A comprehensive framework incorporating deep learning for analyzing fishing vessel activity using automatic identification system data

Abstract

Automatic Identification System (AIS) has emerged as a crucial and cost-effective tool for monitoring ship behavior, widely employed in various fisheries. However, extracting meaningful insights from extensive AIS data to support fishery research remains challenging. In this study, we developed a framework integrating deep learning for marine fishing activity analysis, leveraging AIS data alongside marine environmental factors. Our approach utilized a transformer-based model with a majority vote for classifying fishing vessel types. The model achieved high accuracy, surpassing 90% in vessel type classification using a small subset of AIS records. Our framework employed the Temporal K-Means algorithm to efficiently identify fishing behavior, leveraging the time-series information of AIS data. Subsequently, it mapped fishing hours onto spatial grids to analyze the relationship between fishing activity and environmental factors. Correlation analysis revealed distinct preferences of different vessel types for environmental conditions, influencing their spatial distributions. Trawlers, for instance, exhibited sensitivity to seafloor bottom temperature, whereas seiners were primarily influenced by sea surface density (SSD) and sea surface temperature and gillnetters by SSD. Through this framework, we have established a coherent process to derive valuable insights about fishery resources from AIS data and guide fisheries management.

fishing activity analysis, AIS data, deep learning, fisheries management

Introduction

In 2020, 51% of world fisheries (90 million tonnes) were from capture fisheries, with 78.8 million tonnes from marine waters (FAO 2022). Fishery resources play a crucial role in providing proteins, vitamins, and micronutrients and their sustainable utilization is essential for ensuring future global security, drawing significant attention from public policy (Garcia and Rosenberg 2010). Due to factors like overfishing and ecological environment destruction, marine fishery resources are experiencing a severe decline. The sustainability of fishery resources has remained an essential and controversial issue for over a century (Stokstad 2009). Research on fishery resources predominantly relies on actual catch data (Zhang et al. 2021, Chiarini et al. 2022, Yuan et al. 2022), which is often challenging to obtain and typically exhibits low temporal and spatial resolution. The fishing operations of fishing vessels are intricately linked to fisheries resources, underscoring the importance of studying vessel dynamics for effective fisheries management. To protect the remaining fishery resources and achieve sustainable fishery development, comprehending the dynamics of fisheries through monitoring and managing fishing vessels is imperative.

The three principal methodologies for monitoring fishing vessel dynamics include electronic monitoring (EM), remote sensing, and Automatic Identification System (AIS). The widespread adoption of EM has been hindered by many fishers viewing it as an intrusion into their private workspace (Plet-Hansen et al. 2017), coupled with its generally low acceptance within the fishing industry. Similarly, the primary limitation of satellite remote sensing is its inability to continuously monitor fishing vessels’ navigational and fishing activities over the long term, as satellite images capture only transient snapshots with long revisit periods (Guan et al. 2021). In comparison, AIS has a broader application than EM data and provides finer temporal and spatial resolution than remote sensing imagery. Originally designed to enhance maritime safety by transmitting a ship's identity, position, speed, and course every few seconds to nearby vessels, AIS transmissions are also captured by satellite or ground-based receivers (Kroodsma et al. 2018). The ongoing use of AIS receivers has also generated substantial AIS data, enabling comprehensive monitoring of numerous ships (Zhao et al. 2014).

With the advancement of artificial intelligence, many machine learning methods, particularly deep learning techniques, are being applied to AIS data to enhance fisheries management, including fishing vessel types classification and fishing behavior detection. Various classical machine learning algorithms were employed for fishing vessel types classification, which is the foundational step in fisheries research using AIS data. Guan et al. (2021) utilized a Light Gradient Boosting Machine (LightGBM) approach incorporating 60 features to distinguish between trawlers, gillnetters, and seiners. Huang et al. (2019) employed feature engineering and machine learning techniques, explicitly leveraging XGBoost as the core framework, to classify fishing vessels into nine distinct types. These algorithms relied heavily on manually crafted features. As for deep learning, most studies were based on Convolutional Neural Networks (CNNs) (Acharya et al. 2024). For example, Kroodsma et al. (2018) used a CNN to classify six types of fishing vessels, including trawler, longline, and seiner. Similarly, Ljunggren (2018) applied a CNN to categorize ship into fishing, passenger, and cargo, with features automatically learned by the CNN. Arasteh et al. (2020 ) proposed a CNN for trajectory classification, leveraging a set of invariant spatio-temporal features extracted from AIS data. For other deep learning architectures, Duan et al. (2022) proposed a semi-supervised deep learning approach based on a variational autoencoder, integrating both kinematic and static information from AIS data. However, CNNs are primarily designed for image data and are not well-suited for handling AIS trajectory data of varying lengths. Additionally, they struggle with capturing long-range dependencies. The second aspect extracted information about fishing behavior from AIS data, identifying whether each point in the data indicates that a fishing vessel is engaged in fishing activities. For classical machine learning methods, Gaussian Mixture Model (GMM) (Natale et al. 2015), Hidden Markov Model(HMM) (Peel and Good 2011) and Conditional Random Field (CRF) (Hu et al. 2016) were utilized to identify fishing behavior. However, each of these methods had its drawbacks. For instance, GMM does not consider the contextual information of AIS data, while HMM has a high computational complexity and is highly sensitive to the initial probability distribution. Additionally, CRF is a supervised algorithm requiring a substantial amount of labeled data on fishing vessel status for training. Deep learning methods were also employed for fishing behavior detection, Jiang et al. (2016 ) transformed vessel trajectory patterns into images and employed an autoencoder to classify vessel activities, which was time-consuming.

