The study area is located in the western waters of Sumatra within the eastern Indian Ocean (0.5° N–4° S; 97.5° E–101.2° E; Figure 1), including the Mentawai Islands and the adjacent continental margin of West Sumatra, Indonesia. The selection of the area was based on the presence of a properly documented, economically important tuna fishery with strong, well-characterized oceanographic gradients. These characteristics provided suitable conditions for evaluating the relationships between environmental predictors and tuna catch zones within a GeoAI-based PFZ. Quantitative support for this selection was provided by 2,261 georeferenced yellowfin tuna handline fishing events recorded at the Bungus Ocean Fishing Port during 2022 (Table 1), alongside previous catch and catch per unit effort (CPUE) records showing sustained fishing activity throughout the area (Nurholis et al., 2020; Brown et al., 2023). The data provided showed sustained fishing pressure, supporting this designation with quantitative evidence rather than a purely qualitative description.

The western waters of Sumatra represent a dynamic oceanographic transition zone influenced by several major surface circulation systems. The South Equatorial Current (SEC) and South Equatorial Counter Current (SECC) dominate offshore circulation, while the South Java Coastal Current (SJCC) and the Indonesian Throughflow (ITF) transport warm, low-salinity waters into the eastern Indian Ocean along the western coast of Sumatra (Utari et al., 2019; Rahaman et al., 2020). Interactions among these circulation systems generate significant spatial and temporal variability in SST, salinity, Chl-a, and primary productivity, corresponding to the environmental predictors used in this study and representing key factors influencing prey availability and thermal habitat suitability for yellowfin tuna.

Bathymetric conditions further enhance environmental heterogeneity, as the steep continental slope extending from the Sunda Shelf toward the Wharton Basin promotes upwelling and vertical mixing, increasing nutrient availability and phytoplankton productivity along the continental margin (Dalabehara & Sarma, 2021; Xing et al., 2023). Furthermore, persistent frontal zones and enhanced biological productivity from the interaction between circulation and bathymetry create favorable foraging habitats for yellowfin tuna. These environmental characteristics provide a strong ecological basis for investigating the oceanographic controls on tuna distribution and fishing activity within the study area.

This study adopted a GeoAI-based ML framework to predict PFZs of yellowfin tuna in the western waters of Sumatra, with the overall workflow shown in Figure 2. The method comprised sequential stages including data preprocessing, feature correlation analysis, feature importance assessment, and model optimization to ensure robust predictive performance across spatial datasets. The modeling framework integrated GRRF for feature selection and RF for classification and spatial prediction. Specifically, GRRF was applied to reduce redundancy among correlated oceanographic variables and retain the most informative predictors (Izquierdo-Verdiguier & Zurita-Milla, 2020; Iranzad & Liu, 2025).

The selected features were used in RF algorithm to generate probabilistic spatial predictions of tuna fishing potential. Model performance was evaluated using multiple statistical metrics, including accuracy (ACC), precision (P), recall (R), F1-score, mean absolute error (MAE), and root mean square error (RMSE) (Sahib et al., 2025). The optimized RF model was finally applied to the full dataset to produce spatial probability maps of predicted fishing zones, providing an interpretable, data-driven representation of the nonlinear relationships between oceanographic conditions and tuna fishing patterns.

This study adopted a GeoAI-based ML framework to predict PFZs of yellowfin tuna in the western waters of Sumatra, with the overall workflow shown in Figure 2. The method comprised sequential stages including data preprocessing, feature correlation analysis, feature importance assessment, and model optimization to ensure robust predictive performance across spatial datasets. The modeling framework integrated GRRF for feature selection and RF for classification and spatial prediction. Specifically, GRRF was applied to reduce redundancy among correlated oceanographic variables and retain the most informative predictors (Izquierdo-Verdiguier & Zurita-Milla, 2020; Iranzad & Liu, 2025). The selected features were used in RF algorithm to generate probabilistic spatial predictions of tuna fishing potential. Model performance was evaluated using multiple statistical metrics, including accuracy (ACC), precision (P), recall (R), F1-score, mean absolute error (MAE), and root mean square error (RMSE) (Sahib et al., 2025). The optimized RF model was finally applied to the full dataset to produce spatial probability maps of predicted fishing zones, providing an interpretable, data-driven representation of the nonlinear relationships between oceanographic conditions and tuna fishing patterns.

