Prediction of flood quantiles at ungauged catchments for the contiguous USA using Artificial Neural Networks

Abstract

In this study, we utilise Artificial Neural Network (ANN) models to estimate the 100- and 1500-year return levels for around 900,000 ungauged catchments in the contiguous USA. The models were trained and validated using 4,079 gauges and several selected catchment descriptors out of a total of 25 available. The study area was split into 15 regions, which represent major watersheds. ANN models were developed for each region and evaluated by calculating several performance metrics such as root-mean-squared error (RMSE), coefficient of determination (R2) and absolute percent error. The availability of a large dataset of gauges made it possible to test different model architectures and assess the regional performance of the models. The results indicate that ANN models with only one hidden layer are sufficient to describe the relationship between flood quantiles and catchment descriptors. The regional performance depends on climate type as models perform worse in arid and humid continental climates. Overall, the study suggests that ANN models are particularly applicable for predicting ungauged flood quantiles across a large geographic area. The paper presents recommendations about future application of ANN in regional flood frequency analysis. HIGHLIGHTS Parsimonious ANN models (only one hidden layer and with fewer inputs) have similar performance to more complex models for the study area.; ANN models in arid, semi-arid and humid continental regions have the lowest performance.; Area and standard annual average rainfall have the highest importance for predicting flows in ungauged basins in most regions.;

Keywords

• ann models
• flood frequency analysis
• machine learning
• ungauged basins