Deep Learning and Survival Analysis Reveal Foraging-Driven Habitat Use in Pacific Saury Fisheries

Abstract

Understanding the alignment between fisher behavior and habitat dynamics is essential for data-driven fisheries management. This study analyzed high-resolution Automatic Identification System (AIS) and Vessel Monitoring System (VMS) data, integrated with logbooks from 10 stick-held dipnet vessels targeting Pacific saury (Cololabis saira) in the North Pacific high seas. We developed an optimized CNN-LSTM-SE model to classify vessel trajectories into eight operational states, achieving 91% accuracy. This model generated a high-confidence presence dataset, addressing spatiotemporal data limitations in pelagic species research. A dynamic Ensemble Species Distribution Model (ESDM) mapped habitat suitability index (HSI) for the primary fishing seasons (June–September) of 2023–2024, revealing seasonal northward migrations and an interannual eastward shift in core habitats, primarily driven by sea surface temperature (SST: 6.4–19.1 °C), chlorophyll-a (CHL: 0.2–2.0 mg/m³), mixed layer depth (MLD: 14–30 m), and dissolved oxygen (DO: 220–290 mmol/m³). Receiver operating characteristic (ROC) sensitivity analysis identified an HSI threshold of ≥0.4 for suitable habitats, where 98.4% of fishing effort was concentrated. Kaplan–Meier survival analysis demonstrated that vessels in high-quality habitats (HSI ≥ 0.8) exhibited significantly longer fishing bout durations and lower cessation probabilities (log-rank test, χ² = 20.9, p < 0.001), providing empirical evidence for the Marginal Value Theorem and Optimal Foraging Theory. Although HSI showed a weak direct correlation with catch rates (R² = 0.007), it effectively delineated high-potential fishing grounds (>90% of high-catch days > 30 tonnes in HSI ≥ 0.6). By demonstrating that fishers’ spatial decisions appear to reflect environmental gradients, suggesting that fishing effort may indirectly act as an ecological indicator, this integrated framework bridges fisher behavior with ecological theory, supporting dynamic ocean management in climate-variable fisheries.

1. Introduction

Habitat suitability is a core ecological factor shaping the distribution, abundance, and dynamics of marine species, underpinning resource assessment and biodiversity conservation [1]. Under global climate change, predicting the spatiotemporal evolution of marine habitats poses a significant challenge for ecology and fisheries science [2]. The North Pacific, a highly productive region driven by dynamic oceanographic processes (e.g., the Kuroshio-Oyashio confluence), supports rich biological resources but is sensitive to climate-driven events like marine heatwaves, which profoundly impact key species [3,4]. Pacific saury (Cololabis saira), a pivotal mid-pelagic species in this ecosystem, sustains a major transboundary fishery and serves as prey for top predators. However, its abundance and distribution have exhibited marked fluctuations in recent years [5,6,7,8], introducing substantial uncertainty into fisheries management.

A critical barrier to studying saury habitat dynamics is acquiring reliable species distribution data. Fishery-independent surveys, while ideal, are costly and limited in spatiotemporal coverage for wide-ranging pelagic species like saury [9]. Fishery-dependent data, such as logbooks, offer broader coverage but suffer from biases, low spatial resolution, and restricted access due to data-sharing barriers among fishing nations (e.g., China and Japan) [10]. High-resolution vessel tracking data from Automatic Identification System (AIS) and Vessel Monitoring System (VMS) address these gaps, but extracting accurate fishing events from complex trajectories remains challenging. Traditional threshold-based methods, relying on speed or heading, often misclassify effort due to nuanced operational patterns [11,12].

Recent advances in artificial intelligence, particularly deep learning, provide a solution. Hybrid models combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) excel at capturing spatial and temporal dependencies in trajectories, improving fishing event classification [13,14,15]. Building on recent advances, including the use of attention-based models for behavior recognition and fishing-ground prediction [13,16,17], our study integrates a Squeeze-and-Excitation (SE) attention mechanism to further enhance feature expression. The SE module applies adaptive channel-wise attention weighting [18,19], which effectively emphasizes important feature channels and suppresses irrelevant information, thereby heightening the model’s precision in identifying key behaviors. This integrated framework leverages a dominant commercial fishing fleet as a network of opportunistic ecological indicators. Our analysis uses high-resolution trajectory data from 10 Chinese stick-held dipnet vessels, which account for 30–37% of the total high-seas catch [20]. The framework introduces two key innovations: (1) an Ensemble Species Distribution Model (ESDM) for robust habitat mapping [21], and (2) the application of Kaplan–Meier survival analysis to quantify fishers’ patch residence times, directly linking their “stay-leave” decisions to Optimal Foraging Theory (OFT) [22,23,24,25].

To analyze the relationship between Pacific saury habitat dynamics and fishing vessel behavior, this study pursues three specific objectives: (1) Use an optimized CNN-LSTM-SE model to identify fishing events from multi-source data, constructing a high-resolution presence dataset. (2) Develop and validate a dynamic ESDM to characterize spatiotemporal habitat evolution in 2023–2024, identifying key environmental drivers. (3) Apply Kaplan–Meier analysis to examine fishing activity persistence across habitat quality levels, providing empirical support for the OFT hypothesis in marine fishery ecosystems.

2. Materials and Methods

2.1. Data and Preprocessing

2.1.1. Fishery Data

This study analyzed operational and environmental conditions associated with the Chinese distant-water stick-held dipnet fleet targeting Pacific saury during the 2023–2024 fishing seasons (June–September) in the high-seas region of the North Pacific (35° N–50° N, 145° E–172° E). The fleet contributed substantially to the fishery (30–37% of high-seas catch [20]), operated under tightly regulated seasonal windows, and exhibited strong spatial cohesion; thus, the trajectories of sampled vessels provided a reliable representation of overall fleet behavior.

Vessel trajectories were integrated from Automatic Identification System (AIS) and Vessel Monitoring System (VMS) records of 10 representative vessels. These vessels were specifically selected based on two key criteria: (1) the high quality and standardization of their fishing logbooks, which provided reliable ground-truth data for model validation, and (2) the high homogeneity of their key technical parameters. All 10 vessels operated with identical 1800 kw luring light power and were highly similar in tonnage (1655–1687 t) and length (70.87–72.65 m). This homogeneity ensured comparable fishing power across the sample, making it an ideal proxy for the fleet’s operational patterns.