Once vessel type classification and detection of fishing behavior are accomplished, estimating fishing grounds and fishing effort becomes feasible using AIS data. Fishing effort is related to several factors, including fishing time, the degree of fishing mechanization, the fishing capacity of vessels, horsepower, labor, and the type of nets utilized. Under the same fishing methods and other conditions, fishing time is highly correlated with fishing effort. Natale et al. (2015)’s study indicated that fishing efforts calculated from the AIS and logbooks showed similar distributions. Additionally, some previous studies (Witt and Godley 2007, Lee et al. 2010) used fishing time to estimate fishing effort. Based on this, we used fishing time derived from AIS data to approximate fishing effort. Many studies (Zhou et al. 2020, Vaihola and Kininmonth 2023) also investigated the relationship between fishing efforts and environmental factors. However, limited research examines the relationship between the approximated fishing effort derived from AIS data and environmental factors. Different fishing methods target species with distinct environmental preferences. The distribution of fishing effort, approximated from AIS data, can roughly represent the distribution of target species, enabling us to infer their relationship with the environment. Understanding this relationship helps fill gaps in empirical data and supports predictions of species distributions.

Most importantly, many studies have treated the aspects of vessel type classification, fishing behavior detection, fishing effort estimation, and the correlation between fishing effort and environmental factors as independent, without considering their causal relationships. In light of the above problems, this study aims to establish a coherent and systematic framework for leveraging AIS data and environmental factors in fisheries research. This framework has the following features: (1) Efficiently utilize a transformer-based model to automatically extract features from AIS data and classify different types of fishing vessels, without manual computation of numerous complex features and can efficiently handle the varying lengths of AIS data. (2) Utilize temporal K-Means to detect fishing behaviors and calculate fishing time in an unsupervised manner, fully considering the time-series information of AIS data, while benefiting from low time complexity, suitability for large datasets, and minimal configuration requirements. (3) Integrate environmental data to explore the relationship between fishing effort estimated from AIS data for various vessel types and environmental factors, aiming to understand the potential of AIS-derived fishing effort for fishery resource distribution and management. Moreover, because our framework utilizes time and space independent features, it can be easily applied to other time periods and regions.

Methods

Data

Our AIS data was sourced from Heywhale (https://www.heywhale.com/). The dataset comprised 22 342 trajectories, including three types of fishing vessels: trawlers, gillnetters, and seiners. Specifically, there were 7501 trawler trajectories, 7461 seiner trajectories, and 7380 gillnetter trajectories. Each trajectory message included location (longitude and latitude), time, SOG (Speed Over Ground), and COG (Course Over Ground).

The marine environment plays a crucial role in the distribution of fishery resources. Fish distribution is strongly influenced by species-specific environmental tolerances and optimal conditions (Pörtner and Farrell 2008, Sunday et al. 2010). Changes in specific environmental factors, such as sea surface temperature (SST), can affect the abundance, local distribution, food webs, predator-prey dynamics, and range of fish species in fisheries (Leitão et al. 2018). Additionally, physical and biological processes, shaped by environmental factors, regulate the abundance, distribution, and productivity of marine organisms across various temporal and spatial scales (Solanki et al. 2017). Advances in satellite remote sensing technologies have enabled the acquisition of extensive marine environment data, facilitating the analysis of fishery resources, so we utilized satellite-derived data as our primary source of environmental information, including Chlorophyll-a (Chl-a), SST, sea surface height (SSH), sea surface salinity (SSS), sea surface density (SSD), seafloor bottom temperature (SBT) and dissolved oxygen (DO). Table 1 provides detailed information about these variables. Chl-a and SST were obtained from ocean color measurements with a spatial resolution of 4 km and a monthly temporal resolution. At the same time, depth data were sourced from General Bathymetric Chart of the Oceans (GEBCO). The other variables were sourced from the Copernicus Marine Environment Monitoring Service (CMEMS) with a spatial resolution of 0.25°. After acquiring all the marine environment data, we resampled all datasets to a uniform spatial resolution. To showcase the performance and capabilities of our framework, we selected the Chinese coastal region, specifically focusing on the Yellow Sea. This area is defined by coordinates ranging approximately from 31°N to 39°N in latitude and 119°E to 126°E in longitude. Data from May to July were excluded due to China’s fishing moratorium during these months.

Table 1.

List of marine environment data and the data sources.

Parameters	Data source	Unit	Spatial resolution	Temporal resolution
Chl-a	https://oceancolor.gsfc.nasa.gov/	mg/m³	4 Km	monthly
SST	https://oceancolor.gsfc.nasa.gov/	°C	4 Km	monthly
SSS	https://marine.copernicus.eu/	psu	0.25°	monthly
DO	https://marine.copernicus.eu/	mmol/ m³	0.25°	monthly
SSD	https://marine.copernicus.eu/	kg/m³	0.25°	monthly
SSH	https://marine.copernicus.eu/	m	0.25°	monthly
SBT	https://marine.copernicus.eu/	°C	0.083°	monthly
Depth	https://www.gebco.net/	m	0.25°	_

AIS data and environment factors were processed using the framework illustrated in Fig. 1. We began by preprocessing AIS data, which include feature calculation and segmentation. Afterward, we classified the segments using a transformer-based model to identify different vessel types. Once the vessel types were determined, we applied Temporal K-Means to detect fishing activities for each type and calculate the corresponding fishing times. Finally, we analyzed the relationship between fishing time and environmental data. To facilitate the understanding of the entire process, our source code with machine learning libraries such as scikit-learn and Pytorch is available on GitHub (https://github.com/HFC666/AIS_TS).

