Volume 125, Issue 16 e2019JD032150
Research Article
Free Access

Effects of Convection Representation and Model Resolution on Diurnal Precipitation Cycle Over the Indian Monsoon Region: Toward a Convection-Permitting Regional Climate Simulation

Rakesh Teja Konduru

Rakesh Teja Konduru

Department of Geography, Tokyo Metropolitan University, Tokyo, Japan

Contribution: Methodology, Software, Formal analysis, ​Investigation, Writing - original draft, Writing - review & editing, Visualization

Search for more papers by this author
Hiroshi G. Takahashi

Corresponding Author

Hiroshi G. Takahashi

Department of Geography, Tokyo Metropolitan University, Tokyo, Japan

Correspondence to: H. G. Takahashi,

[email protected]

Contribution: Conceptualization, Methodology, Writing - original draft, Writing - review & editing, Supervision

Search for more papers by this author
First published: 18 July 2020
Citations: 19

Abstract

This study investigated the reproducibility of rainfall characteristics of the Indian summer monsoon using a regional climate model (RCM), focusing on the convection representation and model horizontal grid resolution. To understand the importance of convection representation and model horizontal grid resolution, the study employed the Weather Research and Forecasting model WRFv3.8.1 as the RCM. Two series of experiments, that is, cumulus parameterization on/off (CON/COFF), were performed for the monsoon with three different spatial resolutions. Three contrasting monsoon years were selected for the simulations. The simulated rainfalls were evaluated with the India Meteorological Department data set and Tropical Rainfall Measurement Mission 3B42 Version 7. The results revealed that COFF captured fundamental aspects of the rainfall characteristics, such as diurnal cycle, rainfall intensity, and rainfall frequency. Over central India, in COFF, the diurnal precipitation peaks were simulated in afternoon/evening in close agreement with the observations. However, the diurnal precipitation peak in CON appeared 3–6 hr earlier than the observations. The diurnal precipitation variation was more associated with the horizontal grid resolution of the numerical model than representation of convection. Additionally, COFF could simulate diurnally propagating rainfall systems over the Bay of Bengal. Over the Indian land area, COFF simulated less-frequent intense precipitation systems, whereas CON simulated more-frequent widespread weak precipitation. This study demonstrated that convection representation is very important for simulation of diurnal rainfall systems over the Indian monsoon region, although refinement of model horizontal grid resolution is required over mountainous regions, where precipitation is closely related to orography.

Key Points

  • Diurnal precipitation characteristics of the Indian summer monsoon were simulated realistically without convective parameterization
  • Precipitation characteristics were more dependent on convection representation than model horizontal grid resolution
  • Explicit (parameterized) convection experiments simulate high-intensity localized (high-frequency widespread) precipitation events

1 Introduction

The Indian summer monsoon (ISM), a climate system dominant during June–September, produces 80% of the annual precipitation over South Asia (Jain & Kumar, 2012) and influences both the economy and ecology. The ISM, which can be understood as a land-atmosphere-ocean system, is characterized by lower tropospheric southwesterly that brings moisture into the Indian landmass and causes precipitation (Webster et al., 1977).

Global general circulation models (GCMs) are able to simulate the large-scale features and seasonal precipitation of the ISM to some extent. However, earlier studies on the ISM revealed that GCMs have a massive bias in seasonal precipitation (Goswami et al., 2016). These seasonal precipitation biases are regional and associated with the poorer representation of orography in the coarser grid of GCM (Kumar et al., 2005; Ramu et al., 2016). Murphy et al. (2004) suggested that these regional biases in GCMs are also due to misrepresentation of convective processes, such as cloud formation at subgrid scales. Dirmeyer et al. (2012) showed that current state-of-the-art GCMs with existing convective parameterization and partly resolved convection (like super parameterization) schemes could not properly represent regional convective processes and mentioned that improvement in the horizontal grid resolution could impact these biases. Notable regional biases in precipitation and wind circulation still occurred even with an increase in GCM resolution from ~2° to 0.28° (Sperber et al., 2013).

Regional climate models (RCMs) can provide more details; they span a limited area of the globe and are more useful than GCMs for resolving orography. Several previous studies have tried to simulate the ISM using RCMs (Devanand et al., 2018; Mukhopadhyay et al., 2010; Park & Hong, 2004; Srinivas et al., 2013). All these studies utilized RCMs at a horizontal resolution of 12 to 50 km with the convection parameterized, reproduced the seasonal precipitation, and large-scale monsoon circulation reasonably well with certain ambiguities, which includes biases in the surface temperatures and winds over land and ocean (Lucas-Picher et al., 2011). Other ambiguities includes, misrepresentation of the moist thermodynamic vertical structure of monsoon in an RCM (Raju et al., 2014), which is essential for simulating the thermally direct monsoon circulation (Boos & Hurley, 2013). The major source of such uncertainties associated with the missing physical process and feedbacks over tropics is due to the inclusion of cumulus parameterization in the models (Fersch & Kunstmann, 2014; Rajendran et al., 2013). These cumulus parameterizations introduce a systematic bias in the convection representation in RCM and unrealistically simulate the diurnal cycle of convection (Mukhopadhyay et al., 2010; Takahashi et al., 2010). Therefore, due to the convective nature of the ISM (Romatschke et al., 2010), realistic representation of the diurnal cycle of convective precipitation is crucial in an RCM.

There are limited investigations on the model-simulated diurnal cycle of precipitation during the ISM. The diurnal precipitation system can form the climatological spatial pattern of precipitation over the tropical monsoon region (e.g., Takahashi et al., 2010). Thus, representations of diurnal precipitation variations are required for RCM simulations over tropical monsoon regions. Bhate et al. (2012) compared an RCM using cumulus parameterization with TRMM-3G68; they noted that the RCM could not capture the diurnal phase over central India and the Himalayan foothills, but it was able to simulate the diurnal phase of precipitation over the Bay of Bengal around early morning. In another study on the ISM using the RegCM4.4 model, Bhate and Kesarkar (2019) found that cumulus parameterization is responsible for the out-of-phase uncertainties in the representation of the diurnal cycle of precipitation on comparison with TRMM.