All raw AIS/VMS data underwent systematic preprocessing. Duplicate observations, incomplete timestamps, and corrupted coordinate reports were removed. For timestamp gaps shorter than 1 h, linear interpolation was applied to latitude, longitude, speed, and heading to preserve short-range behavioral continuity while avoiding artificial smoothing of transitions such as ‘Searching → Decelerating → Fishing.’ Gaps exceeding 1 h retained missing dynamic attributes to prevent incorrect inference of vessel behavior. Unrealistic position jumps corresponding to speeds > 16 kn—far exceeding typical transit speeds of this fleet—were identified as drift errors and removed.

Daily logbooks provided fishing start/end times, haul periods, and total catch (t). Records missing time fields, containing inconsistent spatial coordinates (>5 nautical miles from AIS/VMS positions), or exhibiting > 30 min timestamp deviation were removed (~5% of data). Consultation with onboard observers confirmed the alignment between logged fishing periods, gear deployment sequences, and actual operation behaviors, ensuring that logbook-derived events could be used as robust ground-truth labels.

2.1.2. Environmental Data

Environmental variables were derived from monthly Copernicus Marine Environment Monitoring Service (CMEMS) products, aligned to the 2023–2024 fishery periods and resampled to a 0.083° × 0.083° grid. Given Pacific saury’s ecology as a cold-water mid-pelagic species, its distribution correlates strongly with surface temperature and primary productivity [6,26,27,28]. The selected variables were:

Physical Variables: Sea surface temperature (SST), sea surface salinity (SSS), mixed layer depth (MLD), sea surface height (SSH), zonal current (Uo; positive eastward, representing east–west flow component), and meridional current (Vo; positive northward, representing north–south flow component). Positive values indicated eastward/northward advection that enhances nutrient transport and frontal stability, while negative values denoted westward/southward flows that may disperse prey and constrain saury habitats. For analytical convenience, ocean current velocity (CV)—a composite measure of hydrodynamic stability—was derived by combining Uo and Vo into a single vector magnitude (the square root of the sum of their squared values) [29], yielding a non-negative speed that reflected overall current strength while preserving directional influences through the components.

Biochemical Variables: Chlorophyll-a concentration (CHL), a proxy for primary productivity, and Dissolved Oxygen (DO).

MLD inclusion was essential, as it influenced prey aggregation in frontal zones—a key saury distribution driver [25].

To mitigate multicollinearity, predictor variables were screened using Variance Inflation Factor (VIF), retaining only those with VIF < 5. Environmental layers were spatially aligned with fishing event locations and background points for species distribution modeling.

2.1.3. Refined Operational State Classification

Traditional classifications in pelagic fisheries (sailing, fishing, drifting) could not capture the complexity of lift-net vessel behaviors. To accurately represent operational states, we developed an eight-state framework grounded in AIS/VMS data from the ten selected vessels and supported by onboard observer information. The classification integrated multiple dimensions—speed, heading variability, time of day, wave height, and activity context—using empirically derived thresholds (e.g., 95th percentile transit speed). [20]. The eight states are defined as follows (labels 1–8 correspond to operational types), with multi-parameter rules resolving overlaps (e.g., via gear signals or heading irregularity):

(1)

Sailing (State 1): Characterized by high and stable transit speed (10–16 knots, consistent with 95% of AIS transit records from the 10 vessels) and minimal heading fluctuations (±20°). This state primarily occurred during local daytime (06:00 a.m.–14:00 p.m.) for long-distance travel to target fishing grounds, with an average daily displacement exceeding 50 nautical miles (WGS84 coordinate system, 1 nautical mile ≈ 1.852 km). Brief nighttime sailing may occur if the local fish abundance is insufficient, as confirmed by observer logs.

(2)

Fish Searching (State 2): Operated at lower speeds (3–10 knots) than sailing, with larger heading variations (±50°) due to active acoustic detection (fish finders) for saury schools. It mainly occurs between 12:00 a.m. and 16:00 p.m. local time (pre-fishing preparation window) but persists throughout the fishing process. Sudden turns in trajectories were a key identifier, reflecting real-time adjustments to potential prey patches.

(3)

Decelerating (State 3): Triggered by fish school detection, with speed gradually reduced to 3–6 knots and more extreme heading fluctuations (>60°) than fish searching. Most common during 16:00 p.m.–18:00 p.m. (pre-fishing positioning) but may continue during fishing. The narrow activity range (≤5 nautical miles) supports fine-tuning of vessel position for subsequent fish aggregation (distinguished from Positioning Adjustment by larger heading changes and absence of gear deployment).

(4)

Drifting (State 4): Low-speed operation (0–1.0 knots) with minor heading changes, distinguishable from other low-speed states by context: nighttime drifting (18:00 p.m.–22:00 p.m.) uses full fish-luring lights (observer-verified) to attract schools, with duration adjusted by fish density; daytime drifting (06:00 a.m.–16:00 p.m.) occurs for crew rest or transshipment (no luring lights/gear activity).

(5)

Fishing (State 5): The core operational state, occurring at night (19:00 p.m.–06:00 a.m.) with net deployment/hauling alongside the vessel. Speed ranges from 0.1 to 2.6 knots (with occasional 2.0–2.8 knots during single net hauls, per AIS trajectory statistics) and heading fluctuates widely (0–360°) due to gear manipulation. This state is the primary target for fishing event identification in Section 2.2 (distinguished from Drifting by active gear signals and irregular low-speed maneuvers; resolves 20:00 p.m.–22:00 p.m. overlap via light and net activity and logbook-confirmed non-zero catch).

(6)

Weather-Avoidance Sailing (State 6): Activated when wave heights exceed 2.5 m (consistent with NPFC safety protocols), with speed gradually increasing from 2 knots to speeds exceeding 10 knots. Trajectories appear as straight or segmented lines, reflecting active relocation to calmer waters.

