Ecologia e Fisiologia - Pôsteres - 28/11/2023 - 15:30 h às 16:30 h

Photosynthetic Active Radiation (PAR) attenuation is a fundamental variable for water quality, and its continuous monitoring, particularly in urbanized areas, 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 using a 6-year time series of CTD casts obtained between 2014 and 2020. The SSC is a dynamic environment, and measurement of subsuperficial PAR measurements are especially challenging in the upper few meters due to wave action and shadowing effects during sampling. Here we investigated how Random Forest (RF) models (a machine learning technique) would help to estimate PAR using oceanographic variables as predictors, chosen among 12 available in total. Water temperature, salinity, depth, chlorophyll-a fluorescence, dissolved oxygen, turbidity, atmospheric pressure, wind speed, wind direction, air temperature, above-surface PAR and astronomical season. 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 larger 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. Still, we show that it is possible to estimate PAR from a CTD equipped with a turbidity sensor, thus offering an alternative to classic PAR parametric models, although requiring substantially more data. This implies that our models were necessarily exposed to different conditions, which are 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, so this study may be used as a blank for future scientific investigations and local monitoring actions.