Hybrid and Ensemble Machine Learning Models for Accurate Underwater Temperature Prediction in Coastal Islands of Santa Catarina

Predictions of subsurface ocean temperatures are highly relevant to marine science since theycontribute to ocean dynamics understanding and effects of climate change. This work offers an in-depthdiscussion of various machine learning methods used for predicting underwater temperatures fromtemporal data collection. A set of measurements of underwater temperatures was used, furthercomplemented with supplementary variables: geographical coordinates and information on time. Thisdataset was used both to train and to test an Extra Trees Regressor, a LightGBM Regressor, and anensemble consisting of a Stacking Regressor. Data Preprocessing Included feature creation, like extractingthe temporal features from datetime variables and dropping columns by taking an assumption aboutcolumns being irrelevant to the model's training effectiveness. The methodology used in this researchincludes validating the model by using three appropriate indicators: Mean Absolute Error, RMSE, and Rsquared (R²). Based on these three indicators, the strength of each model was indicated in someapplications but weakness in others regarding the predictive output of undersea temperature for each of theapproaches used. The results show the ensemble model performs better than singular models; an R² of0.976 is achieved, thus showing a possibility of using combinations of machine learning techniques toimprove predictive capability. Further from this study, predictions realized using the model can becompared to actual available data in relation to underwater temperatures, therefore acting as a basis forfurther studies in this area. Beyond supplementing the body of knowledge already existing with regard tothe prediction of subaqueous temperatures, it holds very important implications for marine scientistsbeyond this improvement. Results proved to highlight the importance of machine learning in enhancingthe reliability and accuracy of environmental forecasts, hence propelling a deeper understanding of marineecosystems.