Introduction

Global climate change, linked to substantial anthropogenic forcing1,2, has triggered cascading disruptions across Earth’s critical coastal interfaces—where 90% of marine biodiversity converges and 40% of humanity resides. As emphasized by the IPCC Sixth Assessment Report3, this rapid acceleration of climatic forcing has intensified terrestrial-oceanic couplings beyond historical analogs. However, the underlying processes governing these interactions remain poorly understood. Accelerated ice-sheet meltwater4 and intensified hydro-meteorological extremes5,6 have disrupted sediment delivery systems that stabilized nearshore ecosystems over millennial timescales7,8,9. Concurrently, rising sea levels and altered wind regimes10,11 are intensifying coastal erosion and wave-driven sediment remobilization. Together with expanding coastal urbanization12,13 these perturbations are reconfiguring sediment transport pathways and driving emergent degraded water quality hotspots characterized by suspended particulate matter (SPM) accumulation. Crucially, the nonlinear couplings between climate-driven hydrodynamic shifts14 and anthropogenic pressures on nearshore environments15,16 generate feedbacks that may exceed biogeochemical thresholds for coastal water clarity—a vital indicator of ecosystem health. These synergistic forces create intricate and spatially divergent patterns of material flux between land and sea, with some coastal regions experiencing sediment depletion while others exhibit localized accumulation or enhanced resuspension. Such heterogeneity underscores the need to better understand the long-term trends and multifactorial drivers shaping land–ocean exchanges.

Advances in remote sensing big data have transformed the monitoring of land–ocean interactions, providing high spatial and temporal resolution and deep insights into these dynamic exchanges. While agencies like the U.S. Environmental Protection Agency emphasize SPM as a key water quality metric, their understanding of how these fluxes respond to the combined pressures of human activities and climate change on a global scale remains limited. Satellite-derived data products, such as those tracking sediment transport, ocean currents, and land-ocean flux estimates, are indispensable for deepening our understanding of these interconnected systems. Despite advances in satellite-based coastal monitoring17,18, critical knowledge gaps persist in deciphering how these multiscale drivers synergistically govern global patterns of water clarity. The absence of a unified framework to reconcile localized sediment dynamics with planetary-scale transport mechanisms hinders predictive understanding of coastal ecosystem resilience, particularly in quantifying the relative contributions of climate forcing versus anthropogenic pressures on SPM budgets.

We focused on the 100 km marine ecotone along continental coastlines worldwide, leveraging 23 years (2000–2023) of consistent moderate resolution imaging spectroradiometer (MODIS) observations to track decadal-scale trends in coastal SPM, decode multiscale land–ocean coupling mechanisms, and assess their anthropogenic and climatic drivers. A remote sensing dataset for coastal suspended sediment concentration was developed to analyze spatiotemporal patterns of global coastal SPM levels, derived from the National Aeronautics and Space Administration (NASA)’s Terra and Aqua MODIS satellites’ global daily land surface reflectance (SR) products (500-m resolution). A coastal SPM retrieval algorithm, adapted from a global reference method18, was systematically implemented. Validation against in-situ SPM measurements from 1106 coastal observatories confirmed methodological robustness, demonstrating a median absolute percentage error (MAPE) of 21.1% across global heterogeneous marine environments. Our methodology quantifies coastal waters SPM transport using impervious surface area (ISA)19 as the human pressure index, while integrating sea surface height (SSH), wave height, and salinity as natural drivers. The analytical framework is grounded in established sediment transport models20,21,22 but innovates through global-scale big data spatio-temporal coverage and machine learning-enhanced pattern recognition.

Global patterns of coastal waters

Regions with an annual mean SPM concentration exceeding 10 mg L−1 accounted for ~28% of the global continental coastal waters. High SPM concentrations were primarily observed in high-latitude coastal zones, with annual means surpassing 50 mg L−1 in several estuaries and their adjacent waters (Fig. 1a). Among the top 20 countries ranked by the proportion of coastal waters (within 100 km) with annual mean SPM concentrations exceeding 10 mg L−1, most are developing or emerging economies. Belgium, Finland, United States, Estonia, Canada, and United Kingdom are notably the only developed countries appearing on this list (such as Supplementary Fig. 7).

