Short- and Medium-Term Predictions of Spatiotemporal Distribution of Marine Fishing Efforts Using Deep Learning

Abstract

High-resolution spatiotemporal prediction information on fishing vessel activities is essential for formulating and effectively implementing fisheries policies that ensure the sustainability of marine resources and fishing practices. This study focused on the tuna longline fishery in the Western and Central Pacific Ocean (130° E–150° W, 20° S–20° N) and constructed a CLA U-Net deep learning model to predict fishing effort (FE) distribution based on 2017–2023 FE records and environmental variables. Two modeling schemes were designed: Scheme 1 incorporated both historical FE and environmental data, while Scheme 2 used only environmental variables. The model predicts not only the binary outcome (presence or absence of fishing effort) but also the magnitude of FE. Results show that in short-term predictions, Scheme 1 achieved F1 scores of 0.654 at the 0.5°-1-day scale and 0.763 at the 1°-1-day scale, indicating substantial improvement from including historical FE data. In medium-term predictions, Scheme 1 and Scheme 2 reached maximum F1 scores of 0.77 and 0.72, respectively, at the optimal spatiotemporal scale of 1°-30 days. The analysis also quantified the relative importance of environmental variables, with sea surface temperature (SST) and chlorophyll-a (Chl-a) identified as the most influential. These findings provide methodological insights for spatiotemporal prediction of fishing effort and support the refinement of fisheries management and sustainability strategies.

1. Introduction

The Ecosystem Approach to Fisheries Management (EAFM) was derived from the Ecosystem Approach to Fisheries (EAF) introduced by the Food and Agriculture Organization of the United Nations [1]. It requires fisheries data with increasingly detailed temporal and spatial resolution. However, traditional logbook data can no longer meet these new requirements. The absence of refined global marine capture information limits the oversight of high seas fisheries and hampers the implementation of ecological management policies [2]. Moreover, inadequate conservation measures, weak environmental regulations, and overfishing remain serious challenges. These issues highlight the urgent need for detailed capture information to ensure the sustainable development of marine fisheries resources [3,4].

The automatic identification system (AIS) provides real-time, high-precision, and publicly accessible trajectory big data, enabling the direct observation of more than 70,000 fishing vessels worldwide. Initially, AIS data were used to prevent ship collisions [5], but new applications in vessel tracking, maritime traffic safety, marine pollution monitoring, and marine spatial planning were quickly revealed [6,7,8]. AIS data have been applied in areas such as combating illegal, unreported, and unregulated (IUU) fishing; fisheries resource management; and the evaluation of marine ecological pressures [9,10,11,12,13,14,15]. While fishing vessel trajectory data can effectively quantify historical fishing intensity, they are insufficient for predicting future fishing trends. Predicting the distribution of fishing effort can provide guidance for fisheries resource management [16,17], and marine resource management and planning also require refined predictions of future marine capture information [2].

Fishing effort is more challenging to predict than fishery resource conditions. Fishing operations are continuously dynamic spatiotemporal processes that often involve transfers between different grid areas. Fisheries producers make subsequent fishing decisions at sea based on experience and marine environmental conditions [18,19,20]. The spatial distribution of fishing effort is a continuously dynamic process driven by both external factors (e.g., marine environmental conditions) and internal factors (e.g., previous and surrounding fishing activities) [21]. As a result, fishing effort prediction is essentially a spatiotemporal forecasting problem, requiring models that capture both spatial and temporal dependencies to address complex dynamic changes.

Existing studies on fishing ground prediction have primarily focused on classifying central and non-central fishing grounds, typically dividing fishing areas into high-yield and low-yield zones and applying spatial regression or machine learning methods for point-by-point prediction [22,23,24,25]. However, these dimensionality reduction approaches fail to reflect the temporal trends and spatial correlations of fishing information [26,27,28]. In recent years, deep learning methods have been increasingly applied to reveal the relationships between local spatial characteristics of fishing grounds and environmental factors [26,29,30,31]. Nevertheless, current research generally suffers from two major limitations: first, most predictive models consider only the local effects of environmental variables, partitioning the ocean into small units and modeling based solely on local environmental features of neighboring regions, thereby neglecting the continuous spatial dynamics of fishery data [26,30,31]; second, many studies adopt real-time prediction strategies [25,26,30], using environmental data from the current period to predict fishing ground distributions at the same moment, while overlooking the temporal evolution of fishing grounds, thus constraining the models’ forecasting capability. Furthermore, neural networks often exhibit poor interpretability [32]. Understanding the importance of variables represented by the model is often as crucial as, or even more important than, predictive accuracy, and clarifying the role of environmental variables holds significant implications for fisheries research. Various feature importance evaluation methods have been developed in recent years, including gradient-based techniques [33], SHAP (Shapley Additive Explanations) [34], and LIME (Local Interpretable Model-agnostic Explanations) [35]. However, the application of these approaches in the field of fishing ground prediction remains relatively limited. It is also noteworthy that the spatial coverage of fishing effort varies substantially across different spatial statistical scales [36], and inappropriate choices of spatiotemporal scales may impair model performance. Therefore, determining a reasonable spatiotemporal scale is critical for ensuring model accuracy [27,31].

To address the above issues, this study proposes a deep learning model that integrates a Convolutional Long Short-Term Memory network (ConvLSTM) with an Attention U-Net for short- to medium-term spatiotemporal prediction of fishing effort distribution. ConvLSTM can effectively capture spatiotemporal dependencies [37]; however, its convolutional layers primarily extract local features, limiting its ability to fully represent long-range spatial dependencies. Attention U-Net, on the other hand, is renowned for its capability in multiscale spatial feature extraction [38,39] but is insufficient in handling temporal dependencies. By combining ConvLSTM with Attention U-Net, this study simultaneously integrates the spatial and temporal features of fishing intensity together with external environmental factors, thereby achieving high-precision prediction of the spatiotemporal distribution of fishing effort. Furthermore, a gradient-based variable importance evaluation method is introduced to assess the potential influence of each variable on the distribution of fishing intensity [40].

To investigate the effects of historical fishing information and spatiotemporal scales on model prediction performance, this study designed two comparative scenarios: (1) incorporating both fishing effort (FE) and marine environmental data as predictors, and (2) using only marine environmental data as predictors, while modeling under different spatiotemporal scales. The main innovations and contributions of this study are as follows: (1) proposing an integrated ConvLSTM–Attention U-Net model for short- to medium-term prediction of marine fishing effort, which effectively captures complex spatiotemporal dependencies; (2) systematically evaluating the influence of historical fishing information on prediction performance; (3) examining model performance across multiple spatiotemporal scales and identifying the optimal scale for prediction; and (4) establishing a variable importance evaluation framework to enhance the interpretability of deep learning models, thereby offering practical implications for fisheries management and the sustainable utilization of marine resources.

