1. Introduction
Considered one of the biggest threats to marine ecosystems [1], microplastics (MPs) are polymeric, mainly fossil-based particles less than 5 mm in diameter [2,3]. They contribute to an estimated 8 million tons of this type of waste reaching the ocean annually, with astonishing quantities of over 4003 million tons produced and discarded until 2022 [4,5]. On sandy beaches, MPs have emerged as key pollutants, diminishing the aesthetic value of these areas and negatively affecting their appeal and recreational quality [6,7,8]. These pollutants not only reduce the visual and touristic attractiveness of beaches but also reflect broader issues in waste management and the health of marine ecosystems [8,9,10,11].
Chemically, MPs can absorb and concentrate organic contaminants, such as polychlorinated biphenyls (PCBs), dichlorodiphenyltrichloroethane (DDT), and polycyclic aromatic hydrocarbons (PAHs), transferring these substances to many marine species and eventually to us through consumption [12,13]. That is why the increasing presence of MPs in coastal regions raises significant concerns due to their durability, ability to carry harmful substances, and potential entry into the food chain (via bioaccumulation and subsequent biomagnification), affecting both marine organisms and human health [14]. Additionally, by altering the composition of coastal habitats, MPs directly impact ecosystems and the species that rely on them [15]; once MPs are degraded by UV radiation or mechanical abrasion, change the porosity of sediments, affecting water retention capacity and potentially releasing greenhouse gases (GHG) like methane and ethylene [16,17].
In Brazil, a review by Escrobot et al. [18] highlighted that between 2018 and 2023, most studies focused on beaches and marine biota, identifying tourism, fishing, and river discharge as the primary sources of MPs. The São Paulo coast, particularly the Santos Estuary, is a critical pollution hotspot due to industrial activities (mainly production and transportation) and inadequate plastic waste management in densely populated urban areas [2,19,20,21]. Although some studies have addressed this issue, there is still a need for more comprehensive research to understand MPs deposition and accumulation patterns better. These challenges have driven international initiatives such as the elaboration of the United Nations Sustainable Development Goal 14 (SDG 14—Life Below Water) and the G20 Osaka Blue Ocean Vision, aiming to reduce marine pollution by 2025 and eliminate ocean plastic waste by 2050, respectively [22,23]. Furthermore, since 2022 efforts have been made to design a global legally binding agreement among most countries in the UN system [24], but significant advances in this area have not yet been achieved.
Studies by Ferreira et al. [2,20] show that beach slope and orientation directly influence MPs accumulation due to their relationship with hydrodynamic and morphodynamic processes. These variables enhance model accuracy by capturing complex relationships that traditional geomorphological methods, such as grain size distribution studies, often fail to detect. Particularly in regions with high spatial and temporal variability, as observed on the northern coast of São Paulo In this context, machine learning (ML) algorithms have become a widely adopted approach in environmental studies because they can process large datasets and model complex nonlinear interactions between different variables [25,26,27]. Moreover, linear methods are sensitive to multicollinearity and struggle to model complex phenomena, whereas ML techniques demonstrate greater precision and robustness when handling large volumes of heterogeneous environmental data [2,26,28,29].
Algorithms such as Artificial Neural Networks (ANN), Random Forest (RF), Support Vector Machines (SVM), and Gradient Boosting (GB) have shown success in capturing these interactions, generating robust predictive models for issues like water quality, land use, and pollutant dispersion [30,31]. In addition to offering greater accuracy, these models can be programmed to automatically select the most relevant variables, enhancing their efficiency [32]. These features are crucial in such research, where data is often collected across different temporal and spatial scales involving multiple interdependent factors [33]. However, the success of these algorithms depends on the quality of the input data and the appropriate model selection, requiring careful evaluation of the applied methodologies [28].
As remote sensing (RS) technologies and in situ surveys evolve alongside computational capacity, ML in environmental studies is expected to expand, delivering increasingly reliable and deeper insights into complex phenomena [2,20,34,35,36]. This study applies ML algorithms to model and quantify the distribution of MPs/m2, using beach face direction and slope as independent variables, both derived and calibrated from orbital RS images and topographic profiles obtained via the Global Navigation Satellite System (GNSS). This approach innovatively provides an efficient tool for mapping critical areas of MPs pollution, capturing complex deposition patterns, and can be replicated globally. It supports more effective environmental policies and contributes to global plastic waste management in coastal environments, in line with SDG 14 targets, which aims to conserve and promote the sustainable use of oceans, seas, and marine resources [23].
