KR102401337B1 - Method for Identification of Fishing Gear Type using Trajectory Data - Google Patents

Abstract

The present invention prepares track data for learning including the position, course, and speed of a plurality of fishing boats from the automatic ship identification system (AIS), and provides marine environment data including tidal difference, day/night, water temperature, water depth, and seabed type. Prepare, perform a data preprocessing process including interpolating the training track data with respect to a reference value and aligning them at regular time intervals, and combine the training track data and marine environment data to arrange a track window dataset in time series Generates, labels the fishing license information of the deep learning target fishing vessel in the wake window dataset, performs deep learning with a neural network model on the wake window dataset, receives the track data of the discrimination target fishing vessel, and the neural network It provides a fishing vessel fishing discrimination method using the track data, which includes the step of determining the type of fishing of the fishing vessel from the tracking data of the fishing vessel based on the results learned by the model.

Description

{Method for Identification of Fishing Gear Type using Trajectory Data}

The present invention relates to a fishing vessel fishing determination method, and in particular, proposes a new method for accurately determining the type of fishing through deep learning on the tracking data of the fishing vessel.

More than 37% of the world's population consumes protein from fisheries sources. Recently, due to climate change and an increase in the world population, fishery resources have been greatly reduced. In particular, illegal fishing by fishing boats and overfishing of aquatic resources are a direct cause of the depletion of fishery resources.

In order to prevent illegal fishing and overfishing by fishing boats, many countries around the world are promoting fishery monitoring, fishing quota system, and prohibition of fishing. However, in the fishery monitoring system, there was a limitation in that it was not possible to know in which sea area the fishing boat was fishing and what kind of fishing gear was used.

In terms of monitoring fishing vessels, most fishing vessels operate with a fishing vessel license when they are built, but it was not possible to know what type of fishing vessel the fishing vessel was engaged in on the Automatic Ship Identification System (AIS) or radar, which is a device that can check vessel tracking data. . On the other hand, maritime-related organizations can check what type of fishing license the vessel has, but it is difficult to know what type of fishing the vessel is operating if the location transmitter does not operate or it is a foreign fishing vessel.

In addition, in relation to the identification of illegal fishing, the identification of illegal fishing conducted by government offices until now was judged by whether the vessel was inspected or entered a prohibited sea area, and whether there was any fishing activity during the prohibited fishing season. Here, the fishing activities when entering the prohibited sea area or during the fishing season are identified by the ship location data. However, there are still many fishing boats engaged in unauthorized fishing for a large number of fisheries. However, the relevant authorities are only conducting crackdowns on fishing with fishing gear that is not permitted by inspection. It was difficult to determine whether

In the past, attempts have been made to determine whether a fishing vessel is operating using the Vessel Monitoring System (VMS), which is a vessel monitoring system. Even in this attempt, not only did it only judge the three types of fishing due to the deterioration of the VMS data, but it also required at least two weeks of track data to determine the fishing vessel operation, so there is a problem in effectiveness. In addition, there was an attempt to use the speed information of the fishing vessel to determine whether the fishing vessel is fishing. Statistical analysis of the speed of the fishing vessel during fishing activity was used to determine whether the fishing vessel was operating at 4 - 6 knots by throwing net, at 2 - 4 knots by fishing net, or other In the case of speed, a coarse level operation discrimination method to the extent that it is classified into other operations was proposed.

On the other hand, according to Patent Registration No. 10-2113955, as a method for monitoring the periphery of a ship and a port, the first camera and the second camera installed in the port to capture an image, the first including at least one of the ship and the sea A port image and a second port image that is at least partially different from the first port image and the monitoring area are acquired, and through the process of processing these port images, the vessel's surroundings and port conditions are identified when the vessel is berthing or berthing, and berthing or berthing However, it is impossible to determine whether a ship far away from the port is operating or the type of operation.