(7)

Weather-Avoidance Drifting (State 7): Also triggered by wave heights > 2.5 m, but vessels remain stationary with speed < 2 knots (no relocation), as recorded in observer logs during moderate storm events.

(8)

Positioning Adjustment (State 8): Encompassed transitional activities (fish school relocation, net deployment/retrieval, bow direction adjustment, light configuration modification) with speed stabilized at 3–5 knots. It was distinguished from other mid-speed states (e.g., Decelerating) by minimal heading fluctuations (<30°) and short duration (<1 h per event).

These refinements provided unambiguous, data-driven labels for the multi-class deep learning model in Section 2.2, reducing misclassification risks.

2.2. Identification of Fishing Events: Threshold vs. Deep Learning

Having defined the eight operational states in Section 2.1.3, a robust method was required to automatically classify these states. Therefore, to achieve the classification accuracy needed for our analysis, we employed a hybrid deep learning framework (Section 2.2.2).

2.2.1. Threshold Method (Baseline)

A traditional rule-based method was implemented as a benchmark: any AIS record with speed < 2.6 knots during nighttime (19:00 p.m.–06:00 a.m. local time) was classified as a potential fishing event. This threshold aligns with previous pelagic studies [11,12], but could not separate fishing from drifting or weather-avoidance states, necessitating a more refined approach.

2.2.2. Deep Learning Model: CNN-LSTM-SE Architecture

This study employed a hybrid deep learning framework integrating Convolutional Neural Networks (CNN), Long Short-Term Memory networks (LSTM), and a Squeeze-and-Excitation (SE) attention mechanism to classify vessel operational states and extract fishing events with high accuracy. The model was implemented in Python (version 3.10) using TensorFlow/Keras, and was designed to overcome the limitations of threshold-based methods by capturing both short-term kinematic patterns and long-range temporal dependencies within AIS/VMS trajectory sequences. The model distinguished eight operational states, with particular emphasis on separating true fishing activity (State 5) from overlapping low-speed behaviors such as drifting and weather-avoidance drifting [15,30]. The input to the model consisted of 100-step trajectory segments, each containing 11 features that represented a combination of instantaneous kinematic descriptors and derived behavioral indicators. The overall architecture of the proposed CNN-LSTM-SE model was illustrated in Figure 1.

The model began with two one-dimensional convolutional layers (kernel size = 3), which extracted local spatial–temporal patterns from the trajectory input. These layers generated a sequence of feature maps designed to capture fine-scale kinematic variations, such as speed-change gradients and heading fluctuation frequencies that typically precede operational transitions. The convolutional output was then passed into an SE attention block, which performed channel-wise feature recalibration [18,31]. Through global average pooling followed by two fully connected layers with non-linear (ReLU) and sigmoid activations, the SE module selectively amplified feature channels associated with characteristic fishing behaviors (e.g., irregular low-speed movement combined with high heading variability), while suppressing channels dominated by non-informative or noise-related patterns. The enhanced feature sequence entered a two-layer LSTM module, which modeled long-range temporal dependencies within the 100-step window. This step captured behavioral progressions such as Searching → Decelerating → Fishing, which could not be accurately characterized by instantaneous parameters alone. Each LSTM layer output a temporally contextualized representation of feature evolution, enabling the model to distinguish between behaviorally similar but functionally distinct operational states. A multi-head attention layer was then applied to further refine temporal dependencies, allowing the network to assign higher weights to the most influential timesteps within the sequence for identifying fishing activity. The attention output was normalized using a residual Add + Layer Normalization structure to stabilize training and prevent gradient degradation. A Global Average Pooling 1D layer compressed the temporally weighted feature sequence into a 128-dimensional vector representing the overall behavioral signature of the segment. This vector was then passed through a fully connected classification layer with Softmax activation to output the probability distribution across the eight operational states.

The model was trained using data from 10 representative vessels. A stratified training strategy (70% training, 20% validation, 10% testing) was adopted to reduce inter-vessel bias and enhance generalization. The Adam optimizer and categorical cross-entropy loss were used, with training conducted for up to 150 epochs and early stopping based on validation loss (patience = 15). Model performance was evaluated using accuracy, precision, recall, and F1-score metrics, enabling a robust assessment of predictive capability and generalization to unseen trajectory data. This architecture ultimately supported reliable identification of fishing events and provided a foundation for subsequent habitat suitability and fishing strategy analyses [29].

2.3. Ensemble Habitat Suitability Modeling (ESDM)

Once the locations and times of vessel fishing activity were identified (via the State 5 output from the CNN-LSTM-SE model, Section 2.2), the next step was to quantify the ecological quality of those locations. To do this, we developed an Ensemble Species Distribution Model (ESDM). This model used the high-confidence ‘fishing’ events identified in Section 2.2 as the foundational ‘presence data’.

To mitigate algorithmic uncertainty and enhance predictive robustness for Pacific saury habitat mapping, an Ensemble Species Distribution Model (ESDM) was implemented using the biomod2 package (v4.2.6.2) in R (v4.4.1) [21]. This framework aggregated predictions from multiple algorithms to model non-linear species-environment relationships, drawing on high-confidence presence data from verified fishing events and environmental covariates.

Background (pseudo-absence) points were generated using the Surface Range Envelope (SRE) approach, which restricted background selection to the environmental envelope of observed presences [29,32], preventing unrealistic extrapolation. The background-to-presence ratio was set to 5:1 (capped at 5000).

Single models included GLM, GAM, CART, Random Forest (RF), and Boosted Regression Trees (BRT). Models were fitted monthly to capture seasonal habitat variability [33].

Model performance was evaluated with 5-fold cross-validation using AUC and True Skill Statistic (TSS) [34]. Only models exceeding TSS > 0.8 (top ~70% of outputs) were retained for ensembling, ensuring high discriminatory power for habitat prediction.

Monthly ensemble predictions were generated using a TSS-weighted average of retained models, rescaled to a standardized Habitat Suitability Index (HSI) ranging from 0 (unsuitable) to 1 (highly suitable). The resulting monthly HSI rasters captured broad-scale environmental dynamics and formed the basis for subsequent analysis of fishing strategies and behavioral persistence.

