Estimating uncertainty in flood model outputs using machine learning
❄
ABSTRACT
| Numerous studies have extensively employed Monte Carlo-based methods to characterise uncertainty in flood modelling due to their simplicity and flexibility. However, this approach is time-consuming and computationally expensive. Recognizing the need for efficiency, machine learning has emerged as a promising alternative with comparable results. The Bayes Neural Network, which combines Bayes probability theory with deep learning architecture, is widely recognised for this purpose, but its application still needs extensive research. To contribute to this field of study, this paper applied the Bayes Neural Network with Bayes by Backprop (BNNBB) to estimate the variation in flood predictions caused by the uncertainty in flood model inputs. We first employed a Monte Carlo framework to produce flood prediction variations caused by uncertainty in the Digital Elevation Model (DEM) generation process. In particular, multiple DEMs were produced by changing the parameters used to estimate riverbed elevations – river width, flow, and slope – as well as changing the model grid orientation. An analysis of the LISFLOOD-FP model outputs – water depths – produced from these DEMs then created a flood map that shows the proportion of simulations that each cell was flooded. This Monte Carlo procedure simulated the January-2005 Waikanae River flood and the July-2021 Buller River flood in New Zealand. The uncertainty flood map predicted by Monte Carlo simulations at the Waikanae River was used to train BNNBB, and at the Buller River was used to test the accuracy of BNNBB. Two BNNBBs were applied in this study. The first BNNBB labelled pixels as flood, maybe-flood, or no-flood, providing an input for the second BNNBB. The proportion of each pixel being flooded was then predicted by the second BNNBB. Both BNNBBs has identical architectures built based on the fully connected layers of LeNet. The first fully connect layer had 120 hidden nodes and the second had 84 hidden nodes. The inputs for both BNNBBs include the North-South Cartesian DEM, derived from the best-estimated riverbed elevations, flood depth, generated from one-time running the LISFLOOD-FP model height above nearest flood, flood proximity, roughness, slope, and flood edge. Results show that BNNBB effectively learns geometric features in the inputs to identify the boundaries of the flood map. It accurately detects areas with no flood (0%) and complete flood (100%), which is highly consistent with Monte Carlo simulation results. The model also excels at identifying areas of uncertainty – the edges of the flood map where the proportion being flooded ranges between 0% and 100%. This makes it as a promising substitute to traditional Monte Carlo-based methods for characterising uncertainty in flood model outputs caused by variations in flood model inputs. However, given early stage of the model, further work is still needed to improve the efficiency such as adding more informative inputs and a convolutional neural network to account for spatial correlation. |