MSc.

Half of the world's population is living under the highly water-stressed areas,
water availability and management (both in terms of quality and quantity) are going to
be challenging in coming years due to changes in rainfall characteristics and LULC.
Streamflow prediction is important for flood management, reservoir operation, and
agriculture. In the present work, we have trained and tested four data-driven machine
learning models SVM, RF, XGBOOST, and LSTM for streamflow prediction in the
hydrologically similar catchments Meenachil, and Muvattupuzha rivers, situated in
Southern Western Ghats, in the state of Kerala, India. All machine learning models
performed quite well in streamflow prediction with NSE value > 0.56 in both studied
river basins. RF model was found to be outperforming other candidate’s models for
streamflow prediction both during model training and testing periods in daily time scale
and LSTM model outperforms in monthly timescale. We have found that machine
learning performed quite well in hydrologically similar catchments once hyperparameters
are exchanged and resulted in almost similar model accuracy. Machine
learning models found to have significant potential for predictions in ungauged basins,
and thus can be tested in larger sample sets for future potential streamflow modelling
in ungauged basins