Abstract:Photosynthetic active radiation (PAR) attenuation is a fundamental variable for water quality, and its continuous monitoring is essential for coastal management. This work focuses on PAR vertical distribution in the southern portion of the São Sebastião Channel (SSC), São Paulo, Brazil, for the first time. We used a 6-year time series of CTD profiles obtained between 2014 and 2020. The SSC is a dynamic environment, and underwater PAR measurements are challenging in the upper few meters due to wave action and shadowing effects during sampling. We investigated how Random Forest (RF) models (a machine learning technique) would help to estimate PAR using oceanographic variables as predictors. Developed RF models used depth, turbidity, water temperature or above-water solar radiation as predictor variables performed well estimating PAR deeper than 4 m, with RMSE lower than 50 μmol s −1 m −2 . However, they presented more significant errors (RMSE above 200 μmol s −1 m −2 ) near the surface (between 0 and 1 m) and when for PAR above 500 μmol s −1 m −2 . Our approach offers an alternative to classic PAR parametric models. It requires substantially more data, implying that our models were necessarily exposed to different conditions, often mentioned as limiting factors on classic PAR parametrizations. Our results help to partially fill a gap in observed underwater PAR in the SSC, as no preliminary studies are available.

Predicting Underwater Photosynthetically Active Radiation (PAR) Using Random Forests in the São Sebastião Channel, Brazil