Improvement of precipitation estimation at monthly and daily scales for brazil based on remote sensing product and machine learning techniques

Abstract

Precipitation is one of the main components of the hydrological cycle and its accurate quantification is essential to provide information for understanding and predicting physical processes. Occurrence observations based on ground-based devices (manual and automatic rain gauges) are highly accurate but have limited spatial coverage. On the other hand, remote sensing products cover large areas but with lower precision. In this context, this study aims to evaluate machine learning models to create a product with better occurrence estimation, with lower latency than other products and without directly relying on field data. The methodology consists of choosing the best machine learning model (classification and regression) and applying it to satellite-based remote sensing data (IMERG Early Run product) and reanalysis-based variables (MERRA-2). The method was applied throughout the Brazilian territory, on monthly and daily scales, which presents a wide variety of supply regimes. This methodology first resulted in the development of an adjusted IMERG product at the monthly scale (IMERG-BraMaL) and later an improved product at the daily scale with a multiple machine learning technique (IMERG-BraMMaL). Compared to the original IMERG products (Early Run and Final Run) and global estimation products (MSWEP, CHIRPS and PERSIANN-CDR), IMERG-BraMaL improved the analyses evaluated between terrestrial and satellite data in almost all analyses. For example, the KGE (Kling-Gupta Efficiency) went from lower values (0.70, 0.82, 0.09, 0.60 and 0.81 for IMERG Early, IMERG Final, PERSIANN, MSWEP and CHIRPS, respectively) to values above 0.86 in IMERG-BraMal at the monthly scale. On a daily scale, IMERG BraMMAL proved to be more efficient, presenting better results, with a CC of 0.79 compared to 0.68 for IMERG BraMaL. The main conclusions of the study were: (i) much faster availability to end users; (ii) no dependence on any field data, allowing its application in areas where rainfall data are not available or are of low quality; (iii) no correlation of errors with local characteristics; and (iv) much improved estimates in regions of Brazil where, historically, satellite-based products often underestimate the observed data.