Figure 1.

Framework leveraging AIS data for fisheries research.

AIS data preprocessing

Since AIS data was irregularly sampled in time, where time at index t was denoted as |${T_t}$|​, corresponding to the location of a data point in the trajectory, we conducted simple calculations to derive time-invariant and space-invariant features to extend the model across different spatial and temporal contexts. SOG-related features, COG-related features, and location-related features have been used in many studies to differentiate between types of fishing vessels. These studies indicated that these features performed well in distinguishing vessel characteristics (Kroodsma et al. 2018, Guan et al. 2021). Instead of directly utilizing location information, one of our derived features was the Haversine distance (Sofwan et al. 2019) from the previous message:

$$\begin{eqnarray} {d_t} &=& R\arccos \left[ {\cos \left( {{\rm{la}}{{\rm{t}}_{t - 1}}} \right)\cos \left( {{\rm{la}}{{\rm{t}}_t}} \right)\cos \left( {{\rm{lo}}{{\rm{n}}_t} - {\rm{lo}}{{\rm{n}}_{t - 1}}} \right) }\right. \\ &&\left. {+ \sin \left( {{\rm{la}}{{\rm{t}}_{t - 1}}} \right)\sin \left( {{\rm{la}}{{\rm{t}}_t}} \right)} \right]. \end{eqnarray}$$

Here, |${\rm{la}}{{\rm{t}}_{\rm{t}}}$| and |${\rm{lo}}{{\rm{n}}_{\rm{t}}}$| represented latitude and longitude at index t, respectively, and R was the Earth’s radius. Additionally, we introduced |$d{c_t}$| by normalizing |${d_t}$| with the time gap:

$$\begin{eqnarray} d{c_t} = {d_t}/\left( {{T_t} - {T_{t - 1}}} \right). \end{eqnarray}$$

To capture SOG (⁠|${v_t}$|⁠) dynamics, we computed SOG differentiation |${s_t}$|​ and acceleration |${a_t}$|⁠:

$$\begin{eqnarray} {s_t} = {v_t} - {v_{t - 1}} \end{eqnarray}$$

$$\begin{eqnarray} {a_t} = {s_t}/\left( {{T_t} - {T_{t - 1}}} \right). \end{eqnarray}$$

Similar calculations were used to capture the dynamics of COG (⁠|${c_t}$|⁠):

$$\begin{eqnarray} {r_t} = {c_t} - {c_{t - 1}} \end{eqnarray}$$

$$\begin{eqnarray} r{a_t} = {r_t}/\left( {{T_t} - {T_{t - 1}}} \right). \end{eqnarray}$$

Additionally, we calculated the change in distance to the offshore |$d{o_t}$| ​and normalized it as |$do{c_t}$|⁠:

$$\begin{eqnarray} d{o_t} = dt{o_t} - dt{o_{t - 1}} \end{eqnarray}$$

$$\begin{eqnarray} do{c_t} = d{o_t}/\left( {{T_t} - {T_{t - 1}}} \right), \end{eqnarray}$$

where |$dt{o_t}$| represented the distance to offshore at index t.

We utilized these features, along with SOG, to form a 10-dimensional feature vector:

$$\begin{eqnarray} {f_t} = \left[ {{v_t},{d_t}, \,\,d{c_t},{s_t},{a_t}, \,\,{r_t}, \,\,r{a_t}, \,\,dt{o_t},d{o_t},do{c_t}} \right]. \end{eqnarray}$$

Normalization, as suggested by Santurkar et al. (2018), was applied to stabilize and accelerate convergence of our neural network during optimization:

$$\begin{eqnarray} \tilde x = \frac{{x - {\rm{mean}}\left( x \right)}}{{{\rm{std}}\left( x \right)}}, \end{eqnarray}$$

where |$\tilde x$| denoted the normalized feature, |${\rm{mean}}( x )$| was the mean of the feature x, and |${\rm{std}}( x )$| was its standard deviation.

After obtaining spatiotemporally invariant features, we proceed to segment the feature sequence (Fig. 2). Our segmentation method was based on both the time gap between two consecutive AIS messages and the length of each sub-trajectory. Figure 2 illustrates our approach, where we predefined a set of time gap thresholds: 2 h, 1 h, 30 min, and 20 min. Initially, we used the largest time threshold as our segmentation criterion. After segmentation, sub-trajectories longer than 120 units underwent further segmentation using a smaller time gap threshold, while sub-trajectories shorter than 100 units were discarded. Sub-trajectories falling between these lengths were retained. If sub-trajectories segmented using the smallest time gap threshold still exceeded 120 units in length, we further segmented them to ensure their length between 100 and 120.

Figure 2.

Flow chart of segmentation.

After segmentation, we analyzed the distribution of sub-trajectory lengths and the numbers of sub-trajectories within each trajectory (Fig. 3). The length of sub-trajectories ranged uniformly from 100 to 120 units, with a predominant length of 110 units. This is why, when segmenting overly long segments based on length, we divided the segment lengths at 110, the median length of segments ranging from 100 to 120. Regarding the count of sub-trajectories per trajectory, most values were below 50, predominantly falling within the range of 2 to 20.