2. Data and Methods

2.1. Fishing Effort Data and Environmental Data

2.1.1. Fishing Effort Data

This study focuses on the tuna longline fishery in the Western and Central Pacific Ocean, as this region represents the largest global tuna fishing ground, with an overall tuna catch of approximately 2.7 million tons in 2022, accounting for 54% of the global total [41]. Over the past half-century, global tuna populations have declined by 60%, making their conservation and sustainable management a matter of international concern [42]. Tuna longline fishing, as the fishing method with the largest spatial coverage, exerts the greatest pressure on marine ecosystems.

The study area of this research is the western and central Pacific (130° E–150° W, 20° S–20° N; see Figure 1), covering the time period from 2017 to 2023. The fishing type and fishing effort data based on the AIS were sourced from the Global Fishing Watch (https://globalfishingwatch.org/, accessed on 23 April 2024). This website provides daily navigation trajectories and fishing operation times for global fishing vessels, including date, longitude, latitude, Maritime Mobile Service Identity (MMSI), navigation time, and fishing time. Fishing effort was defined as the total fishing hours of vessels within each grid cell, derived from AIS-based detections provided by GFW. GFW classifies whether fishing vessels are engaged in fishing activities by analyzing high-resolution AIS trajectory data in combination with neural network models. The classification results are further used to identify the type of fishing gear employed, and subsequently to calculate the operating time of vessels within each grid cell. The data are spatially gridded at a resolution of 0.01° (or 0.1°), with fishing effort measured in hours within each grid cell [43].

The fishing effort data were downloaded using the R package ($\mathrm{g}\mathrm{f}\mathrm{w}\mathrm{r}$) provided by the Global Fishing Watch website through its API, using R version 4.3.3 (2024-02-29 ucrt). The time range of the data is from 2017 to 2023, with an initial temporal resolution of one day. The spatial range is 130° E–150° W, 20° S–20° N, with an initial spatial resolution of 0.01° × 0.01°. In this study, fishing intensity is quantified by the magnitude of fishing effort, and the distribution of fishing activities is evaluated based on the spatial distribution of fishing effort.

2.1.2. Environmental Predictor Variables

Based on previous research, sea surface temperature (SST, °C) [2], dissolved oxygen concentration at 100 m depth (O2100, mmol/m³) [44], temperature at 400 m depth (Tem400, °C) [2], chlorophyll-a concentration (Chl-a, mg/m³) [26], dissolved oxygen concentration at 400 m depth (O2400, mmol/m³) [2], and salinity at 100 m depth (Salt100, 1 × 10⁻³) [45] were selected as the environmental factors for this study. Marine environmental data were downloaded from the Copernicus website (https://data.marine.copernicus.eu/products, accessed on 19 May 2024).

The time range for the environmental data is from 2017 to 2023, with an initial temporal resolution of one day. The spatial range is 130° E–150° W, 20° S–20° N. The initial spatial resolution for SST, Tem400, and Salt100 was 0.083° × 0.083°, whereas that for O2100, O2400, and Chl-a was 0.25° × 0.25°. A detailed description of the marine environmental variables is provided in Table S1.

2.2. Data Processing

2.2.1. Design of the Scheme

To explore the model’s performance at different spatial and temporal scales, the study area was divided into grids. Given that longline fishing is a method with a large coverage area, large-scale oceanic fishing vessels using longlines can span distances of up to 100 km [46]. Smaller spatial resolutions, such as 0.1° × 0.1° and 0.25° × 0.25°, are inadequate for accurately assigning the fishing activities of individual vessels to the corresponding grid cells. Therefore, two spatial resolutions were established: 0.5° × 0.5° and 1° × 1°. At the 0.5° × 0.5° resolution, the area is divided into 80 × 160 grids; at the 1° × 1° resolution, the area is divided into 40 × 80 grids.

The temporal resolutions were set at 1 days, 3 days, 7 days, 15 days, and 30 days. Different temporal resolutions represent the daily average of fishing effort aggregated over varying time spans. For example, a resolution of 30 days denotes the daily average fishing effort within each grid cell over a 30-day period. Due to the reduction in sample size associated with larger temporal scales, this study established the maximum temporal resolution at 30 days.

By combining different temporal and spatial resolutions, ten spatiotemporal scales were established. For the sake of simplicity, we represent the spatiotemporal scales in the form of N–M days, where N° indicates a spatial scale of N° × N°, and M days denotes a temporal scale of M days. We calculated the daily average of fishing effort for each grid at each spatiotemporal scale, which serves as the fishing effort value for that grid; concurrently, we computed the overall mean of the environmental data within each grid, representing the environmental variable value for that grid.

The prediction of fishing ground distribution based on environmental factors has been well established in numerous studies. However, research incorporating historical fishing information as model inputs remains relatively limited. Xu et al. [21] attempted to use AIS data together with marine environmental variables as inputs to predict future fishing activity distribution, yet they did not perform a comparison with models using only environmental variables. To investigate the impact of historical fishing information on model performance, this study designed two modeling schemes. In Scheme 1, both historical fishing data and environmental data were used simultaneously as inputs to the neural network (Figure 2). In Scheme 2, only environmental variables were utilized as inputs to the neural network, with the network structure based on Figure 2 excluding the input branch for the fishing effort data. By comparing models with and without historical fishing data as inputs, this study quantified the contribution of historical fishing information to predicting the spatial distribution of fishing activities. Each scheme included ten different combinations of spatiotemporal scales, resulting in a total of 20 trained models.

2.2.2. Normalization

Due to the different dimensions and significant range differences between the fishing effort data and the environmental data used in this study, directly using the raw data to train the model may lead to instability in results and slow convergence of the model. Therefore, this study applied min-max normalization to the fishing effort data and each environmental variable separately, with the normalization formula as follows [31]:

$${\mathrm{X}}^{\mathrm{\prime }}={\displaystyle \frac{{\mathrm{X}}_{\mathrm{i}}-{\mathrm{X}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}{{\mathrm{X}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{X}}_{\mathrm{m}\mathrm{i}\mathrm{n}}}}$$

(1)

where ${\mathrm{X}}^{\mathrm{\prime }}$ represents the normalized value of the sample, ${\mathrm{X}}_{\mathrm{i}}$ denotes the initial value of the sample, and ${\mathrm{X}}_{\mathrm{m}\mathrm{a}\mathrm{x}}$ and ${\mathrm{X}}_{\mathrm{m}\mathrm{i}\mathrm{n}}$ refer to the maximum and minimum values of the respective variable, respectively.

Considering the large range of values in the fishing effort data, along with a small portion of data exhibiting abnormally high values, direct normalization may result in most fishing effort data values being very small, which could hinder model training [28]. To address this issue, we applied a logarithmic transformation to the fishing effort data prior to normalization. The formula for the logarithmic transformation is as follows [28]:

$${\mathrm{X}}^{\mathrm{\prime }}={\mathrm{l}\mathrm{o}\mathrm{g}}_{10}({\mathrm{X}}_{\mathrm{i}}+1)$$

(2)

where ${\mathrm{X}}^{\mathrm{\prime }}$ represents the logarithmically transformed fishing effort, and ${\mathrm{X}}_{\mathrm{i}}$ denotes the initial value of the fishing effort.