This study used 2,261 georeferenced fishing zones of yellowfin tuna obtained from the Bungus Ocean Fishing Port (PPS Bungus), West Sumatra, Indonesia, in 2022. Each point represented validated coordinates of handline fishing operations recorded in official port logbooks. Before analysis, records were screened for duplicate coordinates, positions falling outside the study domain or on land, and entries with missing or inconsistent date/position fields. Flagged records were removed, obtaining a cleaned dataset used as the empirical basis for modeling tuna distribution and PFZs. To account for seasonal variability driven by monsoon-related oceanographic processes in the eastern Indian Ocean, the reference data were classified into four climatological seasons, namely DJF (December–February), MAM (March–May), JJA (June–August), and SON (September–November), as well as an annual composite (Table 1). Each fishing record was temporally matched to the corresponding daily Earth observation (EO) composite for recorded date (Section 2.4). Seasonal subsets were formed by pooling the daily-matched records within each monsoonal period. This grouping enabled the model to capture fluctuations associated with monsoon circulation and productivity cycles. For ML implementation, the dataset was divided into training (70%) and validation (30%) subsets using stratified random sampling to maintain balanced spatial and seasonal representation, thereby reducing bias and enhancing model generalization across both seasonal and annual scales (Abdi, 2020; Chabalala et al., 2023). However, the stratified random split did not explicitly account for spatial autocorrelation among nearby fishing points. The fisheries-dependent dataset also showed the spatial distribution of fishing effort, gear type (handline), and fisher behavior rather than a sample of tuna presence. The issues were treated as limitations of the sampling design and were discussed in Section 4. Therefore, the dataset was restricted to a single year (2022), capturing only the seasonal cycle observed rather than interannual variability in tuna distribution or environmental conditions. This temporal limitation was further discussed in Section 5.

Sources: Ocean Fishing Port (PPS Bungus).

To characterize the oceanographic conditions influencing the spatial distribution of yellowfin tuna, this study used multi-parameter Earth Observation (EO) datasets from the Copernicus Marine Environment Monitoring Service (CMEMS) (Lellouche et al., 2018). CMEMS products provide consistent, high-quality, and temporally continuous environmental information suitable for spatially explicit marine habitat modeling. The selected variables included SST, sea surface salinity (SSS), Chl-a concentration, NPP, and surface current velocity components (u and v). These variables serve as key drivers of tuna habitat suitability and fishing ground variability. All EO datasets were extracted for the study domain (0.5° N–4° S; 97.5° E–101.2° E) and matched to each fishing record at daily resolution for the corresponding date in 2022, ensuring temporal consistency between predictors and the reference catch data (Section 2.3). The original spatial resolutions of the datasets ranged from 1/12° (~9 km) to 1/4° (~25 km), as shown in Table 2. To improve spatial compatibility with the densely distributed fishing point observations, all environmental variables were resampled to a uniform grid resolution of approximately 2.7 km² using bilinear interpolation. This resampling represented a three- to nine-fold increase in nominal resolution relative to the native EO products but did not introduce additional spatial information beyond the source grids. However, the procedure established a common analysis grid consistent with the density of the fishing point observations. Consequently, spatial predictions should be interpreted according to the effective resolution of the original EO datasets (~9–25 km) rather than the 2.7 km² analysis grid to avoid an overestimation of spatial detail or predictive capability beyond the limits of the source data.

GRRF was applied to assess the importance of environmental variables influencing the spatial distribution of yellowfin tuna. This algorithm enhanced model performance by selecting the most relevant predictors while reducing computational complexity and addressing high-dimensional data challenges (Breiman, 2001; Izquierdo-Verdiguier & Zurita-Milla, 2020). Through recursive feature elimination and normalization, GRRF was used to evaluate each variable’s contribution to model accuracy and retain predictors with the highest discriminative power for spatial classification (Geerts et al., 2024). Based on the results, variable importance rankings were generated using the default regularization criterion of Izquierdo-Verdiguier and Zurita-Milla (2020). This procedure retained variables with importance scores exceeding a fixed proportion of the maximum observed importance, corresponding to approximately 60% of the extracted data. Since this threshold followed the algorithm's built-in regularization rule rather than an independently tuned or performance-based selection procedure, the resulting subset should be regarded as one plausible configuration rather than a formally optimized solution. Therefore, future investigations should evaluate model sensitivity across a range of selection thresholds. In this study, selected variables were incorporated into RF model to generate high-resolution spatial predictions of PFZs in the western waters of Sumatra. Combining GRRF with RF reduced redundancy, improved computational efficiency, and enhanced classification accuracy, thereby strengthening regional-scale identification of PFZs.