Diurnal cycles of precipitation have been simulated realistically using high-resolution RCMs with explicit convection (no cumulus parameterization) over the tropics (Hara et al., 2009; Jain et al., 2018; Sugimoto & Takahashi, 2016; Takahashi, Yoshikane, et al., 2010). Chen et al. (2018) assessed the intraseasonal oscillation of ISM with the Weather Research and Forecasting model (WRF) as RCM at a single-domain horizontal grid resolution of 9 km, and they mentioned that explicit convection improved ISM simulations. Wu et al. (2018) also studied intraseasonal oscillation of ISM using WRF by using a nested domain with the innermost domain of 4 km, and they mentioned that the explicit convection could improve the ISM simulation. Nevertheless, none of these two studies investigated diurnal convection. Sugimoto and Takahashi (2016) reported that explicit representation of convection in RCM simulations with WRF at a horizontal grid resolution of 5 km (covers Bangladesh) over South Asia could capture certain aspects of diurnal precipitation variations. They also investigated the impact of horizontal grid resolution on the diurnal cycle of precipitation over South Asia, but the use of the nesting technique may have had adverse effects. Willetts et al. (2017) found a later diurnal peak over the Indian land in the U.K. Met Office model simulations integrated for 21 days from 18 August 2011, without convective parameterization at 2.2-, 8-, and 12-km horizontal grid resolutions. In another study with WRF as RCM at 3- and 12-km model horizontal grid resolution, Jain et al. (2018) stated that explicit convection setting could improve the mesoscale propagation over the Bay of Bengal and affects the diurnal convection processes. Thus, the explicit representation of convection over tropics affects the realistic reproducibility of the diurnal cycle. However, the realistic reproducibility of the diurnal cycle has not been explored well over the ISM region.

Therefore, this study investigated the diurnal cycle of ISM precipitation, especially in peak monsoon season, in terms of representation of convective precipitation and horizontal grid resolution of the model. We also explored the reproducibility of precipitation characteristics, such as precipitation amount, intensity, and frequency, as well as the dominant precipitation systems. In section 2 of this paper, we describe the design of the RCM experiment, the data sets, and methods used to evaluate the simulations. In section 3, we evaluate precipitation and then describe the result related to the diurnal characteristics of precipitation in model and observations. In section 4, we discuss possible mechanisms that may explain the diurnal cycle in the model simulations.

2 Data and Methods

2.1 Model and Experiment Design

The reproducibility of the diurnal precipitation cycle over the ISM region was examined using a nonhydrostatic regional atmospheric model, the WRF Version 3.8.1 (Skamarock et al., 2008), as an RCM. Experiments were conducted at various spatial resolutions (25, 12.5, and 6.25 km) to clarify the impact of horizontal grid resolution. To avoid any adverse effects of the nesting technique (Sugimoto & Takahashi, 2016), we selected single-domain experiments for all resolutions. We have conducted several experiments with different single-domain sizes and nested domains in our RCM. We found that different resolution grids produce incoherent convective systems between the domains, for example, when we address a propagating diurnal convective system. Based on our subjective evaluation of model performance for ISM, we decided that we use a single-domain setting in the current study. This setting can reproduce more coherent convective systems over the ISM region than the nested domain settings following a review on RCM by Giorgi (2019). The single domain covers entire India subcontinent (5–38°N, 65.6–100°E; Figure 1a). The model utilized 30 terrain-following vertical levels. Additionally, to investigate the reproducibility related to the convective parameterization, we conducted cumulus parameterization “off” (COFF; explicit convection) and cumulus parameterization “on” experiments (CON). The COFF experiments imply that convective precipitation was calculated explicitly, and model calculation of convection depends on the internal dynamics and microphysics. Indeed, 6.25-km COFF simulation partially resolved the convection, and we have considered this experiment as convection permitting. Nevertheless, in the absence of cumulus parameterization that correctly resolves the subgrid scale convection in this horizontal resolution has left us with no choice to choose COFF as presented by Chen et al. (2018) and Zhang and Smith (2018). Thus, we performed six kinds of experiments for each case.

Details are in the caption following the image
(a) The model domain over the Indian region and (b) the topography (meters) over the domain. The subregions shown in (a) are the foothills of the Himalaya (FoH; 27–29°N, 77–88°E), the central India region (CIR; 20–25°N, 75–88°E), southern peninsular India (SPI; 15–19°N, 77–82°E), and the Bay of Bengal (BoB; 12–18°N, 85–95°E). The green and orange regions in (a) are the all-India land and the northeast complex mountain, respectively.

All of the simulations were driven using the ERA-Interim data set as initial and lateral boundary conditions (Dee et al., 2011). Lateral boundary conditions were updated every 6 hr. The vegetation fraction and land use land cover were considered from the MODIS data set. Only surface conditions like soil moisture and soil temperature were initialized in the simulation from the ERA-Interim data set. Daily Optimum Interpolation Sea Surface Temperature (Reynolds et al., 2007) was prescribed every 6 hr during the simulations. A default value of five rows was used for boundary value nudging. We have not performed any spectral nudging in the model simulation.

We performed simulations for three contrasting monsoon years, 2007, 2009, and 2012, which were above-normal, drought, and normal years, respectively, based on reports from the India Meteorological Department. We used the contrasting monsoon years as three ensembles to obtain robust results that did not depend on a specific monsoon condition. For each year, the six simulations were performed from 1 April to 31 October, and the first 2 months (1 April to 31 May) were excluded as spin-up time and considered only peak or mature monsoon season from 11 July until 10 September as analysis period.