In addition, according to Patent Registration No. 10-1639362, it relates to an illegal fishing vessel monitoring system, which is mounted on a monitoring vessel for monitoring illegal fishing and performs a broadcast requesting vessel information including fishing permission information from the fishing vessel, and A fishing certification processing unit that receives the response signal and processes whether the fishing vessel is permitted to operate, and the database that is connected to the fishing certification processing unit and records the registration information of the fishing permission, and is installed in the fishing vessel and operates with the operation certification processing unit A terminal communicator that performs wireless communication and transmits a response signal in response to a broadcast requesting vessel information including operation permission information of the operation authentication processing unit is provided. Although information on whether or not a fishing license exists can be remotely checked from a monitoring vessel on a fishing vessel, a monitoring vessel operating near the fishing vessel in the sea is required separately, and it is also impossible to accurately determine the type of fishing vessel.

There is an urgent need to develop a fishing vessel identification and remote monitoring technology that can be effectively applied to prevent overfishing or illegal fishing by overcoming such technical limitations.

The present invention was conceived under the above technical background, and an object of the present invention is to provide a method for accurately determining whether a fishing vessel is operated and the type of operation.

Another object of the present invention is to provide a new method capable of determining the type of fishing vessel by using the existing vessel-related equipment and the like and using the general vessel-related data.

Another object of the present invention is to provide a method and system capable of effectively coping with overfishing of water resources or illegal fishing in marine control organizations or fisheries-related organizations.

In addition, other objects and technical features of the present invention will be presented in more detail in the following detailed description.

In order to achieve the above object, the present invention is a method for determining a fishing vessel operation using a computer system that is capable of network communication and includes a deep learning model, wherein the system includes the position, course, Receives track data for learning including speed, the system receives marine environment data including tide difference, day/night, water temperature, water depth, and seabed type, and the system interpolates with respect to a reference value for the learning track data to perform a data preprocessing process including sorting at regular time intervals, the system combines the learning track data and marine environment data, sets several data sets, and arranges them in a time series as a single data form. Creating a dataset, the system labels the fishing license information of the deep learning target fishing vessel in the wake window dataset, the system performs deep learning with a neural network model on the wake window dataset, and the system determines It provides a fishing vessel fishing discrimination method using the track data, comprising the step of receiving the track data of the target fishing vessel, and determining the fishing type of the corresponding fishing vessel from the tracking data of the fishing vessel based on the result learned by the neural network model.
In the present invention, it is preferable to perform deep learning by applying a convolutional neural network (CNN) model to the data on the course and speed of a fishing vessel in the wake window dataset and applying a fully connected neural network (FCNN) model to the marine environment data.

delete

In addition, in the present invention, the wake window dataset is It is preferable to configure a plurality of overlapping datasets at time intervals.

In the present invention, it is sufficient to receive track data for 10 to 24 hours to determine the type of fishing vessel to be identified, and for fishing vessels in operation, 6 including a fishing net, an angang net, a streak, a drift net, a trap, and a trawl It can be determined to be purified by any one of the dog's operations.

According to the present invention, it is possible to determine whether and what kind of fishing the fishing vessel is operating only with the tracking data of the fishing vessel.

The fishing vessel fishing discrimination method of the present invention processes the track data obtained from a commonly used automatic vessel identification system and adds marine environment data to improve the accuracy of fishing vessel identification through deep learning, and provides an effective fishing vessel fishing identification system without excessive cost investment. can be built In addition, even if the fishing vessel does not operate the vessel position for illegal fishing, it is possible to determine whether the fishing vessel is illegally fishing only with the vessel movement pattern information on the radar and satellite light signals.

The fishing boat fishing detection method proposed in the present invention will be able to greatly contribute to the illegal fishing detection of fishing boats as well as fishing boat fishing monitoring in related organizations.

1 is a wake pattern according to the movement of a fishing vessel
Figure 2 is a wake pattern according to the type of fishing boat
Figure 3 is a flow chart showing the outline of the fishing vessel fishing discrimination method of the present invention
4 is a schematic diagram showing a data interpolation method
5 is a schematic diagram showing the outline of a deep learning model
6 is a schematic diagram of the wake window data set and the data set superimposed
7 is a schematic diagram showing a CNN-based feature extraction module
8 is a graph showing the change in fishing vessel fishing identification accuracy (TPI) according to the wake window size
9 is a wake pattern showing the determination result for fishing boat fishing of the learning model

The present invention learns the sailing data pattern of a fishing vessel with a deep learning model, and proposes a method of determining whether to operate and what type of operation is performed using the navigation data of the fishing vessel.