Fig. 1: Global continental coastal waters patterns 2000–2023.
figure 1

a Spatial distribution of the average annual SPM concentration in global continental coastal waters, calculated as the mean of yearly values from 2000 to 2023. b Distributions of annual mean SPM concentrations in global continental coastal waters (within 100 km of the coastline) for the years 2000, 2008, 2016, and 2023. Each violin–box hybrid shows the full data density (shaded area) and summary statistics per year. Boxes represent the interquartile range (25th–75th percentiles), with the horizontal line indicating the median; whiskers extend to the 5th and 95th percentiles. Red dashed lines indicate the annual mean SPM concentration for each year. Sample size for each year is approximately n ≈ 856,000 valid 0.05° grid cells.

Full size image

Generally, SPM concentrations exhibited a gradual decline over the period 2000–2023. The distributions of model-predicted annual mean SPM for the years 2000, 2008, 2016, and 2023 (Fig. 1b) show a progressive shift toward lower concentration values, indicating a consistent decrease in sediment levels across global continental coastal waters over the past two decades. Significant declines were observed in specific coastal regions, including Liaodong Bay (China), Port Hedland (Australia), the northern Amazon River plume (Brazil), and Gironde Estuary (France) (such as Supplementary Fig. 8). The spatial extent of high SPM concentrations has progressively retreated landward over the past two decades, indicating a contraction of turbid zones toward the coastline.

Decline and landward retreat of coastal SPM

To better characterize regional variability, trends in SPM across global coastal waters from 2000 to 2023 were classified into five categories: Further Clearing, Gradually Clearing, Steady, Gradually Turbidifying, and Further Turbidifying (Fig. 2). Although the Further Clearing category comprises ~16% of total coastal pixels, it alone accounted for more than 66% of the global declining trend in SPM concentrations. The Gradually Clearing region contributed an additional 4.5%, and together these two categories explain over 70% of the global SPM decline (Fig. 2c), highlighting the dominant influence of strong clearing zones in shaping global trends.

Fig. 2: Trends and patterns of global continental coastal waters from 2000 to 2023.
figure 2

a Spatial classification of SPM trends in global continental coastal waters based on model-derived annual mean concentrations from 2000 to 2023. Trends are categorized into five classes: Further Clearing (dark blue), Gradually Clearing (light blue), Steady (white), Gradually Turbidifying (light orange), and Further Turbidifying (dark red). Pie charts indicate the proportional distribution of each trend class by continent: NA (North America), SA (South America), EU (Europe), AF (Africa), AS (Asia), and OC (Oceania). Insets (outlined in dark purple) highlight four representative river mouths and delta systems in Asia: (1) the Yellow River Estuary in China, (2) the Yangtze Estuary in China, (3) the Mekong Delta in Vietnam, and (4) the Ganges-Brahmaputra Delta in Bangladesh. b Temporal change in the average distance between the coastline and the spatial boundary where SPM concentrations match the 2000 global mean. Blue line shows the inland retreat of turbid waters; the orange line shows the declining trend in global annual mean SPM concentrations. Shaded bands represent 95% confidence intervals. Both trends are statistically significant (P < 0.01). c Contribution of each trend class to the global change in coastal SPM concentrations. Negative values indicate decreasing SPM (clearing), while positive values indicate increasing SPM (turbidifying). Contributions are calculated as the product of slope magnitude, pixel count, and mean concentration.