Common to COFF and CON, the following physical parameterization schemes were used in the experiments. The Unified Noah Land Surface Physical Scheme (Chen & Dudhia, 2001) was employed to couple the land surface with the atmosphere. A GCM version of the Rapid Radiative Transfer Model for longwave radiation (Iacono et al., 2008) and the updated Goddard shortwave schemes (Shi et al., 2010) was used. We utilized the WRF single-moment six-class microphysics scheme (Hong et al., 2004). The Mellor-Yamada-Janić boundary layer scheme (Mellor & Yamada, 1982) was used to simulate the planetary boundary layer. In the case of CON, we used the KF (Kain, 2004) cumulus parameterization scheme, which was switched off in COFF. The recent studies of Jayasankar et al. (2018) and Ratnam et al. (2017) have become our motivation to choose KF scheme in the simulation settings. Other physics configurations were considered according to Sugimoto and Takahashi (2016); they used these settings over the South Asian tropical regions of Bangladesh.

2.2 Observations

The India Meteorological Department gridded daily rainfall (Pai et al., 2014) at 0.25° (IMD25, hereafter) was employed to evaluate the simulated monsoon rainfall. IMD25 was derived from the daily record of 6,955 rain gauge stations over the entirety of India. In addition to IMD25, we used data from the Tropical Rainfall Measuring Mission 3B42v7 (TRMM-3B42) (Huffman et al., 2007) to evaluate the diurnal cycle of precipitation. TRMM-3B42 data consist of 3-hourly precipitation rates (mm hr−1) on a 0.25° nearly global grid (60°S–60°N). The 3-hourly precipitation rates were used to evaluate the diurnal cycle of precipitation and compute precipitation characteristics. These two precipitation data sets were obtained for contrasting monsoon years 2007, 2009, and 2012 from 11 July until 10 September.

2.3 Evaluation Methods

2.3.1 Preprocessing of Simulated Results

The simulated precipitation was evaluated in a four-step procedure that includes spatial distribution, diurnal cycle, precipitation characteristics, and objective evaluation of precipitation systems. Before evaluating the simulated results statistically at the various resolutions, we unified the horizontal grid resolutions into a 0.25° grid by using bilinear interpolation because the spatial resolution of the observed precipitation was 0.25°. Thus, we could fairly evaluate the simulated results of the various resolutions. The mature monsoon season was utilized to calculate the season mean for contrasting years (2007, 2009, and 2012). Similarly, the IMD25 and TRMM-3B42 precipitations were averaged for those years. All results are presented as ensemble mean of the three contrasting years.

To analyze the diurnal precipitation cycle, we calculated the climatology of the TRMM-3B42 and model-simulated 3-hourly precipitation for the three contrasting years (2007, 2009, and 2012). We extracted the first harmonic to investigate 24-hr diurnal precipitation variation by harmonic analysis, which is the most fundamental and varies with solar insolation. For the first harmonic, diurnal amplitude (maximum precipitation rate) and diurnal phase (diurnal timing of the maximum precipitation rate) in local solar time (LST) were computed.

Similar to the precipitation preprocessing, the ensemble diurnal means of the contrasting monsoon years were calculated for atmospheric variables such as lower atmosphere winds and total precipitable water. The diurnal components of these atmospheric variables were computed by subtracting the daily mean from the diurnal mean at each timing.

2.3.2 Statistical Analysis

Root-mean-square error is often used to analyze the similarity between simulated and observed values. However, by utilizing the mean square difference (MSD: MSD2 = MB2 + Variance2), it is possible to identify which component of spatial variation, such as the mean difference (MB), pattern correlation, or variance, is affecting the model performance. In this study, these statistics were computed by comparing the ensemble means of peak monsoon precipitation from the model simulations with IMD25 over the Indian land region (green area in Figure 1a). We presented these statistics in a Taylor diagram (Taylor, 2001) to attain a pictorial view of the model performances with observations (Table 1). Similar statistics, also shown in Table 1, were computed for precipitation characteristics by comparing the area average (green area in Figure 1a) diurnal variations of the model simulations with TRMM-3B42.

Table 1. Statistical Evaluation Is Shown Between Observation and Model-Simulated Mean Peak Monsoon Precipitation (PMP) Over the India Land Region (Green Area in Figure 1a)
Type Simulations Mean (units) Mean difference (units) Correlation MSD
PMP (mm day−1) (mm day−1)
Observation 8.3 <8.77>
6.25 7.99 (10.23) −0.34 (1.90) 0.82 (0.73) 0.24 (3.92)
12.5 7.57 (10.67) −0.76 (2.34) 0.80 (0.87) 0.72 (5.57)
25 8.14 (10.89) −0.19 (2.56) 0.79 (0.90) 0.27 (6.58)
DCPA (mm day−1) (mm day−1)
Observation [5.86]
6.25 7.49 (9.95) 1.63 (4.09) 0.98 (0.71) 4.02 (19.7)
12.5 7.32 (10.44) 1.46 (4.58) 0.95 (0.65) 3.64 (26.0)
25 7.83 (10.61) 1.97 (4.75) 0.67 (0.54) 5.09 (32.4)
DCPI (mm hr−1) (mm hr−1)
Observation [2.01]
6.25 2.01 (0.42) 0.0 (−1.59) 0.99 (0.91) 0.00 (2.53)
12.5 2.02 (0.44) 0.01 (−1.57) 0.99 (0.90) 0.00 (2.48)
25 2.03 (0.45) 0.02 (−1.56) 0.94 (0.10) 0.03 (2.49)
DCPF (%) (%)
Observation [9.95]
6.25 13.93 (20.2) 3.98 (10.2) 0.97 (0.72) 24.78 (116)
12.5 13.59 (24.0) 3.64 (14.0) −0.23 (0.76) 35.6 (206)
25 13.96 (29.9) 4.01 (19.9) 0.77 (0.74) 28.8 (422)
  • Note. Similar statistics are also shown for the diurnal characteristics of precipitation amount (DCPA), intensity (DCPI), and frequency (DCPF). The mean value of PMP in observation data is shown as IMD <TRMM-3B42>. The statistics of cumulus parameterization off (on) are shown in rows as COFF (CON). For diurnal precipitation characteristics, the mean TRMM-3B42 result is shown in the observation row as “[].” Here, MSD is the mean square difference statistic. Bold MSD shows best performance of the simulation.

