Akhilesh KumarSingh, Saurabh2017-05-092017-05-092016-06http://krishikosh.egranth.ac.in/handle/1/5810010843In this study attempts were made to develop sediment yield prediction models for the upper Ganga basin which comprises an approximate area of 2,31,127 km2. In the development of these models multivariate regression, multilayer perceptron based artificial neural network (MLP-ANN) and single multiplicative neuron based artificial neural network (SMN-ANN) techniques with linear and non-linear input of 20 years (1990-2010) and 30 years (1980-2010) durations have been used. An appropriate value of lag of input data which affects sediment outflow from the basin at any time was determined and following stepwise regression techniques, variables significant at 95% level were identified and used in model development. Large number of models were developed for various combinations of nature of input, number of neurons, epochs and transfer functions. Preliminary comparison lead to identification of six number of models applicable in each category of regression, MLP-ANN and SMN-ANN and three models specifically for monsoon season in each category. In order to distinguish and have an easy reference and better understanding of these models a set of nomenclature was developed and used. Accordingly, RM represents regression based model with prefixes of L, NL and LL which stand for linear, non-linear and log-linear input where as in case of ANN models MLP and SMN with suffixes of LM, NLM and LLM for linear, non-linear and log-linear input model respectively. The number 20 and 30 in a nomenclature signifies the use of data set of 20 years and 30 years respectively and the term MM represents that the model is developed considering monsoon season data set only. Developed models were evaluated quantitatively for their sediment yield prediction efficiency based on selected performance indicators such as, MAPE, RMSE, NRMSE, R2, ISE, CE, d and VE while qualitative performance was evaluated through visual comparison of observed and model predicted sediment yield through graphical representations. Performance comparison was made first within the category to identify a best performing model in a category. Then, performance comparison of category wise best performing models was carried out based on selected performance indices to identify an overall best performing model for sediment yield prediction in the study area both on annual basis and monsoon season basis. The comparative analysis indicated that in general SMN based ANN models provided better total annual sediment outflow prediction as compared to conventional MLP based ANN models in the upper Ganga basin with linear input. The SMN20LM with 20 epochs while SMN30LM with 80 epochs and log-sigmoid as an activation function provided best sediment yield prediction. The RMSE and VE values for these models were obtained as 1.467tons/s & -3.746% and 1.459tons/s & -2.133% respectively. MLP-ANN models were found working satisfactorily for total annual sediment outflow prediction when applied in the study area considering 20 years and 30 years data series with non-linear input. Model architectures of 8-2-1 for MLP20NLM and 5-9-1 for MLP30NLM with log-sigmoid transfer function at hidden layer and tan-sigmoid transfer function at the output layer for both these models provided best prediction of total annual sediment yield. The values of RMSE and VE were obtained as 1.032 tons/s & -3.182% and 0.728 tons/s & -5.348% with index of agreement as 0.996 and 0.998 respectively for these models. With log transformed input SMN20LLM at 3 epochs where as SMN30LLM at 2 epochs provided the best results with respective RMSE and VE values of 2.498 & -4.378% and 2.245 tons/s & 9.975% with coefficient of efficiency 0.90 and 0.919 and index of agreement as 0.975 & 0.977 respectively. In case of monsoon season SMN based ANN models were found to be performing better in each category. On the basis of overall comparison of category wise selected models, it was found that the MLP based ANN model with 20 years non-linear input (MLP20NLM) while SMN based ANN model with 30 years linear input (SMN30LM) provided the best sediment prediction efficiency. The calculated values of RMSE and VE for these two models were obtained as 1.032 tons/s & -3.182% and 1.459 tons/s & -2.133% respectively. Similarly, a comparison of category wise selected monsoon models revealed that single multiplicative neuron based artificial neural network model (SMN-ANN) with 30 years monsoon data in its non-linear form (SMNNLMM30) provided the best sediment outflow prediction of 9,50,48,901.16 tons as against 9,51,82,452.0 tons of observed sediment yield for monsoon season which resulted in the volumetric error value of 0.140%. In both the cases, therefore, it can conveniently be concluded that these models are applicable very effectively in the study area as model predicted total sediment outflow was found to be very close to the total observed sediment outflow.ennullThesis