Full size image

Using the global annual mean SPM concentration in 2000 as a reference threshold, we found that areas exceeding this baseline have progressively retreated landward over the past 23 years, with an average recession rate of approximately 0.014 km year−1 (Fig. 2b). This suggests that clearer waters are extending toward the shoreline in many regions, while turbid zones are shrinking. Regionally, coastal waters in mid and low latitude zones exhibited the most pronounced SPM declines between 2000 and 2023, with the Australian coastline alone contributing over 30 percent to the global downward trend (Fig. 2a). Over the same period, SPM concentrations have decreased at an average rate of 0.28 mg L−1 per year. In Asia, four major river mouths and delta regions, including the Ganges Brahmaputra Delta in Bangladesh, the Mekong Delta in Vietnam, the Yangtze Estuary in China, and the Yellow River Estuary in China also showed substantial clearing signals, with clearing trends exceeding turbidifying. Notably, Chinese coastal waters exhibited one of the fastest rates of SPM decline globally, alongside Singapore, making both regions key hotspots of rapid coastal sediment reduction and shoreline retreat (such as Supplementary Fig. 9).

Anthropogenic and climatic drivers

To assess the influence of human activities on coastal SPM dynamics, we incorporated global coastal ISA data to distinguish between natural, engineered, and transitioning coastlines. In 2000, coastal regions adjacent to ISA exhibited SPM concentrations that were ~9.64 mg L−1 lower than those in non-ISA regions, suggesting that urbanization and industrial development may be directly associated with reduced SPM levels. From 2000 to 2023, SPM concentrations declined across all coastline types (Fig. 3a). Natural and engineered coastlines exhibited nearly identical rates of decline, with slopes of −0.1958 mg L−1 year−1 and −0.1987 mg L−1 year−1, respectively. In contrast, coastlines transitioning from natural to engineered conditions showed a less steep decline, potentially reflecting a lag in the response of SPM dynamics to land use change.

Fig. 3: Influence of anthropogenic and natural factors on coastal SPM trends from 2000 to 2023.
figure 3

a Temporal trends in annual mean SPM concentrations (2000–2023) across natural coastlines, engineered coastlines, and transitioning coastlines. Coastline types were classified using ISA data. b Mean SHAP values indicating the relative importance of physical drivers (SSH slope, salinity slope, and wave height slope) on SPM trend predictions. c SHAP dependence plots showing the nonlinear effects of SSH slope, salinity slope, and wave height slope on SPM slope. Red lines indicate smoothed trends; vertical dashed lines represent SHAP turning points or thresholds.

Full size image

Slopes of SSH, wave height, and salinity on coastal SPM trends between 2000 and 2023 show nonlinear response patterns across these variables (Fig. 3b, c). The Shapley Additive exPlanations (SHAP) dependence plots reveal how each variable contributes to the predicted SPM trend. For the SSH slope, SHAP values exhibit an enhanced decline beyond approximately a critical level (~0.003), suggesting that rising SSH consistently suppresses SPM levels. In contrast, the salinity slope shows a unimodal response, SHAP values peak between −0.05 and +0.05 and decline outside this range. For wave height slope, SHAP values increase within the range of approximately −0.01 to −0.002, drop near zero, and then stabilize or rise slightly, suggesting that moderate wave energy may enhance SPM while very low or high wave intensities are associated with more stable or decreasing trends.

Discussion

This study provides a comprehensive analysis of global trends in coastal waters sediment transport, focusing on the key natural and anthropogenic factors driving changes in coastal waters over the past 24 years. Utilizing a global-scale remote sensing dataset and an enhanced machine learning retrieval model, we quantified the spatiotemporal patterns of SPM over 2000–2023 as a proxy for sediment dynamics in global continental coastal waters, with selected regional examples used to illustrate key trends. Our findings reveal sustained declines in SPM concentrations across global continental coastal waters over the past two decades, reflecting shifts in coastal sediment dynamics in response to both anthropogenic pressures and climatic variability.

The anthropogenic driver of SPM concentrations was represented by coastal ISA and natural factors comprised: SSH, wave height, and salinity. Both ISA and rising SSH were positively associated with long-term declines in SPM concentrations, whereas wave height and salinity exhibit nonlinear relationships, with a mid-range peak, with SPM (Fig. 3).

