Forecasting Discharge and Water Levels of Rivers and Dams using Earth Observation and AI

The exploitation of rivers and hydropower reservoirs involves daily monitoring of  the water resources, the meteorological conditions, the status of  the river banks, the flood areas, etc. As maximum river discharge often results in flooding, it is of importance to provide with timely and reliable forecasts of discharge and water levels. Predicting river discharge and water levels has been a subject of hydrological modelling and a topic of serious research. However, only in recent years scholars and practitioners have turned to consider earth observation data for their studies, mainly to compare evidence of flood mapping. We present an approach of using earth observation data to feed AI architectures – EO4AI – and produce forecasts for discharge and water level with significant degrees of accuracy.  Our starting point is that river discharge and water levels depend on a variety of meteorological and environmental factors like precipitations, snow cover, soil moisture, vegetation index and satellite data offer rich variety of datasets, supplying this information. We adopt a pipeline of deep learning architectures consisting of GAN, CNN, LSTM and EA to actually generate forecasts for river discharge and water level by using historic satellite data of the meteorological features listed above, and in-situ measurements for water level and discharge. The satellite data are provided by ADAM  via the NoR service of ESA. ADAM provides data access to satellite datasets from different satellites with semantic relevance for the construction of sediment transport and deposition forecast model as discussed above.  We explain the purpose of the pipeline components. Our forecast models are calibrated for 3, 5, 7, 30 days ahead, and our experiments provide predictions for one year ahead with each of the calibrated models. We discuss experiment results carried out with data from the Danube and Arda rivers, including three dams from cascade Arda and compare them with predictions derived with other methods. We demonstrate the viability of the approach and the reliability of the forecasting results. We further show how the forecasts can be used in hydrodynamic modelling context, for early warning applications and for routine water resources management and monitoring tasks.