RF algorithm was used to estimate the spatial distribution of PFZs for yellowfin tuna. This algorithm is an ensemble ML method based on bootstrap aggregating, where multiple decision trees are built from randomly selected subsets of training data and combined to produce stable and robust predictions (Zhang et al., 2022a; Salman et al., 2024). The method reduces variance, improves generalization, and effectively captures both linear and nonlinear relationships in complex geospatial datasets without prior assumptions about data distribution (Talebi et al., 2022). Additionally, RF has relatively low sensitivity to variable encoding, allowing diverse numerical predictor datasets to be processed with minimal preprocessing requirements. This characteristic is particularly beneficial for fisheries and oceanographic applications that integrate multiple environmental variables derived from heterogeneous EO products (Song et al., 2023b). RF can also efficiently handle high-dimensional datasets while providing feature-importance estimates that facilitate interpretation of the relative contribution of individual environmental variables to model performance (Kawaguchi et al., 2024). Compared with conventional statistical methods for fish catch prediction (Smoliński, 2017), RF has shown superior performance in handling complex geospatial datasets and achieving higher predictive accuracy (Lattanzi et al., 2026).

In this study, model accuracy and overfitting control were evaluated through internal cross-validation using out-of-bag (OOB) error estimation combined with the 70/30 train–validation split described in Section 2.3. Since both procedures obtained training and test data from the same geographic domain without explicit spatial blocking, nearby fishing points could occur in the subsets. Consequently, the reported performance metrics may overestimate predictive accuracy relative to a spatially independent validation framework and should be interpreted as an upper bound of model performance (Section 4). Although RF is less interpretable than single decision trees, particularly when multiple environmental interactions are included (Wen et al., 2024), systematic hyperparameter optimization was performed to improve robustness and predictive capability. Before model development, essential parameters such as the number of trees (n_estimators), maximum tree depth (max_depth), minimum samples required to split a node (min_samples_split), and bootstrap sampling (bootstrap) were optimized to reduce the risks of overfitting and underfitting while enhancing model generalization (Gladju et al., 2022; Hanif et al., 2026). The optimal configuration consisted of n_estimators = 100, min_samples_split = 10, max_depth = 3, and bootstrap = True. These parameter values were used throughout model training and prediction. The optimized configuration improved predictive stability and generalization for both training and previously unseen data (Nikparvar & Thill, 2021). To capture spatiotemporal variability in tuna distribution, the reference dataset was analyzed under two configurations comprising four seasonal models corresponding to the monsoonal periods (DJF, MAM, JJA, and SON) and one annual model incorporating all reference points. This method generated five prediction maps, providing a comprehensive representation of seasonal and annual PFZs in the western waters of Sumatra. Since all models were developed from observations collected during a single year (2022), the predictions characterize environmental–catch relationships under the oceanographic conditions of that year rather than generalizable multi-year climatological patterns (Sections 4 and 5).

