The study, published in the journal Discover Oceans, was led by Nafisa Ali of the University of Information Technology and Sciences in Dhaka, together with colleagues at several Bangladeshi institutions and the University of New Hampshire. The researchers set out to answer a deceptively simple question: can freely available satellite and reanalysis data, fed into well-designed machine learning models, reliably estimate how many microplastics are floating in a given patch of ocean? Their answer, after training and rigorously testing seven different algorithms, was a qualified but emphatic yes.

The Indian Ocean was a natural choice for the case study. It is the third largest ocean on the planet, bordered by some of the most densely populated and rapidly industrializing coastlines in the world, and it carries the second-highest plastic pollution stress of any ocean basin. Major rivers such as the Ganges, Brahmaputra, and Indus funnel enormous quantities of plastic debris from their crowded catchments into the sea, while coastal regions of South Asia, Southeast Asia, and East Africa contribute through urban runoff, fisheries, shipping, and inadequate waste treatment. Microplastics have even been recovered from deep-sea sediments 5,000 meters down in the Central Indian Ocean Basin, underscoring how thoroughly the particles have penetrated the marine environment.

To build their predictive framework, the team assembled 346 georeferenced microplastic abundance measurements collected between 2012 and 2021, drawn from the National Oceanic and Atmospheric Administration’s NCEI Marine Microplastics Database and restricted to samples taken with the standard Manta net grab method. They then paired each sampling location with twelve oceanographic variables extracted through Google Earth Engine from satellite and reanalysis products. These included sea surface temperature from the NOAA OISST product, chlorophyll-a from MODIS-Aqua, sea surface salinity from the HYCOM ocean model, wind speed, surface roughness, and wind stress components from NASA’s MERRA-2 atmospheric reanalysis, and dissolved oxygen, nitrate, and phosphate from the World Ocean Atlas 2023 climatology. The researchers also replaced raw latitude with a physically meaningful derived variable: the distance from each sampling point to the nearest coastline, which serves as a proxy for land-based pollution inputs.

Seven supervised regression models were trained and compared: random forest, gradient boosting regression, extreme gradient boosting, light gradient boosting machine, support vector regression, k-nearest neighbors, and multiple linear regression. The latter served as a statistical baseline to test whether simple linear relationships could explain the data. Hyperparameters were tuned with the Optuna optimization framework using a tree-structured Parzen estimator, and the team employed a battery of safeguards against overfitting, including leakage-free preprocessing pipelines, tenfold cross-validation, an independent 30 percent holdout test set, a geographic holdout strategy to account for spatial autocorrelation, and bootstrap resampling with 1,000 iterations to quantify uncertainty.

The gradient boosting regression model emerged as the clear winner. During cross-validation it achieved a coefficient of determination of 0.8434 with a root mean square error of 0.2424, and on the independent holdout set it reached an R-squared of 0.8609 with an RMSE of 0.2166 and a mean absolute error of 0.1177. Random forest performed nearly as well, while XGBoost achieved a respectable R-squared of 0.80. In contrast, support vector regression and multiple linear regression lagged far behind, with the linear baseline managing only an R-squared of 0.3924. That gap carries an important message: the environmental controls on microplastic distribution are fundamentally nonlinear, involving threshold effects and interactions among variables that linear models simply cannot capture.

Perhaps the most scientifically valuable part of the work came from the interpretability analysis. Using SHAP values, a technique borrowed from game theory that quantifies each variable’s contribution to individual predictions, the researchers identified which oceanographic factors matter most. Sea surface salinity, surface roughness, dissolved oxygen, distance from the nearest coast, nitrate concentration, and zonal wind stress topped the list. Lower salinity and lower dissolved oxygen both pushed predictions upward, while greater surface roughness and greater distance from the coast showed more complex, threshold-dependent effects. Partial dependence analysis revealed, for example, that predicted microplastic abundance dropped sharply once sea surface salinity exceeded roughly 14 practical salinity units.

These patterns align closely with established oceanographic theory. Wind-driven turbulence and surface roughness control how buoyant plastic particles are mixed vertically through the upper water column, meaning surface measurements can swing dramatically with wind conditions. Temperature and salinity gradients govern stratification, which determines whether particles stay near the surface or sink. Biogeochemical variables such as chlorophyll-a, nitrate, phosphate, and dissolved oxygen reflect the activity of microbial biofilms that colonize plastic particles, increase their density, and eventually drive them from floating to sinking states, a process known as biofouling. The strong negative correlation between dissolved oxygen and microplastic abundance, with a Spearman coefficient of minus 0.63, suggests that oxygen depletion associated with biofilm-driven microbial activity and freshwater inputs marks waters where plastics tend to concentrate. Meanwhile, the importance of coastal distance echoes earlier findings that more than 1,000 rivers are responsible for roughly 80 percent of global riverine plastic emissions to the ocean.

The analysis also documented a statistically significant increase in microplastic abundance across the Indian Ocean between 2012 and 2021, with persistent hotspots in the northern and central basins near heavily populated coastlines and major river outflows, particularly the Bay of Bengal and the Arabian Sea. The authors caution that their model is environmentally informed statistical estimation rather than a causal simulation. Oceanographic variables cannot capture plastic emission rates, waste management quality, shipping density, fishing activity, or polymer-specific degradation, and predictions in sparsely sampled regions carry reduced reliability. Still, the model’s empirical coverage under spatial holdout validation reached 93.27 percent, exceeding the nominal 90 percent confidence level, indicating that its uncertainty estimates remain trustworthy even in unseen geographic areas.

The practical implications extend well beyond academic mapping. Because the framework relies entirely on freely available satellite data and open-source machine learning tools, it can be deployed in data-limited regions where conventional monitoring is financially or logistically impossible. The authors propose a prevention-first policy agenda built on their hotspot maps: prioritizing river-basin interventions such as waste capture and leakage reduction in high-emission catchments feeding the Bay of Bengal and Arabian Sea, institutionalizing regular machine learning and remote sensing monitoring to update pollution maps and guide compliance, and aligning national action plans with international agreements. They also recommend pairing predictive mapping with targeted field validation across oceanic fronts, mesoscale eddies, frontal zones, and river plumes, the very features their SHAP analysis flagged as primary drivers of microplastic distribution. Future work, they suggest, should move toward hybrid frameworks that combine hydrodynamic particle-tracking models and anthropogenic drivers such as coastal population density and river discharge with the satellite-based approach demonstrated here. As marine plastic pollution continues to climb worldwide, tools that can see the ocean from orbit and learn its hidden patterns may prove indispensable for protecting the ecosystems and the billions of people who depend on them.

Subject of Research: Machine learning prediction of marine microplastic abundance in the Indian Ocean using satellite-derived oceanographic variables

Article Title: AI-driven prediction of marine microplastics from space: a case study of the Indian Ocean

Article References: Ali, N., Ahmed, S., Afia, M. H., Islam, A. K. M., Izlal, S., Rahman, S., Arif, M. A., Rahman, M. H., & Akondo, M. R. I. (2026). AI-driven prediction of marine microplastics from space: a case study of the Indian Ocean. Discover Oceans, 3(1), Article 59. https://doi.org/10.1007/s44289-026-00160-2

Image Credits: AI Generated

DOI: 10.1007/s44289-026-00160-2

Keywords: microplastics, machine learning, remote sensing, Indian Ocean, gradient boosting, SHAP, ocean pollution, sea surface salinity, dissolved oxygen, Google Earth Engine, marine debris, NOAA