At the global scale, there was a clear decline in SPM levels over time, particularly in regions where human activities, such as urbanization and industrial development, have led to a significant increase in coastal ISA. Coastal urbanization, particularly in rapidly developing regions such as China, has disrupted land–sea sediment pathways and diminished SPM concentrations in adjacent waters23,24,25. In areas without significant ISA influence, observed declines in SPM may be associated with the gradual and potentially accelerating SSH (Fig. 3). As water depth increases, vertical mixing and resuspension processes are altered, which typically reduces sediment suspension in offshore waters through enhanced sediment settling26,27. These changes suggest a transition toward sediment-starved coastal systems, especially in urbanized or engineered shorelines.

At the regional scale, several major river mouths and delta systems exhibited pronounced reductions in SPM export to adjacent coastal waters. Notably, four prominent systems in Asia—the Ganges-Brahmaputra Delta (Bangladesh), Mekong Delta (Vietnam), Yangtze Estuary (China), and Yellow River Estuary (China)—showed substantial declines, with the most significant reduction observed in the Yellow River Estuary (Fig. 2a). This reduction in sediment delivery from land to sea is likely attributable to the construction of hydraulic engineering structures such as dams and reservoirs, which have been shown to reduce particulate load by nearly 50%28,29,30. These findings underscore the role of anthropogenic modifications to river systems in altering coastal sediment dynamics. The complex and nonlinear nature of these interactions highlights the need for further investigation into localized sediment transport processes in highly regulated deltaic regions.

Monotonic declines in SHAP values with increasing SSH slope indicate that long-term sea level rise has a sustained positive contribution in declining SPM concentrations. This pattern may reflect the physical suppression of sediment transport caused by deeper water depths, reduced land–sea gradient, diminished estuarine outflow, or weaker near-bed turbulence and sediment resuspension as water depth increases. The lack of nonlinearity in the SSH response implies that its influence acts as a gradual yet persistent, and potentially accelerating, forcing mechanism on sediment dynamics.

In contrast, nonlinear SHAP patterns for salinity and wave-height slope point to more complex and threshold-sensitive optimal feedbacks. Salinity changes near zero slope—indicative of stable or low-gradient conditions—may enhance sediment resuspension or inhibit flocculation, leading to increased SPM concentrations. However, larger positive or negative shifts in salinity slope are likely associated with density stratification or reduced turbulence, resulting in greater sediment settling and lower SPM levels. A similar threshold response is evident for wave-height slope: moderate increases tend to promote sediment resuspension, but beyond a certain intensity, stronger wave regimes may drive offshore transport or vertical dilution of sediments, ultimately limiting SPM accumulation. These nonlinear dynamics highlight the dual roles of salinity and wave forcing in both mobilizing and redistributing sediments within coastal waters.

While this study primarily focused on ISA within 5 × 5 km coastal grids and a 5 km landward buffer to assess the localized impacts of shoreline urbanization, additional factors may influence suspended sediment dynamics. Other anthropogenic activities, such as offshore wind farms and oil and gas platforms, can affect SPM levels in certain regions, notably in areas like the northern Yangtze River estuary31. Natural drivers, including wind direction and ocean currents, may also shape localized trends, particularly in enclosed bays32,33,34. Furthermore, broader terrestrial factors such as deforestation, agricultural intensification, and the construction of reservoirs can significantly alter sediment fluxes from inland watersheds, although these factors were beyond the spatial scope of our nearshore-focused framework35,36,37,38. Future work should integrate catchment-scale land use and hydrological connectivity to better capture the full continuum of land and ocean sediment interactions in coastal systems.

By integrating remote sensing observations with systematic analysis of physical and socioeconomic drivers, this study establishes a scalable framework for tracking sediment transport trends under global change, highlighting the key drivers and processes that influence coastal water clarity (Fig. 4). The observed shorelineward shift in regions of high SPM, averaging 0.014 km year−1, reflects the dynamic reorganization of coastal sediment regimes at both global and regional scales. This study primarily focuses on coastal drivers, including long-term sea level rise, increasing coastal ISA, and changing salinity and wave height conditions, which collectively reshape SPM dynamics. We also acknowledge that upstream alterations, such as reductions in river discharge and sediment trapping by dams, may exert substantial influence on coastal SPM trends. Although these riverine factors were beyond the scope of our analysis, the observed patterns underscore the combined roles of climatic and anthropogenic pressures in modulating nearshore sediment behavior, with implications for turbidity, nutrient cycling, and ecological stability.