2.3. Model Design

2.3.1. Convolutional Long Short-Term Memory Network (ConvLSTM)

In this study, the prediction of fishing activities is essentially a spatiotemporal forecasting problem. Due to the continuous evolution of fishing activities and environmental fields, both fishery data and environmental data exhibit temporal and spatial correlations [27]. Convolutional Neural Networks (CNNs) can effectively capture spatial structural information; however, due to their inherent structural limitations, they struggle to capture temporal dependency information and are unable to adapt to the dynamic changes in time series data. Long Short-Term Memory networks (LSTMs) can effectively capture long- and short-term dependencies in time series through their special gating mechanisms, but they have limitations in handling spatial structural information [21]. The Convolutional Long Short-Term Memory network (ConvLSTM) combines CNNs and LSTMs by introducing convolutional operations within the LSTM units, enabling the model to process both spatial and temporal dimensions of the data simultaneously, thereby effectively capturing spatiotemporal correlations [37]. Based on these advantages, we chose to use ConvLSTM to extract spatiotemporal sequence features of fishing intensity and environmental variables. The core equations of ConvLSTM are as follows [37]:

$$\begin{array}{c}{i}_t=\sigma \left(W_{xi}\ast X_t+W_{hi}\ast H_{t-1}+W_{ci}\circ C_{t-1}+b_i\right)\\ f_t=\sigma \left(W_{xf}\ast X_t+W_{hf}\ast H_{t-1}+W_{cf}\circ C_{t-1}+b_f\right)\\ C_t=f_t\circ C_{t-1}+i_t\circ tanh\left(W_{xc}\ast X_t+W_{hc}\ast H_{t-1}+b_c\right)\\ o_t=\sigma \left(W_{xo}\ast X_t+W_{ho}\ast H_{t-1}+W_{co}\circ C_t+b_o\right)\\ H_t=o_t\circ tanh\left(C_t\right)\end{array}$$

(3)

Here, $\ast$ denotes the convolution operation, $\circ$ represents the Hadamard product, $X_t$ is the input at the current time step, $H_{t-1}$ is the hidden state from the previous time step, and $C_{t-1}$ is the memory state from the previous time step. The variables $i_t$, $f_t$, and $o_t$ represent the input gate, forget gate, and output gate, respectively, and $\sigma$ denotes the sigmoid activation function. Through this mechanism, ConvLSTM is able to capture temporal dependencies while simultaneously preserving spatial structural information.

2.3.2. Attention U-Net

The environmental field at each time step can be regarded as a frame of an image, with each grid point on the ocean surface corresponding to a pixel in the image. Given the significant correlation between the marine environment and the distribution of tuna resources, neural networks can infer potential tuna fishing activities by analyzing the distribution of marine environmental factors. Thus, the problem of using marine environmental data to predict the distribution of fishing activities can be considered as an image segmentation task [27].

The ConvLSTM model relies on convolutional layers to capture spatial dependencies; however, these dependencies are local and inefficient. In contrast, long-distance spatial dependencies are significant for the spatial features of fishing effort. As fishing vessels travel at high speeds, they may reach a distant grid after some time, leading to the fishing intensity at that grid being influenced by more distant grids. U-Net has demonstrated superior performance in semantic segmentation and spatiotemporal prediction due to its strengths in capturing multi-scale spatial features [38].

U-Net is a classic fully convolutional network architecture that captures multi-scale features of images through an encoder–decoder structure. The encoder part extracts features at different scales using convolutional and downsampling layers, while the decoder part progressively restores the spatial resolution of the image through upsampling and convolutional layers. Additionally, the skip connections in U-Net preserve high-resolution features by connecting across layers, enhancing the model’s capability for feature extraction and spatial information recovery [38].

Based on this, Attention U-Net further incorporates an attention mechanism, enhancing the model’s ability to focus on important regions. The core idea of the attention mechanism is to dynamically adjust the model’s focus based on the significance of different areas in the input data [39]. This study combines ConvLSTM and Attention U-Net to simultaneously interpret spatial and temporal feature information and output fused expressions, reflecting the potential impacts of all predictive variables on the distribution of fishing intensity. This approach offers a promising avenue for accurately predicting the two-dimensional spatiotemporal distribution of fishing effort.

2.3.3. ConvLSTM Attention U-Net (CLA U-Net)

The CLA U-Net model proposed in this study combines ConvLSTM and Attention U-Net, as illustrated in Figure 2, where different colored rectangles represent various operations. The model receives fishing effort information and environmental variables for the first five time steps and outputs the fishing effort distribution for the subsequent one time step. To accommodate the input data format requirements of ConvLSTM, both fishing effort data and environmental data are processed into five-dimensional tensors, with different environmental variables distinguished through the channels of the tensor.

First, the marine environmental data and fishing effort data are processed through two layers of ConvLSTM to extract spatiotemporal features, which are then concatenated and used as input for the Attention U-Net to achieve the final prediction. The hidden state output from the second layer of ConvLSTM encompasses all spatiotemporal information across the time steps; therefore, we select this hidden state as the extracted spatiotemporal features to input into the subsequent Attention U-Net.

The encoder and decoder of the Attention U-Net consist of three downsampling modules and three upsampling modules, respectively. The downsampling modules reduce spatial resolution by setting a stride of 2 in the two-dimensional convolutional neural network (CNN), while the upsampling modules restore spatial resolution using transposed convolutions. Each downsampling module contains a downsampling layer followed by two two-dimensional convolutional layers, with normalization layers and ReLU activation functions connected after each convolutional layer. The skip connections between the encoder and decoder are facilitated through an attention mechanism gate. Specifically, the attention mechanism gate transforms the feature maps of the current layers in the encoder and decoder using a 1 × 1 convolution, and the transformed feature maps are then summed. Subsequently, the feature maps are further processed with a ReLU activation function and a 1 × 1 convolution, and the attention weights are computed using a sigmoid activation function. These weights are used to weigh the feature maps output by the encoder, which are then fused (concatenated) with the feature maps from the current layer of the decoder before being passed through two two-dimensional convolutional layers and then to the next upsampling layer.

To mitigate overfitting during the early stages of model training, we introduced Dropout2D layers at the end of each downsampling module, as well as at the end of the second and third upsampling modules. Experiments determined that the optimal dropout rate is 0.35. The final part of the model consists of three two-dimensional convolutional layers. Given that this study aims to predict the values of fishing effort, the final output of the model is processed using the ReLU activation function.