Model performance was evaluated using multiple statistical metrics, including test accuracy, recall, F1-score, RMSE, and MAE, to ensure robust and reliable prediction of PFZs. Classification metrics such as accuracy, precision, recall, and F1-score were used to evaluate the binary presence/background prediction of fishing suitability. Meanwhile, RMSE and MAE were additionally reported to quantify error in the underlying predicted probability values, providing a complementary continuous measure of model fit. During the exploratory phase, correlation analysis was conducted to examine interrelationships and multicollinearity among environmental variables, providing diagnostic insight before model training (Atwa et al., 2024). Feature importance analysis was also applied to quantify the relative contribution of each oceanographic variable, thereby identifying key environmental drivers influencing the spatial distribution of yellowfin tuna and enhancing model interpretability. Classification performance was further assessed using a confusion matrix to derive class-level and overall accuracy metrics (Farhadpour et al., 2024). However, AUC-ROC and precision–recall curves, which would provide a threshold-independent assessment of probabilistic performance, were not computed in this study and were recommended for future evaluations. For the classification metrics, the F1-score was used to evaluate the balance between precision and recall. Prediction errors were quantified using RMSE and MAE to capture both systematic and random deviations between observed and modeled outputs. Because model training and evaluation did not use spatially independent (blocked) cross-validation, the reported accuracy, precision, recall, and error metrics could be optimistic relative to predictive performance at genuinely independent locations. This methodological limitation is discussed further in Section 4. Model implementation and performance evaluation were conducted using the Scikit-Learn library in Python 3, ensuring methodological consistency and reproducibility (Miranda et al., 2023).

Correlation analysis of the marine environmental variables showed several strong relationships among key oceanographic parameters (Figure 3), providing insight into the interactions between the physical and biological processes that structure pelagic habitats in the study area. The strongest association occurred between Chl-a and NPP (r = 0.87), indicating a close ecological relationship between phytoplankton biomass and carbon fixation in the water column. Due to this strong relationship, both variables contain substantially overlapping information, thereby increasing the possibility of multicollinearity when used simultaneously as predictors in RF. This effect is partially mitigated by GRRF feature-selection procedure (Section 2.5), which penalizes redundant predictors. Consequently, the importance of Chl-a and NPP should be interpreted jointly rather than as fully independent contributions. These results are consistent with established ecological theory, where Chl-a serves as a proxy for standing phytoplankton biomass, while NPP represents photosynthetic production, causing a greater primary productivity in waters with elevated phytoplankton concentrations (Cloern et al., 2014; Fontaine et al., 2025). In addition to the strong relationship between the biological variables, SST and SSS showed a moderate positive correlation (r = 0.67). This indicated coherent water-mass characteristics associated with monsoon-driven mixing, stratification, and regional advection. However, relationships between the biological variables and vertical current velocity were weak (Chl-a: r = 0.19; NPP: r = 0.11), suggesting that primary productivity was influenced more strongly by surface hydrographic conditions and light availability compared to vertical mixing alone (Dalabehara & Sarma, 2021). Horizontal current velocity also showed only a weak correlation with salinity (r = 0.08), indicating a limited influence of horizontal advection on salinity patterns at the spatial scale examined. These results show that biological variables, particularly Chl-a and NPP, are more closely related than the physical variables, while current velocities exert a comparatively weaker influence on environmental variability. The observed pattern suggests that habitat suitability for pelagic fisheries in the western waters of Sumatra is primarily regulated by surface hydrographic conditions and phytoplankton availability (Lima et al., 2022).

Understanding the relative contribution of environmental predictors is crucial for identifying the ecological drivers influencing the spatial distribution of tuna fishing grounds in dynamic tropical marine systems. The feature importance analysis (Figure 4) shows that oceanographic variables contribute unequally to model performance. This indicates that RF model captures habitat suitability and underlying ecological structures within the dataset. Based on the results, NPP is identified as the most influential predictor (importance = 0.34), underscoring its role as a primary indicator of food availability. NPP integrates nutrient supply and photosynthetic carbon fixation, making it a robust proxy for prey abundance that supports pelagic predators such as yellowfin tuna (Letscher et al., 2023). SST ranked second (0.23), showing the importance of thermal conditions in regulating tuna physiology, metabolism, vertical movement, and prey distribution, thereby acting as a key determinant of habitat selection (Chen, 2022). Salinity also showed a substantial contribution (0.19), reflecting the influence of water-mass characteristics on habitat suitability.

Chl-a contributed moderately (0.14), while vertical (0.07) and horizontal (0.02) current velocities showed smaller but non-negligible effects. These importance values represented single-model point estimates from one training run, although formal uncertainty quantification such as bootstrapped importance scores across repeated RF runs was not performed. Therefore, relative ranking of closely valued predictors, such as Chl-a against current velocity, should be interpreted with caution. The lower importance of current components suggests that circulation processes influence nutrient transport, plankton aggregation, and frontal dynamics. However, the impact of circulation processes on fishing ground suitability is significantly indirect compared to biologically driven variables. These results indicate that biological productivity, particularly NPP, in combination with key physical drivers such as SST and SSS, collectively influence the spatial distribution of potential fishing grounds, with primary production acting as the dominant constraint on habitat quality for higher trophic levels (Franzè et al., 2023).

