Figure 4. Catchment-scale river metabolism estimates can now be developed from distributed sensor networks : (a) GPP measured in the 1200km2 Deva-Cares catchment, northern Spain (Rodríguez-Castillo et al., 2019); (b) GPP measured in the 256km2 Ybbs river, Austria (Segatto et al., 2021). Network outputs such as these can be developed as visualisation tools to aid catchment management decision-making, with dynamic updating in near-real time from linked sensors, telecommunication systems, and computational models.
In the past five years, high-resolution river network data products have become available at both national and global scales, including MERIT-Hydro and the associated GRADES dataset (Lin et al., 2019, Yamazaki et al., 2019) with 35 years of daily flow data from nearly 3 million river segments worldwide. These river network maps have enabled the further development of biogeochemical models that can be used alongside discrete location sensor data to quantify how nutrient and carbon sources, sinks, and transformations vary according to river size, flow, and season in large watershed networks. For example, Maavara et al. (2023) used the US National Hydrology Dataset (NHD Plus HR) product to develop a DOC model for the Connecticut River watershed, NE USA (Figure 1). This model calibrated GPP, terrestrial DOC loading, photo-mineralization, and respiration, partly from a sensor network at 10 locations from 1st-8th order rivers. These continuous dissolved oxygen measurements at 15-min intervals were used to estimate GPP and ER using a Markov Chain Monte Carlo algorithm, which was then scaled to estimate GPP across the entire watershed during all flows and seasons, by calibrating a random forest ML model (Appling et al., 2018)
Efforts in improving process-based metabolism models have focused on expanding estimation to a more diverse set of river environments than previously possible, including estimation in river reaches with large discontinuities (e.g. flow and water quality regulation) or river reaches with significant transient storage (Pathak & Demars, 2023). Progress in this direction is valuable for reducing uncertainties in global estimates of freshwater carbon fluxes. Such process-based models could facilitate large-scale assessments of metabolism and its drivers across river environments, when combined with ML methods (Appling et al., 2018, Bernhardt et al., 2022). Several other physical properties currently overlooked in field studies may significantly impact metabolism and will need to be incorporated into future network models, for example sediment movement (Risse-Buhl et al., 2023, Schulz et al., 2023) and groundwater interactions (Galloway et al., 2019), which can have major impacts on ER.
Quantification of spatial and temporal dynamics of metabolism across river networks is important for estimating regional carbon emissions from rivers (Battin et al., 2023). However, only a few studies have focused on metabolism estimation at the river network scale (Figure 4). Rodríguez-Castillo et al.(2019) utilized the spatial stream network model to identify the factors that govern spatial variations in river metabolism within the Deva-Cares catchment in northern Spain, highlighting benthic biomass, river channel properties, and human activities as important controlling factors. Segatto et al. (2021, 2023) found that ER played a larger role in metabolic stability at the river network scale in the Ybbs River Austria, whereas GPP showed higher sensitivity to flow-induced disturbances and variations in light availability. Mejia et al. (2018) used the BAyesian Single-station Estimation (BASE) (Grace et al., 2015) model to estimate metabolism over a year at ten sites across the Methow River network in Washington State, USA. Their findings indicated that metabolism timing may vary between sites within a river catchment due to the combined influence of local physicochemical conditions, despite having similar regional climates. Metabolism studies at the river network scale are admittedly data-intensive and these approaches need to be evaluated in river systems that are heavily polluted and where water quality often varies significantly over even short distances (Casillas-García et al., 2021). In these systems the implications may be that more dense networks of fixed and robot-mounted sensors are required, alongside additional predictor datasets such as point-source input locations and land use; however, such information is increasingly becoming available with advances in sensor technology, remote sensing products, and modelling techniques including ML. Mobile robots can be used to both increase the range and spatial resolution of the data on which models are trained and validate predictive models by increasing empirical field data collection.