POSTER Session 2

A BIO-OPTICAL ALGORITHM BASED ON MACHINE LEARNING TO OPTIMIZE CHLOROPHYLL-A RETRIEVAL IN TURBID AND CLEAR WATERS: AN APPLICATION TO MODIS

Accurate estimation of chlorophyll-a (Chl-a) concentration across varying aquatic environments presents a considerable challenge in remote sensing, particularly due to the diversity of Optical Water Types (OWTs) implying the necessity to use the whole spectral information available. In this study, we introduce a novel bio-optical algorithm named CONNECT (Combination Of Neural Network models in Estimating Chlorophyll-a over Turbid and clear waters), integrating two Multi-Layer Perceptron (MLP) models, specifically designed for MODIS sensor, to improve the accuracy of Chl-a estimates. Each model was optimized for different turbidity levels: the first model addresses clear to medium turbid waters, the second targets turbid to ultra-turbid waters. These two machine-learning based models were trained using an extensive global dataset including 6747 paired observations of both hyperspectral and multispectral remote sensing reflectance (Rrs) and Chl-a concentrations, ranging from 0.017 to 2587 µg/L with an average of 25.9 µg/L. The combination of these models was achieved through a logistic regression framework that utilizes the probability outputs of each network to yield the most accurate Chl-a concentration. Validation on in-situ and matchup datasets illustrates significant improvements in Chl-a retrieval, especially in coastal waters, where the CONNECT algorithm outperforms existing approaches, such as OC3M and conventional Red/NIR algorithms. This research demonstrates the effectiveness of specialized machine learning and class-based combination techniques in improving Chl-a estimates across varied water conditions, offering valuable tools for environmental management exploiting MODIS long lasting time series.

Manh Tran, Laboratoire d’Océanologie et de Géosciences, France, https://orcid.org/0000-0003-4770-5561

Vincent Vantrepotte, Laboratoire d’Océanologie et de Géosciences, France

Daniel Jorge, Laboratoire d’Océanologie et de Géosciences, France

Cédric Jamet, Laboratoire d’Océanologie et de Géosciences, France

Code of Conduct
Privacy Policy