Forecasting Lakes' Chlorophyll Concentrations Using Satellite Images and Generative Adversarial Networks

• Published in: Water resources research Vol. 60; no. 10
• Main Authors: Nagkoulis, Nikolaos, Vasiloudis, Giorgos, Moumtzidou, Anastasia, Gialampoukidis, Ilias, Vrochidis, Stefanos, Kompatsiaris, Ioannis
• Format: Journal Article
• Language: English
• Published: Washington John Wiley & Sons, Inc 01.10.2024; Wiley
• Subjects: Algorithms; Atmospheric correction; C2RCC; Chlorophyll; Chlorophylls; chlorophyll‐a; Computing costs; data collection; Data sampling; Datasets; Europe; forecasting; generative; Generative adversarial networks; Image quality; Lakes; Natural phenomena; Pattern recognition; remote sensing; Satellite data; Satellite imagery; satellites; Sentinel‐2; surface water; Training; Water bodies; Water monitoring; Water quality; Water quality management; Water quality monitoring; Water resources; Water resources management; Europe
• ISSN: 0043-1397; 1944-7973; 1944-7973
• DOI: 10.1029/2024WR037138

Satellite data are extensively used for water quality monitoring purposes, offering a significantly reduced cost compared to in situ data sampling. Using past measurements to predict future conditions remains a challenging task, because of the complexity of the natural phenomena that are involved, with great potential in terms of water resources management. This paper proposes a model that can be used to forecast Chlorophyll‐α $\alpha $ (Chl‐α $\alpha $) values in water bodies, which are a common water quality indicator. The operation of the model lays on the fact that typically Chl‐α $\alpha $ increases and decreases periodically. First, we apply C2RCC, which is a common atmospheric correction algorithm, to Sentinel‐2 images to get Chl‐α $\alpha $ maps for 15 lakes for 12 consecutive months around Europe. Then, we use this data set (∼ ${\sim} $1,000 Sentinel‐2 images) to train a Generative Adversarial Network (GAN) to recognize spatiotemporal patterns. To accomplish this task, pix2pix algorithm is employed, matching consecutive past and current Chl‐α $\alpha $ maps to future Chl‐α $\alpha $ maps. This model has been applied to 3 water bodies around Europe that are not included in the 15‐lakes training data set and has been found to perform accurately, achieving high Pearson and Spearman correlations and low RMSE values. Overall, the model can be used to make Chl‐α $\alpha $ maps' predictions with low computational cost and without using any in situ data and without the requirement of training for every water body. Key Points Use of Sentinel 2 data for creating inland water bodies' Chlorophyll‐α $\alpha $ time‐series Generative Adversarial Networks are trained to recognize Chlorophyll‐α $\alpha $ spatio‐temporal patterns A continental‐scale model is created for Chlorophyll‐α $\alpha $ short term predictions