2.4. Fishing Strategy and Behavioral Response Analysis

This final section integrated the outputs from the preceding steps to perform the central statistical analysis. We used the ‘fishing bouts,’ precisely delineated by the CNN-LSTM-SE model (Section 2.2), and spatially joined them with the monthly Habitat Suitability Index (HSI) rasters generated by the ESDM (Section 2.3).

2.4.1. Fishing Bout Definition and Aggregation

Fishing State (State 5) outputs from the CNN-LSTM-SE classifier were aggregated into continuous “fishing bouts.” A bout was defined as a continuous period of fishing behavior, ending when the interval between consecutive fishing points exceeded 2 h. This threshold was validated using fishing logbooks, where >95% of logbook-recorded hauls matched the same bout segmentation, confirming consistency with operational practice.

2.4.2. Habitat Classification

HSI values were categorized into four habitat-quality levels to facilitate ecological interpretation and management relevance:

• Non suitable: HSI < 0.4
• Low suitability: 0.4 ≤ HSI < 0.6
• Moderate suitability: 0.6 ≤ HSI < 0.8
• High suitability: HSI ≥ 0.8 [35].

These thresholds were determined through ROC-based sensitivity analysis and reflect ecological preferences for productive frontal zones. Fishing bouts were spatially assigned to habitat classes based on their centroid positions within monthly HSI maps [36,37].

2.4.3. Bout Duration Analysis

To evaluate how habitat quality influences fishing persistence, bout durations were compared across the four habitat categories [15,17,38]. Normality was tested using the Shapiro–Wilk test. When data met normality assumptions, one-way ANOVA with Tukey’s HSD was applied; otherwise, the Kruskal–Wallis test with Dunn’s post hoc comparisons was used. This analysis assessed whether fishers tend to remain longer in higher-quality habitats, consistent with optimal foraging theory expectations [39,40].

2.4.4. Survival Analysis of Fishing Persistence

To quantify fishers’ “stay-leave” decisions as proxies for patch residence time in Optimal Foraging Theory [41], Kaplan–Meier (KM) survival analysis was applied [22,23]. Fishing bouts were defined as survival objects: bout duration was “survival time,” and transition from State 5 (fishing) to any non-fishing state was treated as the “event” (i.e., termination of fishing persistence). KM curves plotted the probability of persistence across habitat levels, with flatter curves indicating lower cessation risk. Weighted log-rank tests compared curves across the four levels, testing the null hypothesis (H₀) that fishing duration was independent of habitat quality (p < 0.05). Analyses were conducted using Python’s lifelines package.

$$E_{ij}=O_j\times {\displaystyle \frac{N_{ij}}{N_j}}$$

(1)

$$O_i={\displaystyle \sum _j{O}_{ij}}$$

(2)

$$E_i={\displaystyle \sum _j{E}_{ij}}$$

(3)

$$X^2={\displaystyle \sum _{i=1}^k\frac{{\left(O_i-E_i\right)}^2}{E_i}}$$

(4)

Define related variables as follows: let k denote the total number of groups (in this study, k = 4); i represent a specific group (where i = 1, …, k); and j denote a specific time point at which an event (fishing cessation) occurs; let O_ij be the number of observed events in group i at time j; N_ij denote the number of at-risk individuals (vessels still engaged in fishing) in group i prior to time j; O_j and N_j, respectively, represent the total number of events and at-risk individuals across all groups at time j. At each event time point j, compute the expected number of events E_ij for group i (Equation (1)). Next, sum the observed and expected values across all event time points to obtain the total observed events O_i (Equation (2)) and total expected events E_i (Equation (3)) for each group i. Finally, calculate the total chi-square statistic χ² (Equation (4)) by aggregating the weighted squared differences between observed and expected values over all groups. This statistic followed a chi-square distribution with k − 1 degrees of freedom, through which the p-value can be calculated for the final hypothesis test.

3. Results

3.1. CNN-LSTM-SE Model Performance and Validation of Fishing Effort Metrics

3.1.1. Model Performance Validation

The multi-class deep learning model demonstrated robust performance in classifying the eight operational states, achieving an average 5-fold cross-validation accuracy of 90.83% (SD = 0.0077) across data subsets, with a test set accuracy of 91%. The weighted F1-score reached 0.91, while the macro F1-score was 0.90, indicating balanced recognition across imbalanced classes without bias toward majority states. Overall, the model classified the preprocessed trajectories into the following proportions: Sailing (20.4%), Fish Searching (7.6%), Decelerating (6.0%), Drifting (35.5%), Fishing (16.6%), Weather-Avoidance Sailing (1.4%), Weather-Avoidance Drifting (5.4%), and Positioning Adjustment (7.0%). This distribution, ranging from 1.4% to 35.5%, confirmed that the training data did not exhibit extreme class imbalance. Training and validation losses converged stably (Figure 2A,B; accuracy rising from ~0.6 to 0.91, losses falling from ~1.2 to 0.15/0.18), with early stopping triggered after 85 epochs (patience = 15), confirming no overfitting as per the protocol in Section 2.2.

This performance directly reflected the architecture’s design tailored to trajectory nuances: the CNN layer effectively extracted local temporal patterns (e.g., heading fluctuations in State 5 fishing), LSTM captured temporal transitions (e.g., from State 2 searching to 3 decelerating), and SE attention re-weighted low-speed + high-turn features, reducing confusion between overlapping states like drifting (State 4) and fishing (State 5). Detailed class metrics showed near-perfect F1 scores for State 1 (Sailing/Transiting, 0.96) and State 7 (Weather-Avoidance Drifting/Storm-drifting, 0.96), reliable performance for States 2–4 (F1 ≥ 0.90), and a relative weakness in State 8 (Positioning Adjustment/Repositioning, precision = 0.72)—likely due to its transitional nature and lower sample prevalence (4.6% of trajectories). Critically, for fishing events (State 5), the model achieved precision = 0.92, recall = 0.89, and F1 = 0.90, ensuring high-confidence presence points for habitat modeling with minimal false negatives (~11% missed true events).

