Poster Session 4, Thursday, October 6, 11:00–12:40
Machine-learning methods to correct optical data affected by biofouling
With technological advancements in high capacity data and energy storage solutions and real-time data communications, the biggest remaining challenge with long-term ocean observations is biofouling. Marine growth on oceanographic sensors is inevitable. This is a problem, particularly for optical sensors, where micro-, macro-growth or biofilms on optical windows yields useless data. We evaluated the efficacy of machine-learning techniques for filling data gaps in long time series optical observations caused by biofouling (or other equipment issues). Three different machine-learning methods were explored: supervised (partial least squares [PLS] regression), artificial neural network (ANN), and artificial recurrent neural network (RNN). Model predictors were spectral absorption coefficients measured by an ac-9, with simulated periods of biofouling. Model responses were current velocity, depth, temperature, salinity, backscattering, and chlorophyll fluorescence, which are usually unaffected by biofouling. Training and testing datasets were created, assuming between 20% and 30% of data were affected by biofouling. Results indicate that the RNN (Long Short-Term Memory [LSTM]), is more accurate in filling data gaps for spectral absorption coefficients as compared to PLS or the ANN (Nonlinear Autoregressive Network [NARX] with exogenous inputs); NARX performs the worst. LSTM modeled versus measured R 2 values average 0.66; higher R 2 are found for blue wavelengths. Model II slopes range between 0.93 and 1.04. The coefficient of variation of root mean-square error is lowest for blue wavelengths (<0.06) and highest for red (0.09). These results indicated that machine-learning methods show promise for filling data gaps that are compromised due to biofouling.
Galen Egan, Sofar Ocean Technologies, [email protected]
Frank Spada, Integral Consulting Inc., [email protected]
Stephen Monismith, Stanford University, [email protected]
Oliver Fringer, Stanford University, [email protected]