Four headwater river sections were instrumented in the Upper Aragón river catchment (located in the Central Spanish Pyrenees). One river reach has snow-melt dominated floods and the other three river reaches have a combination of snow-melt and precipitation dominated floods. We recorded in each river section streamflow values, suspended sediment concentration values and water temperature values. In parallel, we ran aerial topographic surveys with an Unmanned Aerial Vehicle (UAV) and we collected sediment data such as superficial sediment grain size distribution (through ground surveys) and sediment size in suspension (through an ISCO sampler and lab analysis). This monitoring occurred before and after flood events capable of mobilizing sediment transport. Furthermore, the field measurements included the release of active Radio Frequency Identification (RFID) tags, which allowed to infer travelling distances, sediment velocities and sediment diffusion characteristics. We also deployed geophones on river banks to capture the ground vibrations caused by bedload particles interacting with the river bed. This allows to relate streamflow values and river seismic signals to infer bedload transport estimates. Finally, we implemented a robust, quasi-automatic, protocol to collect aerial photos taken by UAVs and to post process that information in order to infer topographic data and sedimentological data (e.g. superficial grain size distribution).
Conversely, we used data-driven methods to compare their performance with traditionally lumped process-based modeling approaches for streamflow prediction. In particular, we used Long Short-Term Memory (LSTM) machine-learning models for punctual streamflows prediction and Convolutional Neural Networks (CNN) for including spatial features within a catchment area. We used as a benchmark a rich database from the Ebro River Water Authorities (CHE), with daily information since before 1990. Data was curated and homogenized in order to be used with this type of data-driven models. Climatic forcing (precipitation, air temperature, radiation), atmospheric indices (NAO, WeMO) and drought indices (SPEI) were also taken into account. These data-driven methods require less spatially distributed information, less computational demand and can outperformed traditional process-based hydrological models. Furthermore, we tested a physical constraint imposed in the data-driven model to guarantee the mass conservation of the results. This physical constraint helps to better predict the timing and magnitude of flood events, with particular focus on extreme events. This in turn ease the computation of sediment transport estimates conveyed by each flood event.
We defined socio-economic scenarios (which include both climatic forcing and land use-land cover characteristics) at the spatial scale required to infer the impact of Global Change at headwater river systems. Climate data and land use information was downscaled to a regional (finer) resolution based on neighborhood rules and stochasticity in the placement of areas with significant changes. This allows to infer water and sediment fluxes future evolution at headwater river systems.