2.3.3 SAL Method

The upscale propagation of convection-related error leads to a “double-penalty” problem in evaluations of a high-resolution simulation with observations (Prein et al., 2015). To address this problem, Wernli et al. (2008, 2009) introduced a qualitative structure (S), amplitude (A), and location (L) error analysis of precipitation. In a SAL analysis, a precipitation system is defined over the selected region as precipitation greater than a threshold over the entire region. In this study, we utilized the threshold of greater than the 98th percentile of precipitation over the area under consideration. The SAL precipitation system consists of S, A, and L component errors computed by comparing TRMM-3B42 with simulated precipitation; these are presented on the SAL plot (for a pictorial view, see supporting information Figure S1). Here, the “S” component describes the error in the weighted volume average of all selected precipitation systems over the considered area. The “A” component denotes the error in the area-mean precipitation value, and the “L” component explains the error in the center of gravity of the precipitation system over the area. To compute these components, we followed the procedure described by Wernli et al. (2008). In a SAL plot, points located near the origin indicate that the simulated precipitation system agrees with observation, otherwise shows some error. The SAL method is equivalent to subjective visual judgment of the simulated precipitation system compared with observations.

2.3.4 Definitions of Precipitation Characteristics

The definitions of precipitation characteristics were determined based on previous work by Takahashi (2016), Chen and Dai (2018), and Takahashi and Polcher (2019). Briefly, 3-hourly precipitation rates of TRMM-3B42 and model simulations were used to compute precipitation characteristics. Precipitation amount (mm day−1) was defined as the total precipitation accumulated during the period greater than 0.1 mm hr−1. Precipitation frequency (%) was defined as the number of observations with precipitation greater than 0.1 mm hr−1 compared to the total number of observations. Precipitation intensity (mm hr−1) was defined as the ratio of the rainfall amount to the rainfall frequency.

3 Results

3.1 Spatial Distribution of Peak Monsoon Precipitation

This subsection assesses the performance of the simulated precipitation by the various settings (Figure 2). The detailed spatial pattern of precipitation was obtained from IMD25, which was used as an observation. Spatial patterns of simulated precipitation were generally reproduced in all settings. In IMD25 (Figure 3a), major spatial peaks in seasonal precipitation were observed over the western coast of India (western slope of the Western Ghats), central India, the foothills of the Himalayan mountains, and northeastern India. These spatial peaks were generally simulated in all settings. The spatial pattern correlations between simulated seasonal precipitation and observation (IMD25) were larger than 0.70 in all settings (Figure 2 and Table 1). Normalized standard deviations in simulated seasonal precipitation were much lower than observations. These results indicate that the simulations could generally simulate the spatial pattern of seasonal precipitation amount. However, spatial contrasts between dry and wet climates were weakly simulated. In the cases of low-resolution (25 and 12.5 km) simulation, large differences were observed in spatial correlations and normalized standard deviations.

Details are in the caption following the image
Ensemble mean peak monsoon precipitation evaluated between IMD observations and model simulations at different mesh sizes (25, 12.5, and 6.25 km) with cumulus parameterization off (COFF) and on (CON) over all Indian land (green area in Figure 1a) by Taylor diagram. REF points to a spatial variance of reference IMD data set. Correlations were calculated by employing a pattern correlation between IMD and model simulations. Dashed arcs represent normalized standard deviation, and solid half circles indicate centered root-mean-square error. The ensemble mean was calculated based on the three contrasting monsoon years of 2007, 2009, and 2012.
Details are in the caption following the image
Spatial distribution of ensemble mean of peak monsoon precipitation (mm day−1) displayed for IMD observations (a) and model simulations at different mesh sizes (25, 12.5, and 6.25 km) with cumulus parameterization on (CON; b–d) and off (COFF; e–g). The ensemble mean was calculated based on the three contrasting monsoon years of 2007, 2009, and 2012.

3.2 Different Simulated Monsoon Precipitation Biases Between COFF and CON

This subsection focuses on the differences in the horizontal grid resolution of the model and the representation of moist convection in depictions of simulated precipitation. The spatial pattern of seasonal precipitation was commonly simulated overall (Figures 3b3g). Heavy precipitation spatial patterns were simulated over central India, where monsoon precipitation was observed climatologically. Additionally, a large amount of precipitation was simulated over mountains of northeastern India and along the western slope of the Western Ghats. Dry regions were simulated over northwestern India and the eastern plain of the Western Ghats.

The spatial patterns of simulated precipitation in the COFF and CON (Figures 3b3g) differed significantly in some aspects, although both simulated the general spatial pattern of seasonal precipitation. Over central India, significant wet biases occurred in CON. These biases remained even when we used the high-resolution setting. Along the foothills of the Himalaya, large wet biases were observed in CON; these biases could be reduced by an increase in horizontal grid resolution. The wet biases over central India and the foothills of the Himalaya were not clear in COFF. However, over northeastern India, large wet biases were simulated in all settings. For these regions with very complex topography, much higher resolution may be required for representation of precipitation. Over the western slope of the Western Ghats, all settings yielded wet biases, particularly in COFF. These wet biases in COFF were improved as the horizontal grid resolution became higher.