3.1.2. Validation of Fishing Effort Metrics

Model outputs were validated against logbook ground truth, revealing superior alignment compared to the threshold baseline. In temporal allocation (Figure 3A: state proportions), the DL model estimated fishing (State 5) at 16.6% of time versus 20.8% for threshold method—a 4.2% reduction in overestimation, primarily from distinguishing drifting (State 4, 35.5% vs. 31.6%) via multi-feature rules. The confusion matrix on the test set quantified this: threshold misclassified 18.2% of State 4 as State 5, while DL reduced it to 5.3%—yielding 30% fewer false positives overall and confirming logbook alignment in fishing durations.

Monthly fishing days (Figure 3B: monthly days per vessel) further confirmed fidelity: DL deviations averaged 1 day (e.g., June: DL = 18, logbook = 17; September: DL = 12, logbook = 13), within statistical error for the sample size. Threshold deviations exceeded 3 days in all months (e.g., July: 5 days), underscoring the DL model’s nuanced handling of low-speed ambiguities. These metrics validated the 8-state granularity’s role in reconstructing operational reality, providing reliable effort proxies for spatiotemporal habitat analysis.

3.1.3. Relationship Between Fishing Duration and Catch

Linear regression revealed a significant positive correlation between fishing duration and daily catch across sources and seasons (all p < 0.05; Figure 4), but with low explanatory power (R² = 0.04–0.09), consistent with Pacific saury’s epipelagic schooling ecology: catch depended more on encounter rates in high-density patches than prolonged effort, as short, targeted hauls yielded high returns once schools were located, while extended searching in low-density areas yields diminishing marginal gains [6,26]. Off-peak season (August–September) showed stronger association for logbooks (R² = 0.21, coefficient = 0.3 tonnes/hour, 95% CI [0.2, 0.4]), reflecting dispersed schools requiring longer persistence for viable yields; peak season (June–July) distributions were more scattered (R² = 0.04 for DL, coefficient = 0.1 tonnes/hour, 95% CI [0.05, 0.15]), as concentrated schools allowed efficient, shorter bouts. DL and threshold durations yielded similar weak R² (0.04), but DL’s alignment with logbooks (>85% of >5 h bouts linked to high catch > 30 tonnes) underscored its utility for effort reconstruction despite definitional nuances (State 5 excludes prep time present in logbooks).

3.2. Spatiotemporal Dynamics of Pacific Saury’s Suitable Habitat and Its Environmental Drivers

The ESDM generated monthly HSI rasters revealing pronounced spatiotemporal habitat evolution for Pacific saury, with core areas (HSI ≥ 0.8) exhibiting seasonal northward expansion followed by contraction, modulated by interannual variability in environmental drivers (Figure 5). This pattern aligned with the model’s integration of multiple algorithms and covariates, where VIF < 5 ensured multicollinearity control and robust predictions.

3.2.1. Spatiotemporal Evolution of Suitable Habitats

In 2023, the spatiotemporal pattern of saury suitable habitats showed a typical seasonal trajectory (Figure 5a–d): in June (Figure 5a), the core habitat (HSI ≥ 0.8) formed a broad band distributed between 41–45° N and 160–175° E; in July (Figure 5b), the core habitat moved northward to reach its northernmost position, spanning 45–48° N and 163–170° E with high aggregation intensity; in August (Figure 5c), the core habitat contracted toward the southwest; by September (Figure 5d), it had retreated to the west of 160° E.

In 2024, the habitat trajectory generally mirrored the 2023 pattern of “northward expansion followed by retreat” but exhibited a notable eastward shift (with the longitude centroid at a mean of 165° E in 2023 versus 168° E in 2024; Figure 5e–h): in June (Figure 5e), the core habitat started at a more eastern location compared to 2023; in July (Figure 5f), the core habitat showed a dispersed distribution; from August to September (Figure 5g,h), the core habitat persisted within the range of 160–175° E and did not exhibit a pronounced westward migration.

3.2.2. Key Environmental Drivers of Habitat Dynamics

The variable importance analysis of the Ensemble Species Distribution Model (ESDM) (Figure 6) revealed that the key environmental factors driving Pacific saury habitat dynamics exhibited significant monthly and interannual variations. Sea Surface Temperature (SST) was a consistently important driver, often accounting for the largest share of the total relative importance, peaking at 44.7% in July 2024. However, the model results uncovered a clear pattern of alternation among the dominant environmental factors across different periods. In specific months, other variables surged to the top: Mixed Layer Depth (MLD) was prominent in both July 2023 and July 2024, with relative importances of 31.4% and 29.6%, respectively. Dissolved Oxygen (DO) saw its influence increase during the 2024 summer, becoming the primary factor in June and August with importances of 31.0% and 30.4%. Furthermore, Sea Surface Salinity (SSS) experienced a significant surge to 21.8% in September 2024, indicating a marked shift in the late-season drivers. In summary, while SST provided a constant baseline influence, the habitat drivers dynamically alternated between MLD, DO, and SSS, highlighting the complex and seasonal nature of the saury’s habitat controls.

3.2.3. Response Curves and Optimal Ranges

Response curves (Figure 7) and optimal environmental ranges (

Table 1. Optimal ranges of environmental variables driving habitat dynamics across monthly and interannual scales.

