Regionalizing precipitation allows capturing regional-scale variability into manageable smaller regions (Abatzoglou et al., 2009), which is useful to make information-based resource management decision-making about agriculture, drainage, watershed basins, hydropower generation, and vulnerability under natural threats. The regionalization of precipitation is also relevant to identify the physical processes responsible for the spatio-temporal variability in each region (Badr et al., 2015). Such processes are determinant for the understanding of ecosystems functioning, development, and vulnerability.

Although regionalization of precipitation is a common task within climate research, it is not a trivial issue specially under sparse monitoring (Zhang et al., 2016) and within regions with complex spatio-temporal dependencies of precipitation (Hagenauer and Helbich, 2013). The precise spatial delineation of regions would only be possible if rain gauges were densely and uniformly deployed (Abatzoglou et al., 2009). However, this is not a suitable approach under sparsely monitoring or even ungaged regions, which is usually the case in developing countries. Classical regionalization methods addressed mainly bidimensional problems (Hagenauer and Helbich, 2013). Thus, for spatio-temporal problems, such as regionalizing precipitation, dimensionality reduction is usually performed. According to Satyanarayana and Srinivas (2011), the classical regionalization approaches include elementary linkage analysis, spatial correlation analysis, common factor analysis, empirical orthogonal function analysis, principal component analysis (PCA), cluster analysis, and PCA in association with cluster analysis. These approaches, though, apply dimensionality reduction by using PCA (Hagenauer and Helbich, 2013) or by simply using temporal statistics for precipitation such as annual mean (Satyanarayana and Srinivas, 2011). Thus, the precipitation time series are reduced or summarized by a single value. The main limitation of such reduction is that the inherent spatio-temporal structure of precipitation is not taken into account (Jacques and Preda, 2014). There are two challenging issues when regionalizing precipitation in a sparsely monitored region. First, precipitation data are mainly available from remote sensing sources, and, second, dimensionality reduction may hinder the discovery of the underlying spatio-temporal dependencies of precipitation.