Carbon dioxide in streams
Streams play a key role in the freshwater carbon budget where carbon is transformed, transported and released to the atmosphere as CO2. The magnitude of stream carbon emissions became apparent when constructing a national freshwater carbon budget for Denmark. However, the uncertainty in current estimates are high. One way to overcome this is by improving our ability to predict the CO2 partial pressure (pCO2) in streams.
Assembling the data
In order to train a predictive model observational data is needed. While stream greenhouse gasses are rarely measured directly, pCO2 can be estimated from alkalinity, pH and temperature. This is often available from national monitoring of surface waters and we open data from Denmark, Sweden and Finland. As predictor variables we found suitable remote sensing derived products covering EU. We gathered high resolution (25 meter) data on elevation, land cover (water and wetness, forest, grassland) and coarser scale climate data such as air temperature and precipitation and runoff.
Leveraging machine learning and GIS
The EU stream network obtained from very high resolution imagery was used to delineate a realistic stream network for flow modeling. TauDEM software was used to process the elevation model and calculate catchment characteristics throughout the network. Using the annual mean pCO2 as the response and catchment characteristics as predictor variables, the performance of eight different machine learning models were assessed using nested cross-validation. The random forest algorithm provided the best accuracy and was used to predict stream pCO2 through the entire stream network in a 25 meter resolution grid. Runoff data and stream channel slope was used to estimate gas transfer velocity from empirical relationship and in turn stream CO2 flux.
Zoom in to explore the map, which shows a section of the final grid of predicted stream pCO2 (unit micro-atmosphere) which covers the stream network in Denmark, Sweden and Finland.
The results of this project have been published as a research article in the Global Biogeochemical Cycles journal. The data products, and the scripts used to create them, resulting from the analysis is openly available.
