Chlorophyll-a and total suspended solids retrieval and mapping using Sentinel-2A and machine learning for inland waters

• Published in: Ecological indicators 2020-06, Vol.113, p.106236, Article 106236
• Main Authors: Saberioon, Mohammadmehdi, Brom, Jakub, Nedbal, Václav, Souc̆ek, Pavel, Císar̆, Petr
• Format: Article
• Language: English
• Subjects: Cubist modelling; Fish ponds; Monitoring; Remote sensing; Small inland waters; Water quality
• ISSN: 1470-160X; 1872-7034
• DOI: 10.1016/j.ecolind.2020.106236

•Various spectral band and spectral indices were used to model Chl-a and TSS.•Sentinel-2 coupled with machine learning is a promising tool for fast and cheap monitoring of inland waters.•Data-driven methods for predicting water quality parameters improves the overall predictive accuracy for complex spectral relationships and interactions. Chlorophyll-a (Chl-a) and Total Suspended Solids (TSS) are both key indicators of the biophysical status of inland waters, and their continued monitoring is essential. Existing conventional methods (e.g., in situ monitoring) have shown that they are impractical due to their time and space limitations. The recently operated Sentinel-2A satellite offers the potential to have higher temporal, spatial, and spectral resolution images with no cost for monitoring water quality parameters of inland waters. The main aim of this study was to develop a semi-empirical model for predicting water quality parameters by combining Sentinel-2A data and machine learning methods using samples collected from several water reservoirs within the southern part of the Czech Republic, Central Europe. A combination of 10 spectral bands of the Sentinel-2A and 19 spectral indices, as independent variables, were used to train prediction models (i.e., Cubist) and then produce spatial distribution maps for both Chl-a and TSS. The results showed that the prediction accuracy based on Sentinel-2A was adequate for both Chl-a (R2=0.85,RMSEp=48.57) and TSS (R2=0.80,RMSEp=19.55). The spatial distribution maps derived from Sentinel-2A performed well where Chl-a and TSS were relatively high. The temporal changes in both Chl-a and TSS could be seen in the distribution maps. The temporal changes are showing that The values of TSS dramatically changed in fishponds compared to sand lakes over time which might be due to indifferent management practices. Overall, it can be concluded that Sentinel-2A, when coupled with machine learning algorithms, could be employed as a reliable, inexpensive, and accurate instrument for monitoring the biophysical status of small inland waters like fishponds and sandpit lakes.