Artificial Intelligence–Based Ensemble Modeling for Correction of GPM IMERG Precipitation Product over the Brahmaputra River Basin

Accurate quantitative precipitation estimation is vital for improving hydrometeorological applications. The integrated multi-satellite retrievals for GPM (IMERG) generates precipitation estimates from passive microwave sensors in satellites are associated with a number of uncertainty sources. This study aims to provide a comprehensive investigation of multiple Artificial intelligence (AI) techniques (Random Forest, Bayesian Additive Regression Trees, Boosted Decision Trees, and Neural Networks), to generate stochastically an error corrected improved IMERG precipitation product over Brahmaputra basin. In this work, we have used four features (elevation, soil type, land type, and temperature) as well as IMERG product to produce improved precipitation prediction in Brahmaputra basin. We have trained, tested, and optimized AI algorithms on precipitation information from march 2015 to march 2019. Outputs from individual models are stacked together forming an ensemble method to generate a single prediction (improved IMERG). Evaluation is performed at daily time scale and 0.1-degree spatial resolution. The accuracy of the predicted precipitation product is assessed using 4 years (within the period of 2015–2019) of reference rainfall data derived from rain gauge. In terms of quantification, the magnitude of systematic and random error for the AI generated precipitation product was noticeably decreased compared to original IMERG for the study area, which is a strong indication for using the proposed scheme in retrieving global precipitation across the globe. We conclude that the proposed AI based ensemble framework has the potential to quantify and correct the sources of the error for improving and promoting the use of satellite and -precipitation estimates for water resources applications.