Predicting dissolved organic carbon concentration in a dynamic salt marsh creek via machine learning

Abstract

Dissolved organic carbon (DOC) is a master variable in aquatic systems. Resolving DOC dynamics requires high‐temporal resolution data. However, DOC concentration cannot be directly measured in situ, and discrete sample collection and analysis becomes expensive as temporal resolution increases. To surmount this problem, an option is to predict site‐specific DOC concentration with linear modeling and optical data predictors collected from high‐cost, high‐maintenance in situ spectrophotometers. This study sought to improve upon the accuracy and field costs of linear predictive DOC methods by using machine learning modeling coupled to low‐to‐zero cost predictors. To do this, we collected 16 months of in situ data (e.g., spectrophotometer attenuation, salinity, temperature), assembled freely available predictors (e.g., point in year, rainfall), and collected samples for DOC analysis, all in a salt marsh creek. At seasonal timescales, machine learning (coefficient of determination [R2] = 0.90) modestly improved upon the accuracy of linear methods (R2 = 0.80) but offered substantial instrumentation cost reductions (~ 90%) by requiring only cost‐free predictors (online data) or cost‐free predictors paired with low‐cost in situ predictors (temperature, salinity, depth). At intertidal timescales, linear methods proved ill‐equipped to predict DOC concentration compared to machine learning, and again, machine learning offered a substantial instrumentation cost reduction (~ 90%). Although our models were developed for and applicable to a single site, the use of machine learning with low‐to‐zero cost predictors provides a blueprint for others trying to model DOC dynamics and other analytes in any complex aquatic system.

Self‐diagnosis of model suitability for continuous measurements of stream‐dissolved organic carbon derived from in situ UV–visible spectroscopy

“…This can improve the model by capturing additional samples, particularly those with lower and high S R values, in a new training dataset. Furthermore, using a site and temporal variable, such as a point‐in‐year metric (Codden et al 2020), can help the model distinguish site‐specific and temporally‐specific signatures. Following the implementation of these new variables, model performance can further be evaluated with and SPE to find out if they are likely to improve the model application across sites and periods.…”

“…Partial least squares (PLS) regression is commonly used in the field of chemometrics (Wold et al 2001) and, in recent years, has been successfully used to train models to monitor in situ DOC concentration from UV–Vis absorption of CDOM (Langergraber et al 2003; Avagyan et al 2014; Etheridge et al 2014; Vaughan et al 2017; Codden et al 2020; Zhu et al 2020; Table 1). The PLS method allows for the whole spectrum of UV–Vis absorption to be used to best predict DOC concentration without problems due to multicollinearity while providing consistency checks for new data used to inform the model's performance.…”

Deep learning of estuary salinity dynamics is physically accurate at a fraction of hydrodynamic model computational cost

“…Alizadeh et al (2018) showed that ML models could accurately forecast estuarine water quality constituents, including salinity, up to 2 h in the future (Alizadeh et al 2018). Other studies have demonstrated how ML models can reduce cost in data collection and computation (Codden et al 2021). The flexibility and computational efficiency of such models raises exciting possibilities about their capacity to harness ever‐growing data availability to develop accurate models at speeds and resolutions not possible with traditional modeling approaches.…”

Seasonal drivers of dissolved oxygen across a tidal creek–marsh interface revealed by machine learning

“…The Random Forest algorithm is becoming increasingly popular as a simple-to-use machine learning method that allows for estimation of predictor importance (Breiman 2001). Recent examples of Random Forest implementations in aquatic biogeochemistry, including developing high-frequency proxies for parameters with limited temporal resolution (Castrillo and García 2020;Codden et al 2021;Green et al 2021), and linking in situ and remotely sensed parameters (Gu et al 2020), show great promise for understanding complex coastal biogeochemical processes. However, applying Random Forest to time series requires careful consideration of temporal dependence and other model parameterization decisions (Regier et al 2022).…”