A Novel Multichannel Deep Learning Method Driven By Data and Process Representation for Flood Forecasting

A Strategy to Balance Model-Dependency (process-based) and Model-independency (data-driven) to Magnify the Long Term Predictive Potential of Deep Learning Models

Abstract

Over the last decades, floods have become the most common and deadly natural disaster on the planet. While many countries currently lack effective early warning systems and alerts. Flood forecasting is a very exploratory project because there is substantial peer-reviewed evidence that flood early warning can prevent between 30% to 50% of both fatalities and economic harms. Existing challenges in flood forecasting include how to establish relations between more variables, extend the lead-time for actual warning with improved accuracy. Given the competitive performance deep learning models have shown in temporal variation modeling, temporal dependency learning, and multivariate representation, this paper aims to explore the potential of deep learning model in forecasting floods and focuses on the adaptation and training of the deep learning network, namely AdapFAN, for flood forecasting. Building on the data-driven basis, AdapFAN incorporates process algorithms in separate paths to further its predictive length and reliability. Specifically, AdapFAN transforms complex 1D time series into 2D tensors, tackling the representation limitation of 1D series. Then, AdapFAN adopts channel independence to avoid the loss due to the average of multi periodicity between different dimensions. To take the influence of dimensional dependencies into account, dimensions are aggregated into 3D space and allow conv3d to analyze the hidden inner relations between multivariates. Using AdapFAN, this project achieves reliability in predicting extreme riverine events in ungauged watersheds at up to a 4-day lead time that is better than the reliability of nowcasts (0-day lead time) from a current state-of-the-art global modeling system.