Possible Roles of Artificial Neural Networks in Hydraulic and Hydrological Models: Fid-Bau Portal

Possible Roles of Artificial Neural Networks in Hydraulic and Hydrological Models

Liong, Shie-Yui / Kim, Dongeon / Liu, Jiandong / Gourbesville, Philippe / Andres, Ludovic

This paper presents 2 past and 1 on-going works on the applications of Artificial Neural Networks (ANN). The 2 past works are river flow forecasting in Bangladesh (a country with vast delta and severe floods perennially) and forecasting of maximum wave height and its time of arrival of the devastating Aceh tsunami in 2004. The on-going work is on deriving a cost effective and high accuracy Digital Elevation Model (DEM) from publicly accessible satellite data.

The paper first demonstrates the ANN application on the river stage forecasting at Dhaka, Bangladesh, from 1 to 7 lead day forecast horizons. Although the input nodes used only the water levels at most upstream reaches of 3 main transboundary rivers (Ganges, Brahmaputra, Meghna), the goodness-of-fit R² values are very high ranging from 0.99 (for 1 lead day) to 0.91 (for 7 lead days). The reason, even without rainfall data in the input nodes, is that 80–95% of the catchments of these 3 large rivers lie in India; thus, the main flow contributions come from India. The high degree of accuracy, accompanied with very short computational time (less than 1 min), makes ANN a desirable advanced warning flow forecasting tool.

The paper continues with a second ANN application demonstrating its effectiveness and efficiency as a forecasting tool for devastating Indian Ocean/Aceh tsunami in 2004. The ANN was trained with simulation output data of a widely used process-based tsunami propagation model, TUNAMI-N2. The input nodes comprised, among others, the earthquake magnitude and epicenter with spatial values of maximum tsunami heights and tsunami arrival times (snapshots) for the most probable ocean floor rupture scenarios as its target. Validation tests demonstrated that with a given earthquake magnitude and location, the ANN method provides accurate and near instantaneous forecasting of the maximum tsunami heights and arrival times for the entire computational domain covering South China sea (the Philippines inclusive) and the Indian Ocean (India inclusive).

The 3rd ANN application shows the on-going DEM improvement scheme, which significantly improves DEM originating from a publicly accessible satellite SRTM (Shuttle Radar Topography Mission). The scheme uses the DEM data of SRTM and the multispectral data of another publicly accessible satellite Sentinel-2 as the input to the ANN while the target is the high spatially resolution and high accuracy DEM from German Aerospace Center (DLR). Thus far the present improvement scheme manages to reduce the Root Mean Square Error up to 42.3%. Equally interesting is that the trained ANN can also be used to provide DEM in another part of the world with accuracy much higher than the raw DEM from SRTM.