Figure 3.

Summary of the distribution of sub-trajectories.

Vessel types classification

Trawlers, seiners, and gillnetters are three main types of fishing vessels in the world. Due to the depth where trawlers drag or pull can be classified as bottom trawling and midwater trawling (Zeeberg et al. 2006). With a wide fishing range and high efficiency, trawling is the active pursuit of fish (Guan et al. 2021). Seiners are horizontal nets that hang vertically from floats around schooling fish (Gaertner et al. 2011). Gillnetter lay long ribbon-shaped nets vertically in the sea (Zain 2013). Most countries worldwide have adopted these three fishing methods for a long history, so we focused on identifying these three fishing vessel types.

Many studies have applied different neural network architectures to AIS data, with CNNs (Kroodsma et al. 2018, Ljunggren 2018, Arasteh et al. 2020) and recurrent neural networks (RNNs) (Nguyen et al. 2018, 2022, Yuan et al. 2022) being the most commonly used architectures in these studies. The model we employed was based on the Transformer architecture, a significant and recently developed class of deep learning models. Transformers were initially proposed for natural language translation (Vaswani et al. 2017) but have since achieved state-of-the-art performance across virtually all NLP tasks (Raffel et al. 2020). Transformer-based models offer several advantages over CNNs and RNNs in AIS data modeling. Unlike RNNs, which process data sequentially and struggle with long-range dependencies (Pascanu et al. 2012), Transformers use a self-attention mechanism that can capture dependencies across the entire sequence in parallel. This allows for better modeling of long-term relationships without the vanishing gradient problem. Compared to CNNs, which are limited by local receptive fields and require deeper architectures to capture long-range patterns, Transformers can dynamically attend to different parts of the sequence, making them more flexible and efficient in handling complex AIS data. Rather than utilizing the entire Transformer architecture, we exclusively used the encoder module, as the decoder module is unsuitable for classification tasks (Zerveas et al. 2021). The architecture of our model is illustrated in Fig. 4. Unlike the typical Transformer architecture, we utilized learnable parameters for position encoding instead of fixed parameters. Additionally, we replaced layer normalization with batch normalization. These changes can mitigate the impact of outlier values in the time series (Zerveas et al. 2021). Despite variations in sub-trajectory lengths, as depicted in Fig. 3, our model effectively handled this issue. We set the maximum sequence length to 120, padding shorter sequences with zeros. Before computing the self-attention distribution using the softmax function, we generated a padding mask, which assigned a significant negative value to the attention scores for the padding positions (Zerveas et al. 2021). To apply our model for classification, we first obtained the representation vectors |${z_t} \in {R^d}$| for each trajectory at all time steps after passing through the encoder. Next, we concatenated these vectors along the time dimension to form a single vector |$\bar z \in {R^{d \cdot 120}}$|⁠, where 120 represents the number of time steps in the trajectory. This concatenated vector, |$\bar z = {\rm{ \,\,}}[ {{z_1}; \cdots ;{z_{120}}} ]$|⁠, then serves as the input to a multilayer perceptron (MLP), which maps it into a three-dimensional space to differentiate among the three types of fishing vessels. Finally, after applying the softmax function, the model outputs three-dimensional vectors that indicate the probability of each vessel type. In classification tasks, a commonly used loss function is typically the cross-entropy loss, which is represented by the following equation:

$$\begin{eqnarray} {\rm{Loss}} = - \frac{1}{N}\mathop \sum \limits_{i = 1}^N \mathop \sum \limits_{j = 1}^3 {y_{ij}}\log \left( {\widehat {{y_{ij}}}} \right). \end{eqnarray}$$

Figure 4.

Transformer based architecture for vessels types classification. The representation in the gray box indicates the encoders.

With |${y_{ij}}$| representing the true label for sample i concerning class j (1 if the sample belongs to that class, 0 otherwise), and |$\widehat {{y_{ij}}}$| denoting the predicted probability that sample i belongs to class j.

Following segmentation, we obtained a total of 175 840 segments: 45 590 for trawlers, 78 286 for seiners, and 51 964 for gillnetters. For model training, 70% of the segments from each type were utilized, comprising 35 000 segments for trawlers, 55 000 for seiners, and 40 000 for gillnetters. The remaining 20% were designated as the test set, and 10% were used as the validation set. We implemented an early stopping strategy based on validation loss to prevent overfitting, halting training if no improvement was observed for 10 consecutive epochs. The validation set was also used for hyperparameter selection, ensuring optimal model performance before final evaluation.

To optimize our deep neural network, we employed Adaptive Moment Estimation (ADAM), an adaptive optimizer that combines the characteristics of RMSprop and momentum optimizer. ADAM is effective in adapting to system changes, minimizing the computational load for parameter optimization, and eliminating the need for complex hyperparameter tuning (Wakitani et al. 2017).

Since sub-trajectories within the same trajectory had the same vessel type, we employed a majority voting mechanism for our prediction process. This approach determined the classification result of an entire trajectory based on the aggregated results of its sub-trajectories. Nevertheless, it was still possible to identify the vessel type from a single sub-trajectory. Additionally, we investigated the correlation between classification accuracy and the number of sub-trajectories per trajectory.

