the training loss of its own codec model based on the decoded information, and feeds back to the corresponding UAV terminal to update the parameters of the codec model. Repeat the training multiple times to obtain multiple optimal models.
[0095] In this embodiment, the calculation formula of the training loss is:
[0096]
[0097] Where i is the drone terminal index, k is the voxel index, L is the total number of voxels, c is the number of categories, and c=2, m i,k,c represents the kth voxel of the i-th drone terminal and is a one-hot encoded vector, represents the prediction of the kth voxel of the ith UAV terminal.
[0098] In this embodiment, the optimal model is obtained by the following method:
[0099] The voxel-level IOU values between the voxel-level labels predicted by each buoy terminal and the corresponding true labels of the UAV terminal are compared. When the voxel-level IOU value is greater than 0.5 and the average predicted accuracy is the highest, the optimal model is obtained.
[0100] S2, multiple drone terminals respectively obtain the ecological environment information of the local area and perform preliminary processing, and then transmit it to the corresponding buoy terminal. Each buoy terminal constructs the obtained ecological environment information through the corresponding optimal model obtained in step S1 to obtain multiple local ecological environment information, which specifically includes the following steps:
[0101] S21, multiple UAV terminals respectively obtain ecological environment information of local areas, and use their own encoders to encode and compress data respectively, obtain multiple intermediate features, and send them to corresponding buoy terminals respectively;
[0102] S22, each buoy terminal uses its own decoder and the corresponding optimal model obtained in step S1 to predict and decode the received intermediate features to obtain multiple local ecological environment information;
[0103] S3. Due to the influence of interference and other factors, the local ecological environment information may not be complete, so the multiple local ecological environment information obtained in step S2 is uploaded to the cloud server, and the cloud server integrates and splices the multiple local ecological environment information to form a sparse global environmental information;
[0104] S4. The sparse global environmental information obtained in step S3 is supplemented by interpolation and other methods to generate complete environmental information, and the marine ecological environment system is stored in the cloud server for reference in command decision-making.
[0105] In an embodiment of the present invention, the global ecological environment information may become sparse due to incomplete local information and loss during the splicing process, but it can be converted into complete global ecological environment information through a supplementary model and constructed into a global ecological environment scene graph to provide a global reference for command decisions.
[0106] Example 5
[0107] Considering the following scenario, the distributed device can only obtain the ecological environment information of the local area, but needs the global ecological environment information to support its own decision-making actions. Therefore, in this case, the embodiment of the present invention provides a method for constructing a marine ecological environment system, which is implemented based on the training method of the drone terminal and the buoy terminal provided in Example 1, such as Figure 8 and Fig. 9 The method specifically comprises the following steps:
[0108] S1. Each UAV terminal and the corresponding buoy terminal (edge server) train their own codec models to obtain multiple optimal models, which specifically includes the following steps:
[0109] S11, each UAV terminal collects its own position and ecological environment information within the field of view, and uses its own encoder to encode and compress the data to obtain multiple intermediate features, and sends them to the corresponding buoy terminal respectively;
[0110] S12, each buoy terminal decodes the received intermediate features using its own decoder. Based on the decoded information, each buoy terminal calculates the training loss of its own codec model and feeds it back to the corresponding UAV terminal to update the parameters of the codec model. The training is repeated multiple times to obtain multiple optimal models.
[0111] In this embodiment, the calculation formula of the training loss is:
[0112]
[0113] Where i is the drone terminal index, k is the voxel index, L is the total number of voxels, c is the number of categories, and c=2, m i,k,c represents the kth voxel of the i-th drone terminal and is a one-hot encoded vector, represents the prediction of the kth voxel of the ith UAV terminal.
[0114] In this embodiment, the optimal model is obtained by the following method:
[0115] The voxel-level IOU values between the voxel-level labels predicted by each buoy terminal and the corresponding true labels of the UAV terminal are compared. When the voxel-level IOU value is greater than 0.5 and the average predicted accuracy is the highest, the optimal model is obtained.
