Global Chlorophyll- a Retrieval algorithm from Sentinel 2 Using Residual Deep Learning and Novel Machine Learning Water Classification

Yotam Sherf 1 , Bar Efrati, Gabriel Rozman, and Moshe Harel BlueGreen Water Research Center, HaMiktsot Blvd. 16, Modi’in Makabim-Re’ut, 7178096, Israel · 2025

Read Global Chlorophyll- a Retrieval algorithm from Sentinel 2 Using Residual Deep Learning and Novel Machine Learning Water Classification by Yotam Sherf 1 , Bar Efrati, Gabriel Rozman, and Moshe Harel BlueGreen Water Research Center, HaMiktsot Blvd. 16, Modi’in Makabim-Re’ut, 7178096, Israel (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-24124.txt.

Introduction

Yotam Sherf and three co-authors build a global system to measure chlorophyll a in inland waters from the Sentinel-2 satellite. Chlorophyll a signals algae and water health, but telling water from clouds, glint, snow, ice, plants, and land in satellite images is hard. Their classifier learns from nearly one hundred water bodies spread around the world.

On top of that classifier, they train a model on more than thirteen thousand matched points with United States Geological Survey field data. Scenes the classifier marked as true water gave more accurate chlorophyll estimates than those it rejected. The two step system aims for stable performance across regions and water types.

Catalog record

CREATORYotam Sherf 1 , Bar Efrati, Gabriel Rozman, and Moshe Harel BlueGreen Water Research Center, HaMiktsot Blvd. 16, Modi’in Makabim-Re’ut, 7178096, Israel
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-24124

