Every now and then, water reservoirs fill up and lots of water are spilled, surprising the downstream inhabitants with very sudden and unexpected floods. This is a relatively common phenomenon. Therefore, in the TWIGA project, we decided to start designing a service that determines how full the reservoirs are based on satellite imagery.
We estimate the water level (altitude with respect to the sea level) by using radar ESA Sentinel-1 imagery, Deep Learning (DL), and a surface elevation map. A DL algorithm is designed and trained to detect water from Sentinel-1 imagery by using large reference datasets generated at the water reservoirs of interest. The outcome is a prediction of water level every ~6 days, whenever there is a Sentinel-1 image available.
Figure 1 shows an example of water levels estimated from Sentinel-1 imagery at one single site of analyses of ~2.5×2.5 km2 at the Bagre reservoir (Ghana).
Figure 2 shows the temporal evolution of the water level at the Bagre reservoir, according to our algorithm and according to the Copernicus Global Land Service Water Level product. The Copernicus product is based on a sensor that only “sees” the earth’s surface in cloud-free conditions. Therefore, on the cloudy season, on the rainy season when the reservoirs fill up more quickly, the Copernicus product does not provide estimates (from June onwards). Conversely, the Sentinel-1 radar sensor is a little sensitive to clouds and hence we can provide estimates systematically every ~6 days.
Figure 3 shows an example of the temporal evolution of the water level with associated uncertainties. In this case for the Bui reservoir (Ghana). A steady increase in the water level can be appreciated from mid-August onwards. Copernicus Water Level product is not available for this reservoir.