Fig. 4: Diagram illustrating the main anthropogenic and natural drivers affecting coastal suspended sediment dynamics.
figure 4

Land-based influences include sediment supply from upstream dams and coastal ISA. Oceanic factors include SSH, wave-induced resuspension, salinity changes, and offshore industrial activities. Light orange arrows indicate pathways associated with sediment reduction.

Full size image

These findings advance understanding of how anthropogenic pressures and climate variability reshape sediment fluxes at the land–sea interface. Recognizing the nonlinear and spatially heterogeneous nature of these dynamics is critical for informing adaptive coastal management. As sediment-starved conditions become more widespread, integrated monitoring frameworks—combining satellite observations, in-situ data, and land-use models—will be essential for sustaining ecosystem resilience. Future progress will depend on the synthesis of multi-scale big datasets and scalable analytical approaches to support locally tailored, evidence-based responses to accelerating climate change and intensifying human impacts.

Methods

Data sources

We utilized daily surface reflectance (SR) products from MODIS Terra (MOD09GA) and Aqua (MYD09GA), with a spatial resolution of 500 m, to estimate SPM. Atmospheric correction and cloud removal were performed on MOD09GA and MYD09GA data using Google Earth Engine (GEE) tools (https://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MOD09GA; https://developers.google.com/earth-engine/datasets/catalog/MODIS_061_MYD09GA#description). These datasets encompass surface spectral reflectance across MODIS bands 1–7, corrected for Rayleigh scattering, as well as gas and aerosol signals. Residual errors were mitigated through band subtraction, and SR values were further adjusted to the SRland scale to minimize inaccuracies from excessive atmospheric correction in coastal regions39. To reduce the influence of short-term extreme events and episodic anomalies in SPM, we based our analysis on annual mean SPM values computed from GEE-derived cloud-free surface reflectance composites. These annual averages inherently smooth over daily and seasonal variability, including spikes caused by storms, monsoons, or tidal resuspension. Additionally, GEE employs standardized cloud and shadow masking algorithms, ensuring that reflectance data used in model prediction is largely unaffected by persistent cloud cover. Using these corrected SRland datasets, we then developed a global SPM inversion model by matching the remote sensing data to field-measured observations.

We used four databases from in situ field observations: the China Coast and Estuary database (including the Yellow River Estuary, Yangtze River Estuary, Pearl River Estuary); the Sea-Viewing Wide Field-of-View Sensor (SeaWiFS) Bio-optical Archive and Storage System (SeaBASS); and the United Nations Environment Programme Global Environment Monitoring System (GEMStat) (such as Supplementary Fig. 1). These databases cover coastal areas, estuaries, and bays, representing a wide range of turbidity and nutrient conditions, which enabled a thorough assessment and validation of the retrieval algorithm’s accuracy. A total of 48 observation points along the coast and estuaries of China from 2001 to 2021 were selected, with SPM concentrations ranging from 2 to 1890 mg L−1. For SeaWiFS, SeaBASS, and GEMStat, 15 coastal areas and 7 estuaries were chosen, with sampling concentrations ranging from 0.0001 to 1740.45 mg L−1. In total, 1106 stations globally were matched with MODIS daily products using GEE.

We utilized the 30-m time-series global ISA (Version 2.0) dataset, which maps annual increases in global ISA from 1972 to 201919. This dataset serves as a crucial indicator of urbanization and is used to characterize the scale of impact of human activities along coastlines. We defined coastal ISA as impervious surface coverage exceeding 10% within 5 × 5 km coastal grid cells and analyzed their influence within a 5 km landward buffer zone. While we focused on shoreline-adjacent areas, we acknowledge that broader terrestrial drivers such as land clearing and agriculture may also influence sediment dynamics but are beyond the spatial resolution of this framework. For detailed processing methods, please refer to the Supplementary Materials.