2.4. Deep Learning Model Fitting

The implementation of the models is based on the PyTorch 2.3.0 deep learning framework and runs in a Python 3.12 environment on the AutoDL cloud platform (https://www.autodl.com/, accessed on 27 July 2024) equipped with an NVIDIA RTX 4090 GPU, running Ubuntu 22.04. The training process relies on CUDA 12.1. The dataset used encompasses fishing effort data and marine environmental data for the Central Western Pacific region (130° E–150° W, 20° S–20° N) from 2017 to 2023, with data from 2017 to 2021 used for training, data from 2022 serving as the validation set, and data from 2023 as the test set. The experimental results indicate that Scheme 1 achieved the best training performance when using the 1:1 combination of MAE and MSE as the loss function, while Scheme 2 performed optimally with the MSE loss function. The model training utilized the Adam optimizer with a learning rate set at 0.0002 and a batch size of 16, training each model for 100 epochs. To ensure the stability of the conclusions, this study conducted six independent experiments and ultimately reported the mean and standard deviation of each metric across these six experiments. Since this study considers the F1 score as the primary evaluation metric, the models saved during training correspond to those with the highest F1 scores on the validation set.

2.5. Evaluation Metrics

A regression prediction of the distribution of longline tuna fishing effort in the WCPO was conducted in this study. Traditional regression evaluation metrics, such as the mean absolute error (MAE) and root mean square error (RMSE), were employed to measure the accuracy of numerical predictions, thereby reflecting how well the predicted fishing effort values matched the observed values [21]. However, low regression errors do not necessarily indicate that the model can accurately predict the spatial distribution of fishing effort. Therefore, we introduced classification metrics to assess the model’s accuracy in predicting the locations of fishing activities.

When evaluating the model’s accuracy in predicting fishing activity locations, grid cells with fishing effort greater than 0 were labeled as 1 (fishing activity present), while all other cells were labeled as 0 (no fishing activity). We then used accuracy, recall, precision and F1 score as classification metrics. Accuracy reflects the proportion of correctly classified grid cells among all cells, providing an overall measure of prediction correctness; precision quantifies the proportion of grid cells predicted as fishing activity that are indeed correct; and recall emphasizes the model’s ability to identify grid cells where fishing activity actually occurred. Since accuracy, precision, and recall alone do not comprehensively capture classification performance, the F1 score is adopted as a composite metric to evaluate the model’s overall classification capability [47]. Additionally, to assess the model’s precision in numerical predictions, MAE and RMSE were retained as supplementary regression metrics.

The formulas for calculating the regression metrics are as follows [48]:

$$\mathrm{M}\mathrm{A}\mathrm{E}={\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}} \left\{\left|\mathrm{F}\mathrm{E}\left({\mathrm{x}}_{\mathrm{i}}\right)-\widehat{\mathrm{F}\mathrm{E}}\left({\mathrm{x}}_{\mathrm{i}}\right)\right|\right\}$$

(4)

$$\mathrm{R}\mathrm{M}\mathrm{S}\mathrm{E}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}} {\left\{\mathrm{F}\mathrm{E}\left({\mathrm{x}}_{\mathrm{i}}\right)-\widehat{\mathrm{F}\mathrm{E}}\left({\mathrm{x}}_{\mathrm{i}}\right)\right\}}^2}$$

(5)

where $\mathrm{n}$ is the total number of grid points within the region, $\mathrm{F}\mathrm{E}\left({\mathrm{x}}_{\mathrm{i}}\right)$ is the true fishing effort value at grid point $\mathrm{i}$ and $\widehat{\mathrm{F}\mathrm{E}}\left({\mathrm{x}}_{\mathrm{i}}\right)$ is the predicted fishing effort value at grid point $\mathrm{i}$.

The formulas for calculating the classification metrics are as follows [26]:

$$\mathrm{A}\mathrm{c}\mathrm{c}\mathrm{u}\mathrm{r}\mathrm{a}\mathrm{c}\mathrm{y}={\displaystyle \frac{{\mathrm{N}}_{\mathrm{T}\mathrm{P}}+{\mathrm{N}}_{\mathrm{T}\mathrm{N}}}{{\mathrm{N}}_{\mathrm{T}\mathrm{P}}+{\mathrm{N}}_{\mathrm{F}\mathrm{P}}+{\mathrm{N}}_{\mathrm{T}\mathrm{N}}+{\mathrm{N}}_{\mathrm{F}\mathrm{N}}}}$$

(6)

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{{\mathrm{N}}_{\mathrm{T}\mathrm{P}}}{{\mathrm{N}}_{\mathrm{T}\mathrm{P}}+{\mathrm{N}}_{\mathrm{F}\mathrm{P}}}}$$

(7)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{{\mathrm{N}}_{\mathrm{T}\mathrm{P}}}{{\mathrm{N}}_{\mathrm{T}\mathrm{P}}+{\mathrm{N}}_{\mathrm{F}\mathrm{N}}}}$$

(8)

$$\mathrm{F}1-\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \frac{2\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}\times \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}{\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}+\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}}}$$

(9)

where ${\mathrm{N}}_{\mathrm{T}\mathrm{P}}$ denotes the number of grid points where fishing effort actually exists and is predicted to have fishing effort, ${\mathrm{N}}_{\mathrm{F}\mathrm{P}}$ represents the number of grid points where fishing effort actually exists but is predicted to have none, ${\mathrm{N}}_{\mathrm{T}\mathrm{N}}$ denotes the number of grid points where fishing effort does not exist and is predicted to have no fishing effort, and ${\mathrm{N}}_{\mathrm{F}\mathrm{N}}$ indicates the number of grid points where fishing effort does not exist but is predicted to have fishing effort.

2.6. Variable Importance Assessment

Due to the “black box” nature of deep neural networks and their limited interpretability, it is often challenging to explain the contribution of each variable to the model’s performance [32]. Existing variable selection methods are primarily suited for shallow networks or incur high computational costs on large datasets. To address this issue and assess the importance of each variable on the model’s output, we employed a gradient-based variable importance measurement method [40]. This method quantifies the influence of each variable on the model output by calculating the average absolute gradient of the loss function with respect to that variable.

Specifically, we first utilized the standard backpropagation algorithm to compute the gradient values of the loss function relative to each input variable. The gradient values reflect the sensitivity of the loss function to the input variables; a larger absolute gradient indicates a more significant impact of that variable on the model’s loss. Subsequently, we ranked the average absolute gradient values for each variable, with variables exhibiting larger gradient values considered to contribute more significantly to the model’s loss and thus deemed more important; conversely, variables with smaller gradient values were regarded as having lower importance for the model’s predictions.

Taking into account that the gradients of different models may vary in magnitude, we calculated the percentage of each variable’s gradient relative to the total gradient to provide a more intuitive representation of their importance. The calculation formula is as follows:

$${\mathrm{i}\mathrm{m}\mathrm{p}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}}_{\mathrm{i}}={\displaystyle \frac{{\mathrm{G}}_{\mathrm{i}}}{{\sum }_{\mathrm{j}}^{\mathrm{n}}{\mathrm{G}}_{\mathrm{j}}}}\times 100\mathrm{\%}$$

(10)

In this equation, ${\mathrm{i}\mathrm{m}\mathrm{p}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}}_{\mathrm{i}}$ represents the importance of variable $\mathrm{i}$, while ${\mathrm{G}}_{\mathrm{i}}$ denotes the average absolute gradient value of variable $\mathrm{i}$.