Month	Variable	Peak HSI	Optimal Value (Range)	Unit
2023-06	SST	0.322	6.7(3.1~9.2)	(°C)
2023-06	SSS	0.3649	32.5 (32.0~32.9)	(‰)
2023-06	SSH	0.3791	1.1 (0.3~1.1)	(m)
2023-06	MLD	0.4247	28.4 (20.2~36.6)	(m)
2023-06	DO	0.3118	211.3 (211.3~398.4)	(mmol/m3)
2023-06	CV	0.4486	0.9 (−0.6~0.9)	(m/s)
2023-06	CHL	0.8424	1.3 (0.6~2.0)	(mg/m3)
2023-07	SST	0.5392	10.5 (6.3~10.5)	(°C)
2023-07	SSS	0.4523	34.7 (33.9~35.4)	(‰)
2023-07	SSH	0.2287	−0.3 (−0.3~−0.2)	(m)
2023-07	MLD	0.3477	14.3 (13.7~14.9)	(m)
2023-07	DO	0.4248	229.3 (202.1~256.4)	(mmol/m3)
2023-07	CV	0.1307	−0.8 (−0.8~0.9)	(m/s)
2023-07	CHL	0.2199	0.5 (0.4~0.6)	(mg/m3)
2023-08	SST	0.8554	19.1 (17.9~20.3)	(°C)
2023-08	SSS	0.8581	33.6 (32.0~35.3)	(‰)
2023-08	SSH	1.0431	−0.2 (−0.4~−0.0)	(m)
2023-08	MLD	0.8605	14.6 (12.7~16.5)	(m)
2023-08	DO	0.8729	277.8 (228.5~327.1)	(mmol/m3)
2023-08	CV	0.8919	0.4 (−0.2~0.9)	(m/s)
2023-08	CHL	0.8571	1.1 (0.1~2.0)	(mg/m3)
2023-09	SST	0.8384	22.3 (19.7~26.9)	(°C)
2023-09	SSS	0.6281	32.6 (32.0~33.2)	(‰)
2023-09	SSH	0.8704	−0.1 (−0.4~−0.1)	(m)
2023-09	MLD	0.8041	15.9 (13.6~18.2)	(m)
2023-09	DO	0.6987	288.4 (250.0~326.7)	(mmol/m3)
2023-09	CV	0.5863	0.3 (−0.3~0.8)	(m/s)
2023-09	CHL	0.6556	0.4 (0.1~0.4)	(mg/m3)
2024-06	SST	0.6783	6.4 (5.5~7.4)	(°C)
2024-06	SSS	0.4283	34.4 (33.6~35.2)	(‰)
2024-06	SSH	0.4274	1.2 (−0.4~1.2)	(m)
2024-06	MLD	0.4104	29.9 (24.8~35.0)	(m)
2024-06	DO	0.4537	251.3 (210.2~292.5)	(mmol/m3)
2024-06	CV	0.397	0.5 (0.5~0.9)	(m/s)
2024-06	CHL	0.5876	0.6 (0.4~0.7)	(mg/m3)
2024-07	SST	0.3539	7.7 (5.5~9.1)	(°C)
2024-07	SSS	0.1197	35.4 (32.0~35.4)	(‰)
2024-07	SSH	0.2892	1.2 (1.2~1.4)	(m)
2024-07	MLD	0.2993	16.4 (15.6~17.2)	(m)
2024-07	DO	0.1906	243.0 (200.8~243.0)	(mmol/m3)
2024-07	CV	0.1347	0.4 (0.4~0.9)	(m/s)
2024-07	CHL	0.0816	0.1 (0.1~0.1)	(mg/m3)
2024-08	SST	0.8839	16.8 (14.9~18.6)	(°C)
2024-08	SSS	0.9067	34.4 (32.0~34.4)	(‰)
2024-08	SSH	0.9248	−0.2 (−0.4~0.1)	(m)
2024-08	MLD	0.9532	13.6 (13.6~23.1)	(m)
2024-08	DO	0.9641	223.5 (197.0~250.0)	(mmol/m3)
2024-08	CV	0.8677	0.2 (−0.5~0.9)	(m/s)
2024-08	CHL	0.9193	1.1 (0.2~2.0)	(mg/m3)
2024-09	SST	0.5338	18.2 (18.2~25.5)	(°C)
2024-09	SSS	0.5872	32.2 (32.0~32.5)	(‰)
2024-09	SSH	0.5295	−0.2 (−0.4~−0.2)	(m)
2024-09	MLD	0.1572	26.3 (17.9~34.7)	(m)
2024-09	DO	0.1793	285.5 (262.7~328.4)	(mmol/m3)
2024-09	CV	0.3749	−0.4 (−1.0~−0.4)	(m/s)
2024-09	CHL	0.3983	0.8 (0.3~1.4)	(mg/m3)

) revealed significant plasticity in Pacific saury’s habitat preferences, which varied on both monthly and interannual scales. This dynamic was particularly evident for sea surface temperature (SST), where the optimal range shifted from a cool 6.4–10.5 °C during the early-season northward foraging migration (June–July) to a warmer 16.8–19.1 °C in late-summer high-latitude feeding grounds (August). Similarly, the optimal range for chlorophyll-a (CHL), a proxy for prey availability, showed high variability. It narrowed significantly in some months (e.g., 0.4–0.6 mg/m³ in July 2023), suggesting precise prey concentration requirements, while widening in others (e.g., 0.2–2.0 mg/m³ in August 2024), likely indicating that other factors like SST became more limiting when prey was abundant. Despite this seasonal flexibility, consistent preferences were observed for a shallow mixed layer depth (MLD) of 14–30 m, which facilitates prey aggregation in frontal zones, and moderately high dissolved oxygen (DO) levels of 220–290 mmol/m³, aligning with the physiological tolerances of mid-pelagic species [6]. Collectively, these dynamic, non-monotonic responses underscore the species’ adaptability to shifting oceanographic conditions and validate the inclusion of non-linear models in the ESDM framework.

3.3. Fishing Strategy as a Response to Habitat Quality

Spatial analysis of fishing effort (Figure 8A) revealed strong coupling with ESDM-predicted habitat suitability: 98.4% of total time in suitable habitats (HSI ≥ 0.4), with 77.8% in highly suitable (HSI ≥ 0.8)—over 90% of high-catch days (>30 tonnes/day) in HSI ≥ 0.6. This validates ESDM discrimination and sets up behavioral metrics.

3.3.1. Fishing Effort Allocation and Bout Duration

Bout durations further revealed habitat influence (Figure 8B; n = 1805 bouts). One-way ANOVA confirmed significant differences (F = 6.63, p = 0.0002). Means increased with quality: highly suitable (3.85 h) > moderate (3.50 h) > low (2.59 h) > non-suitable (2.50 h). Tukey’s HSD post hoc identified key contrasts: highly vs. non-suitable (meandiff = −1.35 h [i.e., highly > non-suitable by 1.35 h], p = 0.0094); highly vs. low-suitable (meandiff = −1.26 h [i.e., highly > low-suitable by 1.26 h], p = 0.0338)—other pairs nonsignificant (e.g., highly vs. moderate p = 0.1242; all p > 0.05).

