Comparative study of performance for real-time flash flood forecasting in the upper Meghna basin

Abstract

Flash flood in the pre-monsoon season (March-May) is one of the main natural disasters of the Upper Meghna Basin, which frequently destroys Boro rice, the primary agricultural product of the northeast Bangladesh. Forecasting of flash floods has provided an opportunity to reduce these damages by giving early warnings and providing adequate time to farmers for harvesting at least a part of their crops. Unfortunately, flash flood forecasting is an inherently complex process mainly because flash floods can occur very rapidly after an intense rainfall event. There are two types of methods available for flood forecasting, the physically-based models and the data-driven models. Physically-based models require large amounts of data and are computationally expensive, while data-driven models require less data can be quickly developed. This study investigated with Artificial Neural Network (ANN) and Support Vector Machine (SVM), two data-driven models. Forecasting was done at the Bijoypur, Laurergarh, Muslimpur and Sunamganj gauging stations of Bangladesh Water Development Board with lead times of 6, 12, 24 and 48 hours. As input data, 3-hourly satellite-based TRMM rainfall and 3-hourly observed river stage data at the selected stations were used to calibrate (1999-2009) and validate (2010-2014) the models. As most of the drainage area of the selected stations are located outside Bangladesh, TRMM was chosen as it is a global product available in real-time. Three types of inputs were investigated: i) rainfall, ii) river stage and iii) both rainfall and river stage combined to develop ANN models. Results show that the third type of inputs give the most optimum performance. However, the lead time of these forecasting models were increased, their performances gradually decreased. When the performance of the models was calculated for the overall annual data, R2 values for Bijoypur in Someswari River, Laurergarh in Jadukata River, Muslimpur in Jhalukhali River and Sunamganj in Surma River were 0.869, 0.958, 0.986, 0.987 for ANN and 0.775, 0.914, 0.973, 0.986 for SVM respectively. However, when the performance of the models was calculated only for the pre-monsoon season, SVM models performed better than ANN models because the hydrological characteristics of the study area are markedly different between the pre-monsoon and monsoon seasons A significant part of the Upper Meghna Basin becomes inundated in monsoon causing the streams to become interconnected. ANN models function better under these conditions than SVM models. Also, model performances in Sunamganj were found to be better than other areas, as Sunamganj is flatter and has a larger time of concentration. This study shows that data-driven modeling with freely available satellite-based rainfall data can be a viable alternative for forecasting flash floods in a data-poor basin with a reduced computational time.