3. Result

3.1. Spatiotemporal Characteristics of Tuna Longline Fishing Vessel Operations in the WCPO

Analysis of fishing effort based on AIS data revealed significant spatial and temporal variations in tuna longline fishing activities in the western and central Pacific. Specifically, from 2017 to 2019, the total annual fishing effort in this region exhibited a year-on-year increase, reaching its peak in 2019 (Figure 3a). However, fishing effort declined sharply in 2020 and reached its lowest point in 2021. Subsequently, fishing effort partially recovered in 2022 and 2023, indicating a trend of gradual rebound. Fishing activity was most frequent between March and August, with April to July showing significantly higher effort compared to other months (Figure 3b). In contrast, fishing effort during January, November, and December was relatively low, markedly below the levels observed in other months.

In terms of spatial distribution, the total annual fishing effort from 2017 to 2023 exhibited some variation, yet the overall operational area displayed a consistent distribution pattern. The annual fishing hotspots—defined as grid cells with fishing effort exceeding 500 h at a spatial resolution of 0.5° × 0.5°—demonstrated similar spatial patterns across years (Figure 4). These hotspots were primarily concentrated in the following regions: 130° E–140° E, 10° N–20° N; 150° E–170° E, 20° S–0°; and a longitudinal strip within 18° W–150° W, 10° N–20° N. The distribution characteristics of these areas indicate that high-frequency fishing activities are strongly concentrated within specific marine regions.

3.2. Optimal Spatiotemporal Scale Assessment

To ensure the reliability of the results, six independent experiments were conducted for each spatiotemporal scale, and the mean ± standard deviation of each evaluation metric was calculated (

Table 1. Accuracy of Scheme 1 at different spatiotemporal scales.

Spatiotemporal Scale	Accuracy	F1	Precision	Recall	MAE	RMSE
0.5°-1 days	0.961 ± 0.000	0.654 ± 0.002	0.651 ± 0.009	0.656 ± 0.011	0.372 ± 0.014	1.93 ± 0.04
0.5°-3 days	0.917 ± 0.004	0.578 ± 0.007	0.547 ± 0.022	0.615 ± 0.020	0.437 ± 0.034	1.89 ± 0.06
0.5°-7 days	0.875 ± 0.006	0.592 ± 0.005	0.562 ± 0.023	0.626 ± 0.034	0.427 ± 0.018	1.65 ± 0.02
0.5°-15 days	0.835 ± 0.003	0.638 ± 0.002	0.603 ± 0.009	0.677 ± 0.012	0.420 ± 0.007	1.40 ± 0.00
0.5°-30 days	0.802 ± 0.008	0.689 ± 0.004	0.636 ± 0.023	0.753 ± 0.037	0.392 ± 0.012	1.15 ± 0.01
1°-1 days	0.943 ± 0.000	0.763 ± 0.001	0.767 ± 0.002	0.759 ± 0.004	1.14 ± 0.03	4.31 ± 0.10
1°-3 days	0.893 ± 0.002	0.697 ± 0.002	0.684 ± 0.011	0.711 ± 0.015	1.40 ± 0.03	4.82 ± 0.04
1°-7 days	0.843 ± 0.003	0.693 ± 0.001	0.661 ± 0.009	0.728 ± 0.011	1.50 ± 0.02	4.86 ± 0.05
1°-15 days	0.803 ± 0.003	0.726 ± 0.001	0.673 ± 0.008	0.787 ± 0.011	1.52 ± 0.01	4.46 ± 0.04
1°-30 days	0.790 ± 0.007	0.770 ± 0.002	0.715 ± 0.017	0.835 ± 0.022	1.39 ± 0.01	3.71 ± 0.04

Note: The bold values highlight the highest F1 scores under the same spatial scale across different temporal scales.

and

Table 2. Accuracy of Scheme 2 at different spatiotemporal scales.

Spatiotemporal Scale	Accuracy	F1	Precision	Recall	MAE	RMSE
0.5°-1 days	0.811 ± 0.041	0.245 ± 0.012	0.161 ± 0.017	0.542 ± 0.092	0.478 ± 0.022	2.62 ± 0.01
0.5°-3 days	0.736 ± 0.062	0.324 ± 0.021	0.216 ± 0.027	0.673 ± 0.100	0.490 ± 0.027	2.24 ± 0.01
0.5°-7 days	0.754 ± 0.046	0.432 ± 0.011	0.333 ± 0.035	0.648 ± 0.128	0.478 ± 0.010	1.92 ± 0.02
0.5°-15 days	0.712 ± 0.029	0.534 ± 0.008	0.412 ± 0.024	0.768 ± 0.059	0.475 ± 0.022	1.61 ± 0.01
0.5°-30 days	0.720 ± 0.025	0.626 ± 0.011	0.513 ± 0.028	0.805 ± 0.046	0.458 ± 0.036	1.34 ± 0.08
1°-1 days	0.708 ± 0.024	0.375 ± 0.006	0.253 ± 0.009	0.722 ± 0.045	2.03 ± 0.10	7.47 ± 0.14
1°-3 days	0.672 ± 0.028	0.458 ± 0.011	0.321 ± 0.015	0.802 ± 0.035	1.95 ± 0.03	6.76 ± 0.06
1°-7 days	0.653 ± 0.017	0.547 ± 0.006	0.401 ± 0.011	0.861 ± 0.022	1.98 ± 0.04	6.03 ± 0.06
1°-15 days	0.695 ± 0.006	0.648 ± 0.006	0.524 ± 0.006	0.850 ± 0.010	1.98 ± 0.08	5.45 ± 0.20
1°-30 days	0.715 ± 0.010	0.720 ± 0.005	0.615 ± 0.013	0.870 ± 0.015	1.89 ± 0.04	4.85 ± 0.25

Note: The bold values highlight the highest F1 scores under the same spatial scale across different temporal scales.

), thereby providing statistical significance. To intuitively illustrate the model performance across different spatiotemporal scales, the key metrics F1, Precision, Recall, MAE, and RMSE were visualized (Figure 5 and Figure 6).

