MACHINE LEARNING CLASSIFICATION OF TROPHIC STATE INDEX FROM SENTINEL-2 MSI IMAGE

Trophic State Index (TSI) is a widely used metric to assess water quality and identify potential health risks for water resources use. This study tested the Random Forest (RF) and Decision Tree (DT) algorithms to classify the TSI in the Billings and Ibitinga reservoirs, São Paulo, Brazil using Sentinel-2 images. The model inputs were derived from in situ spectral data collected in the Billings reservoir, and the validation was performed with the Ibitinga/SP data. The results achieved an overall accuracy of 0.74. The Hypereutrophic TSI class was predominant and confirmed the hypothesis that the urban mechanism of São Paulo state impacts the reservoir’s water quality. The accuracy metrics can be improved by a more extensive and balanced field training dataset. The results suggest that the DT algorithm may become a valuable tool for recovering the TSI in productive turbid waters.