The overall spatial reproducibility of the COFF and CON simulations was evaluated by using an MSD statistic calculated for peak monsoon precipitation (Table 1; PMP). When comparing the MSD values for all simulations, it was evident that the COFF and CON simulations had values less than 0.75 and greater than 3.75, respectively. Especially, 6.25-km COFF simulations showed the least MSD of 0.24 among all COFF experiments. These MSD statistics reveal the higher reproducibility of seasonal monsoon precipitation over India in COFF that fixes significant biases seen in CON over India land and discloses the added value of 6.25-km COFF simulation.

3.3 Diurnal Cycle of Peak Monsoon Precipitation

This subsection analyzes the diurnal cycle of precipitation over the Indian monsoon region because tropical monsoon precipitation is basically convective, which can be associated with diurnally varying precipitation systems. Tropical precipitation is known to occur in the afternoon/evening hours over land and in the midnight/morning hours over ocean, with some exceptions attributable to coastal and topographic effects. Afternoon/evening precipitation was observed over central India, and early morning precipitation was observed over the western Bay of Bengal (Figure 4a). Nocturnal precipitation was observed along the southern slope of the Himalaya. These observational facts are consistent with the more accurate observations of previous studies (e.g., Hirose & Nakamura, 2005; Takahashi, Fujinami, et al., 2010). The propagating precipitation system over the Bay of Bengal in the late morning and early afternoon has also been studied (Zuidema, 2003). Note that diurnal variations in precipitation were weak along the western coast of India and the western slope of the Western Ghats.

Details are in the caption following the image
Ensemble mean of diurnal phase and amplitude of precipitation rate (mm hr−1) during peak monsoon season for TRMM-3B42 observations (a) and model simulations of different mesh sizes (25, 12.5, and 6.25 km) with cumulus parameterization on (CON; b–d) and off (COFF; e–g). The ensemble mean was calculated based on the three contrasting monsoon years of 2007, 2009, and 2012. The color of the wheel represents the diurnal phase in local solar time (LST) from 030 LST to 2330 LST. The intensity of color shows the diurnal amplitude, which increases from the center to the outside of the wheel from 0 to 1.6 mm hr−1 with an interval of 0.2.

The observed precipitation peaks were generally simulated in all settings of experiments (Figures 4b4g). Daytime precipitation over central India and nocturnal precipitation along the foothills of the Himalaya were simulated in all of the settings. However, the reproducibility of morning and daytime precipitation over the Bay of Bengal was largely different between COFF and CON. CON produced earlier peaks over the Bay of Bengal than observations, which can be associated with the reproducibility of the propagating precipitation systems (also see Figure 5). Weak diurnal variations were simulated along the western coast of India, due to orographic lifting by the Western Ghats.

Details are in the caption following the image
Ensemble mean diurnal precipitation rate (mm hr−1) at different diurnal timings during peak monsoon for TRMM-3B42 observations and model simulations at mesh size of 6.25 km with cumulus parameterization off (COFF) and on (CON). The ensemble mean was calculated based on the three contrasting monsoon years of 2007, 2009, and 2012. The diurnal timings of night (2330 LST; a–c), morning (0530 LST; d–f), noon (1130 LST; g–i), and evening (1730 LST; j–l) are shown in local solar time (LST).

Concerning the detailed spatial pattern of diurnal precipitation variations, Figure 5 presents the diurnal evolution of precipitation in 6.25-km COFF and CON. The diurnal amplitude and phase in COFF simulations were spatially diverse over land in India and the Bay of Bengal, whereas CON simulations had a relatively homogeneous spatial distribution of diurnal variations and stronger precipitation (Figure 5). Over central and southern peninsular India, a realistic diurnal peak of 0.6–0.8 mm hr−1 at around 1430–1730 LST was simulated in the COFF experiments. However, a noon to afternoon diurnal peak of 0.8–1.0 mm hr−1, which was somewhat earlier and stronger than the diurnal peak in COFF, was dominant in CON. Over the foothills of the Himalaya, both COFF and CON simulated an early morning peak (0230–0530 LST) of more than 1.0 mm hr−1, which originated from afternoon precipitation over the southern slope of the Himalaya. The simulated precipitations over the southern slope and foothills of the Himalaya were basically the same between the COFF and CON experiments. Additionally, the reproducibility of the diurnal variation over these complex mountain regions was more associated with the horizontal grid resolution of the numerical model than the representation of convection (Figure 4). Along the windward side of the Western Ghats, diurnal precipitation variations were very weak in both COFF and CON. In contrast, diurnally varying precipitation was clearly simulated over the eastern coast of India and western Bay of Bengal in COFF. Specifically, the coastal precipitation systems propagated offshore to the center of the Bay of Bengal until morning, thus showing a nocturnal propagation of precipitation systems (e.g., Zuidema, 2003). However, in CON, the precipitation over the eastern coast of India was very weak, but stronger precipitation occurred over the whole open ocean of the Bay of Bengal, which was largely different from the observation and COFF.

3.4 Objective Evaluation of Hourly Precipitation Systems

The diurnal precipitation cycles simulated by 6.25-km COFF and CON were studied objectively using SAL (Wernli et al., 2008, 2009) over three regions (see Figure 1a): central India (20–25°N, 75–88°E), southern peninsular India (15–19°N, 77–82°E), and the Bay of Bengal (12–18°N, 85–95°E). These regions were identified based on the dissimilar diurnal phase differences between modeled and TRMM-3B42 results (Figure 5d). To carry out the analysis, each hourly map of TRMM-3B42 and model-simulated 3-hourly precipitation was collocated for a region throughout the season for 3 years (out of 183 samples 180 were used and 3 were omitted due to very little precipitation), and SAL was applied over those samples. The explanation of the SAL plot analysis was shown in Figure S1.