Fishing behavior detection

After classifying each fishing vessel, we detected the fishing activities for each type of vessel. Given the significant role of vessel speed in detecting fishing activities (Peel and Good 2011, Natale et al. 2015), we used it as the basis for our detection. The K-Means clustering algorithm, one of the most influential data mining algorithms in the research community (Ahmed et al. 2020), also has many advantages in AIS data mining. K-Means is relatively straightforward in implementation, making it easy to apply to AIS data analysis. Additionally, K-Means is computationally efficient, with a linear time complexity, and converges quickly (Arthur and Vassilvitskii 2007 ), making it practical for applications involving millions of AIS data points. To incorporate the contextual information of speed, we combined K-means with delay embedding, which we referred to as the Temporal K-means algorithm. Temporal K-means, compared to HMM and other time series modeling methods, offers fast computation times, the ability to handle large volumes of AIS data, and does not require making strong assumptions about the underlying data distribution. Given a time series of speed of length n:

$$\begin{eqnarray} {s_1},{s_2},{s_3}, \cdots ,{s_n}, \end{eqnarray}$$

by choosing a window size d, we can perform a delay embedding to form the design matrix X (Kantz and Schreiber 2004),

$$\begin{eqnarray} X \,\, = \,\,\left[ {\begin{array}{@{}*{3}{l}@{}} {\begin{array}{@{}*{3}{l}@{}} {{s_1}}& \quad {{s_2}}& \quad \cdots \\ {{s_2}}& \quad {{s_3}}& \quad \cdots \\ \vdots & \quad \vdots & \quad \ddots \end{array}}&{\begin{array}{@{}*{1}{l}@{}} {{s_d}}\\ {{s_{d + 1}}}\\ \vdots \end{array}}\\ {\begin{array}{*{20}{l}} {{s_{n - d + 1}}}& \quad {{s_{n - d}}}& \quad \cdots \end{array}}& \quad {{s_n}} \end{array}} \right] \,\, = \,\,\left[ {\begin{array}{@{}*{1}{c}@{}} {{x_1}}\\ {\begin{array}{@{}*{1}{c}@{}} {{x_2}}\\ \vdots \\ {{x_{n - d + 1}}} \end{array}} \end{array}} \right]. \end{eqnarray}$$

Each row of X can be interpreted as a vector in |${R^d}$| (Eirola and Lendasse 2013 ). We chose |$d \,\, = \,\,3$| and applied the K-means algorithm to cluster these vectors in |${R^3}$|⁠. This dimension allows us to capture the short-term dependencies of each point concerning the two preceding and following points while detecting fishing activities. Moreover, conducting calculations in a lower-dimensional space reduces computational complexity and minimizes the risk of overfitting. After clustering, the resulting sequence had a length of |$n - 2$|⁠, which was shorter than the original AIS series. To identify the status of each point in the series, we used the mode of |$[ {{x_{i - 2}},{x_{i - 1}},{x_i}} ]$|as the representation of ​|${s_i}$|⁠. This method functions similarly to a weighted moving average, ensuring that the temporal dynamics are accurately captured while providing a clearer understanding of whether a vessel is engaged in fishing activities.

Fishing effort estimation

After using Temporal K-Means to detect fishing activities, we obtained the locations and times where fishing occurred and estimated the fishing effort by calculating the fishing time. The time spent in fishing operations was estimated for each grid cell with a spatial resolution of |${0.25^ \circ } \times {0.25^ \circ }$|⁠. For each fishing AIS record k, the variable Duration was calculated in seconds as follows (Galparsoro et al. 2024):

$$\begin{eqnarray} \it{Duratio}{n_k} = {t_k} - {t_{k - 1}}, \end{eqnarray}$$

where |${t_k}$| was the time registered in the AIS record, and |${t_{k - 1}}$| was the time registered in the previous AIS record. Since we were calculating fishing time, it is essential that the vessel was engaged in fishing at both |${t_k}$| and |${t_{k - 1}}$|⁠. The fishing operation time for each grid cell was estimated by summing the durations of all AIS fishing records within that cell.

Correlation analysis

Generalized Additive Models (GAMs) are a flexible approach used to analyze the relationship between response and predictor variables (Wang et al. 2020), particularly in fishery studies. Many studies utilized GAMs to investigate the relationship between satellite-derived environmental data and fishery resources (Murase et al. 2009, Hua et al. 2019). Unlike traditional linear models, GAM allows for the exploration of nonlinear relationships through smooth functions, making them particularly suitable for modeling complex environmental data (Solanki et al. 2017).

In our study, we aimed to assess the influence of various environmental variables on the abundance and distribution of fishery resources. The logarithm of fishing time was considered the response variable, while environmental factors as well as the month in which fishing occurred were utilized as the predictor variables. The formulation of GAM is:

$$\begin{eqnarray} Y = {\rm{\alpha }} + \mathop \sum \limits_{j = 1}^n {f_j}\left( {{x_j}} \right) + \varepsilon , \end{eqnarray}$$

where Y represented the logarithm of fishing time, |${x_j}$| denoted the environmental variables, |${f_j}$| were the smooth functions of the predictors, |${\rm{\alpha }}$| was the intercept, and ε was the error term. The smooth functions |${f_j}$| enable the model to capture nonlinear trends and interactions among predictors, providing insights into how these factors influence fishing time.

GAM's interpretability and ability to model complex relationships make it a valuable tool in fisheries research. It allows researchers to understand the effects of environmental variables on fish resource distribution.