In Scheme 1, which incorporated fishing effort data, the models achieved consistently robust results across all spatiotemporal scales, with relatively small standard deviations for all metrics, indicating high consistency and statistical stability of the experiments. Regarding classification performance, the model at the spatiotemporal scale of 1°-30 days demonstrated the best results (Table 1), achieving an F1 score of 0.770 ± 0.002, which reflects a high level of accuracy in predicting the presence or absence of fishing activities within the study area. The Recall reached 0.835 ± 0.022, suggesting that the model correctly identified 83.5% of the actual fishing activity locations, while the Precision of 0.715 ± 0.017 indicates that 71.5% of the predicted fishing activity locations corresponded to true fishing activity. At the same spatial scale, the key classification metrics (F1, Precision, and Recall) exhibited a trend of initially decreasing and then increasing with longer temporal scales, highlighting the influence of temporal resolution on model performance (Figure 5a). At the same temporal scale, the models with a spatial resolution of 1° × 1° consistently outperformed those with a resolution of 0.5° × 0.5° in terms of classification metrics, indicating that a coarser spatial scale was more advantageous for accurately predicting fishing activity locations. Notably, the model also demonstrated strong classification performance at the shortest temporal scale. At the 1°-1-day spatiotemporal scale, the F1 score reached 0.763 ± 0.001, underscoring the model’s potential for short-term prediction.

The regression metrics were used to evaluate the accuracy of the models in predicting the numerical values of fishing effort. The results revealed that, at the same temporal scale, the models with a 0.5° spatial resolution consistently exhibited lower errors compared to those with a 1° resolution (Figure 5b), suggesting that finer spatial resolution was more beneficial for accurately predicting fishing effort values. Among these, the 0.5°-30-day model achieved the lowest RMSE (1.15 ± 0.01), while the 0.5°-1-day model obtained the lowest MAE (0.372 ± 0.014). Further analysis showed that, at the 0.5° spatial scale, RMSE decreased progressively with increasing temporal scale. In contrast, at the 1° spatial scale, both MAE and RMSE followed the same trend as the classification metrics, first increasing and then decreasing.

In Scheme 2, the standard deviations of the F1 scores were relatively small across all spatiotemporal scales. However, at the scales of 0.5°-1 days, 0.5°-3 days, and 0.5°-7 days, the standard deviations of Recall were comparatively larger, indicating that the model exhibited certain instability at these scales. In terms of classification metrics, Scheme 2 achieved its best performance at the spatiotemporal scale of 1°-30 days, with an F1 score of 0.720 ± 0.005, a Precision of 0.615 ± 0.013, and a Recall of 0.870 ± 0.015 (Table 2). This indicates that, at this scale, the model successfully identified 87% of the actual fishing activity locations. However, only 61.5% of the locations predicted by the model as fishing areas corresponded to true fishing activities, resulting in a false alarm rate of 38.5%, which contributed to the lower F1 score compared with Scheme 1. At the same spatial scale, the model’s ability to predict fishing activity locations improved with increasing temporal scale; under the same temporal scale, models with a 1° × 1° spatial resolution outperformed those with a 0.5° × 0.5° resolution (Figure 6a).

Regarding regression metrics, both MAE and RMSE in Scheme 2 generally decreased with increasing temporal scale at the same spatial resolution. At the same temporal scale, models with a spatial resolution of 0.5° × 0.5° consistently achieved higher regression accuracy than those with a resolution of 1° × 1° (Figure 6b). These findings suggest that, when only environmental variables are used as inputs, longer temporal scales and finer spatial resolutions are more favorable for improving the accuracy of fishing effort value predictions. Nevertheless, compared with Scheme 1, the regression accuracy of Scheme 2 was consistently lower across all spatiotemporal scales.

Overall, at the same spatiotemporal scale, the F1 values of Scheme 2, which only used environmental variables, were consistently lower than those of Scheme 1, which incorporated fishing effort data. This demonstrates that Scheme 1 had stronger overall predictive capability for fishing activity locations. Except for the scale of 1°-1 days, the Recall of Scheme 2 was higher than that of Scheme 1 across all scales, whereas the Precision of Scheme 2 was consistently lower. This suggests that Scheme 2 tended to predict more areas as fishing zones, thereby increasing the detection rate of fishing activities but also introducing a higher false alarm rate. In terms of regression metrics, Scheme 1 achieved lower MAE and RMSE than Scheme 2 across all spatiotemporal scales, further confirming that incorporating historical fishing effort data is beneficial for improving the accuracy of fishing effort value predictions.

3.3. Model Prediction Performance Evaluation

We visualized the prediction results of the two schemes across different spatiotemporal scales (Figure 7 and Figure 8). The visualization results were consistent with the aforementioned evaluation metrics. In predicting the location and intensity of fishing activities, Scheme 1 outperformed Scheme 2 at all spatiotemporal scales, demonstrating a closer alignment with the actual distribution of fishing activities. At the same temporal scale, the model with a spatial resolution of 1° exhibited better predictive performance than the 0.5° model in both schemes. At the same spatial scale, the similarity of predictions in Scheme 1 to the actual distribution of fishing activities initially decreased and then increased with increasing temporal scale, whereas the similarity of predictions in Scheme 2 continuously improved as the temporal scale increased.

Scheme 1 achieved an F1 score of 0.770 ± 0.002 at the optimal spatiotemporal scale (1°-30 days), whereas the F1 score at the 1°-1-day scale was 0.763 ± 0.001, which was only slightly lower than that at the optimal scale. Scheme 1 demonstrated a more effective ability to differentiate fishing intensity across all spatiotemporal scales. The visualization results indicated that the predictions of Scheme 1 at the 1°-1-day scale closely aligned with the actual distribution of fishing activities (Figure 8a), indicating significant potential at small temporal scales.

Scheme 2 tends to predict more areas as potential fishing zones, enabling it to forecast a greater amount of actual fishing activity, albeit with a higher false positive. At smaller temporal scales, the predicted values of Scheme 2 were generally low, making it difficult to distinguish fishing intensity effectively. However, as the temporal scale increases to 15 days and 30 days, Scheme 2 showed improvement in differentiating fishing intensity. Therefore, the optimal scale for Scheme 2 was (1°-30 days), with the model’s predictions closely aligned with the actual distribution of fishing activities and effectively distinguished fishing intensity.

3.4. Assessment of Variable Importance

This study evaluated the importance of each variable in model predictions by calculating the average absolute gradient values with respect to the loss function. In Scheme 1, the contributions of historical fishing intensity data and environmental data exhibit opposing fluctuation trends. Historical fishing intensity data contribute the most to the model across all spatiotemporal scales (Figure 9a). At the spatiotemporal scales of 0.5°-1 days and 1°-1 days, the importance of fishing effort (FE) reached 64.6% and 61.9%, respectively. At the same spatial scale, the importance of FE exhibited a decreasing–increasing trend as the temporal scale increased. In contrast, the importance levels of the remaining six oceanic environmental variables were relatively similar, with their contributions ranging from 5% to 12% across different spatiotemporal scales. Although these variables influenced the model predictions, distinguishing their relative importance was challenging.

