A simple machine learning approach to model real-time streamflow using satellite inputs: Demonstration in a data scarce catchment

Abstract

Abstract Real-time streamflow modeling is a challenging endeavor in regions where real-time ground-based hydro-meteorological observations are scarce. Nevertheless, with the emergence of satellite remote sensing, some of the essential hydro-meteorological datasets such as rainfall and soil moisture are available in near real-time, which can circumvent the problem. In such scenario, machine learning (ML) approaches can be used to model streamflow even with limited real-time data. In view of this, the effectiveness of integrating satellite-based near real-time rainfall, i.e., Tropical Rainfall Measuring Mission based (3B42RT V7), and Advance Scatterometer (ASCAT) based soil moisture observations through Support Vector Machine (SVM) for simulating streamflow in a poorly gauged catchment in India has been investigated. Results show that 3B42RT forced streamflow performed poorly compared to streamflow simulated using the observed gridded rainfall. This inferior performance is mainly attributed to errors in 3B42RT rainfall product. Hence, a recently developed SVM based error reduction model using ASCAT soil moisture data is used to correct 3B42RT. The obtained corrected satellite rainfall outperformed 3B42RT in streamflow simulations but still it is affected by notable errors. To further improve the streamflow estimates, ASCAT soil moisture is used as a predictor variable along with 3B42RT and the corrected satellite rainfall separately into SVM based streamflow modelling. Results indicate that this integration considerably improves the streamflow estimates with respect to the use of 3B42RT and corrected satellite rainfall individually. This is attributed to the crucial role of antecedent soil moisture for streamflow generation. Among all the tested scenarios, streamflow derived by the integration of soil moisture and corrected satellite rainfall showed the highest performance. Hence, it can be said that the proposed approach has considerable potential to model real time streamflow in the chosen poorly gauged catchment and possibly beyond. However, the approach needs detailed evaluation before applying in other such catchments.