Deep Learning Approach to Estimate Streamflow from Remote Sensing Data
No Thumbnail Available
Date
2024-02
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Streamflow measurements that are reliable, precise, and continuous are essential for water resource analysis, planning, and management. However, there are relatively limited available hydrological data globally. The limited number of gauging stations worldwide is a common challenge, even worse in developing countries such as Ethiopia. As a result, novel research methodologies that can simulate river flow using multiple data sources are crucial. Remote sensing is one of the data sources that is considered a promising alternative. It has shown remarkable progress to date and in the future, such as the Surface Water and Ocean Topography (SWOT) mission, a satellite built to precisely observe nearly all the water on our planet's surface launched on December 16, 2022.
Artificial intelligence, analogically "the new electricity" as a tool, has enormous potential to support the big challenge we face in monitoring our water resources. Hence, this study aims to integrate remote sensing and deep learning approaches for continuous streamflow time series generation. This will be achieved by collecting remote sensing-based precipitation products, vegetation indices and ground-based hydrometeorological data. Various types of single, hybrid, and ensemble deep learning models with the conceptual hydrological model are applied as a tool to model these univariate and multivariate types of data in different agroclimatic conditions, including the upper Tiber River basin (Italy), Abay River basin (Gummera subcatchment, Ethiopia), Awash River basin (Borkena subcatchment, Ethiopia), and Baro-Akobo River basin (Sore and Masha subcatchment, Ethiopia).
Finally, by simulating streamflow with consistent top performance in all case study catchments, the results demonstrate the tremendous potential of the ensemble deep learning model. Although vegetation indices also showed much potential as an input for machine learning models, more study is needed to optimize performance with various data assimilation techniques. The findings of this research also provide further evidence for the significance of input data preprocessing, selection, length, and variability in conjunction with the kind of machine learning model we use to simulate streamflow accurately. The study also provides recommendations for further study and development in data-driven rainfall-runoff modelling.
Description
Keywords
Streamflow, Remote Sensing Data, water resource analysis, SWOT