Machine Learning-Based Quality Control for Oceanographic Sensor Data

Kurzfassung

Ensuring the quality of oceanographic underway data is critical for accurate environmental monitoring and climate research. Traditional quality control (QC) methods often require expert knowledge and manual intervention. This study explores the potential of using Random Forest models to automate and enhance QC processes for oceanographic sensor data. A Random Forest Regressor is trained to model key variables, including temperature and salinity, while a Balanced Random Forest Classifier is used to predict QC flags, distinguishing data points of good and bad quality. Experimental results demonstrates strong regression performance on a held-out test set, with R2 values between 0.84 and 0.95 and good generalizability to previously unseen geographic regions. Classification results varies, with salinity quality flags achieving a balanced accuracy of 93%, while temperature quality flag predictions achieves 81% balanced accuracy, suffering from a higher false-negative rate. Analyzing the models using Explainable AI methods shows that the temperature quality flag classification overfits to the location of the vessel, which the regression model and salinity quality flag classification do not. These findings indicate that machine learning can effectively replicate traditional QC methods and improve efficiency by reducing manual workload, though further refinement is needed to improve the performance.