Severe atmospheric phenomena that culminate in human, material, and financial disasters have occurred more frequently in recent years, according to the Human Cost of Disasters 2000–2019 (MIZUTORI, M. and GUHA-SAPIR, 2020) report of the United Nations (U.N.): more than 4,000 occurrences were recorded worldwide between the years 1980 and 1999, leading to about 1.19 million deaths and costs of 1.63 trillion dollars. However, during the following twenty years, the number of records almost doubled (about 7,300), 1.23 million human lives were lost, and the financial losses reached 2.97 trillion dollars, according to the same document.
In Brazil, adverse effects associated with meteorological systems that favor the occurrence of extreme rainfall events have also been the cause of material, financial, and often human losses (Machado 2022; G1 ES 2022; Kruse 2022; BBC News Brasil 2021; Treinish et al. 2013). In this context, an episode that marked the country's history was the one that occurred between January 11 and 12, 2011, in the Mountain Region of Rio de Janeiro (MRRJ). Considered the most significant catastrophe in Brazil associated with high volumes of rainfall, the event caused more than 900 deaths, thousands of homeless, and hundreds of missing (Busch and Amorim 2011; Pinheiro et al. 2011). The region continued to be the scene of disasters of this nature in subsequent years (Marques et al. 2022; Mello 2013). However, the 2011 episode is still the largest, considering the number of deaths and material and financial losses.
In light of these occurrences, there is a growing need for decision-making by government officials and civil defense agents with greater anticipation. Therefore, tools are necessary to help them plan and execute actions to mitigate the damage associated with natural disasters. Among the existing ones is the ability to forecast the weather numerically: once the possibility of occurrence of a severe phenomenon is known, such as storms, it is possible to alert people allocated in risk regions, trigger contingency plans, etc.
Atmospheric modeling is based on applying physical equations to represent the future state of the atmosphere from its initial conditions (Palmer and Hagedorn 2006; Warner 2010). Inevitably, sources of uncertainties are incorporated into the process, such as those related to (1) the numerical resolution of the governing equations; (2) the imperfection of the initial and boundary conditions data; (3) the discretization of the atmosphere; (4) the spatial and temporal resolutions chosen in the model used and (5) the parameterization of microscale physical processes, among others (Palmer and Hagedorn 2006; Warner 2010). Regarding the spatial resolution, it is common to use nested domains so that smaller areas of greater interest are modeled at a higher resolution than the surroundings, avoiding a relatively large domain (which is essential even when the region of interest is smaller than it) is configured with high resolution entirely without necessity. Thus, model processing time and computational cost decrease. The parameterizations make the modeling realistic because they consider the effects of decisive physical processes at the microscale (Palmer and Hagedorn 2006). Cloud development, for example, depends on this type of process, and without the parameterizations, there would be no cloud simulation, forecast, or rainfall. Some processes of this nature are cumulus convection, cloud microphysics, those that occur in the planetary boundary layer (PBL), at the Earth's surface, etc. (Palmer and Hagedorn 2006; Warner 2010). Depending on the atmospheric phenomenon, the representation of one may be closer to the actual scene than the other. Although essential, such processes are parameterized, and how a parameterization works are sources of uncertainty. Moreover, such parameterizations use physical parameters that should be calibrated according to the region where the modeling is performed (Palmer and Hagedorn 2006; Warner 2010). It is not the case for simulations or forecasts performed for Brazil (Webster, 2013) due to the lack of measured local data.
Given the importance of the reliability associated with weather forecasts, research investigating the accuracy and sensitivity of atmospheric numerical models to parameter variation is fundamental. Several authors have worked with the variation, simultaneous or not, of initial and (or) boundary conditions (Taraphdar et al. 2021), physical parameterizations (Njuki et al. 2022; Jacondino et al. 2021; Taraphdar et al. 2021; Yoon et al. 2021; Yang et al. 2021; Wang et al. 2021; Luz Barcellos and Cataldi 2020; Chen et al. 2020; Spiridonov et al. 2020; Wang et al. 2020; Sousa et al. 2019; Zittis et al. 2017), domains (Taraphdar et al. 2021; Zittis et al. 2017), parameterizations' internal parameters (Yang et al. 2022; Chinta et al. 2021; Afshar et al. 2020), etc. This work was inspired by research that studied variations in physical parameterizations and spatial resolutions. All those cited below used the Weather Research and Forecasting (WRF) model.