Results

Performance of transformer-based model

The model selected using the best hyperparameters from the validation set converged after 400 training epochs, indicating that there was no significant improvement in accuracy over the last 10 epochs. It ultimately achieved an accuracy of 0.84 for the sub-trajectories. By applying the majority vote criterion to determine the type of each trajectory containing various numbers of sub-trajectories, the accuracy improved to 0.90. The statistical performance of our model for each fishing type is presented in Table 2. The gillnetter classification exhibited the highest accuracy at 0.93, indicating that trawlers and seiners were rarely misclassified as gillnetters. Conversely, seiners were more frequently misclassified as trawlers, and trawlers were often misclassified as seiners.

Table 2.

Statistical results of the real types of each fishing vessel in the classification results.

	trawler	seiner	gillnetter	Accuracy
trawler	5472	669	28	0.89
seiner	418	5410	267	0.89
gillnetter	116	318	5511	0.93

Given that different trajectory consisted of varying sub-trajectories, we evaluated the accuracy based on the number of sub-trajectories. In Fig. 5(a), we plotted the accuracy of trajectories with the number of sub-trajectories exceeding the values on the x-axis. The overall accuracy trend increased with the number of sub-trajectories, indicating that the majority vote mechanism enhanced accuracy.

Figure 5.

Results of vessel classification model.

Additionally, we randomly sampled a fixed number of sub-trajectories to analyze the accuracy trend. Figure 5(b) showed that accuracy improved as randomly sampled sub-trajectories increased. When the number of sampled sub-trajectories reached 10, corresponding to a length of 1000, the accuracy achieved 0.90. We repeated this sampling process 100 times to calculate the standard deviation (std), as depicted in Fig. 5(c). The overall downward trend in std suggested that our model’s results improved stability as the number of sampled sub-trajectories increased.

Temporal K-means for fishing behavior identification

We employed temporal K-means clustering on the speed data of three different types of vessels due to their distinct operational behaviors. We set two clusters for all models: one representing fishing behavior and the other non-fishing behavior. K-means was chosen for its flexibility in modeling data from nearly any distribution. Table 3 presents the results of Temporal K-means clustering for trawlers, seiners, and gillnetters based on the three-dimensional delay embedding of their speeds. The “fishing centroid” represents the cluster center during fishing activities, while the “non-fishing centroid” refers to the center during non-fishing activities. For the fishing centroid, the differences between the dimensions were minimal, indicating relatively stable speeds during fishing. In contrast, the non-fishing centroid showed greater variation across dimensions, suggesting more significant speed fluctuations when vessels were not fishing. Additionally, while differences existed between vessel types, the speed behaviors of trawlers and seiners were notably more similar than those of gillnetters. The fishing centroid for gillnetters was smaller than that for trawlers and seiners. However, the non-fishing centroids for all three vessel types were nearly identical. These differences likely stemmed from the operational characteristics of the different types of fishing gear.

Table 3.

Center centroid of temporal K-means models.

Vessel type	Fishing centroid	Non-fishing centroid
trawler	[1.83, 1.77, 1.83]	[7.24, 7.70, 7.19]
seiner	[2.03, 1.99, 2.05]	[7.24, 7.57, 7.19]
gillnetter	[0.96, 0.91, 0.96]	[7.42, 7.65, 7.41]

Figure 6 illustrates the results of applying temporal K-means clustering to the speed data of trawlers, seiners, and gillnetters. The clustering results were smooth, indicating minimal change points within short time intervals. This smoothness demonstrated that temporal K-means effectively preserved the stationary properties of vessel dynamics.

Figure 6.

Temporal K-means results of examples of trawler, seiner, and gillnetter.

Spatial characteristic of fishing time

After determining the fishing behavior of the vessels, we mapped the locations where the vessels engaged in fishing onto grids. We then calculated the fishing time within each grid for the three vessel types. As shown in Fig. 7, different vessel types occupied distinct fishing areas. Both trawlers and seiners were distributed in the southern waters of the Shandong Peninsula, whereas gillnetters were not. The high fishing zones of trawling were mainly distributed near the coast, while those of seining were further offshore compared to trawling, but the high fishing zones of gillnetting were distributed more evenly. Although gillnetters had the most extensive distribution scope, their fishing time was smaller than that of trawlers and seiners (Fig. 8). The histograms of fishing time for trawlers and seiners were similar, centered around 14. Seiners had the highest magnitude of fishing time and a larger distribution area than trawlers.

Figure 7.

Fishing time (unit: second) distribution of three types of vessels.

Figure 8.

Histograms of fishing time of three types of fishing vessels.

GAMs were employed in this study to investigate the correlation between the distribution of fishing time of different vessels and environmental variables. The Akaike Information Criterion (AIC) was utilized for model selection, where a lower value indicates a better fit of the model (Venables and Dichmont 2004). All variables except Month, Chl-a, and SST were statistically significant in the trawler model. SBT emerged as the most influential factor among these variables, explaining 37.47% of the variance. DO and Depth also had substantial impacts, contributing more than 20% to the explained variance (Table 4). As shown in Fig. 9, trawling activities tended to yield higher fishing time in areas where SBT ranged between 20°C and 30°C. The effects of SSH and SSS on trawling fishing time exhibited roughly linear trends. Regarding Depth, both excessively deep and shallow waters were unfavorable for trawling, the highest fishing times were found in regions with depths ranging from 30 m to 60 m. Additionally, the month of the year did not significantly impact trawling fishing time.