3.3.2. Fishing Persistence: Kaplan–Meier Survival Analysis

Kaplan–Meier (KM) survival curves quantified persistence probability (probability of continued fishing) across habitat levels, based on 1805 bouts (non-suitable n = 15; low-suitable n = 26; moderate n = 387; high n = 1377) and defining bout termination (transition to non-fishing states 1–4 or 6–8) as the survival event (Figure 9). Despite natural sample imbalance reflecting effort bias toward high-HSI zones, curves showed clear gradients: in highly suitable habitats, 50% of bouts persisted > 5 h with gradual decline, signaling low abandonment risk in optimal patches; in marginal habitats (low/non-suitable), <20% exceeded 5 h, with sharp drops after ~2 h—indicating rapid shifts to better areas.

These gradients provide empirical support for the Marginal Value Theorem: vessels maximize foraging efficiency by extending persistence in high-quality habitats and minimizing time in low-quality ones (to reduce opportunity costs), bridging environmental dynamics to adaptive decisions.

3.3.3. HSI Correlation with Catch and Integrated Visualization

Linear regression confirmed a significant positive correlation between HSI and daily catch (p < 0.001; Figure 10), albeit with weak explanatory power (R² = 0.007). This weak explanatory power may be influenced by changes in resource abundance, reflecting a “constraining wedge” pattern: low HSI (<0.4) limited catches to <10 tonnes (environmental filter); high HSI (≥0.6) captured >90% high-yield events (>30 tonnes), but without guaranteeing yields—consistent with limiting factor principle, where HSI envelopes suitability rather than predicts output.

Three-dimensional scatter plots (Figure 11) synthesized environmental drivers, behavioral metrics, and catch outcomes: >75% of bouts (n = 1805) and catch records (n = 1578) aggregated in an optimal niche (HSI > 0.8, mean = 0.83; SST 10.5–18.1 °C; CHL 0.3–0.4 mg/m³), with larger, yellower bubbles denoting extended bout durations (mean = 3.85 h) or elevated yields (mean = 9.60 tonnes/day). This visualization empirically links high HSI to prolonged effort and higher potential returns, consistent with the limiting factor principle. Overall, these patterns underscore HSI’s value for hotspot identification and behavioral forecasting, integrating spatiotemporal dynamics with adaptive foraging strategies.

4. Discussion

4.1. Model Accuracy, Data Source Innovation, and Reliability

Understanding the habitat dynamics of Pacific saury demands datasets that integrate ecological realism with fine spatiotemporal resolution—an enduring challenge in pelagic ecosystem research. Conventional survey-based datasets provide valuable environmental context but are temporally sparse, while fishery-dependent data offer temporal continuity but often lack standardized observation protocols [9]. This study bridged that gap by refining large-scale vessel trajectory data into verified operational records through a CNN-LSTM-SE model, capturing behavioral transitions such as searching, drifting, and fishing. The resulting model achieved an overall accuracy exceeding 91% and a TSS consistently above 0.8 across ensemble runs, demonstrating robust predictive capacity. Such performance parallels high-precision ensemble habitat models like those developed for tuna and mackerel in similar oceanic regions [29,38]. By integrating multi-source datasets—including remote-sensing environmental variables, vessel trajectories, and validated fishing outcomes—the framework enhanced both ecological interpretability and statistical reliability.

Validation against daily catch records revealed that while the HSI exhibited a limited direct correlation with instantaneous yield (R² = 0.007), it effectively constrained the environmental boundary within which high catch rates occurred—a manifestation of the “limiting factor” principle in ecological modeling [1]. The wedge-shaped pattern between catch and HSI observed here aligns with habitat-based assessments for chub mackerel (Scomber japonicus), where high habitat suitability delineates potential fishing hotspots rather than guaranteeing high yields [29]. This weak correlation is ecologically expected: our monthly HSI model defines the large-scale potential for good fishing grounds, but instantaneous catch rates (CPUE) are ultimately determined by finer-scale, unmodeled factors such as the real-time presence of local prey aggregations, inter-vessel competition, or the individual fisher’s skill. Collectively, the integration of behavioral modeling, ensemble forecasting, and environmental diagnostics forms a coherent framework that could be extended to other pelagic fisheries across dynamic ocean regions.

It is noteworthy that our analysis is based on data from the Chinese distant-water stick-held dip net fleet, a dominant contributor to the fishery responsible for 30–37% of the total high seas catch in the North Pacific [20]. While our study utilizes data from 10 of these vessels, we acknowledge that this sample size may limit the full spatial and behavioral representativeness of our results. However, this sample serves as a robust proxy for the fleet’s core operational patterns for two key reasons. First, as justified in Section 2.1.1, the vessels were selected for their high-quality logbooks and high homogeneity in technical parameters (e.g., identical 1800 kw luring light power), ensuring comparable fishing power. Second, the Pacific saury fishery is characterized by strong inter-fleet “following” dynamics. Vessels from different national fleets or fishing parties actively track competitors and rapidly converge on productive fishing grounds, often within 24 h of a high-catch event being identified. This aggregative behavior implies that the fishing locations of our 10-vessel sample are highly indicative of the core fishing grounds being exploited by the international fleet as a whole.

The use of high-resolution, integrated AIS and VMS trajectory data from this well-defined sample therefore allows for a fine-scale analysis of habitat use and behavioral dynamics. This approach captures short-term responses to environmental changes often missed by traditional survey methods, thus ensuring that our findings are not only ecologically grounded but also highly relevant to the operational realities of this major fishery.