In Scheme 2, the importance of the six oceanic environmental variables exhibited relatively pronounced differences (Figure 9b) and varied with changes in the spatiotemporal scale. At the optimal spatiotemporal scale of 1°-30 days, sea surface temperature (SST) was identified as the most important variable, accounting for 23.8%, which is significantly greater than the contributions of the other variables. In contrast, O2400 was considered the least important variable, with an importance of 12.4%. SST exerted the greatest influence at the spatiotemporal scales of 0.5°-30 days and 1°-30 days, whereas Chl-a was identified as the most important variable at the remaining eight spatiotemporal scales, highlighting the substantial impacts of SST and Chl-a on fishing activities. At all spatiotemporal scales, the importance of most variables exceeded 10%, indicating that the six selected oceanic environmental variables had a significant impact on the model.

4. Discussion

4.1. Spatiotemporal Prediction of Fishing Effort

This study employed the CLA U-Net model to predict the short- and medium-term spatiotemporal distribution of tuna longline fishing intensity in the western and central Pacific. Compared with Crespo et al. [2], Scheme 1 and Scheme 2 showed slightly lower accuracy at the 1°-30-day scale, primarily due to differences in study design: Crespo et al. predicted fishing effort within the same month using contemporaneous environmental variables, whereas this study predicted future distributions from historical data, a more challenging task. Their global study domain included many regions with little or no longline activity, making predictions easier and facilitating higher accuracy. However, high accuracy does not necessarily indicate accurate capture of actual fishing activity; in regions with many non-fishing grids, accuracy can be inflated. For example, under Scheme 2 at 0.5°-1 days, accuracy reached 0.811, but the F1-score was only 0.245, reflecting the influence of negative samples and potential overestimation of predictive capacity. To address this, F1-score was employed for a more comprehensive performance assessment. By comparison, Su et al. [26] achieved a maximum F1-score of 0.888 in a binary classification model, substantially higher than this study, likely due to predicting within the same month using contemporaneous data and a coarser spatial resolution (5° × 5°). These results indicate that study area, prediction targets, model design, and spatial resolution significantly affect performance, and comprehensive evaluation using multiple metrics (e.g., F1, Recall, Precision) is essential under different research frameworks.

In Scheme 1, the CLA U-Net model achieved strong performance across all spatiotemporal scales. In Scheme 2, although the model performed suboptimally at finer temporal resolutions, it demonstrated satisfactory results at coarser temporal scales. Overall, the CLA U-Net model effectively captures the dependencies in both temporal and spatial dimensions, demonstrating significant application potential in predicting marine fishing intensity. Compared with traditional regression models and machine learning methods, this model extends the input and output dimensions from one dimension to multiple dimensions. The spatial distribution of fishing intensity is a continuous dynamic evolution process that is driven by the coupling of internal and external factors [21]. When data conditions are sufficient, deep learning methods for spatiotemporal prediction should be prioritized for forecasting the spatial information of fisheries.

Although the models developed in this study achieved satisfactory results, several limitations remain. First, the fishing effort data employed were inferred from AIS trajectory data, which may be incomplete due to factors such as deliberate deactivation, equipment malfunctions, or signal obstruction, potentially affecting data integrity and accuracy. Nevertheless, the extensive, continuous, and wide-coverage vessel behavior information provided by AIS has been widely applied in fisheries research, demonstrating high utility and reliability, particularly in vessel activity classification, fishing effort estimation, and spatial behavior analysis [6,7,8,9,10,11,12,13,14,15]. Second, the models primarily focused on the driving effects of historical fishing data and oceanic environmental factors, without incorporating more complex socioeconomic variables related to operational costs, such as distance from shore (DSH), fuel prices, port distribution, and fish market prices, which may limit a comprehensive understanding of longline fishing activities. Moreover, at larger spatiotemporal scales, model stability and generalization may be constrained by limited sample sizes. Future research could enhance predictive accuracy and practical applicability by integrating longer time series, additional environmental variables, and a broader set of socioeconomic factors.

4.2. Impact of Historical Fishing Information on Model Outcomes

Currently, studies incorporating historical fishing information as model input remain relatively limited. Xu et al. [21] developed a multi-task tuna fishery prediction model based on ConvLSTM, integrating historical fishery information with environmental variables. However, they did not compare their model with one using only environmental variables, and thus the specific contribution of historical fishery information was not fully demonstrated. It is worth noting that, although studies directly incorporating historical fishing data for prediction remain limited in the fisheries field, similar approaches have been widely adopted in other domains [49,50]. For instance, in short-term rainfall forecasting, researchers often combine historical radar echo data or meteorological observations with environmental variables as model inputs to improve prediction accuracy and stability [49,50]. These cross-disciplinary studies demonstrate that leveraging historical dynamic information to predict future activities is an effective and reliable modeling strategy, thereby supporting the rationale of Scheme 1 in this study.

A comparative analysis of Scheme 1 and Scheme 2 indicates that the introduction of historical fishing information significantly enhances the model’s performance in predicting fishing locations and intensity. The habitat distribution of tuna is profoundly influenced by marine environmental factors [26,48,51]. Areas where oceanic environmental conditions satisfy the habitat requirements of tuna may serve as potential fishing grounds. Neural networks trained solely on environmental predictors can learn the ecological niche of tuna longline operations and thus forecast likely operating areas. Due to the existence of areas not covered by fishing vessels or regions without fishing activities (such as marine protected areas), the model in Scheme 2 exhibits a high recall rate, enabling it to predict most potential fishing zones; however, this also leads to an increased number of false-positives (Figure 7 and Figure 8). Moreover, Scheme 2 performs poorly in predicting fishing intensity because of the similar environmental conditions in different marine areas. Fishermen typically track and catch fish based on their experience and real-time oceanic environmental conditions, making the movement of fishing vessels at sea a dynamic and continuous process [18,19,20]. Therefore, leveraging historical fishing information can effectively constrain the model’s tendency to over predict potential fishing areas, thereby reducing the false-positive. Additionally, fishermen’s behavior often exhibits directional trends. By integrating historical fishing intensity data across multiple time steps, the model can better capture vessel movement patterns, thereby improving the accuracy of predictions for tuna longline fishing activities. Furthermore, the interaction between historical fishing intensity data and environmental factors allows the model to identify and predict fishing intensity more effectively.

4.3. Variability Analysis of Model Performance Across Different Spatiotemporal Scales

Numerous studies have demonstrated that the choice of spatiotemporal scales significantly affects the prediction of fishery resource distributions [27,31,52,53]. In this study, at the same temporal scale, coarser spatial resolution (1°) improved the accuracy of predicting fishing locations, whereas the precision of fishing intensity predictions was relatively lower. This discrepancy primarily arises from the spatial aggregation effect: larger grid cells aggregate and average data, thereby mitigating local fluctuations and observational noise, which enhances model performance in classification tasks [31,52,53,54]. In contrast, at finer spatial resolutions, the fishing effort within each grid cell is generally lower than that at coarser scales, and this numerical difference contributes to improved model performance in predicting fishing intensity. Furthermore, smaller grid cells retain more detailed local environmental information, facilitating more precise numerical fitting of fishing intensity [55].

