Mapping the challenges of deep-learning-powered flood forecasting: from data-rich region to data-scarce regions

Abstract:

Deep learning methods, especially long short-term memory (LSTM) networks, have emerged as a mainstream tool for flood forecasting. However, given that past case studies have mainly focused on smaller reference basins in data-rich regions, or individual gauging stations, it is unclear if this data-driven method is applicable on a worldwide scale. This presentation showcases efforts that probe at the boundaries of these barriers. While models trained on a widely-accepted dataset may perform poorly in settings different from the benchmarks, we show that LSTM models can be conditioned to make forecasts in data-scarce regions by either migrating knowledge across continents or integrating “soft” data. We also show that some basins with major reservoirs can be surprisingly well simulated despite their radically different behaviors, but some other reservoirs are much more difficult. With continued innovations, it is likely that deep-learning models will be able to handle more and more cases so there will not be dead corners, offering higher accuracy at lower cost.