Instrument Bias Correction With Machine Learning Algorithms: Application to Field-Portable Mass Spectrometry

Abstract

Read online

In situ sensors for environmental chemistry promise more thorough observations, which are necessary for high confidence predictions in earth systems science. However, these can be a challenge to interpret because the sensors are strongly influenced by temperature, humidity, pressure, or other secondary environmental conditions that are not of direct interest. We present a comparison of two statistical learning methods—a generalized additive model and a long short-term memory neural network model for bias correction of in situ sensor data. We discuss their performance and tradeoffs when the two bias correction methods are applied to data from submersible and shipboard mass spectrometers. Both instruments measure the most abundant gases dissolved in water and can be used to reconstruct biochemical metabolisms, including those that regulate atmospheric carbon dioxide. Both models demonstrate a high degree of skill at correcting for instrument bias using correlated environmental measurements; the difference in their respective performance is less than 1% in terms of root mean squared error. Overall, the long short-term memory bias correction produced an error of 5% for O2 and 8.5% for CO2 when compared against independent membrane DO and laser spectrometer instruments. This represents a predictive accuracy of 92–95% for both gases. It is apparent that the most important factor in a skillful bias correction is the measurement of the secondary environmental conditions that are likely to correlate with the instrument bias. These statistical learning methods are extremely flexible and permit the inclusion of nearly an infinite number of correlates in finding the best bias correction solution.

Keywords

• neural network
• long short-term memory
• mass spectrometry
• generalized additive model
• bias
• ocean carbon