The SAL evaluation (Figure 6 and Table 2) yielded more peaked, localized-type precipitation (S < 0 and A ~ 0 or A < 0) over central India and southern peninsular India during evening in COFF but widespread-type (S > 0 and A > 0) precipitation systems in CON. The concentrations of points around/left of the origin in COFF resemble/underestimate the precipitation systems of TRMM-3B42; however, in CON, the points are scattered to the right of the origin, indicating the overestimation of structure and amplitude. The shift in the interquartile range window position from right of the origin in the 6.25-km CON to the left of it in the 6.25-km COFF displays a straight decrease in the structure and amplitude errors of precipitation systems. For example, on 29 July 2012, a mesoscale precipitation system formed in the evening over central India, as seen in TRMM-3B42 (Figure S2). A similar precipitation system appeared to be simulated realistically in 6.25-km COFF than in CON simulation. This reproducibility of precipitation system was evaluated with the SAL components that show the least error considering structure (S = 0.002) and location (L = 0.001) of the system in 6.25-km COFF and large error in the structure (S > 0.6) in 6.25 and 25-km CON that shows large precipitating systems than observation.

Details are in the caption following the image
Three-hourly precipitation rates (mm hr−1) during peak monsoon compared objectively between TRMM-3B42 and model simulations of the 6.25 km using SAL analysis. This analysis was done at evening diurnal time over central India (a, b; 20–25°N, 75–88°E) and southern peninsular India (c, d; 15–19°N, 77–82°E) and at early noon over the Bay of Bengal (e, f; 12–18°N, 85–95°E). The first and second columns show model simulations at the mesh size of 6.25 km with cumulus parameterization off (COFF) and on (CON), respectively. The data set includes 180 samples for the 3 years (2007, 2009, and 2012), which are represented as points. Squares, filled circles, and stars are points indicating July, August, and September of monsoon season, respectively. Please refer SAL plot example in Figure S1.
Table 2. The Medians of Structure (S), Amplitude (A), and Location (L) Errors Between 3-Hourly TRMM-3B42 Data and Model-Simulated Monsoon Precipitation Over Central India (CIR; 20–25°N, 75–88°E), Southern Peninsular India (SPI; 15–19°N, 77–82°E), and the Bay of Bengal (BoB; 12–18°N, 85–95°E)
Type Simulation S A L Precipitation system physical nature
CIR 6.25 (COFF) −0.41 −0.10 0.03 Peaked localize
12.5 (COFF) −0.28 −0.09 0.02 Peaked localize
6.25 (CON) 0.20 −0.10 0.01 Widespread
12.5 (CON) 0.51 −0.17 0.06 Widespread
25 (CON) 1.02 −0.01 0.10 Widespread
SPI 6.25 (COFF) 0.06 0.20 0.00 Uniform precipitation rates
12.5 (COFF) −0.14 −0.33 0.00 Peaked localize
6.25 (CON) 0.81 0.43 0.11 Large widespread
12.5 (CON) 1.02 0.40 0.06 Large widespread
25 (CON) 1.34 0.75 0.10 Large widespread
BoB 6.25 (COFF) 0.24 0.89 0.09 Large widespread
12.5 (COFF) 0.31 0.90 0.10 Large widespread
6.25 (CON) 1.22 1.19 0.23 Large widespread
12.5 (CON) 1.27 0.95 0.22 Large widespread
25 (CON) 1.38 0.85 0.21 Large widespread
  • Note. The SAL analysis results are shown in rows for different resolution simulations with cumulus parameterization off (on) COFF (CON). The SAL statistics explain the physical nature of the error in simulated precipitation system in comparison with observation data (Wernli et al., 2008).

In contrast to the case for land, over the Bay of Bengal, the precipitation systems in all simulations were the large, widespread type during the early afternoon, as opposed to the small precipitation systems in TRMM-3B42. This is due to increased structure, amplitude, and location errors. In all CON simulations, the SAL metric showed that precipitation systems were in the order of 2–3 times the observation (A > 1.0, S > 0.80, and L > 0.20), which can be seen from the concentration of points near the top-right quadrant corner. However, COFF simulated much smaller systems (S ~ 0.3, L ~ 0.10, and A ~ 0.80) than CON, as can be seen from the approximately 50% of points around/left of the origin in COFF, which agrees with TRMM-3B42 (similar example like Figure S2 is shown over Bay of Bengal in Figure S3).

From the SAL analysis, we could objectively analyze the precipitation system types in COFF and CON. The peaked, localized precipitation in COFF corresponded to higher precipitation intensity, whereas CON simulated widespread precipitation systems over land and the Bay of Bengal (see Figure S4, a cross section over Bay of Bengal showing precipitation systems). Furthermore, the widespread precipitation in CON corresponded to a greater number of precipitation systems generation, which could aggregate, resulting in higher precipitation amounts and frequency. SAL analysis reveals the added value in simulating the precipitation systems by 6.25-km COFF than CON (see Figure S5, which shows precipitation system types based on cloud top temperatures [°C]). These results imply that the reproducibility of 6.25-km COFF than CON can be associated with their precipitation characteristics.

3.5 Precipitation Characteristics in COFF and CON

To reconfirm the statistical analysis of SAL in the previous subsection, we analyzed precipitation characteristics, such as precipitation intensity and frequency (Figure 7). The definitions of the precipitation characteristics are provided in section 2.3.4. The diurnal cycle of precipitation characteristics was calculated and area averaged over four subregions (shown in Figure 1a): the foothills of the Himalaya, central India, southern peninsular India, and the Bay of Bengal.

Details are in the caption following the image
Area average diurnal cycles of precipitation amount (a–d; mm day−1), frequency (e–h; %), and intensity (i–l; mm hr−1) for TRMM-3B42 (black line) and model simulations of different mesh sizes (25, 12.5, and 6.25 km) with cumulus parameterization off (COFF; thick lines) and on (CON; thin lines). The ensemble mean was calculated based on the three contrasting monsoon years of 2007, 2009, and 2012. The area average of the ensemble mean was performed over central India (20–25°N, 75–88°E), southern peninsular India (15–19°N, 77–82°E), the foothills of the Himalaya (27–29°N, 77–88°E), and the Bay of Bengal (12–18°N, 85–95°E). The abscissa shows diurnal timings in local solar time, and ordinate of each panel is different.