As detailed in Table 1, we selected several global environmental reanalysis datasets, including: Wave height, SSH, Salinity, EEZ, and Global coastline.

Table 1 List of datasets used in the study along with their source and link
Full size table

For the global study area, we selected the Marine Regons’ continents dataset based on previous research40 and defined the study area as the global continental coastal waters, extending approximately 100 km seaward from the shoreline to match the spatial coverage of MODIS land products and previous coastal studies (such as Supplementary Fig. 1)40.

Global coastal waters SPM retrieval algorithm

Following methodology established previously18, the XGBoost model was used to estimate global SPM. XCBoost incorporates regularized boosting of gradient-based decision trees, which helps reduce model variance41. The “XGboost” package in Python was used to train the model.

Four bands from the atmospherically corrected MODIS data (SRland(469), SRland(555), SRland (645), SRland (859)) were selected as optimal bands for estimating SPM \(({P}_{{SPM}})\). Latitude and longitude were included to reduce errors from spatial heterogeneity in the estimation as follows:

$${P}_{{SPM}} = {XGboost}({{{\rm{SR}}}}_{{land}}\left(469\right),{{{\rm{SR}}}}_{{land}}\left(555\right),{{{\rm{SR}}}}_{{land}}\left(645\right),\\ {{{\rm{SR}}}}_{{land}}\left(859\right),{{{\rm{SR}}}}_{{land}}(859)/{{{\rm{SR}}}}_{{land}}(645),\\ {{{\rm{SR}}}}_{{land}}(645)/{{{\rm{SR}}}}_{{land}}(555),{{\rm{Latitude}}},{{\rm{Longitude}}})$$
(1)

Here, the ratios SRland (859)/SRland (645) and SRland (645)/SRland (555) represent two effective band ratios previously used for estimating SPM in certain water bodies42,43,44,45. The SWIR bands were not used due to their low signal-to-noise ratios and negligible water-leaving signal at these wavelengths46. The model requires the adjustment of several hyperparameters during training, including the type of booster, learning rate, maximum depth, and regularization coefficients.

To accommodate the skewed distribution of SPM concentrations in the training data, we first applied a logarithmic transformation to reduce heteroscedasticity and improve model stability. We then constructed two separate models—one for high SPM and one for low SPM conditions. These models are dynamically selected based on an adaptive threshold and blended through a smooth transition zone to minimize blocky artifacts during prediction.

To evaluate model accuracy, we use the mean absolute error (MAE) (Eq. 2) and MAPE (Eq. 3). MAE values are calculated numerically and expressed as relative differences (i.e., as percentages). Generally, relative differences such as MAPE within ~35% are considered satisfactory for assessing estimation performance relative to measured values, while MAPE > 60% indicates unsatisfactory performance. Refer to the Supplementary Materials for details.

$${MAE}=\left|10\frac{{\sum }_{i=1}^{N}|{\log }_{10}\left({y}_{i}\right)-{\log }_{10}\left({x}_{i}\right)|}{N}-1\right|\times 100 \%$$
(2)
$${MAPE}=100 \% \times {median}\left[\frac{{y}_{i}-{x}_{i}}{{x}_{i}}\right]{for}i=1,2,\ldots ,N$$
(3)

Global coastal waters SPM trends

To calculate trends for SPM values in the coastal waters of the study area for the period 2000–2023, a linear regression model was used to estimate the annual mean trend for each grid cell, using a spatial resolution of 0.05° (~5 km at the equator), as shown in Eq. (4):

$$y\left(t\right)={k}_{1}t+{a}_{1}$$
(4)

where \(y(t)\) is the value of the SPM trend, which is a function of time, \(t\) (i.e., year), coefficients \({k}_{1}\) and \({a}_{1}\) were determined by standard least-squares approaches.

