Volume 123, Issue 22 p. 12,621-12,646
Research Article
Open Access

Sensitivity of the Amazonian Convective Diurnal Cycle to Its Environment in Observations and Reanalysis

Kyle F. Itterly

Corresponding Author

Kyle F. Itterly

Science Systems and Applications, Inc., Hampton, VA, USA

Correspondence to: K. F. Itterly,

[email protected]

Contribution: Conceptualization, Methodology, Validation, Formal analysis, ​Investigation, Writing - review & editing

Search for more papers by this author
Patrick C. Taylor

Patrick C. Taylor

NASA Langley Research Center, Climate Science Branch, Hampton, VA, USA

Contribution: Conceptualization, Methodology, ​Investigation, Resources, Writing - review & editing, Supervision, Funding acquisition

Search for more papers by this author
Jason Brant Dodson

Jason Brant Dodson

Science Systems and Applications, Inc., Hampton, VA, USA

Contribution: Conceptualization, Methodology, Writing - review & editing

Search for more papers by this author
First published: 31 October 2018
Citations: 7

Abstract

Atmospheric model parameterizations of tropical deep convection struggle to reproduce the observed diurnal variability of convection in the Amazon leading to climatological biases in the energy budget and water cycle. To identify the physical process contributions to these biases, we analyze the relationships between the convective diurnal cycle and atmosphere state variables relevant to convection in the Amazon using satellite observations and reanalysis data sets for wet and dry seasons between 2002 and 2016 and two Green Ocean Amazon periods. The analysis first stratifies the diurnal cycle into convective and nonconvective days using a daily maximum rain rate threshold of 0.5 mm/hr. Second, the population of days is constrained by requiring reanalysis and observations to agree on the occurrence of convective rain rates, controlling for frequency-dependent biases in convection. The model-generated precipitation phase in Modern-Era Retrospective Analysis for Research and Applications-2 is closer to observations than ERA during 2002–2016, which exhibits a systematic noontime bias and exaggerated diurnal amplitude. Despite the systematic noontime precipitation bias, ERA produces better agreement with Green Ocean Amazon observations due to the frequent midmorning arrival of the coastal front acting to shift the observed diurnal cycle closer to noon. Model disagreement between middle-tropospheric vertical velocity is largest overnight during the dissipation stage of convection, acting to sustain biases through radiative effects. Specifically, the slower dissipation of convection in Modern-Era Retrospective Analysis for Research and Applications-2 acts to reduce morning surface fluxes and increase convective inhibition, whereas enhanced nocturnal midtropospheric subsidence and higher boundary layer humidity in ERA reduce morning convective inhibition leading to an earlier initiation of afternoon deep convection.

Key Points

  • Higher morning convective inhibition shifts the precipitation phase later and increases the amplitude for GOAmazon observations
  • The radiative diurnal cycle is less biased in ERA than MERRA-2 for the GOAmazon domain, but MERRA-2 is less biased over the entire Amazon
  • Differences in nocturnal middle tropospheric vertical velocity in models influence diurnal biases in column humidity and radiative fluxes

1 Introduction

The diurnal cycle is an important climatological mode of variability in the Amazon intertwined with the frequent and intense deep convection in the region. The convective diurnal cycle (CDC) is defined as the response of convection and its related processes to the daily cycle of solar insolation regulating the timing and intensity of clouds and convective rainfall, thus impacting the energy budget and water cycle. Modern convective parameterizations struggle to simulate the complex diurnal and spatial variability of tropical deep convection (Suhas & Zhang, 2014), leading to considerable biases in the simulated climatological water cycle and energy budget in the Amazon (e.g., Itterly & Taylor, 2014). Improving the simulation of the CDC requires an improved understanding of the factors that modulate its behavior.

Three types of convective systems dominate the Amazonian climatology (Greco et al., 1990). Coastally generated convective systems are common during all seasons indicated by westerly propagating moisture flux convergence from the formation of a coastal front fueled by land-ocean temperature contrasts (Adams et al., 2015; Burleyson et al., 2016). Eastward propagating basin-wide systems are common during the wet season resulting from large-scale dynamical forcing originating from tropical waves such as the Madden-Julian Oscillation (Song & Zhang, 2017). Locally generated convective systems are the dominant type of convection in the dry season, arising from solar heating or local features such as orography and river-breeze circulations (Fitzjarrald et al., 2008; Song & Zhang, 2017; Tanaka et al., 2014).

Many local-scale processes (including radiation, surface boundary layer turbulence, cloud dynamics, orographic lift, and microphysics) influence the evolution and diurnal cycle of convection and the large-scale atmospheric state. In models, subgrid-scale processes are parameterized; thus, the behavior of convection is accounted for using a theory of its statistical relationship with the resolved large-scale conditions (Randall et al., 2003). Therefore, to improve the representation of convection in models, we must better understand the statistical relationships between convective processes and the large-scale atmospheric state. While many studies investigate the relationships between convection and its environment (e.g., Adams et al., 2013, 2015; Giangrande et al., 2017; Schiro et al., 2016; Strong et al., 2005; Xie et al., 2014; Zelinka & Hartmann, 2009; Zhang & Klein, 2010), relatively few studies systematically analyze the relationship between the CDC and atmospheric conditions.