Most of the fishing vessels sailing are transmitting the position, course, and speed information of the fishing vessel through the Automatic Identification System (AIS). The fishing vessel fishing identification method proposed in the present invention receives track data for a certain period of time, for example, 10 to 24 hours, from AIS, pre-processes this track data, and processes it time-series together with marine environment data to learn as a deep learning model. It is the learned model that accurately determines whether or not fishing is performed from the track data of other fishing boats. As a deep learning model, a convolutional artificial neural network (CNN) model can be used, and the variables used for machine learning are the position, speed, course (change) of the fishing vessel, which is wake data, and the tide, day/night, and the marine environment data. It includes data such as water temperature, water depth, and seabed type.

Fishing vessel fishing and wake patterns

In general, fish migrate periodically or seasonally according to appropriate temperature, pH, dissolved oxygen, and food (plankton). For example, if the tidal difference is small, plankton are on the sea floor, so the algae stay near the sea floor. Therefore, depending on the tidal difference, the fish have various patterns, and it also affects the fishing pattern of the fishing vessel. 1 is a wake pattern of a fishing vessel appearing in AIS data. As a non-fishing pattern, (a) is sailing, (b) is a standby state, and as a tracking pattern for fishing, (c) throwing net and (d) fishing net are shown, respectively.

On the other hand, the type of fishing gear is related to the method of fishing. Fishing methods mainly used in coastal waters include purchase seine, stow net, longline, drift gill net, trap, and trawl. Each type of fishing gear shows different fishing patterns because the shape and required time of the throwing net and the netting are different (refer to FIG. 2). On the other hand, the patterns of sailing and waiting conditions without fishing appear similar in all types of fishing vessels. In spite of the different fishing patterns for each type of fishing vessel, the fishing vessel has a very complex form because the fishing section and the non-fishing section are mixed with each other. Due to the complexity and irregularity of such tracks, it is difficult for fisheries (fishery) monitoring personnel to determine what kind of fishing vessel the fishing vessel is engaged in.

The fishing vessel sends a ship location message, and the nearby fisheries resource authorities or maritime traffic control center collects the location signal of the fishing vessel and uses it for monitoring. If the type of fishery is identified in the AIS message sent from the fishing vessel, the fisheries authorities will be able to manage illegal fishing (IUU) and overfishing. The present invention proposes a method for identifying a fishery type from AIS message data.

There are a total of 34 types of fishing vessels licensed in Korea. Excluding double carriers and farm fishing vessels, the remaining 21 types of vessels can be broadly classified into 6 types: seine nets, angang nets, longline nets, drift nets, traps, and trawls (see Table 1). In the present invention, in determining the operation of the fishing vessel, the determination result for a total of six operations is presented based on the track data, and the type of operation of the fishing vessel (or the fishing license information of the fishing vessel) is labeled when generating the track window data to be described later. In this case, information of any one of the six operation types is reflected in the wake window data.

[Table 1]

Overview of Fishing Vessel Fishing Identification

The present invention receives track-related information including the location, course, and speed of a plurality of fishing vessels from Automatic Identification System (AIS) data, and, along with this, tidal cars, day/night, water temperature, water depth, and seabed type Receive marine environment data from related organizations. Wake data and marine environment data are used as machine learning variables to discriminate fishing boats. From the track data, a vessel whose type is a fishing vessel is extracted and the data is stored for each vessel.

The track data extracted for each fishing vessel in AIS contains error information unnecessary for operation determination or the message reception interval is non-uniform, so a pre-processing process is performed to apply it to the learning model. For example, a data pre-processing process including interpolating the tracking data for training with respect to a reference value and aligning it at regular time intervals is performed. For deep learning, the track data for learning and the marine environment data are combined to create a track window dataset arranged in time series, and the fishing gear type of the deep learning target is labeled in the wake window dataset.

