Long-term element budgets from small watershed have provided insights into how forested ecosystems regulate biogeochemical cycles. Advances in sensing technology have provided higher resolution streamwater concentration data, which has improved estimates of small watershed element budgets and facilitated the generation of new hypotheses on ecosystem controls on streamwater solutes. Additionally, machine learning algorithms have aided the analysis and utility of new large data sets produced by advanced sensing. These algorithms demonstrate great skill in transforming multi-dimensional data into robust predictions or novel patterns.
Results/Conclusions
We applied machine learning algorithms to advanced sensor data to estimate high-resolution concentrations of streamwater water quality constituents at the Hubbard Brook Experimental Forest. Sensed dissolved oxygen, pH, specific electrical conductivity, turbidity, absorbance at 254 nm, and fluorescent dissolved organic matter were used in a Support Vector Machine model to produce 15-minute estimates of streamwater concentrations of non-measures solutes (e.g., Ca, Na, K) between 2013 and 2017. The predictions were reasonably strong with coefficient of determinations ranging from 0.5 to 0.8. This allowed us to estimate element fluxes at the storm event scale and evaluate the hydrologic and environmental conditions that influence these fluxes. The application of machine learning with sensed data provided new insights into catchment ecosystem processes that are likely influencing catchment element budgets.