Deep Learning for Probabilistic Forecasts Using Features from Rainfall Objects: A Case Study in the Amazon Basin

Hydrological forecasting is of global importance, especially with the spotted increasing trend of flood-related disasters as seen in the last two decades.  The causative rainfall events of these extreme events are primarily analysed in a one-dimensional method. However, through an object-based approach, more data on these rainfall fields can be generated and studied to link them to the hydrological response observed. Through an object-based methodology ST-CORA, features from rate of change of rain intensity in space and time can be extracted by simple visual inspection. Every side of an object provides time variations that can be used as images that contain features not easy to extract. In general, rainfall events in previous studies have used aggregated information, like the duration, area, volume, maximum intensity, and the centroid. In this work, more information is captured that describes the spatial and temporal properties of the event. The main objective of this research is to use these 3D objects and their features with a deep learning model to produce a 15-day hydrological probabilistic forecast for flood prediction.

A calibrated version of a large-scale hydrological model (MGB) is used to study an Amazon subbasin. The model is forced with the 50-member perturbed forecast from the TIGGE dataset for the period 2006 to 2014 (from ECMWF). The purpose of using the large-scale model is to better capture the spatio-temporal characteristics over a wider area in an effort to reduce the uncertainty in the analysis. For data-driven models, there is a need for sufficiently large databases, in this case for both the causative rainfall events and the observed hydrological responses. As such, the first two steps relate to the data generation. The first database is developed from the daily streamflow which is generated from the calibrated hydrological model at specific locations of interest with the known higher performance metrics. Second, the ST-CORA methodology is applied to extract the features from the rainfall events in order to develop a database of the rainfall objects. Third, an analysis on the statistics of the features of the objects to understand the rainfall which occurs within the study area. The final part of the research involves the effective use of these features and objects with a deep learning model. From the average annual rainfall from 2001 to 2020, three distinct precipitation patterns are observed. For the streamflow, the subbasin shows a relatively fast response which is captured within a 15-day window.

A convolutional LSTM deep learning model is developed to handle 3D rainfall objects as sequences of images representing space time sequences. The outcome of this research contributes to the end-to-end deep learning model which receives the forecasted rainfall as objects and generates a corresponding hydrograph at the area of interest for which it has been trained. A potential contribution of this Conv-LSTM network is that it may provide an efficient and automated approach for streamflow forecasting in basins where there is known complexity and non-linearity, which is especially useful for early warning systems.