At the same spatial scale, the optimal temporal scale for both Scheme 1 and Scheme 2 was 30 days, consistent with the majority of previous studies [29,56]. However, with increasing temporal scale, the models exhibited different trends: the prediction accuracy of Scheme 1 initially declined and then increased, whereas that of Scheme 2 continuously improved. For Scheme 2, at smaller temporal scales, the number of observed fishing locations was far lower than that of non-fishing locations, while the area meeting fishing conditions remained relatively large. This led to a severe imbalance between positive and negative samples (presence versus absence of fishing), substantially affecting the performance of the neural network [57] and causing the model to overpredict areas with fishing activity. In practice, fishers do not necessarily exploit all available environmental information but often select fishing locations based on specific environmental factors [58,59]. Consequently, the environmental conditions at fishing and non-fishing locations may be highly similar [31], further increasing the difficulty of accurate model prediction. From another perspective, the overprediction by Scheme 2 at small temporal scales may also reflect the model’s identification of potential fishing areas. As the temporal scale increases, data aggregation helps eliminate noise and background variability, clarifying the relationship between fishing activity and environmental conditions [31]. Simultaneously, fishers’ spatial distribution gradually expands, with more locations exhibiting fishing activity, thereby mitigating sample imbalance and improving model accuracy. The effect of temporal scale on model performance differed between the two schemes because Scheme 1 incorporated historical fishing information. At a 1-day temporal scale, each fishing point in Scheme 1 could be regarded as the operational location of a single vessel, with variations in fishing locations reflecting the trajectories of individual vessels. The neural network effectively captured these movement patterns, resulting in high prediction accuracy. However, when the temporal scale extended to 3 or 7 days, a single vessel might operate at multiple locations within that period, making it difficult for the model to extract clear movement patterns and causing a decline in predictive performance. As the temporal scale further increased to 15 or 30 days, the operational range of individual vessels could cover the entire study area, allowing the model to learn overall trends in fishing activity and thereby improving predictive accuracy. Across both schemes, the MAE remained relatively stable, whereas the RMSE significantly decreased with increasing temporal scale, indicating that data aggregation played a crucial role in reducing noise, mitigating the influence of outliers, and improving RMSE performance [53].

The selection of spatiotemporal scales has a significant impact on model performance. Xie et al.’s [27] study on squid fishing grounds in the northwestern Pacific indicated that larger temporal scales lead to higher model accuracy, a finding consistent with the results of Scheme 2 in this study. However, their study did not observe performance improvements with increasing spatial scale, which contrasts with our findings. In contrast, Han et al.’s [31] research on purpleback flying squid fishing grounds in the northwestern Indian Ocean demonstrated that, for both 2D CNNs and 3D CNNs, model performance declined as spatiotemporal scales increased, directly opposing the conclusions of this study. These comparisons indicate that differences in study species and methodological approaches can lead to varying effects of spatiotemporal scale on model performance. Therefore, when selecting appropriate spatiotemporal scales, it is essential to consider both oceanic environmental conditions and the biological characteristics of the target species.

4.4. Analysis of Variable Importance in the Model

The spatial distribution of fishing activities is a continuously dynamic process influenced by both external factors (e.g., marine environmental conditions) and internal factors (e.g., prior and surrounding fishing activities) [18,19,20,21]. The average importance of variables at different spatial and temporal scales in Scheme 1 indicates that the importance of historical catch data is markedly greater than that of environmental variables, whereas the differences in importance among the six environmental variables are relatively minor. Notably, the model also exhibited strong classification performance at a 1-day temporal scale, highlighting its potential for short-term prediction. At this temporal scale, the importance of historical fishing data reaches its maximum. This is because, within this temporal resolution, the number of fishing points in the study area is relatively limited, making it difficult to accurately identify fishing activity locations using only environmental variables. At the same time, vessel movement trajectories are clearer at this scale, enabling the neural network to effectively capture their movement patterns, thereby substantially enhancing the importance of historical catch data and improving the model’s predictive accuracy.

Numerous studies have demonstrated a close relationship between marine environmental conditions and the distribution of fishing grounds [26,44,48]. In Scheme 2, the influence of environmental variables on fishing activity distribution varies across different spatiotemporal scales. The six environmental variables employed in this study have been recognized in previous research as having significant or high impacts on tuna longline fishing activities [2,26,44,45]. Crespo et al. identified sea surface temperature (SST) as the most influential factor affecting fishing activity distribution [2], which aligns with the present findings in Scheme 2, where SST exhibits the highest importance at the 0.5°-30-day and 1°-30-days scale. In the remaining eight spatiotemporal scales, chlorophyll-a (Chl-a) consistently ranks highest in importance, indicating its key role in fishing activity, consistent with Su et al.’s results showing optimal model performance when Chl-a is used [26]. Overall, these findings suggest that fishing activities respond consistently to key environmental factors, reflecting the ecological adaptation strategies of tuna within the marine ecosystem. Environmental factors influence fishing activities by affecting the spatial distribution of tuna. In this study, SST governs the spatial distribution of tuna, Chl-a concentration determines their food availability, and salinity plays a critical role in growth and metabolism [60,61]. Moreover, due to tuna’s high oxygen demand, dissolved oxygen levels in seawater also affect their distribution [43]. The model results confirm that these variables significantly influence the spatial distribution of fishing effort, providing a scientific basis for constructing accurate predictive models of fishing activity.

5. Conclusions

This study applied the CLA U-Net model to predict short- and medium-term spatiotemporal distributions of tuna longline fishing activities in the WCPO. Two modeling schemes were evaluated: Scheme 1, incorporating both historical fishing effort (FE) and environmental variables, and Scheme 2, using environmental variables only.

The results support the following conclusions: (1) The CLA U-Net model effectively predicts the spatiotemporal distribution of fishing intensity for both short- and medium-term periods. At the optimal 1°-30-day scale, Scheme 1 achieved an F1 score of 0.770, a Precision of 0.715, and a Recall of 0.835, indicating a high accuracy in predicting the locations of fishing activities. (2) Historical fishing intensity data substantially enhance prediction performance. Scheme 1, which includes these data, better captures spatiotemporal trends, while Scheme 2 identifies a broader set of potential fishing areas but with higher false-positive. (3) Spatiotemporal scale significantly affects model performance. Larger spatial scales (1° × 1°) tend to improve location prediction, whereas finer scales (0.5° × 0.5°) provide more accurate intensity estimation. The optimal scale for both schemes is 1°-30 days. (4) Variable importance analysis shows that historical FE data dominate predictions in Scheme 1, while sea surface temperature and chlorophyll-a are the most influential predictors in Scheme 2.

In summary, this study offers new avenues for predicting future spatial patterns of global fisheries activities and assessing the impacts of climate change. Future research could benefit from integrating longer time series, additional environmental variables, and socio-economic drivers to further improve prediction accuracy and enhance the practical applicability of the model.