In COFF, diurnal peaks of precipitation amount and frequency were observed in the evening over the central and southern peninsular regions of India, similar to observation data. Additionally, in COFF, diurnal peaks of precipitation intensity were observed in the evening or late evening over central and southern peninsular India, although precipitation intensity in COFF was largely overestimated. CON simulated peaks in precipitation amount and precipitation frequency at noon/afternoon, approximately 3 hr earlier than in COFF, over the same regions. Moreover, precipitation intensity in CON was much weaker over the central and southern peninsular regions than observation data. It is noteworthy that these tendencies of precipitation characteristics over the central and southern peninsular regions are basically consistent with the results of the SAL analysis, although the COFF simulations also had some biases.

COFF and CON simulated twofold larger precipitation amounts along the foothills of the Himalaya and over the Bay of Bengal than observation data. With regard to the overestimation biases of COFF and CON simulations, an overestimation of precipitation intensity significantly contributed to the overestimation of precipitation amount in the COFF experiments, whereas an overestimation of precipitation frequency contributed to the overestimation of precipitation amount in the CON experiments. However, because the diurnal phases in precipitation characteristics over the foothills and southern slope of the Himalaya regions are complex due to the complicated topography, biases could be canceled by averaging. Moreover, along the foothills of the Himalaya, the effects of model horizontal grid resolution can be very significant, in addition to the representation of convection. Over the Bay of Bengal, CON simulated earlier diurnal precipitation peaks, as shown in Figure 5.

In summary, the diurnal amplitude of precipitation amount over the central and southern India regions in CON was overestimated due to higher precipitation frequency. In contrast, although the COFF and CON experiments had notable biases in precipitation characteristics along the foothills of the Himalaya, higher resolution of model grid spacing improved the simulated precipitation characteristics. Over the Bay of Bengal, the representation of convections was significantly associated with the reproducibility of diurnally propagating precipitation systems. Further, statistical (Table 1 and Figure S6) analysis shows higher (unrealistic) reproducibility of the precipitation amount, frequency, and intensity by 6.25-km COFF (CON), which can be ascribed from the lowest (highest) MSD.

4 Different Local Circulations Between COFF and CON

The previous sections discussed how the simulated precipitation characteristics, particularly the diurnal variations, differed in some aspects between COFF and CON. These differences between COFF and CON can be associated with the mechanisms of the diurnal precipitation systems. Generally, diurnally varying convective systems are associated with local circulations. These features over the tropical Asian monsoon regions have been examined in previous studies (Fujinami et al., 2017; Takahashi, Fujinami, et al., 2010; Takahashi, Yoshikane, et al., 2010). This section investigates possible mechanisms that could explain the differences in precipitation characteristics between COFF and CON. It focuses only on the 6.25-km COFF and CON experiments.

The spatial patterns of the ensemble mean diurnal components of total precipitable water and surface winds (Figure 8) and low-level winds at 925 hPa (Figure 9) were generally associated with the diurnal pattern of precipitation in both COFF and CON.

Details are in the caption following the image
Diurnal components during peak monsoon are shown for the ensemble mean of total precipitable water vapor (mm; shaded) and 10-m winds (m s−1; vectors) of simulations at 6.25 km with cumulus parameterization off (COFF; a, d, g, j) and on (CON; b, e, h, k). The difference between 6.25-km COFF and CON is displayed in (c), (f), (i), and (l). The ensemble mean was calculated based on the three contrasting monsoon years 2007, 2009, and 2012. The diurnal components of night (2330 LST), morning (0530 LST), noon (1130 LST), and evening (1730 LST) are shown in local solar time (LST). These were calculated by removing the daily mean at each diurnal time. The reference vector of 1 m s−1 was considered to plot 10-m winds.
Details are in the caption following the image
Similar to Figure 8, but for the ensemble mean diurnal components of 925-hPa low-level wind direction (vectors; reference vector is 1 m s−1) and wind speed (m s−1; shaded).

In the late evening (around 2330 LST), positive precipitable water anomalies were simulated along the inland area of South India (Figure 9). It shifted along the eastern coast of India late at night, and some positive anomalies were also found over the entirety of India. It is noteworthy that positive and negative signals were stronger in CON than in COFF. Anomalous southerlies were simulated at the near-surface and lower troposphere along the southeastern coast of the Indian subcontinent that strengthens water vapor toward the land. These atmospheric conditions were favorable for the evening precipitation over central and southern peninsular India and the nocturnal precipitation along the foothills of the eastern Himalaya in both simulations (see Figure 5).

In the early morning (around 0530 LST), positive precipitable water anomalies were simulated over the same regions in both COFF and CON, along the eastern coast of the Indian subcontinent and over the western Bay of Bengal. Weak positive anomalies were also found in the northern and eastern Bay of Bengal. With regard to late evening, precipitable water anomalies were stronger in CON than in COFF. Additionally, positive anomalies were simulated along the foothills of the Himalaya. Negative anomalies were simulated over most of the Indian subcontinent. Anomalous low-level winds changed from blowing toward land in the late evening to blowing toward the ocean in the early morning. Anomalous northwesterly low-level winds were simulated over the eastern coast of the Indian subcontinent and offshore over the western Bay of Bengal. These atmospheric conditions can explain the precipitation in the early morning over the eastern coast of the Indian subcontinent and offshore over the western Bay of Bengal. Concerning precipitable water, low-level winds were also stronger in CON than in COFF, which can explain the stronger precipitation signals in CON over the Bay of Bengal (Figure 5). The low-level easterlies along the foothills of the Himalaya, which are stronger in CON than COFF, and the higher positive precipitable water anomalies may support the morning precipitation there (Figure 7c).

