Mean slope, precipitation seasonality, and mean elevation were also significantly correlated with the deviation between the topographic and effective areas of the Brazilian catchments (Figure 5, the other influencing factors based on RFA are available in Supplement S3). Our ECI results were positively correlated with mean slope and did not vary with decreasing slope contrasting the results of Liu et al.(2020), who observed positive and negative ECI in lower slope degrees. On the contrary, a clear pattern shows a tendency of flat areas to lose water (A_eff < A_topo) and hilly areas to gain water (A_eff > A_topo). The ECI variability decreased with increasing elevation, showing that elevated areas tend to gain water (A_eff > A_topo). Furthermore, the catchments with summer precipitation (P seasonality close to +1, Figure 5c) are prone to have effective areas larger than their corresponding topographic areas. We found positive ECI in areas with a well-defined precipitation seasonality, mainly in the Cerrado and Atlantic Forest biomes, while the northeast region — Caatinga biome — endures long drought spells leading to an unbalanced timing between temperature seasonal dynamics, seasonality close to 0 (negative ECI). We provide further detail about the relationship between those most influencing attributes and ECI in section 4.1 (i.e., aridity index, mean slope, mean elevation, and precipitation seasonality).

It is important to mention that, besides the four most relevant attributes, WTD and HAND also explain the variance in the attribute’s dataset according to PCA (variance-based test). A significant non-linear correlation (Spearman’s p-value < 0.05) corroborates the results from PCA (Figure 5e and f). Nevertheless, they showed less influence on the ECI than the previously mentioned attributes after the RFA (Supplement S3). The fact that HAND is closely related to slope and mean elevation may have contributed to its lower score after the machine learning process in the RFA.