Evaluation of Machine Learning Models to Improve Daily Precipitation Estimations From Orbital Remote Sensing and Reanalysis Data

This study focuses on the evolution of the monthly IMERG (Integrated Multi-satellitE Retrievals for GPM) BraMaL (Brazilian Machine Learning) product, the so-called IMERG BraMaL-M, to produce accurate daily precipitation estimations in Brazil (IMERG BraMaL-D) without dependence on ground-based local or regional data. IMERG BraMaL-D uses the satellite-based precipitation product IMERG Early Run and 53 re-analysis variables from MERRA-2 (Global Modelling and Assimilation Office) as inputs for calibration. To achieve the main goal, we evaluated the performance of single (SML), double (DML) and multiple (MML) machine learning methods, using combinations of 6 regression and 8 classification models. The evaluation showed that the K-nearest neighbours (KNN) and the random forest (RF) were the best regression and classification models, respectively. The MML method, combining 5 regression models, was chosen to produce the IMERG BraMaL-D product because it performed statistically better when compared with the observed data from 3227 rain gauges. Compared with the original calibrated product of IMERG (i.e., the Final Run) and three other global satellite-based precipitation products (i.e., PERSIANN-CDR, MSWEP and CHIRPS), IMERG BraMaL-D statistically presented a better performance for almost all analyses. For instance, IMERG BraMaL-D exhibited a KGE (Kling-Gupta Efficiency) of 0.70 for daily estimations, against values ranging from 0.05 (PERSSIANN-DCR) to 0.66 (IMERG Final Run) for the other analysed global products. The monthly accumulated estimations of IMERG BraMaL-D also presented better performance, with smaller data dispersion and KGE rising from 0.86 to 0.95 when compared with IMERG BraMaL-M. Like IMERG BraMaL-M, the main advantages of the IMERG BraMaL-D product are the non-dependency on ground-based datasets after the model's calibration, the improvement of precipitation estimations where the satellite-based products usually underestimate the rain gauge data, and the faster availability to the end-users.