Preprocessing satellite rainfall products improves hydrological simulations with machine learning

A new preprocessing methodology for gridded satellite precipitation products (SPP) is developed to improve the performance of Machine Learning (ML) algorithms using this type of input data for runoff prediction. A preprocessing approach was applied to capture the rainfall patterns across a watershed, and to select relevant input data for runoff prediction. This approach was tested using the FeedForward Neural Network (FFNN) and the Extreme Learning Machine (ELM) given their flexibility and ability in hydrological modeling. The methodology has been tested in a semi-arid transboundary watershed located in North Africa (Algeria, Tunisia) with the Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (GPM-IMERG) and the Precipitation Estimation from Remotely Sensed Information using Artificial Neural Networks (PERSIANN) satellite rainfall products. The results demonstrated the effectiveness of the proposed approach when using all employed SPPs. In terms of Nash criterion, the suggested pre-processing technique improved the prediction ability of FFNN by 13%, and the ELM model by 15%, which highlights how preprocessing techniques significantly enhance ML models with SPPs data.