The wake window dataset is set to reflect the complex wake pattern that varies depending on the type of operation. Multiple datasets are constructed by overlapping datasets at time intervals. The size is important in composing the track window data, and in the present invention, the track data for learning and the track data of the fishing vessel to be discriminated can receive the track data for 10 to 24 hours to constitute the track window data set.

Deep learning is performed with a neural network model on the generated wake window dataset. As a machine learning model, a convolution model, an LSTM model, an SVM model, etc. can be used. In the embodiment of the present invention, a convolutional neural network (CNN) model is applied to the data on the course and speed of a fishing vessel in the wake window dataset, and Deep learning is performed on marine environment data by applying a fully connected neural network (FCNN) model.

Thereafter, the tracking data of the fishing vessel to be determined is received, and the fishing type of the fishing vessel is determined from the tracking data of the fishing vessel based on the result learned by the neural network model. It is discriminated as any one of six operations including drift nets, traps, and trawls.

The sequence of such an operation determination method is schematically shown in FIG. 3, and below, the technical characteristics of each step will be described in more detail.

resistance of data Pretreatment

The AIS message includes vessel-specific information and real-time movement-related information of the vessel. This information is transmitted to neighboring ships at different transmission intervals of 2 to 10 seconds depending on the ship's sailing status in the VHF band frequency. AIS data may contain noise due to interference from signals of adjacent ships or geographic obstacles, and may also temporarily display remote locations due to GPS errors. In order to determine the type of operation from the fishing vessel tracking data, error information must be removed, and data with a constant receiving cycle interval are required to perform learning by deep learning.

To this end, first, the Euclidean distance between the received data of the same previous vessel and the current received data is calculated, and error information greater than the predefined distance is removed. Second, the fishing vessel also removes information on areas where vessel fishing is not performed, such as port areas and marinas. Third, the position, course, and speed information of AIS data transmitted at irregular intervals is converted at regular time intervals.

4 schematically shows a conversion process of received information according to a change in time, and red indicates actual non-regular received data. Blue color indicates information generated at a reference time and a constant time interval (sampling time Δt , for example, 60 seconds). For computational efficiency, the non-regular real ship AIS information is converted into AIS information (position, course, speed) at regular intervals by linear interpolation.

The interpolation process is as follows. Assuming that t _k is the k-th sampling time and time( M _i ) is the time at which the AIS message M _i is received, if the AIS message M _i appears between time k and time k+ ₁ , the location (latitude lat _i , longitude lon _i ), course, and speed are converted to values of position and course speed at t _k+1 . When Δt = t _k+1 - time( M _i ) , the conversion formula for the position is as follows.

Fishing vessels usually use low-cost, low-quality equipment, so route or speed information may not be accurate, and this information is often lost. In addition, since the AIS packet is measured at a specific moment, speed and route information may not be accurate. In the present invention, speed and route information are calculated from interpolated location data instead of AIS packet information.

Assuming that the latitude and longitude differences between the interpolated positions pp( t _i ) and pp( t _{i +1} ) are Δlat and Δlon , respectively, the angle θ of the ship’s moving direction and the course Co( t _i ) are calculated as follows .

The speed Spd( t _i ) is calculated as follows by dividing the distance traveled by the elapsed time.

The cruising speed of a vessel is related to whether it is operating or not. In an embodiment of the present invention, a fishing vessel with a cruising speed of 6 knots or more is regarded as not operating.

deep learning Model

Fishing vessels are characterized by providing a unique wake pattern according to the movement data of the vessel and the marine environment variables at the time of operation. In order to discriminate the fishing vessel operation, it is necessary to investigate the partial movement and long-term movement fluctuation of the fishing vessel, and there is a limit to the improvement of accuracy in mathematical modeling and statistical analysis methods for this. On the other hand, since machine learning creates a model based on data, it is possible to model in consideration of various non-linear variables, and this limitation can be effectively improved. there is.

The present invention proposes a method for determining fishing boat operation based on a deep learning model. A convolutional neural network and a fully connected neural network are used in the front part of this deep learning model, and a fully connected neural network is used again in the rear part. This model receives the tracking data and marine environment data of the fishing vessel and determines the type of fishing vessel involved. 5 shows the overall structure of the artificial neural network model proposed in the present invention.

wake window dataset produce

The learning data of the deep learning model proposed in the present invention uses time series data of ship course change and speed, and marine environment data of water depth, water temperature, day/night, tide difference, and sea bed quality at the center of each time series.

