This paper is an attempt to predict one-day ahead flood levels of a river using Neural Networks (NNs), when only available records are flood levels of previous days. The Basua of Kosi river, India, is selected as the focus site. Several NNs are constructed and floods are forecasted. The results found that the feed forward back propagation network with input layer consisting of past two days’ flood levels and the hidden layer with four neurons to be the best performing NN model for forecasting the current flood level. The same datasets were utilized for prediction of flood levels from the five developed Autoregression models (ARs). The predicted results of flood levels by NNs are very satisfactory and acceptable as compared to ARs and seems to be a good alternative for flood forecasting.

Knowledge of flood levels at a site in a continuous time frame is important for developing the flood fighting and water diversion strategies. The importance of accurate and real time flood level forecasting may be appreciated by recalling monsoonal floods in Kosi river of north Bihar every year which causes great destruction to life and property. On August 18, 2008, Kosi breached its embankment at Kushaha and the flood fury left at least 2.5 million people marooned in eight districts of Bihar and inundated 650 km². The Prime Minister of India declared it a national calamity. The Indian Army, National Disaster Response Force (NDRF) and non-government organizations have been managing the biggest flood rescue operation in India for more than 50 years.

The problem of river flooding is getting more devastating due to poor and late flood forecasting. Authorities and people of the area are not forewarned of the incoming high flood, thereby, not giving them enough time for taking up appropriate flood fighting measures. Though many excellent literatures are available on modeling flood levels, which are primarily conceptual or heuristic, yet modeling of river flood level at a site is a difficult task. Conceptual models, which are based on comprehensive theories, may be more accurate and robust, but defining parameters of a hydrological system which depends on too many variables and their nonlinear interrelationships make these models complex and difficult to be adopted by field engineers. Moreover, the occurrence of river flood levels being probabilistic may not be conducive to deterministic modeling.

Soil And Water Sciences Journal, Soil Health Management, Tree Plantations, Soil Organic Carbon, Rainfed Ecosystem, Biological Productivity, Soil Salinity Development, Low Productivity System, Traditional Farming System, Water Management, Agroforestry System, Agriculture Diversification, Decision Support System, Pest Management.