An application of artificial neural networks to carbon, nitrogen and phosphorus concentrations in three boreal streams and impacts of climate change

We designed artificial neural networks to model daily total organic carbon (TOC), total nitrogen (Ntot) and total phosphorus (Ptot) concentrations in streamwater and used the simulated concentrations to predict future fluxes under scenarios of climate change. The streams drain two forested catchments located in southern and eastern Finland. In the period 1990–2000, observed TOC, Ntot and Ptot concentrations were in the range 2–60 mg C L−1, 0.1–1.4 mg N L−1 and 1–60 μg P L−1 with mean discharge per unit area in the range of 10–15 L s−1 km−2 for the three streams. cial neural networks consisting of 13 input variables, 1 hidden layer with 7 nodes and 1 output variable each were trained with the backpropagation algorithm to estimate the concentration of TOC, Ntot and Ptot in streamwater. Daily air temperature, precipitation and runoff observations were included in the input variables as well as catchment characteristics such as catchment area and the area of lakes and peatland within the catchment. The networks performed well in comparison with the alternative method, i.e. flow-weighted average concentrations. Artificial neural networks are a useful method for creating black-box models of streamwater quality in cases where the involved processes are too complex to simulate directly. ted changes in monthly temperature and precipitation under a changing climate were used to generate daily temperature and precipitation series for the 2050s. Daily runoff values for this hypothetical year were produced with an operational runoff model. Carbon and nitrogen loads were calculated for the 2050s using the neural network models of TOC, Ntot and Ptot concentrations. The low change scenario resulted in annual fluxes close to present fluxes, while the high change scenario gave an increase of approximately 26% of annual TOC, Ntot and Ptot fluxes.