Técnicas computacionais aplicadas na estimativa da pressão parcial de Co2 em águas oceânicas

Abstract

Coastal regions have little data regarding pCO2in marine waters. In part, this is due to the great difficulty of modeling these biogeochemically active regions, due to the complexity of coastal processes, the possible environmental changes and their effects on ecosystems and the high logistics costs required for these measurements in the ocean. Therefore,estimates based on a group of measurement variables that are easier to performareeventually necessary. The estimate of pCO2is recorded in several studies, however its application has been directed tothe open ocean and not to continental shelves. Multiple linear regression (MLR) and artificial neural network (ANN) techniques have been confronted several times.However, the use of the Levenberg-Marquardt algorithm, commonly used in computational vision when it is desired to fit a model to a set of data, isto date, not used for the estimation of pCO2inthe Brazilian coast. Thus, in this study, machine-learningalgorithms were used to estimate pCO2on the equatorial northeast coast of Brazil, using statistical and/or artificial intelligence techniques. For the training of the computer techniques, the data of the MS Monte Olivia ship were used, in order to certify them in international databases, and later applied these techniques to the Cruzeiro do Sul (NHOCS) data of the Brazilian Navy obtained by Coastal Biogeochemistry Research Group. Data were analyzed for pCO2, sea surface temperature (SST) and sea surface salinity (SSS). We obtained the Residual Standard Deviations (RSD) to justify the best technique based on the smallest error. The RNA presented better results in relation to MLRwith the obtaining of RSDs equivalent to 0.1034µatm for the total data of Monte Olivia and 0,8394µatm for the total data of the Cruzeiro do Sul, that is, smaller values than those published in the literature. In addition, for the open ocean,the SST and SSSmeasuresare more significant and represent better the pCO2than inthe continental shelf.By means of these two only variables, the estimate of the pCO2was not satisfactory. However, even with a lower number of data, adding a new variable, chlorophyll a, ANNestimation showed a satisfactory result, with RSD equivalent to 0.5983µatm. When only Salinity and Chlorophyll are combined, the RSD via RNA drops to 0.3514µatm for the estimated data after normalization.