Application of hydrometeorological indices for hydrologic forecasts within an artificial neural network framework

Abstract:

For many years, application of Machine Learning (ML) modeling approaches in hydrologic forecasts have become popular. Among all ML methods, Artificial Neural Network (ANN) has received special attention because of its high performance on watersheds with different hydrologic regimes and for short- to long-term forecasts. Moreover, the use of ANN eliminates the need of geomorphologic data such as land use and soil type, with hydro-meteorological data as their sole input variables. In the context of streamflow forecasts, hydro-meteorological inputs can represent both meteorological conditions and hydrological state of the watershed.

Our case study is the Androscoggin watershed located in Maine (USA), typically representative of northern hydrologic regimes. Watershed is mostly covered by forest (83.5%), with open water (8.5%) and agriculture (3.5%) covering the remaining areas. Focused on summer season and over a ten-year period, a short-term (1 to 7 days) streamflow forecast model is developed using ANN with different inputs: the modified Antecedent Precipitation Index (modAPI) and more classic variables such as daily precipitation, temperature, evapotranspiration and soil moisture.

Special consideration is given to modAPi as it attempts to capture soil moisture temporal variations due to combined effects of precipitation, evapotranspiration and saturation excess overland flow: the index is derived from the preceding daily rainfall with a dependence on daily mean temperature and a soil wetness threshold. Therefore, compared to the classic API, the modAPI adds information about physical process governing runoff and is directly related to the soil moisture state of the watershed.

In order to test the performance of ANN model, hydrological forecasts are compared with forecasts produced using Hydrotel, a physically based distributed hydrological model, and with flow observations from a USGS station at the watershed’s outlet. Training/ validation of the ANN covers a 10-year period with 8 summers for training and 2 summers for validation.

The present work is a preliminary step for a more general study about the use of remote sensing data for hydrologic forecasts within an ANN framework. Thus, the next steps will be to determine which remote sensing products to use, and to develop the ANN model with these data sets.