Machine Learning Oceanographic Data for Prediction of the Potential of Marine Resources

Abstract

Marine data and information are very important for human survival, therefore this data and information is attractive to investors because of the potential economic value. This data and information has been difficult to obtain, the solution to overcome this is by analyzing oceanographic data for 2009-2019 collected from the marine database belonging to the Agency for the Study and Application of Technology (BPPT). The data is the result of a collaborative marine survey between Indonesian and foreign researchers from various countries who sailed in various Indonesian waters. Raw oceanographic data is converted and classified into Conductivity, Temperature, and Depth (CTD) data as oceanographic data parameters identified as predictor variables (X) that are correlated with each other. CTD data is processed into numeric data attributes that have been labeled for input and training. The data was modeled using the Machine Learning (ML) type Supervised Learning (SL) method with the Decision Tree (DT), Linear Regression (LR) and Random Forest (RF) algorithms which were interpreted according to the characteristics of the CTD data. ML will learn data models to understand and store. Next, the model is evaluated using accuracy metrics by measuring the difference between the predicted value and the actual value to obtain a good prediction model. The prediction results show a salinity level of 34.0 parts per thousand (ppt), meaning that in this area of marine waters salinity will affect the solubility of Oxygen (O₂) and play a major role in the sustainability and growth of the fertility level of biological resources which is supported by sea surface temperature conditions 29.2°C. So the salinity values obtained using ML techniques and marine resource potential can be assumed to have a strong correlation. The research results show that the RF model has the lowest level of prediction error based on the values: Mean Square Error (MSE) = 0.007; Root Mean Squared Error (RMSE) = 0.082; Mean Absolute Error (MAE) = 0.007 compared to DT model: MSE = 0.008; RMSE = 0.088; MAE = 0.012 and LR model: MSE = 1.008; RMSE = 1.004; MAE = 0.281. The equivalent RF and DT models have a Determination Coefficient (R²) = 0.999, meaning that a model is created that is good at predicting, compared to the LR model with a value of R² = 0.914. The correlation between variables shows that the LR model is very linear with a Correlation Coefficient (r) = 1.000 compared to the DT model (r) = 0.621 and the RF model (r) = 0.379. Therefore the algorithm that has a value of (r) +1 has the best level of accuracy. The use of ML to predict marine resource potential is a relatively new research field, so this research has the potential to contribute data and information as a reference for innovative studies and investment decision material for investors.