[0116] S2, multiple drone terminals respectively obtain the ecological environment information of the local area and perform preliminary processing, and then transmit it to the corresponding buoy terminal. Each buoy terminal constructs the obtained ecological environment information through the corresponding optimal model obtained in step S1 to obtain multiple local ecological environment information, which specifically includes the following steps:
[0117] S21, multiple UAV terminals respectively obtain ecological environment information of local areas, and use their own encoders to encode and compress data respectively, obtain multiple intermediate features, and send them to corresponding buoy terminals respectively;
[0118] S22, each buoy terminal uses its own decoder and the corresponding optimal model obtained in step S1 to predict and decode the received intermediate features to obtain multiple local ecological environment information;
[0119] S3. Due to the influence of interference and other factors, the local ecological environment information may not be complete, so the multiple local ecological environment information obtained in step S2 is uploaded to the cloud server, and the cloud server integrates and splices the multiple local ecological environment information to form a sparse global environmental information;
[0120] S4, supplementing the sparse global environmental information obtained in step S3 by interpolation and other methods to generate complete environmental information, i.e., the marine ecological environment system;
[0121] S6. The cloud server distributes the marine ecological environment system obtained in step S4 to multiple buoy terminals for storage according to access requirements, and updates it locally and provides data support for multiple drone terminals. The embodiment of the present invention provides two access strategies: one is global ecological environment access, and the other is on-demand access.
[0122] Although the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.
Claims
1. A method for constructing a marine ecological environment system, characterized in that: The following steps are involved: S1. Each UAV terminal and the corresponding buoy terminal train their own codec models to obtain multiple optimal models; S2, multiple UAV terminals respectively obtain the ecological environment information of the local area and perform preliminary processing, and then transmit it to the corresponding buoy terminal. Each buoy terminal constructs the obtained ecological environment information through the corresponding optimal model obtained in step S1 to obtain multiple local ecological environment information; S3, uploading the multiple local ecological environment information obtained in step S2 to the cloud server, and the cloud server integrates and splices the multiple local ecological environment information into sparse global environmental information; S4. Supplement the sparse global environmental information obtained in step S3 to generate complete environmental information, namely, the marine ecological environment system.
2. The method for constructing a marine ecological environment system according to claim 1, characterized in that: Step S1 comprises the following steps: S11, each UAV terminal collects its own position and ecological environment information within the field of view, and uses its own encoder to encode and compress the data to obtain multiple intermediate features, and sends them to the corresponding buoy terminal respectively; S12. Each buoy terminal uses its own decoder to decode the received intermediate features. Based on the decoded information, each buoy terminal calculates the training loss of its own codec model and feeds it back to the corresponding UAV terminal to update the parameters of the codec model. Repeat the training multiple times to obtain multiple optimal models.
3. The method for constructing a marine ecological environment system according to claim 1, characterized in that: Step S1 comprises the following steps: S111, each UAV terminal collects its own position and ecological environment information within the field of view, and uses its own encoder to encode and compress the data to obtain multiple intermediate features, and sends them to the corresponding buoy terminal respectively; S122. Each buoy terminal receives the sent intermediate features respectively, and shares the received intermediate features with multiple adjacent buoy terminals. Then each buoy terminal uses its own decoder to decode the multiple intermediate features obtained. Each buoy terminal calculates the training loss of its own codec model based on the decoded information, and feeds back to the corresponding UAV terminal to update the parameters of the codec model. Repeat the training multiple times to obtain multiple optimal models.
4. The method for constructing a marine ecological environment system according to claim 2 or 3, characterized in that: In step S12 and step S122, the calculation formula of the training loss is: Where i is the drone terminal index, k is the voxel index, L is the total number of voxels, c is the number of categories, and c=2, m i,k,c represents the kth voxel of the i-th drone terminal and is a one-hot encoded vector, represents the prediction of the kth voxel of the ith UAV terminal.
5. The method for constructing a marine ecological environment system according to claim 2 or 3, characterized in that: In step S12 and step S122, the optimal model is obtained by the following method: The voxel-level IOU values between the voxel-level labels predicted by each buoy terminal and the corresponding true labels of the UAV terminal are compared. When the voxel-level IOU value is greater than 0.5 and the average predicted accuracy is the highest, the optimal model is obtained.
6. The method for constructing a marine ecological environment system according to claim 1, characterized in that: Step S2 comprises the following steps: S21, multiple UAV terminals respectively obtain ecological environment information of local areas, and use their own encoders to encode and compress data respectively, obtain multiple intermediate features, and send them to corresponding buoy terminals respectively; S22. Each buoy terminal uses its own decoder and the corresponding optimal model obtained in step S1 to predict and decode the received intermediate features to obtain multiple local ecological environment information.
7. The method for constructing a marine ecological environment system according to claim 1, characterized in that: The step S5 further comprises the following steps: The cloud server distributes the marine ecological environment system obtained in step S4 to multiple buoy terminals for storage according to access requirements, and updates it locally and provides data support for multiple drone terminals.