The input data to be applied to the deep learning model is processed into a dataset in the preprocessing module to be effectively used for learning while including the dynamic change of the wake pattern of the fishing vessel. First, the time series data of course and speed is a set of several data, and the trajectory window ( TW(i)) ) is converted to A reference point is needed to extract TW . If the reference point is t _i and the length of TW is s , TW(i) can be expressed as

One TW contains sufficient information for the fishing vessel to determine the type of fishing vessel fishing, such as throwing nets, sheep nets, and atmospheric air. extracted. The wake window dataset is set to reflect the complex wake pattern that varies depending on the type of operation. It is preferable to configure a plurality of data sets by overlapping data sets at time intervals. FIG. 6 shows individual wake windows including data for 24 hours, and each track window in course and speed change is generated by overlapping while moving at a constant step (time interval).

Course change (CC( t _i )) in course data is defined as the absolute value of the course difference (Cd( t _i )) at a specific time point, and is obtained as follows.

In addition, in the deep learning learning and operation discrimination model proposed in the present invention, the training dataset FAD(i) is t _{i +s- 1} at the time t _i as follows. It consists of the values of course change and speed.

Marine environmental factors are factors that affect fishing activities, such as day/night, tidal difference, water temperature, water depth, and seabed type. In the present invention, the marine environment dataset of the deep learning model includes all these factors. Specifically, as the day/night variable has the characteristic that some fish cluster according to the light, the fish gathers especially when the light is lit at night, and the daytime is set to 0 and the nighttime is set to 1. The tidal difference (tidal velocity) is closely related to the movement of plankton, which is food for fish, and the water depth and water temperature differ depending on the environment depending on the preferred water temperature because fish are cold-blooded animals, and the corresponding values were used as data. . The bottom quality of the seabed affects the hatching or living of fish, and in the embodiment of the present invention, it is divided into five types: mud, shell, rock, sand, and gravel. , were classified as 0, 1, 2, 3, and 4, respectively.

Feature Extraction and Prediction module

In the deep learning model of the present invention, a feature extraction and prediction process based on input data will be schematically described. A well-known CNN algorithm may be referred to for matters related to the application and learning process of the deep learning model.

The feature extraction extracts the features of meaningful data from the preprocessed data, and the prediction module determines the fishing type of the fishing vessel. In the feature extraction module, the sub-module of 1D CNN is a deep learning model architecture that extracts k number of features by preprocessing the course change and speed data of s-1 fishing boats. The structure of this model is as follows.

In the above architecture, 1D Conv(30) means that a one-dimensional data vector is a feature extracted with 30 convolutional layers. 1D Conv uses input data of course change and speed.

The convolution operation extracts features from the input data using the kernel as follows: Let the one-dimensional input data be C ^{n - 1} and the one-dimensional array kernel W ⁿ . In the formula below, * denotes a convolution operation, and f denotes an activation function. In the embodiment of the present invention, the ReLU function was used as the activation function. When the convolution kernel W ⁿ is applied to the input data C ^{n -1} , the output C ⁿ is calculated as where b ⁿ represents the model bias. When learning, the convolutional kernel parameters are automatically learned according to the training data.

Max pooling extracts the maximum value from the feature values and reduces the feature values to one value. The convolutional layer uses a 3×1 convolution kernel, and maxpooling is used with a 4×1 window size. The input is data of ship course change and speed, which is one channel and is given in the form of (s-1)×1 size. Here, s represents the number of samples sampled by interpolating the track data. In an embodiment of the present invention, the length of the trajectory window is 24 hours, and the sliding trajectory window is set to overlap by moving every 3 minutes. Therefore, the total number of interpolation times is 480. That is, it has 480 data samples.

On the other hand, the output is generated as an array of k×1 . where k is the number of nodes in the output layer of the feature extraction module. In the embodiment, k is 3570 (119×30) pieces. 7 schematically shows a result in which a result input from the feature extraction module is displayed as a one-dimensional feature vector.

