| Popis: |
Conventional physically based models have long yielded promising results, as they have been the main tool to depict the underpinnings of the physics governing the hydrological events. These models, however, suffer from certain issues such as the intense calibration time or the uncertainty in the estimation of hydrological variables. The development of the sophisticated data-driven techniques, and machine learning models in particular, combined with rapid increases in computational abilities (graphics processing units, computer clusters. etc.), has enabled hydrologists to utilize the data driven models in tandem with the well-established hydrological models to simulate miscellaneous environmental processes nimbly, and therefore circumvent the aforementioned conundrums associated with the physically based models. To this end, the present study aims at exploring a sophisticated neural network called variational Bayesian neural network, to improve the accuracy of physically based predictions such as runoff. Our neural network was able to accurately forecast the runoff rates with the mean Pearson correlation coefficient of 86.27%+0.0599 within a randomly selected subset of cells in the Brazos River Basin. As these cells are selected randomly across the basin, we exclude the possibility of biasing our neural network by any specific cell. Moreover, this work for the very first time, to the best of our knowledge, suggests a similarity-based solution to transfer the learning model developed in a basin to be deployed across a different basin. In other words, there would be no need to develop a learning model for each basin from scratch. We, instead, utilize the models learnt from the previously studied basins. We cross-validated our proposed transfer learning solution via leave-one-out strategy within the grid cells of the Brazos River basin achieving a mean Pearson correlation coefficient of 85.83%+0.0592. |
