AI Maps the Seafloor of Egypt’s Largest Lagoon From Space With Centimeter Precision

Mapping the depth of a shallow, murky, vegetation-choked lagoon the size of Manzala in Egypt’s Nile Delta has long been a slow, expensive, and sometimes dangerous undertaking. Conventional echo-sounder surveys require boats to crisscross the water again and again, and in a lagoon dotted with fishing structures, dense aquatic plants, and shifting sediments, that work is both logistically punishing and financially draining. Now a team of Egyptian researchers has shown that a carefully designed machine-learning pipeline, fed by nothing more exotic than freely available Landsat 8 satellite imagery and a set of field measurements, can reproduce the lagoon’s bathymetry with an accuracy that rivals the surveys themselves.

The study, published in Environmental Monitoring and Assessment by Hesham M. El-Asmar and Mahmoud Sh. Felfla of Damietta University and Hussein M. Rashad of Egypt’s Ministry of Environment, comes at a pivotal moment for the lagoon. Between 2017 and 2022, Manzala underwent extensive restoration and dredging, reshaping its channels and circulation patterns. Managers suddenly needed up-to-date, high-resolution depth maps to feed hydrodynamic models, track sediment movement, and plan further interventions. Traditional surveying over such a large, shallow water body was impractical, so the researchers turned the problem over to algorithms.

The core idea behind satellite-derived bathymetry is elegant: light penetrates clear water, and the deeper the bottom, the dimmer and more color-shifted the reflected signal becomes when it finally reaches a sensor in orbit. Blue and green bands carry most of the depth information, while red light is absorbed quickly. But Manzala is anything but clear. Turbidity, suspended sediments, algal activity, and bottom vegetation all scramble the optical signal, which is precisely why simple empirical formulas developed for tropical coral reefs tend to collapse in such environments. Machine learning, the researchers reasoned, could learn the messy, nonlinear relationships between spectral reflectance and actual depth if it was given enough trustworthy ground truth.

That ground truth came from extensive echo-sounder measurements collected across the lagoon, which were paired with reflectance values extracted from Landsat 8 multispectral imagery. Seven machine-learning algorithms were then pitted against one another, spanning tree-based ensembles and boosting-based regressors. The evaluation was deliberately thorough: rather than relying on a single correlation coefficient, the team examined regression metrics, depth-class performance, uncertainty estimation, and, crucially, an interpretability analysis using SHAP values, a technique borrowed from explainable artificial intelligence that reveals how much each input feature contributes to each prediction.

The results were striking. Tree-based ensembles consistently outperformed the boosting-based regressors, with Decision Tree and Extra Trees models achieving the highest test-set accuracy under the adopted validation design, reaching a correlation coefficient of 0.996, an R-squared of 0.992, and a root-mean-square error of just 0.073 meters. In practical terms, the satellite-based predictions deviated from the echo-sounder measurements by less than eight centimeters on average, a level of precision that would have seemed implausible for a turbid lagoon only a few years ago.

Yet the researchers were careful not to crown a winner on accuracy alone. Two models that score identically on a test set can behave very differently when deployed across a full lagoon, particularly in areas where the optical conditions differ from the training data. When the team quantified prediction uncertainty, Extra Trees emerged as the more dependable choice, with uncertainty of roughly 0.09 meters, substantially lower than its equally accurate competitor. That distinction matters for operational mapping, where a map that is confidently wrong is far more dangerous than one whose errors are known and bounded. For the purposes of managing a living lagoon, the Extra Trees model was deemed the most reliable.

Perhaps the most scientifically interesting finding came from the SHAP analysis, which identified the near-infrared band as the single most influential predictor. On the surface, this is counterintuitive: near-infrared light is absorbed within the first few tens of centimeters of water, so it cannot directly sense the bottom of a lagoon several meters deep. The authors interpret its dominance as evidence that the model is exploiting surface and turbidity-related effects rather than true depth sensing, effectively using the near-infrared signal as a proxy for the optical state of the water column. This is a valuable caution for the rapidly growing field of satellite-derived bathymetry: a model can achieve spectacular statistics while leaning on correlations that may not generalize to other water bodies or seasons. The transparency offered by SHAP-based interpretation allows practitioners to spot such behavior before it leads them astray.

The study also highlighted the value of feature engineering rooted in physics. Logarithmic band ratios, a transformation descended from classical empirical bathymetry methods dating back to Lyzenga’s foundational work in the late 1970s and Stumpf’s ratio-based algorithms of the early 2000s, improved model robustness by capturing nonlinear spectral relationships while partially canceling out variable bottom reflectance. In other words, the best results came not from throwing raw pixels at a black-box algorithm, but from combining decades of optical oceanography with modern ensemble learning, then interrogating the result with explainability tools.

The final bathymetric maps resolved depth gradients spanning roughly negative 0.6 to negative 4.0 meters across the lagoon, capturing the shallow fringes, dredged channels, and deeper basins that govern how water moves through the system. Such maps are not merely cartographic trophies. They are the geometric foundation for hydrodynamic models that simulate circulation and flushing, for sediment transport studies that predict where dredging will need to be repeated, and for ecological assessments of habitats that depend on specific depth and light regimes. For a lagoon that supplies much of Egypt’s fish catch and buffers the densely populated Nile Delta coast, that information carries real economic and environmental weight.

Beyond Manzala, the framework’s significance lies in its scalability and cost profile. Landsat data are free, the algorithms are open and interpretable, and the only expensive ingredient, the echo-sounder calibration campaign, can be far smaller than a full survey because the machine-learning model interpolates between calibration points. The authors position the approach as a transparent, efficient alternative to exhaustive field programs, one that can be repeated as conditions change, whether through further restoration work, sedimentation, or sea-level-driven alterations to the Delta’s fragile coastal wetlands. As turbid shallow waters around the world face mounting pressure from climate change and development, the Manzala study offers a template: pair trusted field measurements with ensemble machine learning, demand uncertainty estimates alongside accuracy statistics, and never trust a model you cannot explain.

Subject of Research: Satellite-derived bathymetry mapping of a turbid coastal lagoon using comparative machine-learning models

Article Title: Comparative machine-learning retrieval of high-precision bathymetry from multispectral remote sensing: a case study of Manzala Lagoon, Egypt

Article References: El-Asmar, H. M., Felfla, M. S., & Rashad, H. M. (2026). Comparative machine-learning retrieval of high-precision bathymetry from multispectral remote sensing: a case study of Manzala Lagoon, Egypt. Environmental Monitoring and Assessment, 198(10), Article 1049. https://doi.org/10.1007/s10661-026-15878-x

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15878-x

Keywords: satellite-derived bathymetry, machine learning, Landsat 8, Manzala Lagoon, Nile Delta, remote sensing, SHAP interpretability, Extra Trees, coastal lagoon management, echo-sounder calibration, turbid waters, Environmental Monitoring and Assessment