Based on the geomorphological influence on MPs deposition, as highlighted in previous studies, this research hypothesizes that beach face slope and orientation are key predictors of MPs accumulation patterns on São Paulo’s northern coast. Specifically, beaches with lower slopes and orientations toward the SSW are expected to show higher MPs deposition due to their reduced sediment transport dynamics and greater exposure to storm waves and anthropogenic inputs.
4. Discussion
The choice of ML models used in this study (GB, RF, Lasso, Ridge, SVR, and PLS) is justified based on their documented performance in similar environmental studies and their capacity to address the specific characteristics of the dataset. The GB model was selected for its ability to capture complex and nonlinear relationships between predictor variables and the target variable, as demonstrated in environmental predictions involving high data variability [32,62,63]. RF, known for its robustness and resistance to overfitting, performs well with multicollinear and high-variability datasets, which are common in these same contexts [26,30,74].
The Lasso and Ridge Regression models were included as baseline techniques due to their computational efficiency and effectiveness in handling multicollinearity or reducing irrelevant predictors [28,64,65]. While SVR was chosen for its ability to model nonlinear patterns in complex and heterogeneous datasets, which has proven useful in coastal and environmental studies [31,66,67]. Last, PLS method was applied for its suitability in managing multicollinear variables and small sample sizes while efficiently reducing dimensionality without compromising predictive power [29,68].
The SHAP plots play a crucial role in assessing the relative influence of the independent morphometric variables (tanβ and Aspect) on predicting MPs deposition. These plots enabled the identification of each variable’s individual contribution to the performance of ML models, assisting in both model selection and interpretation [71]. In this study, SHAP results showed that beach slope exerts a significant influence on MPs accumulation, emerging as the most impactful variable, while beach face orientation exhibited a smaller but still relevant effect during storm surge events [2,20]. Furthermore, the plots highlighted the GB model’s ability to capture complex patterns by integrating environmental variables [75,76,77], providing a holistic understanding of the processes governing MPs distribution (Friedman, 2001 [32]; Ke et al., 2017 [63]).
The results indicate that the GB model achieved the best precision and generalization capability for predicting MPs/m2 deposition. This performance suggests that the GB model captured the relationship between predictor variables and the dependent target variable (MPs/m2 deposition) more efficiently than other tested models, aligning with previous studies pointing the potential of GB in environmental contexts due to its robustness with noisy data and fine-tuning capability. In this regard, GB demonstrated good fit and lower complexity, reflecting its ability to balance fit and simplicity [28,32,62,63], which is essential given the geomorphological characteristics of São Paulo’s northern coast’s (irregular, with pocket beaches of various sizes and orientations) [2,40,78].
In turn, although the RF model also demonstrated satisfactory performance, it was inferior to GB, consistent with the literature describing RF as a robust model but with precision limitations when compared to more complex algorithms, especially in highly variable data [30]. Linear models, such as Lasso and Ridge, also showed intermediate performance, demonstrating a lower capacity to capture the non-linearity of variables, a characteristic of MPs deposition conditions [64,65]. The SVR model also presented intermediate performance, indicating a good fit without overfitting. This aligns with studies showing SVR’s ability to capture non-linear patterns, especially in complex and heterogeneous data [31,66,67]. In contrast, PLS stood out as a viable alternative when model simplicity is preferable (e.g., in the case of more homogeneous coastlines, considering the beach slope and orientation), confirming its traditional use in scenarios that seek to explain variance in multivariate datasets effectively reducing dimensionality without compromising predictive power [29].
Emphasizing again, the RF, although effective for datasets with high variability, often fails to detect subtle interactions [26,30]. Similarly, Ridge Regression does not adequately represent environmental phenomena characterized by nonlinearity [28,65]. While, GB outperforms RF and Ridge due to its sequential structure, which iteratively adjusts predictions to correct residual errors. This flexibility allows GB to identify intricate patterns that simpler models often fail to capture [28,32,62]. Nevertheless, RF remains valuable in scenarios where interpretability is less critical, and computational efficiency is a priority, as it trains faster and resists overfitting in noisy data [30]. Ridge Regression remains applicable when linear relationships dominate the dataset, offering a computationally efficient and straightforward solution [65]. Combining RF and Ridge Regression with GB can create a more comprehensive analytical framework. While GB excels in accuracy and generalization, RF and Ridge can serve as complementary tools for baseline analysis or cross-validation [33,79].
