Comparing different deep learning architectures for S2S forecasting of widespread floods in the Yangtze River Basin

Widespread floods are floods that co-occur in space over a large geographical area, typically caused by prolonged or extreme rainfall, which affects extensive regions and even whole countries. The improvement of early warning systems, particularly improving the skill of seasonal and sub-seasonal (S2S) forecast is imperative to improve our preparedness and reduce loss of life, property damage, and environmental disruption caused by spatially co-occurring floods. The aim of this study is to forecast the (sub-)seasonal probabilities of widespread flooding across the highly anthropogenically regulated Yangtze River Basin in China using deep learning techniques. For that we test three contrasting state-of-the-art deep learning architectures for predicting sequential time series: recurring (i.e., Long short-term memory, LSTM), convolutional (i.e., dilated convolutional neural network, dCNN) and transformer-based networks (i.e., Informer). We use monthly antecedent precipitation and large scale climatic indices to forecast widespread floods severity index for different lead times at S2S timescale. In our study the widespread flood severity is estimated as the sum of daily maximum streamflow that exceeds local (i.e., gauge-specific) 2-year return period within the given months for the period 1961-2018 across 40 sub-catchments in the Yangtze River Basin. The three deep learning models are trained on the whole Yangtze River basin and four distinct hydroclimatic regions, intending to provide a deeper understanding on regional variability of large-scale atmospheric drivers of widespread flooding. Our preliminary results for the LSTM-based models indicate that in the case one-month-ahead forecasts, the seasonal patterns of widespread flooding are captured accurately for the whole Yangtze Basin and for the four individual regions. However, the models tend to underestimate flood severity under extreme conditions. In the next steps, we plan to extend lead time to three months and compare the performance of three different architectures mentioned above with the aim to enhance the accuracy of early warning systems for widespread floods.