Njuki et al. (2022) evaluated the influence of seven planetary boundary layer physical parameterizations on modeling temperature at 2 m, relative humidity at 2 m, wind speed at 10 m, and precipitation in the Kenyan highlands. They realized that only the latter two were sensitive to variations, and one parameterization stood out positively. Taraphdar et al. (2021) studied spatial and temporal distributions of rainfall over the Middle East and the UAE by conducting seven experiments with variations in spatial resolutions, cloud microphysics, PBL parameterizations, and initial and boundary conditions.
Wang et al. (2021) performed 120 simulations for the Amazon region using various parameterizations for cloud microphysics, land surface, PBL, surface layer, and cumulus processes. They concluded that the physical processes influenced certain variables more than others. The cumulus parameterizations interfered more with soil moisture, moisture at 2 m, latent heat, and net radiation. In comparison, the land surface parameterizations were more critical for the temperature at 2 m, sensible heat, condensation, and elevation level. Yoon et al. (2021) sought to improve sea breeze circulation modeling in South Korea by testing combinations of PBL, land surface, longwave radiation, and shortwave radiation parameterizations, evaluating the results for the temperature at 2 m, relative humidity at 2 m, wind speed and direction at 10 m, and vertical wind profile. The authors found the optimal combination of the experiment, which contributed to a 29% improvement.
Yang et al. (2021) studied the variation of cloud microphysics and cumulus parameterizations and their influence on the simulation of extreme precipitation events in a watershed in China. The authors found that the model's sensitivity to cumulus parameterizations was higher than cloud microphysics parameterizations and identified a combination that led to the best and most satisfactory performance.
Wang et al. (2020) compared the performance of precipitation modeling with different cloud microphysics, cumulus, and PBL parameterization schemes for the Central Asian region. These authors also identified an optimal group of parameterizations. However, unlike the work of Yang et al. (2021), they concluded that the cloud microphysics process was more sensitive than cumulus and PBL.
Brazilian researchers also rely on varying parameterizations to investigate and improve the performance of atmospheric modeling. For example, Sousa et al. (2019) studied the performance of two PBL parameterizations and evaluated the modeled height under each influence for the Northeast region. Using several statistical tools, they identified the most efficient scheme. Jacondino et al. (2021) varied parameterizations of cloud microphysics, longwave radiation, shortwave radiation, cumulus, PBL, and land surface to analyze the sensitivity of the model in predicting wind speed in wind farm areas also in Northeast Brazil, seeking to find the combination that would lead to the best performance.
These works show that the sensitivity of the models to specific parameterizations and the quality of the predictions are not universal, as they depend, for example, on the location of the study, the atmospheric variable of interest, and the type of phenomenon analyzed. Therefore, research such as this must be customized for locations and event types that have not yet been investigated.
Luz Barcellos and Cataldi (2020) studied the sensitivity of rainfall forecast to the variation of parameterizations of the physical processes of cloud microphysics, cumulus, land surface, surface layer, and planetary boundary layer for the extreme event of January 2011 in MRRJ. Their research represents a significant step towards improving weather forecasting in this region, and one of the possibilities of advancing in this direction would be to verify the accuracy of forecasts considering the configurations used by these authors. Therefore, in this work, it was chosen to use the same grid configurations and physical parameterizations of Luz Barcellos and Cataldi (2020) for the construction of numerical rainfall forecasts to proceed with the generation of knowledge about the accuracy of weather forecasts for the study area. Qualitative checks of rainfall forecasts are performed, as in their work, but considering more predictions, domains, and spatial resolutions. This work's differential is that the verification of the forecasts occurs through four quantitative indexes, two using observed hourly data and two applying daily observed data. The focus is to measure the accuracy of all possible forecasts quantitatively and qualitatively with the settings considered.