In this coastal area, MPs deposition on sandy beaches was found to be strongly influenced by beach slope (tanβ) and beach face direction (Aspect), corroborating other studies that emphasize the importance of geomorphological parameters in MPs retention [2,20,80,81,82,83]. Beach faces oriented toward directions with greater wind and current exposure exhibit higher debris deposition, including MPs [4,84,85]. This finding is consistent with ASR values and with Ferreira et al. [20,35], who observed on the São Paulo coast that beaches facing SSW are particularly susceptible to MP deposition. These observations suggest that adjusting predictive models based on beach slope and orientation could enhance understanding of plastic debris accumulation in coastal zones [2,79,86,87,88].
In this context, beaches facing SSW are more predisposed to coastal stability and/or accretion processes that make them susceptible to MPs accumulation. This accumulation occurs due to the approximate 90° angle between storm surge waves from the south and the coastline, which reduces sediment transport rates alongshore, thereby promoting the deposition of sediments and anthropogenic debris, including MPs [2,16,20,21,35,89,90]. The SW-NE orientation of this portion of São Paulo’s coast also makes SE-facing beaches more vulnerable to erosion, as storm waves strike at an angle of approximately 45°, maximizing sediment transport during the austral autumn and winter [2,20,35,45,89,91]. The seasonal incidence of storms promotes sediment resuspension and increases MPs mobility, especially in areas near estuaries, such as Bertioga (BET), which serve as critical entry points for MPs into the ocean due to urbanization and human activities [92,93,94,95,96].
In compartment C5, near São Sebastião (SSB), high population density and industrial and port activity promote MPs dispersion in the area [21]. Ivar do Sul et al. [19] MPs concentrations correlate with industrial and port activities, particularly in Southeastern Brazil. Although compartments C4, C5, and C6 are not as densely populated as southern portions of the São Paulo coast, their beaches receive many visitors on weekends, holidays, and vacations. As a result, these beaches do not undergo frequent mechanical cleaning, and waste from insulating containers (i.e., Styrofoam and other materials) for storing beverages and/or food left by tourists and street/beach vendors tends to accumulate mainly in the higher beach portions [2,20,97,98,99,100].
5. Conclusions
The study highlights the importance of incorporating variables like beach slope (tanβ) and beach face orientation (Aspect) into predictive multivariate techniques and the central role of ML in better understanding the patterns of MPs pollution on sandy beaches and generating information. GB model emerged as the most efficient in predicting MPs deposition, with the highest coefficient of determination (R2 = 0.9771) and the lowest error values (MAE and RMSE). The SHAP effectively captured complex relationships between predictor and dependent variables, making it suitable for environmental studies involving high data variability. While, the PLS model performed well in simplicity, with low AIC and BIC values, making it a viable option in contexts prioritizing lower complexity.
It is essential to conduct a literature review on the generalization capacity of models under different environmental conditions to deepen the understanding of MPs predictions. The geometric shape, density, and size of MPs were not directly analyzed in this study. Information such as polymer density can be investigated using additional methods, such as Raman spectroscopy. However, the study has other limitations, including the absence of additional metoceanographic variables, such as waves and winds, which may influence MPs deposition. It is also suggested to expand the study area to cover the entire São Paulo coast, increase in situ sampling points, and explore more advanced ML models, such as deep learning methods.
Expanding the use of these models in other regions can increase understanding of the dynamics of this pollutant. GB has demonstrated a great capacity to capture complex patterns, making it suitable for analyzing large data sets in different environments, demonstrating its potential for broader applications along the coast of São Paulo. Once again, morphometric analysis revealed that beaches with Sloping profiles and facing SSW are more prone to MPs accumulation due to the interaction between storm waves from the south and the south-facing coastline. This angle of incidence reduces sediment transport and promotes debris deposition, including MPs, besides other anthropogenic factors, such as industrial and port activities (near São Sebastião), that also contribute to MPs dispersion.
These results help to create more effective environmental policies that consider local characteristics to mitigate marine pollution, allowing proper management and conservation of coastal ecosystems. It also contributes to achieving SDG 14, precisely 14.1—to prevent and significantly reduce marine pollution, particularly from land-based activities, including marine debris (Indicator 14.1.1—density of plastic debris) by 2025—and Goal 14.a, which seeks to enhance scientific knowledge, develop research capacities, and transfer technology to improve ocean health [23].
Source link
Anderson Targino da Silva Ferreira www.mdpi.com