Operating environment data such as tides, day/night, water temperature, water depth, and seabed type were created in the form of one-dimensional vector data, and features were extracted by learning it on the following fully connected neural network (FCNN) model.

Here, FCL(p) means a fully connected layer with p nodes.

The output values of the one-dimensional CNN-based feature extraction module of fishing vessel tracking data and the FCNN-based feature extraction module of marine environment data are combined and transmitted to the prediction module, and the prediction module has the following structure.

Based on FCNN, the prediction module consists of a hidden layer of 50 and 30 nodes, respectively, and an output layer of 6 nodes (the same as the number of operation types in the present invention), and the final output value is calculated by applying the softmax process to the output node result do.

performance rating index

In order to evaluate the fishing vessel operation discrimination performance according to the deep learning learning model, the track data of the fishing vessel to be discriminated, not the data for learning, was generated. For the test data for fishing identification, a total of 480 wake window datasets, TW(i), were created by moving and overlapping the wake window dataset for 24 hours for a fishing vessel in 3 minute units. information was excluded.

The trajectory window-wise performance index (TPI), defined by the following equation, was used for performance evaluation.

는 조업일의 어선에 관한 항적데이타 수,

는 조업일의 어선에 대한 n 번째 항적데이타,

는

에 대한 조업 종류,

는 실제 조업 종류이고,

는 a=b 일 때는 1, 다른 경우는 0 으로 결정되는 함수이다. TPI 는 0 과 1 사이의 값을 가지며, TPI 가 클수록 학습 모델의 정확도가 우수하다.Here, D _te is the number of fishing days from which the track data is extracted, S _d is the number of fishing boats operating on a specific day,

is the number of track data on fishing boats on a working day,

is the nth track data for the fishing boat on the working day,

Is

for the type of fishing,

is the actual type of operation,

is a function determined to be 1 when a=b and 0 otherwise. TPI has a value between 0 and 1, and the higher the TPI, the better the accuracy of the learning model.

In addition, in the present invention, as another performance evaluation index, a daily performance index, DPI (day-wise performance index) defined by the following equation was used.

는 최대 빈도를 나타내는 조업 종류를 나타낸다. 선박의 조업은 당일에 하나의 방식만 수행하는 것으로 추정할 수 있는데 한번에 여러가지 조업 방식을 수행하기 어렵기 때문이다. DPI 가 클수록 학습 모델의 정확도는 우수한 것으로 판단할 수 있다.here,

indicates the type of operation representing the maximum frequency. It can be estimated that the operation of a ship is carried out in only one method on the same day, because it is difficult to perform several operation methods at once. It can be judged that the higher the DPI, the better the accuracy of the learning model.

Example

AIS track data were collected in the southern seas of the Korean Peninsula including Jeju Island. The daily data for the region was 2.5 gigabytes, and the total amount of data was about 900 gigabytes. In the collected data, the data related to fishing vessels is one-third, and 5,865,696,000 pieces of data including static and dynamic information of 1380 ships were classified by fishing vessel, and the track data was interpolated according to the method described above.

For learning and discrimination application as a deep learning model, track data for each fishing vessel was interpolated in units of 1 minute and a track window dataset was created with a length of 24 hours. The wake window dataset was constructed by moving it to overlap at 3-minute intervals, and the total number of the wake window datasets was 52,560,000. Of these, 80% of data (292 days) was used for the learning model, and the remaining 20% of data (73 days) was used for operation discrimination. In the dataset, a total of 24,966,000 wake window datasets were used as data for deep learning learning and operation discrimination, excluding information related to non-fishing activities such as high-speed navigation or anchoring near ports.

The training data has data imbalance by label as follows: envy class 21.2% (4,234,234 records), eye network class 15.7% (3,135,730 records), continuous class class 12.4% (2,476,627 records), drift net class 9.2% (1,837,498 records), trap Class 18.4% (3,674,995 records), troll class 23.1% (4,613,717 records). In order to resolve this data imbalance, only 1,000,000 data were randomly extracted for every training from each label. In the test data set, the performance was evaluated by extracting 300,000 data as well.

