Enhancing the rain/no-rain identification in remote sensing-based rainfall products: what impact on streamflow simulation

Abstract

The daily remote sensing-based rainfall estimates have often been problematic in several regions around the globe. This is particularly prevalent in semi-arid regions where, in addition to misestimating the magnitude of the rain events, the spatial rainfall products (SRP) often fail to detect many events correctly. Whether missed or falsely detected, misestimating many events is a real constraint in running (calibrating/validating) hydrological models. Thus, here we are attempting to enhance the capability of some of the well-known SRPs (GPM IMERG, PERSIANN-CDR, and CHIRPS) in rain/no-rain identification (using ancillary data) and how that can impact predicting the hydrological response. To this end, the SRPs were used to drive the HBV and GR4j conceptual hydrological models in watersheds from different climatic contexts. Using the raw SRPs, the performance of the HBV and GR4j models was relatively poor and temporally unsteady. This was primarily due to uncertainties associated with the SRP estimates. Even the best-performing product (GPM IMERG), was found to largely misestimate rainfall up to 50%. In particular, a prevalence was also observed in terms of detection capacity with non-negligible missed events (according to POD; Probability Of Detection) and many rainfall events detected as false alarms (according to FAR; False alarm Ratio). However, the SRPs blended with remote sensing-based ancillary data allowed us to relatively enhance the streamflow simulation, particularly using the HBV model. This enhancement was possible as using ancillary data allowed us to reduce the number of false alarms and recover some of the missed events. Still, some bias persists in the SRPs, which can be addressed by incorporating in-situ observations employing conventional (e.g., Scaling Factor, CDF matching…) and AI-based bias correction techniques.