The spatial distributions of mean SST, SSS, and NPP during 2022 showed significant seasonal variability across the western waters of Sumatra (Figure 5), indicating the combined influence of monsoonal forcing, regional circulation, and upper-ocean processes on the physical–biogeochemical environment supporting pelagic fish. Based on the results, SST ranged from 29.0 to 30.0°C, with warmer conditions during MAM (March–May) and JJA (June–August), particularly along the coastal shelf and the Mentawai Island chain. This warming during the southeast monsoon is consistent with weakened convective mixing and reduced entrainment of cooler subsurface waters, leading to enhanced surface stratification (Clift et al., 2022). In comparison, cooler SST during DJF (December–February) and SON (September–November) are associated with stronger convection and vertical mixing under the northwest monsoon regime. These seasonal thermal variations have been shown to influence habitat compression and the vertical distribution of tuna within the upper euphotic zone, thereby affecting accessibility to surface fishing gear (Matsubara et al., 2024; Aoki et al., 2025). Seasonal salinity varied between 32.0 and 34.0 PSU, with the lowest values observed during DJF in response to increased rainfall and terrestrial runoff. Higher salinity during JJA and SON corresponded to reduced precipitation and enhanced offshore advection (Zhang et al., 2022b). These patterns are consistent with the alternating influence of the SJCC and ITF in regulating regional water-mass exchange (Purba et al., 2021; Li et al., 2023). Since the descriptions are based on a single annual cycle (2022), the observed differences should be regarded as indicative rather than statistically confirmed. Therefore, formal statistical comparisons, such as ANOVA or non-parametric alternatives, were not performed (Section 5).

NPP varied between 200 and 500 mg C m⁻² day⁻¹, showing pronounced seasonal contrasts across the western Sumatra waters. High values of NPP during DJF and SON were primarily concentrated along the continental slope and offshore regions, consistent with periods of enhanced nutrient supply and upwelling-related processes. In comparison, lower NPP during MAM and JJA was consistent with warmer and more stratified surface conditions, indicating reduced vertical nutrient flux. The correspondence between seasonal NPP enhancement and large-scale circulation signatures suggests that basin-scale dynamics play a major role in structuring spatial bioproductivity gradients, in line with observations from the southeastern Indian Ocean (Navarro et al., 2025). In this study, the spatial distribution of yellowfin tuna catch points showed a strong association with moderate SST of 29.2–29.8°C and elevated NPP of >350 mg C m⁻² day⁻¹. The observed pattern suggests that yellowfin tuna preferentially occupy thermally favorable and highly productive habitats (Debiyanti et al., 2025). This observation confirms the ecological principle that pelagic predators track regions where physical forcing enhances prey availability, supporting the value of bio-physical indicators for ML–based fisheries habitat modeling. However, the relationship is descriptive rather than statistically confirmed because formal spatial analyses are not performed. Furthermore, the fisheries-dependent catch records reflect the spatial distribution of fishing effort in addition to tuna occurrence, and the potential influence of effort-related bias is not explicitly modeled. The methodological limitations are discussed further in Section 4.

Temporal analysis further showed that intra-annual variability in yellowfin tuna catches during 2022 closely tracked seasonal fluctuations in SST, SSS, and NPP (Figure 6). Catch levels peaked during April–June and October–December, coinciding with the monsoon transition periods characterized by elevated primary productivity and favorable thermal conditions. This seasonal association is consistent with the dynamics of monsoon-dominated marine systems, where mixed-layer processes regulate nutrient availability and prey production (Ducklow et al., 2022). In comparison, lower catches during periods of elevated SST (>30 °C) and reduced productivity, particularly during JJA, suggest a redistribution of yellowfin tuna to maintain favorable thermal conditions and access prey resources (Waller et al., 2024). Seasonal boxplot analysis further supports these observations by showing higher median catches during DJF and SON under moderate SST and elevated NPP. Meanwhile, MAM and JJA have lower catch levels despite relatively stable salinity conditions. Seasonal salinity ranged from 32.8 to 34.0 PSU, with slightly higher values during JJA and SON, reflecting the influence of monsoon-driven circulation. Since the temporal patterns are derived from a single annual cycle (2022), the observed seasonal relationships should be regarded as indicative rather than representative of long-term climatological variability. These results indicate that yellowfin tuna distribution is jointly regulated by biological productivity and monsoon-driven hydrodynamics. Consequently, predictable seasonal environmental forcing provides a strong foundation for ML–based operational prediction of PFZs in the western waters of Sumatra (Ismail & Al-Shehhi, 2023).

