Estimation of Mean Annual Flood in Indian Catchments Using Backpropagation Neural Network and M5 Model Tree

This paper explored the potential of Backpropagation Neural Network (BNN) and M5 model tree based regression approach to estimate the mean annual flood. Data used in this study were taken from an earlier study by Swamee et al. (J Water Resour Plan Manage 121:403–407, <CitationRef CitationID="CR23">1995</CitationRef>) for 93 Indian catchments spread over the entire country. The relationship proposed by Swamee et al. (J Water Resour Plan Manage 121:403–407, <CitationRef CitationID="CR23">1995</CitationRef>) was compared with the predictive accuracy of a BNN and M5 model tree approach. The data were analyzed using a tenfold cross-validation. Comparison of the results showed that predictions with the backpropagation neural network fell within a scatter line of ±30% with a correlation coefficient of 0.975. Furthermore, predictions with the M5 model tree fell well within a scatter line of ±15% with a correlation coefficient as high as 0.994. The results also showed that predicted values with neural network and M5 model tree were within about 1.25 times the actual values. However, the predicted values obtained using the Swamee et al. (J Water Resour Plan Manage 121:403–407, <CitationRef CitationID="CR23">1995</CitationRef>) approach fell much beyond the scatter line of ±50% and the predicted mean annual flood values were sometimes as high as eight times the actual values. The correlation coefficient with this approach was 0.897. The results from this study suggest that backpropagation neural network and M5 model tree-based modeling approaches are superior in accuracy to the model proposed by Swamee et al. (J Water Resour Plan Manage 121:403–407, <CitationRef CitationID="CR23">1995</CitationRef>). This study also suggests that M5 model trees, being analogous to piecewise linear functions, have certain advantages over neural networks as they offer more insight into the generated model, are acceptable to decision makers and are very efficient in training, and always converge. Copyright Springer Science+Business Media B.V. 2010