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Global Chlorophyll- a Retrieval algorithm from Sentinel 2 Using Residual Deep Learning and Novel Machine Learning Water Classification Preprint: APS/123-QED Yotam Sherf 1 , Bar Efrati, Gabriel Rozman, and Moshe Harel BlueGreen Water Research Center, HaMiktsot Blvd. 16, Modi’in Makabim-Re’ut, 7178096, Israel Abstract We present the Global Water Classifier (GWC), a supervised, geospatially extensive Machine Learning (ML) classifier trained on Sen2Cor corrected Sentinel-2 surface reflectance data. Using nearly 100 globally distributed inland water bodies, GWC distinguishes water across Chlorophyll- a (Chl a ) levels from non-water spectra (clouds, sun glint, snow, ice, aquatic vegetation, land and sediments) and shows geographically stable performance. Building on this foundation model, we perform Chl a retrieval based on a matchup Sentinel-2 reflectance data with the United States Geological Survey (USGS) AquaMatch in-situ dataset, covering diverse geographical and hydrological conditions. We train an XGBoost regressor on 13626 13626 matchup points. The positive labeled scenes by the GWC consistently outperform the negatives and produce more accurate Chl a retrieval values, which confirms the classifier’s advantage in reducing various interferences. Next, residual analysis of the regression predictions revealed structured errors, motivating a residual CNN (RCNN) correction stage. We add a CNN residual stage trained on normalized residuals, which yield substantial improvement. Our algorithm was tested on 867 water bodies with over 2,000 predictions and Chl a values up to 1000 mg / m 3 /m^{3} , achieving R 2 = 0.79 R^{2}=0.79 , MAE = 13.52 mg / m 3 /m^{3} , and slope = 0.91, demonstrating robust, scalable, and globally transferable performance without additional tuning. Harmful algal blooms 0 0 footnotetext: Corresponding author: [email protected] (HABs) have become increasingly prevalent across inland and coastal waters, threatening drinking water quality, producing toxins and greenhouse gasses, and disrupting aquatic ecosystems with mounting economic and public health costs gobler2020climate ; paerl2016controlling ; kaplan_harel_2025_bloomGHG . Their growing frequency and intensity underscore the urgent need for continuous, spatially extensive monitoring. Traditional field sampling, while accurate, is costly and geographically limited, prompting the rapid expansion of satellite-based monitoring as an indispensable complement to in situ observations. Satellite remote sensing enables global-scale HAB surveillance with frequent revisit times and high spatial resolution. Instruments such as Sentinel-2’s MultiSpectral Imager (MSI) provide imagery at 10–20 m resolution every five days, making them particularly valuable for small and optically complex inland waters ioccg2006report5 ; mishra2013phycocyanin ; llodra2023review . Yet, accurate Chl a retrieval, the principal proxy for phytoplankton biomass, remains challenging, as observations are prone to natural and instrumental perturbations that can substantially alter the expected optical response of Chl a pahlevan2019seamless ; moses2012evaluation ; llodra2023review . Ideally, we would expect a predictable linear (or at least near linear) relationship of Chl a to the spectral imprint. However, in practice, the signal deviates due to non-algal particles such as turbidity, colored dissolved organic matter (CDOM), sediments, and others, which modify the optical signature. At high Chl a levels (often referred as scum), optical saturation and multiple non-linear scattering further distort the spectrum moses2012evaluation ; pahlevan2019seamless . In addition, imaging artifacts such as sun glint, cloud shadows, snow or ice, and adjacency effects frequently degrade data quality. Collectively, these optical and atmospheric perturbations introduce nonlinear and nonstationary distortions that complicate model transferability across lakes, seasons, and sensors he2022deep ; llodra2023review . Despite these challenges, empirical algorithms remain widely used in operational monitoring systems due to their transparency, computational simplicity, and compatibility with long-term archives matthews2014eutrophication ; lobo2021algaemap . Over the past two decades, diverse strategies have been proposed to improve Chl a estimation moses2012evaluation ; lobo2021algaemap ; pahlevan2019seamless ; gilerson2010retrieval . Index-based and band-ratio algorithms remain the most interpretable and computationally efficient options but often fail under optically variable conditions moses2012evaluation ; lobo2021algaemap . Semi-analytical inversion models, which couple reflectance with radiative-transfer theory, improve physical interpretability but depend on precise atmospheric correction and reliable prior knowledge of optical properties zhang2022bayesian . Increasingly, machine-learning (ML) approaches, including Random Forest (RF), gradient-boosted trees, and deep neural networks, are demonstrating strong potential to model nonlinear relationships between reflectance and Chl a across diverse optical regimes he2022deep ; joshi2024cross ; li2025mlretrieval . Recent advances in interpretable and hybrid learning architectures have further improved model transparency and physical consistency zhang2022bayesian ; zhong2022interpretable , enabling more robust generalization across optical conditions. These models have been successfully applied to multisensor datasets (e.g., Sentinel-3 OLCI, MODIS, and MSI) to predict algal biomass and detect bloom dynamics at regional to continental scales pahlevan2019seamless ; he2022deep . However, domain shift, inconsistent atmospheric corrections, and limited labeled training data remain critical barriers to reliable generalization wynne2010detecting ; kudela2015habsummary ; gilerson2010retrieval ; yacobi2011phytoplankton ; dev2022a ; dev2022b ; dasilva2024chlorophyll , as recent multi-lake studies have demonstrated steep performance declines outside the training domain dasilva2024chlorophyll . Additionally, many of these prior studies optimized for a single lake or region, which is not applicable across diverse inland waters without site-specific tuning joshi2024cross ; dasilva2024chlorophyll . Recent operational systems such as AlgaeMAp for Latin American inland waters demonstrate the feasibility of near-real-time bloom monitoring at large scales, but still rely primarily on empirical indices and thresholding rather than advanced data-driven frameworks lobo2021algaemap . At the same time, emerging deep-learning methods that integrate physical constraints and spectral–spatial fusion have shown notable improvements in capturing complex optical variability and cross-sensor consistency he2022deep ; zhang2022bayesian . Our framework leverages this paradigm by coupling global water classification with interpretable residual learning, integrating physical understanding with data-driven inference zhong2022interpretable ; zhang2022bayesian . These developments collectively emphasize the need for robust, generalizable frameworks that unify data cleaning, feature learning, and predictive modeling across heterogeneous aquatic environments. To address these challenges, this study introduces a three-stage ML framework for global Chl a mapping from Sentinel-2 MSI imagery. The first stage employs a Global Water Classifier (GWC) trained on globally distributed spectra to probabilistically retain reliable water pixels and remove environmental artefacts such as clouds, ice, aquatic vegetation, and sediment. The second stage applies an XGBoost regression model trained on 13,626 georeferenced in situ Chl a samples from the USGS AquaMatch dataset brousil2024_aquamatch to learn nonlinear spectral relationships. Finally, a Residual Convolutional Neural Network (RCNN) refines predictions by modeling structured residual errors between observed and predicted Chl a , effectively reducing bias while preserving cross-lake generalization. This residual-learning design conceptually parallels recent deep fusion networks that blend physical and data-driven principles for improved optical retrievals he2022deep ; zhang2022bayesian but extends their application to the residual domain. Together, these stages deliver a globally scalable, physically informed, and data-driven solution for robust Chl a estimation across diverse inland waters. The novelty of our framework lies in its integrated three-stage workflow. Stage 1 (GWC) acquires analyzable water spectra by removing optical artifacts and isolating true bloom signatures, providing clean inputs for downstream modeling. Stage 2 (XGBoost) learns interaction-rich relationships among bands and indices using 5,000 AquaMatch in situ points, yielding accurate base Chl a predictions consistent with evidence that tree-based learners perform robustly across sensors and optical regimes joshi2024cross ; he2022deep ; li2025mlretrieval . Stage 3 (residual 1-D CNN) identifies and corrects structured residuals across Chl a regimes, improving accuracy and preserving cross-lake generalization. The result is a scalable pipeline suitable for diverse geographic regions and environmental conditions, reducing the need for site-specific tuning

Read the complete text