After performing learning with the deep learning model as described above for the learning data, the performance of the learning model was evaluated using the test data for operation discrimination. The performance of the learning model was particularly affected by the wake window size, and FIG. 8 shows the change in accuracy (TPI) for each ship type when the wake window size is 1, 5, 10, 15, 20, 25h.

In relation to the fishing vessel fishing discrimination method of the present invention, a discrimination model was generated based on another machine learning model SVM (support vector model) for comparative evaluation of performance. In this comparative model, the course change and speed were separated by 10 intervals, a histogram was generated, and it was used as an input value together with the marine environment data. There are three kernel functions applied to this model: sigmoid, polynomial, and linear. As a result of checking the performance of the learning model for each type, the linear case was excellent.

a) TPI index comparison

By determining the type of fishing for 24 hours of fishing vessel tracking data, each TPI index was evaluated based on the SVM model and the CNN model of the present invention, and the results are shown in Table 2.

[Table 2]

The present invention showed a performance of 0.92 or more in the identification of envy, trap, and trawl operations, and the SVM model had a lower performance index than the present invention except for the determination of winning streak, and the discriminating performance by operation was also severe. This result is considered to be due to considering the continuous change of the speed and course of the fishing vessel when constructing the wake window dataset in the present invention.

b) DPI index comparison

DPI is related to the performance of discriminating the type of operation from the track data in consideration of the main operation performed on a specific day, and the DPI index according to the present invention and the SVM model was evaluated and the results are shown in Table 3.

[Table 3]

It can be seen that the overall discrimination performance was improved by 0.062 compared to the TPI, and the operation discrimination of the angang net, streak, and drift net was significantly improved to more than 0.5.

Fig. 9 shows the results of fishing boat fishing by the deep learning model of the present invention. You can see the track pattern for each fishing vessel operation of the learning model, red corresponds to accurate discrimination, blue means a wake of 6 knots or more, and other colors correspond to discrimination errors. In general, it is judged that the result of the determination is similar to the wake pattern of each fishing vessel shown in FIG. 2 above.

The present invention is expected to be able to discriminate a total of 6 types of fishing only with track data for 24 hours, and to prevent overfishing or illegal fishing due to excellent identification accuracy, and to greatly contribute to the protection of water resources.

In the above, the present invention has been exemplarily described through preferred embodiments, but the present invention is not limited to such specific embodiments and may take various forms within the scope of the technical idea presented in the present invention, specifically, the claims. may be modified, changed, or improved.

Claims (5)

As a method for determining fishing boat fishing using a computer system capable of network communication and including a deep learning model,
The system receives the tracking data for learning including the position, course, and speed of a plurality of fishing vessels from the automatic ship identification system (AIS),
The system receives marine environment data including tide difference, day/night, water temperature, water depth, and seabed type,
The system performs a data pre-processing process including interpolating the track data for learning with respect to a reference value and aligning it at regular time intervals,
The system combines the learning track data and marine environment data and sets several data to create a track window dataset arranged in time series as one data form,
The system labels the fishing license information of the deep learning target fishing vessel in the wake window dataset,
The system performs deep learning with a neural network model on the wake window dataset,
The system includes the step of receiving the tracking data of the fishing vessel to be determined, and determining the type of fishing of the fishing vessel from the tracking data of the fishing vessel based on the result learned by the neural network model,
The wake window dataset comprises a plurality of datasets by overlapping datasets at predetermined time intervals.
A method of identifying fishing boats using track data.
According to claim 1,
A convolutional neural network (CNN) model is applied to the data on the course and speed of a fishing vessel in the wake window dataset, and a full-connected neural network (FCNN) model is applied to the marine environment data to perform deep learning. A method for identifying fishing boats using
delete According to claim 1,
The tracking data of the fishing vessel to be determined is a fishing vessel fishing determination method using the track data, characterized in that receiving the track data for 10 to 24 hours.
According to claim 1,
The step of determining the type of fishing vessel fishing is a fishing vessel fishing discrimination method using track data, characterized in that it is determined as any one of six operations including a fishing net, an angang net, a streak, a drift net, a trap, and a trawl.

Images