Around noon, anomalous northwesterly and anomalous cyclonic circulations were dominant in the low-level winds along the eastern coast of the Indian subcontinent and the offshore western Bay of Bengal, with positive precipitable water anomalies, in both COFF and CON, which were consistent with the daytime precipitation over the Bay of Bengal. In this diurnal phase, the strength of the surface and low-level circulations largely differed between COFF and CON. Stronger circulations with larger anomalies in precipitable water were simulated in CON than COFF, which is consistent with the overly strong and earlier daytime precipitation in CON over the Bay of Bengal (Figure 7d). Additionally, CON simulated larger positive anomalies in precipitable water and stronger local circulations over the coastal regions of India around noon. These differences in atmospheric circulation can explain the stronger precipitation over the coastal and plain regions of India during the hours before noon and the early afternoon, which had earlier precipitation peaks than observations.

A precipitation system from early morning to noon that travels southeastward from the western coast of the Bay of Bengal to the central Bay of Bengal, which can be associated with the low-level southeastward winds, was simulated in COFF (see Figure S7, a snapshot from Figure S5 that shows offshore-propagating precipitation systems). The precipitation system simulated in CON over the Bay of Bengal during the midnight or early morning was not likely to travel southeastward, but a stronger precipitation system developed within a shorter number of hours. This type of difference in diurnally propagating precipitation systems and characteristics is an important aspect differentiating COFF and CON.

Around evening, anomalous winds toward land were simulated over the southern Indian subcontinent with positive anomalies in precipitable water. Compared with the late evening, the notable winds toward land were simulated only over the southern Indian subcontinent, not over central India. However, in the late evening, low-level winds toward land developed into central India. This could be a portion of land-sea breeze circulations at the scale of the Indian subcontinent. Around central India, the development of local circulations is likely to take longer than those over southern India, due to the larger spatial scale.

Importantly, contrasting diurnal variations of precipitable water over central India and the Bay of Bengal were striking features during late evening/night to morning/early noon. These indicate low-level wind convergence (divergence) over central India in late evening/night (morning/early noon) and synchronous divergence (convergence) over the Bay of Bengal. Both COFF and CON captured these contrasting features, but the diurnal amplitude and phase are stronger and earlier, respectively, in CON than in COFF, which may closely related with the representativeness of the diurnal convective activity. The strengths of the local circulations are likely associated with the diurnal moist convection over this region, and deeper understanding will be needed on this point.

To improve understanding of the mechanisms behind variations in diurnal precipitation, we also need to analyze various processes including microphysical growth, land surface processes, and detailed orographic effects.

5 Conclusions

To understand the reproducibility of diurnal precipitation cycle over South Asia, this study investigated precipitation characteristics of the Indian summer monsoon during the mature monsoon season using a series of climate simulations by WRF. We ran simulations at three different resolutions and with two kinds of representation of convection by switching cumulus parameterization on/off (CON/COFF). Particularly, we focused on diurnal precipitation cycles over and around India and evaluated these based on TRMM-3B42 observations.

The evaluation of model-simulated precipitation over the landmass of India with IMD25 revealed the lowest MSD among all simulations of 0.24 in the 6.25-km COFF, but CON simulations had MSD values of more than 3.0. This low MSD in the 6.25-km COFF indicated better spatial reproducibility of monsoon precipitation over India.

We found that the diurnal cycles of precipitation amplitude and phase over central India in the COFF simulations were closely related to TRMM-3B42 observations. CON showed a 3–6 hr earlier diurnal peak than TRMM-3B42 with overestimated diurnal amplitude. Over the Bay of Bengal, southeastward-propagating precipitation systems were simulated in COFF, which was consistent with observation data. However, CON could not reproduce such propagation and instead simulated short-lived systems in the early morning.

A SAL analysis revealed that COFF calculated higher intensity, was more localized, and had lower-frequency precipitation systems than CON. CON simulated low intensity, was more widespread, and had higher-frequency precipitation systems. The overestimation of precipitation amount in CON was associated with the higher frequency of widespread precipitation systems. Further analysis revealed that the higher frequency of precipitation in CON was due to strong anomalous diurnally varying local circulation at the lower troposphere. The simulated precipitation characteristics in COFF were closer to observation data than CON in some aspects, although COFF still had some biases in precipitation characteristics.

Our analysis revealed that the reproducibility of monsoon mean precipitation and diurnal cycle is both dependent on the convection representation in the model, rather than on the model mesh size. Nevertheless, refinement of mesh size is very important for regions including the foothills of the Himalaya and the western slope of the Western Ghats, where orographic precipitation is dominant. Further numerical simulations are required to clarify the actual control on evening precipitation, late-night precipitation, and early morning precipitation over the Indian monsoon region.

Acknowledgments

This work was partly supported by the Japan Society for the Promotion of Science (JSPS) KAKENHI Grant 19H01375, Japan Aerospace Exploration Agency (JAXA) EO-RA2 (PI ER2GPF012), and “Tokyo Human Resources Fund for City Diplomacy” from the Tokyo Metropolitan Government, Japan. We thank the India Meteorological Department (IMD) for providing the gridded precipitation data sets.

    Data Availability Statement

    The Weather and Research Forecasting model Version 3.8.1 is available from the repository of National Center for Atmospheric Research (http://www2.mmm.ucar.edu/wrf/src/). The Era-Interim data are obtained via the European Center for Medium-Range Weather Forecasts public reanalysis-data sets web interface (https://www.ecmwf.int/en/forecasts/datasets). The OISST v2 data are obtained from the online website of the NOAA Earth System Research Laboratory's Physical Sciences Division (https://www.ncdc.noaa.gov/oisst). The Tropical Rainfall Measuring Mission precipitation data are available online (https://pmm.nasa.gov/data-access/downloads/trmm).