Evaluation of statistical and ML performance is essential for assessing model generalization beyond the training dataset and the ability to capture the nonlinear ecological relationships that characterize pelagic fisheries environments. In this study, RF evaluation showed strong predictive capability for identifying potential fishing areas from oceanographic variables (Figure 7), and the corresponding performance metrics are summarized in Table 3. The classifier showed high discriminative ability, indicating that the environmental variability contained sufficient spatial and temporal structure to support reliable learning of habitat patterns associated with yellowfin tuna. Furthermore, the annual model achieved an accuracy of 0.91, precision of 0.86, recall of 0.97, and an F1-score of 0.91, with low prediction errors (RMSE = 0.29; MAE = 0.08). As discussed in Sections 2.6 and 2.7, these metrics were derived without spatially independent validation and might overestimate predictive performance at genuinely independent locations. Therefore, the results should be interpreted as measures of relative model fit under the current train–validation design rather than as unbiased estimates of operational predictive skill. The high recall further indicates that the classifier effectively identified true positive fishing grounds, a characteristic that is particularly valuable in operational fisheries applications where missed detections have greater consequences than false-positive predictions.

Seasonal performance was similarly consistent, with the highest accuracy obtained during SON at 0.92 and recall at 0.99, while the lowest accuracy was obtained during MAM at 0.87 with a precision of 0.80. The reduced precision during MAM indicates high spatial noise and low environmental contrast, particularly when stratification and productivity gradients are weaker compared to transitional monsoon periods. Despite slight seasonal variations, the model showed reliable classification capability and stable predictive accuracy throughout the year. This stability across seasons shows the robustness of RF in handling heterogeneous and multivariate oceanographic datasets, outperforming many conventional statistical methods when dealing with high-dimensional nonlinear processes. Furthermore, the performance consistency strengthens the feasibility of adopting RF-based PFZ models for continuous or near-real-time operational fisheries forecasting in dynamic Indonesian waters, particularly where environmental signals related to tuna presence show strong seasonal modulation.

Spatial interpretation of the predicted PFZs is essential to evaluate the correlation of the modeled habitat suitability with regional oceanographic processes and the ecological behavior of yellowfin tuna. The seasonal prediction maps (Figure 8) for DJF, MAM, JJA, SON, and the annual composite show the combined influence of primary productivity, SST, salinity, and current structure on habitat suitability in the western waters of Sumatra. Based on observations, high-probability fishing zones were concentrated mainly along the continental slope and offshore frontal areas. This indicates the importance of mesoscale circulation features, including upwelling cells, current shear zones, and retention structures, that enhance prey aggregation and create energetically favorable feeding habitats for large pelagic predators (Braun et al., 2022). The prediction maps represent spatial patterns of habitat suitability but do not include measures of prediction uncertainty or confidence. Therefore, future applications should incorporate pixel-level uncertainty estimates, such as the variance of predictions across RF trees, to strengthen confidence in operational PFZ forecasting.

Seasonal differences in habitat suitability were evident, with DJF and SON showing the most extensive high-suitability areas, corresponding to elevated NPP and greater observed catches. In comparison, MAM showed more fragmented and spatially restricted suitable habitats, indicating stronger water-column stratification and reduced nutrient supply that limit prey availability. The annual prediction identified persistent core habitats that remained favorable across the monsoonal cycle, suggesting the presence of stable oceanographic features such as eddy boundaries and coastal–current interaction zones. These persistent habitats represent priority fishing grounds for monitoring and sustainable management, emphasizing the value of GeoAI-based predictions for supporting long-term fisheries management in the western waters of Sumatra.