To extract decadal SPM trends, we removed temporal autocorrelation, computed linear slopes from annual model outputs (2000–2023), excluded extreme outliers, interpolated missing values, and aggregated results to 0.05° resolution for global coastal analysis. To identify significant trends, we employed the Mann–Kendall (MK) test, a widely used non-parametric method for detecting trends and assessing randomness in time series data. All analyses were performed using Python.

Distance from coastline

We quantified the distance of SPM concentrations from the coastline by calculating the spatial shift of a specific concentration level relative to the coast. First, the annual mean SPM concentration for the year 2000 was chosen as the baseline. We then counted the number of grid cells corresponding to this concentration extending from the coastline and determined the total area (\({S(i)}_{2000{mean}}\)). The length of the coastline associated with this baseline concentration (\({d(i)}_{2000{mean}}\)) was also measured. Finally, we calculated the relative distance of the SPM concentration, as shown in Eq. (5):

$${D}_{i}=\frac{{S(i)}_{2000{mean}}}{{d(i)}_{2000{mean}}}(i=2000,2001,\ldots 2023)$$
(5)

In the above equation, \(i\) represents the year, and \({D}_{i}\) denotes the offshore distance in a given coastal waters area relative to the annual mean SPM concentration for 2000 for year \(i\). For further details, please refer to the Supplementary Materials.

Trend contribution analysis

To quantify the contribution of different regions or trend classes to the overall global change in SPM, we calculated a trend contribution score by weighting the slope of each grid cell by its spatial extent and SPM magnitude. This approach considers both the direction and magnitude of change, as well as the ecological significance of the SPM baseline. For a given grid cell \(i\), its contribution to the global SPM trend is defined as shown in Eq. (6):

$${C}_{i}={S}_{i}\times {A}_{i}\times \overline{{{SPM}}_{i}}$$
(6)

where \({C}_{i}\) is the weighted contribution of grid cell \(i\) to the global trend. \({S}_{i}\) is the annual SPM trend (slope) at grid cell \(i\), in units of mg L−1 year−1. \({A}_{i}\) is the area of the grid cell (pixel count). \(\bar{{{SPM}}_{i}}\) is the mean SPM concentration over the study period at grid cell \(i\), in mg L−1.

The total contribution of a given class or region (e.g., Class 1: further clearing) is calculated as shown in Eq. (7):

$${C}_{{class}}={\sum}_{i\in {class}}{C}_{i}$$
(7)

To estimate the relative contribution (%) of each class or region to the total global trend, we use the method shown in Eq. (8):

$${Contribution}=\frac{\left|{C}_{{class}}\right|}{{\sum}_{k}\left|{C}_{k}\right|}\times 100 \%$$
(8)

where \(k\) is the summation over all classes or regions. The absolute value is taken to reflect the relative weight of trend magnitude, regardless of direction. This formulation allows the identification of regions that disproportionately drive global trends due to their strong SPM gradients or high baseline concentrations, even if their spatial extent is limited.

XGBoost and interpretability

We employed the robust and widely used XGBoost Regression model to relate variations in SPM concentrations to the primary drivers of these changes. This machine-learning approach is well-suited for handling long-term, multi-scale datasets with complex structures and multiple predictors.

The model identified strong relationships between coastal SPM concentrations and key environmental drivers, including wave height, SSH, and salinity. To improve the interpretability of each variable’s contribution, we employed Shapley Additive exPlanations (SHAP)47, a game-theoretic method that assigns optimal contribution scores to individual predictors based on classical Shapley values. To minimize the influence of extreme outliers while preserving dominant trend patterns, we restricted the SHAP analysis to the 5th–95th percentile range of SPM slope values. This filtering step excluded rare and potentially noisy extremes, enabling a more robust focus on core transition regimes—particularly those corresponding to Further Clearing and Further Turbidifying trends.

To interpret model outputs, we used SHAP bar plots to quantify the relative contributions of wave height, sea level, and salinity to annual mean SPM concentrations. Trend lines were overlaid to illustrate the direction and form of each relationship, indicating whether variables exerted positive or negative effects on SPM. All modeling and interpretation were performed in Python, using the XGBoost package for training and the SHAP package for post hoc variable attribution.