Figure 9.

Correlation analysis between fishing time of trawler and environment factors.

Table 4.

Variance explained rate of environment factors (%).

Vessel type	Chl-a	SST	SSH	SSS	SSD	DO	Depth	SBT	Month
trawler	0.07	0.00	4.17	3.61	10.63	23.19	20.87	37.47	0.00
seiner	6.89	19.80	3.77	5.68	28.75	6.83	9.41	18.60	0.28
gillnetter	0.15	1.14	1.30	1.04	52.18	-	1.27	2.03	40.89

All environmental variables demonstrated significant relationships with fishing time for the seiner model. SSD, SST, and SBT were influential, explaining more than 10% of the variance (Table 4). SSD was identified as the most critical factor, accounting for 28.75% of the explained variance. The highest fishing time for seiner occurred in areas where SST ranged between 25°C and 30°C, and SBT ranged between 20°C and 25°C. Unlike trawling, the effects of SSH and SSS on seiner were not linear. The regions with the highest fishing time for seiner were found at SSH levels of 0.6–0.7 meters and SSS levels of 32–33 psu. Like trawling, the highest fishing time for seiner was observed at a depth of around 50 meters. Additionally, the peak fishing time for seiners occurred from August to October (Fig. 10).

Figure 10.

Correlation analysis between fishing time of seiner and environment factors.

In the case of gillnetters, we removed DO due to AIC, leaving all remaining environmental variables significantly influenced fishing time. SSD and Month pronounced influenced gillnetter fishing time (Table 4). For SST, the distribution range of high fishing time for gillnetters spanned from 15°C to 30°C, which was broader than the ranges for trawlers and seiners. The effects of SSH and SSS on gillnetter fishing times were similar to those for seiner, with high fishing time occurring at SSH levels of 0.6–0.7 m and SSS levels of 32–33 psu. Additionally, the impact of the Month on gillnetter fishing time was similar to that of seine, with high fishing times observed from August to October (Fig. 11).

Figure 11.

Correlation analysis between fishing time of gillnetter and environment factors.

Discussion

Classification performance of transformer-based model

Our framework incorporated a Transformer-based model with a majority vote for classifying vessel types, leveraging the advantages of deep learning, including high precision, efficiency, and intelligence (Dong et al. 2022). While AIS data have been extensively used for vessel type identification in numerous studies, these studies often utilized the entire AIS trajectory and required manual feature extraction, such as calculating summary features like mean speed and maximum speed (Huang et al. 2019, Guan et al. 2021). These manually extracted features can be susceptible to the effects of outliers. Recent advancements in deep learning architectures, such as the Transformer, have shown significant promise in AIS data mining. The primary reason for the growing popularity of these architectures is their ability to automatically extract features, eliminating the need for manual feature extraction and selection (Shaheen et al. 2016).

In addition to implementing a deep learning model based on the Transformer architecture, we introduced a majority vote mechanism. We utilized segmentations rather than using the entire AIS data sequence as input for our model. There are two primary reasons for this approach. First, the length of AIS data sequences varies significantly, and a single AIS data segment is sufficient to reflect the operational characteristics of a fishing vessel. Second, processing long time series increases the computational burden. Although Transformer is a robust neural network architecture widely used in various applications, such as machine translation, image segmentation, and classification (Wang et al. 2023), its memory footprint increases quadratically, and computations increase cubically with longer sequence lengths. Therefore, using AIS segmentations instead of entire AIS trajectories helps to conserve computational memory (Beltagy et al. 2020). The majority vote mechanism also offered two key advantages. First, we can classify only the new segments if additional AIS data for a particular vessel becomes available. This allows us to leverage the majority vote to make our results more reliable without recalculating the features of the entire AIS dataset. Second, in terms of accuracy, while using AIS segments for classification might reduce individual classification accuracy, the majority vote can enhance overall classification precision without increasing the computational burden. Most misclassified trajectories had a relatively small number of sub-trajectories, with the majority having fewer than 20 (Fig. 12 a). Additionally, as shown in (Fig. 12 b), more than 80% of the misclassified trajectories contained fewer than 20 sub-trajectories. This indicated that the majority vote mechanism significantly improved classification accuracy as the number of sub-trajectories increased. However, the classification accuracy variation differed among the fishing vessel types (Fig. 13). For gillnetter and seiner vessels, the classification accuracy remained consistently high and improved as the number of sub-trajectories increased. In contrast, the accuracy for trawler vessels exhibited a different pattern: it initially started at a high level, then quickly dropped to 0.7, and gradually rose again.

Figure 12.

Histogram and cumulative probability density plot of the number of sub-trajectories in misclassified trajectories.

Figure 13.

Classification accuracy of three different types of fishing vessels.

We employed a straightforward and time-efficient segmentation method based on time intervals and segment lengths. However, the chosen segmentation method, segment length, and the quality of the resulting segments can significantly affect the model's accuracy, especially when the AIS data quality is poor. Therefore, it is essential to select appropriate segment lengths and time thresholds tailored to the characteristics of the AIS data. When necessary, selecting representative segments for classification can enhance accuracy.