4.2. Ecological Driving Mechanisms of Habitat Spatiotemporal Dynamics

The spatial and temporal evolution of Pacific saury habitats reflects a dynamic equilibrium between temperature, oxygen availability, mixed-layer structure, and primary productivity. The present results show a clear seasonal migration pattern—northward expansion in early summer and southward retreat in autumn—consistent with historical satellite and survey analyses [7,26]. The optimal temperature window identified here (6.4–10.5 °C in early summer and 16.8–19.1 °C in late summer–autumn) agrees with a previously reported empirical range [42], suggesting that saury’s thermal preference is largely stable across years but shifts latitudinally in response to basin-scale anomalies. The 2024 eastward displacement of core habitats toward 160° E–175° E marks a significant deviation from the climatological norm. Similar spatial shifts have been observed during anomalous warm years, particularly under the weakening of the Oyashio intrusion and the northward expansion of the Kuroshio Extension [5,6]. In 2024, a shallower mixed-layer depth (13–23 m) and elevated dissolved oxygen (>250 mmol/m³) suggest enhanced stratification and upper-layer retention of nutrients, fostering localized phytoplankton blooms and prey aggregation. Such environmental configurations are typically associated with delayed southward migration and prolonged feeding activity [6]. Interestingly, the dominant environmental driver exhibited temporal switching: SST dominated habitat distribution in early summer, while MLD and DO became more influential in late summer, and SSS contributed marginally during autumn cooling. This switching reflects adaptive habitat selection under shifting frontal structures and indicates that saury dynamically optimizes its spatial niche according to short-term physical–biological coupling. Compared with species with narrower ecological niches, such as Japanese sardine or anchovy, saury exhibits remarkable flexibility in exploiting transient frontal habitats [43].

When contrasted with Japanese and Korean fleets, the Chinese fleet’s operational footprint reveals broader ecological coverage across the transition zone between the Kuroshio and Oyashio systems. This enhances our capacity to capture habitat variability across multiple oceanographic regimes, providing stronger statistical evidence for cross-frontal migration behavior. The observed eastward habitat core aligns with recent findings that identified a persistent eastward shift in the main fishing grounds after 2015, a change likely driven by increased stratification and a westward contraction of subarctic waters [7]. From a long-term perspective, these trends may signify an ongoing regime transition analogous to the 1990s warm phase, when Pacific decadal oscillation (PDO) shifts altered nutrient transport and frontal stability [5]. If the current warming and stratification trends continue, the Pacific saury’s foraging grounds could increasingly migrate eastward into the central Pacific, potentially modifying the spatial overlap between national fleets and altering future management boundaries.

4.3. Fishing Behavior as Empirical Evidence for Optimal Foraging

Fishing behavior mirrors predator–prey interactions in natural ecosystems. The concentration of more than 98% of fishing activity within suitable habitats (HSI ≥ 0.4) provides direct behavioral validation for the Optimal Foraging Theory (OFT), which posits that foragers preferentially allocate effort to high-reward patches [41,44]. This pattern highlights the capacity of human fishers to align with environmental cues—such as temperature fronts or productivity gradients—and to reflect decision-making processes analogous to ecological optimization.

Kaplan–Meier survival analysis further quantified this adaptive behavior. Vessels operating within high-HSI zones (≥0.8) displayed significantly prolonged residence times, with 50% of fishing events persisting beyond five hours, compared to less than 20% in marginal habitats. This pattern corresponds closely to the Marginal Value Theorem (MVT), suggesting that skippers dynamically adjust operation duration based on diminishing returns [40]. Such alignment between empirical vessel decisions and theoretical ecological models supports the use of behavioral metrics (e.g., patch persistence probability) as proxies for ecosystem productivity. Moreover, the spatial distribution of operation density reveals a behavioral hierarchy: initial searching occurs along the thermal front (SST ≈ 17 °C), while longer-duration fishing concentrates in zones of stable chlorophyll gradients. This demonstrates that fishing effort not only responds to environmental suitability but also reinforces the identification of productive microhabitats. The observed behavioral feedback thus offers a unique real-world analog for testing ecological foraging theories under human-mediated decision frameworks.

4.4. Implications for Dynamic Ocean Management and Future Outlook

The integration of high-resolution vessel trajectories and habitat suitability modeling provides a foundation for Dynamic Ocean Management (DOM), an emerging paradigm emphasizing flexible, data-driven regulation of marine resources. Recent developments in artificial intelligence have further supported this direction, as deep-learning architectures have been increasingly applied in fisheries research for behavior recognition, fishing-ground prediction, and adaptive management [13,16]. These studies highlight the scalability and near real-time potential of neural frameworks for tracking fleet dynamics and identifying productive zones, which could assist in the development of operational monitoring tools.

In this context, the proposed framework demonstrates how habitat suitability modeling and behavioral metrics can be combined to detect shifting high-HSI zones and analyze fleet responses to environmental variability. Comparable methods have been successfully applied in large pelagic fisheries such as tuna [37]. For the Pacific saury fishery, dynamic identification of suitable habitats could provide insights into balancing fishing efficiency with ecological sustainability. The inclusion of behavioral indicators—such as operation duration or mobility range—offers additional information to evaluate how fishing activity adapts to environmental changes in real time.

The observed eastward habitat shift, enhanced stratification, and long-term warming trends suggest that the spatial distribution of suitable habitats for Pacific saury may continue to evolve [45]. Such ecological changes could alter fleet overlap patterns and affect future fishing behavior. While these findings have potential management implications, they should be interpreted cautiously and within the limits of ecological inference. Future studies integrating environmental, behavioral, and socio-economic data could help evaluate how this modeling framework can contribute to adaptive and near real-time management strategies in practice. From a scientific perspective, the present work establishes a reproducible, data-rich framework linking environmental dynamics, species distribution, and human behavioral patterns. This approach can be extended to other pelagic species exhibiting flexible habitat use across dynamic oceanographic regimes.

5. Conclusions

By combining ensemble species distribution modeling with behavioral analysis, this study enhances understanding of how physical conditions and fishing behavior jointly shape habitat utilization in the Pacific saury fishery. The results provide quantitative evidence that behavioral adaptation and environmental forcing interact to determine fishing patterns and habitat persistence. The proposed integrated framework offers a transparent and mechanistic basis that can inform future assessments and support the development of evidence-based, adaptive management approaches for pelagic fisheries, while maintaining an emphasis on ecological inference rather than prescriptive policy conclusions.