MODELLING FOR RUN-OFF & SEDIMENT YIELD - A CASE STUDY

Abstract

The rapid growth in population, urbanization and industrialization, and economic and social changes during last few years has resulted in an increased and diversified demand of water. At the same time, quantity of available water from surface and ground water sources has remained constant. So water has progressively emerged to become one of the most pressing national issues of present time on the development and environment. Thus there is need to use recent advancement in technology and modeling tool to develop optimal water resources system and manage sediment problems. Rainfall-runoff and sediment yield models are the key components of the overall modeling framework for sustainable watershed development and management. The process of runoff and sediment erosion is influence by the spatial controls exerted by the land surface such as elevation, slope, drainage network and vegetation cover. So the knowledge of geomorphology and vegetation of watershed can enhance one's understanding and capacity to model the processes. Soil erosion is a serious environmental problem in Nepal where more than 80 percent of the land area is mountainous and tectonically active. Anthropogenic causes such as deforestation, overgrazing and intensive farming have accelerated the erosion problem. Unscientific cultivation, haphazard construction and intensive monsoon have further aggravated the situation. Thus, it is very important to understand the erosion process and assess the magnitude of the problem so that effective counter measures and appropriate sediment management method can be implemented. The present study has been taken up in the above background to develop rainfall-runoff and sediment yield models to predict runoff and sediment yield from Kankaimai watershed in eastern Nepal. Geomorphologic and vegetational analysis was carried out using remote sensing and geographic information system. The study has revealed that the Kankaimai watershed is fairly good with moderately high peak flow of shorter duration giving quick response of sediment yield and runoff. With this knowledge of watershed characteristic and statistical analysis, Mekhnath et al., (2005) related the runoff and sediment yield with various parameters and finally found that runoff depends on rainfall of three rain gauge stations (Ilam Tea Station, Soktim iii Station & kanyam Station) and previous day's runoff at Mainchuli Station of time step t- 1 (Qt_l) and time step t-2 (Qt_2) parameters, they similarly found that suspended sediment concentration at Mainchuli station in the Kankaimai_ River depends on runoff (Qt) parameter. In the present study Support Vector Machine models have been developed using above combinations of data as input and above data were preprocessed such as scaling (Normalized with range 0 to 1). RBF Kernel has been selected as kernel mapping function and Sequential Minimal Optimization (SMO) as a SVM algorithm. Nash coefficient (R2), coefficient or correlation (CC), root mean square error (RMSE) and recover ratio (RR) were estimated to assertion the model performance. SVM model validation statistics resulted in R2= 0.68, CC= 0.85, RMSE=134.77 m3/s and RR=0.83 for runoff prediction, R2= 0.95, CC= 0.9792 RMSE =17281 t/day and RR=0.78 for sediment yield prediction ANN model used by Mekhnath et.al 2005 for same combination of input variables and statistical validation resulted in R2= 0.82, CC = 0.91, RMSE= 103.67 m3/s and RR =0.89 for runoff prediction and R2 =0.93, CC =0.97, RMSE =19058 .t/day and RR= 1.18 for sediment yield prediction. Using regression model, CC = 0.78, RMSE =157.45 m3/s and RR = 0.62 for runoff and R2 values= -.85, CC = 0.97, RMSE =106607 t/day and RR = 1.65 for sediment yield prediction (By Mekhnath et.al 2005). Furthermore SVM model performed not better than the ANN model but performed better than regression equations for rainfall - run off. However, SVM model performed better than the ANN model and regression equations for sediment yield. Efforts were made to develop efficient model but due to constraints like huge running time for optimizing the parameters, unavailability of high computing machine and limited available time etc, the present model is proposed, which can be further improved in future with more runs in high computing system.