Fishing behavior characteristic of three types of vessels

The fishing time of vessels is generally related to their total navigation time at sea. Figure 14 showed a linear relationship between the fishing time, calculated using temporal K-Means, and the total navigation time. There was a significant linear correlation between fishing time and total time for all three fishing vessels. However, this correlation was more pronounced for trawlers and seiners than gillnetters. The slope of this relationship can be considered as the proportion of fishing time to total sailing time, serving as a measure of fishing efficiency. Accordingly, trawlers and seiners demonstrated significantly higher fishing efficiency than gillnetters, possibly due to the passive nature of gillnetting (Sales Henriques et al. 2023).

Figure 14.

Linear regression between fishing time and total time.

Trawlers and seiners use active fishing methods, while gillnetters do not. This distinction was also evident in our model results. In the vessel classification model, trawlers and seiners were more likely to be confused by the model, whereas gillnetters were rarely misclassified (Table 2). Similarly, the speed distributions of trawlers and seiners during fishing and non-fishing states were similar in the fishing activity identification model. In contrast, the speed distribution for gillnetters during fishing activities differed significantly from the other two (Table 3).

Current advances in AIS data processing allow near real-time updates on vessel locations and activities. By leveraging these capabilities, along with the efficiency of the Temporal K-Means algorithm, we can implement real-time detection of the spatial characteristics of fishing time. This combination enables fast and accurate large-scale data processing, ensuring timely monitoring. By incorporating gear characteristics, such as selectivity and fishing efficiency, we can assess the fishing pressure and environmental impact in different areas, refining management strategies.

Impacts of environment factors on the distribution of three types of fishing vessels

The distribution of fishing times also showed significant differences among the three fishing methods. For trawling, areas with high fishing times were primarily concentrated near the coast, consistent with the findings of (Zhang et al. 2016)’s study. The high fishing time distribution for seining was similar to that of trawling but covered a larger area, with both methods showing concentrated high fishing time regions. In contrast, gillnetting exhibited a more dispersed and evenly distributed pattern of high fishing times. However, the fishing time in each area was noticeably lower than that of trawling and seining. This discrepancy may be due to the passive gear's lower efficiency than the active gear.

SBT was the most significant environmental factor affecting trawl fishing times, as bottom trawling is the predominant method used in coastal China (Wang et al. 2022). In contrast, SST had a more pronounced impact on seiner fishing because the primary target species in coastal China are mostly pelagic fish. The influence of month in a year rather than temperature on gillnetter fishing was highly significant, likely because the main target species for gillnetters in coastal China, such as Pampus argenteus and Scomberomorus niphonius (Wang et al. 2022), exhibited migratory behavior. Consequently, the fish catch varies with the seasons for gillnetters.

The distribution of fishing vessels with different operation modes reflected the distribution of their target fish species, so the relationship between these vessels and environmental factors can also, to some extent, indicate the relationship between target fish species and environmental factors. Larimichthys polyactis is one of the main target species of trawls in the Yellow and East China Seas (Wang et al. 2022). In April, trawler fishing locations were concentrated in areas where the SBT ranged from 10°C to 18°C, with the highest fishing time occurring between 13°C and 17°C (Fig. 15 A). In August, the fishing locations were areas where the SBT ranged from 18°C to 23°C, with the highest fishing time occurring between 21°C and 23°C (Fig. 15 B). These temperature ranges included the suitable habitat for Larimichthys polyactis identified in (Zhang et al. 2019)’s study, which was 15°C to 18°C in spring and 19°C to 23°C in autumn. Scomber japonicus is one of the vital fishery resources in coastal China and a primary target species for seiner fishing. (Li et al. 2009)’s study indicated that the optimal temperature range for Scomber japonicus was between 28°C and 31°C, which aligned with the seiner fishing distribution observed in Fig. 15 C. The main target fish for gillnetting in Chinese seas are Scomberomorus niphonius and Pampus argenteus. Pampus argenteus has a broad SST tolerance range of 6°C to 20°C, while Scomberomorus niphonius has a narrower range of 9°C to 14°C. Most of these temperature ranges fall within the 10°C to 18°C range shown in Fig. 15 D. The broad temperature tolerance of these target species may contribute to the wide distribution of gillnetters. These findings suggested that the spatial and temporal distribution of fishing activities inferred from AIS data can approximate the spatiotemporal distribution of target species. Therefore, AIS-derived fishing distribution can be a valuable proxy in the absence of direct catch data. A substantial body of research (Yuan et al. 2011, Zhu et al. 2017) has utilized environmental factors to predict fishing ground distribution, thereby guiding fisheries production. In this context, the spatial distribution of fishing time derived from AIS data can also serve as a valuable indicator of fishing ground distribution. AIS data enables continuous monitoring and reveals trends over time, facilitating the identification of hotspots of fishing activity at finer spatial and temporal scales.

Figure 15.

Distribution of temperature and fishing time.

Author contributions

F.H.: conceptualization, methodology, data analysis, writing; Y.L.: conceptualization, methodology, supervision; H.T.: conceptualization, writing—review & editing; J.L.: writing—review & editing; Y.T.: writing—review & editing.

Conflict of interest

Authors declare no conflict of interest.

Funding

This research was funded by the National Key R&D Program of China (2023YFD2401303) and “R&D and Industrialization Project of Key Technologies for Smart Detection System of Digital Marine Fisheries Satellite” by China's Ministry of Education (202482230112900600).

Data availability

AIS data are sourced from Heywhale (https://www.heywhale.com/), and marine environmental data are referenced in Table 1.