The mean state and variability of the CDC is linked to atmospheric state anomalies on various timescales (e.g., Adams et al., 2017; Betts et al., 2015, 2017; Derbyshire et al., 2004; Dodson & Taylor, 2016; Itterly et al., 2016; Lintner et al., 2017; Taylor, 2014a, 2014b; Zelinka & Hartmann, 2009; Zhang & Klein, 2010; Zhao et al., 2017). For example, Zhang and Klein (2010) show robust relationships between preconvective environmental parameters and afternoon deep convection in the U.S. Southern Great Plains indicating that higher convective available potential energy (CAPE) is associated with a later onset time and shorter duration of precipitation, whereas higher humidity above the boundary layer leads to an earlier onset time and longer duration of precipitation due to reduced entrainment of developing cumulus clouds (Derbyshire et al., 2004; Zhang & Klein, 2010). In the Amazon, Song and Zhang (2017) use radar-observed precipitation along with European Centre for Medium-Range Weather Forecasts (ECMWF) model fields constrained by Green Ocean Amazon (GOAmazon) observations to demonstrate the importance of CAPE to the CDC. They find that thresholds of undilute dCAPE (CAPE generation rate from large-scale forcing in the free troposphere) between 60 and 110 J/kg · hr provide the highest skill in the prediction of convective initiation for all seasons due to the inclusion of large-scale forcing terms associated with propagating mesoscale systems (Suhas & Zhang, 2014). Additionally, they find that the inclusion of entrainment dilution on dCAPE (dilute dCAPE threshold exceeding 55 J/kg · hr with an entrainment rate of 2.5 × 10−4 m−1) results in the highest skill for all convective systems (Song & Zhang, 2017).

Our previous work uses observations from satellites, radiosondes, and field campaigns to show sensitivities of CDC statistics to variations in morning atmospheric state; however, the results are spatiotemporally limited due to a lack of reliable in situ data (Itterly et al., 2016). Taylor (2014b) and Dodson and Taylor (2016) show statistically significant sensitivities of the Amazonian CDC to reanalysis atmospheric state conditions on monthly timescales. The use of daily data offers enhanced temporal resolution of convective processes and enables stricter case selection.

The objectives of the present study are to advance our understanding of the Amazonian diurnal cycle and to evaluate the diurnal cycle of top-of-the-atmosphere (TOA) fluxes and precipitation in reanalysis models for a convective (CON) regime and a stable (STA) regime. Moreover, we assess the sensitivity of the Amazonian CDC to atmospheric state conditions using composite analysis and conditional averaging. A description of the observations and models is provided in section 2, a summary of the methodology in section 3, results in section 4, and a discussion and conclusion in section 5.

2 Data

2.1 Observations

2.1.1 CERES

The Clouds and the Earth's Radiant Energy System (CERES) Ed4a SYN1deg data set contains TOA radiative fluxes, surface radiative fluxes, and cloud fraction retrievals extending from July 2002 to present (with a 6-month processing lag) at 1° × 1° spatial resolution and 1-hourly temporal resolution (Doelling et al., 2013, 2016; Loeb et al., 2009; Minnis et al., 2011; Wielicki et al., 1996).

The cloud radiative effect is traditionally defined as the longwave cloud forcing (LWCF) and shortwave cloud forcing (SWCF), where LWCF = OLRCLR − OLR and SWCF = RSWCLR − RSW (e.g., Loeb et al., 2009; Taylor & Loeb, 2013). OLR refers to the outgoing longwave radiation at TOA. RSW refers to the reflected shortwave radiation at TOA, and the subscript CLR refers to the TOA flux in the absence of clouds, called the clear-sky flux. From the shortwave variables, we calculate cloudy-sky albedo (αCLD) by dividing SWCF by incoming solar radiation at TOA. Shortwave retrievals are only included if the solar zenith angle at the middle of time interval for the midpoint of each grid box is less than 82° in order to avoid spuriously large albedos.

Additional CERES products include layered cloud properties, for example, cloud fraction retrieved from the passive sensor cloud retrieval algorithms applied to the Moderate resolution Imaging Spectroradiometer and geostationary (GEO) satellite retrievals (Minnis et al., 2011). Due to the limitations of a passive sensor in a region of frequent multilayer cloud scenes, cloud properties are not shown in the present study.

The CERES SYN1deg product fuses CERES Terra and Aqua Sun-synchronous fluxes with GEO-observed radiances to achieve 1-hourly resolution (Doelling et al., 2016). The merging algorithm includes (1) a calibration of each GEO instrument against Moderate resolution Imaging Spectroradiometer imagers, (2) a conversion from narrowband to broadband radiance, (3) an integration of broadband radiance to flux, and (4) a normalization of GEO-derived flux to the CERES flux. To convert GEO radiances to CERES longwave fluxes, Edition 4 uses a radiance-based algorithm with a regional normalization technique, which significantly reduces errors at higher temporal resolutions and for convective regions relative to the column relative humidity (RH) normalization algorithm used in Edition 3 (Doelling et al., 2016). Root-mean-square errors in CERES SYN1deg Edition 4 1-hourly longwave fluxes are estimated to be 2.8% (compared to 3.6% for the column RH normalization) in January 2010 averaged over the Meteosat domain (60°S–60°N, 60°W–60°E) using the Geostationary Earth Radiation Budget instrument as truth (Doelling et al., 2016). Uncertainties are greater than the domain average for land convective regions; however, convective regions exhibit the largest improvements using the new algorithm. Doelling et al. (2013, 2016) provide a detailed description of the CERES SYN product and algorithms.

2.1.2 GOAmazon Field Campaign Data

Field campaigns provide key insights on the complex spatial variability and physical mechanisms of convection across the Amazon region (Greco et al., 1990; Machado et al., 2002, 2014). Most recently, the Observations and Modeling of the GOAmazon campaign spanned 2014–2015 collecting observations of the energy, water, and chemical cycles from a suite of modern sensors in the central Amazon (Martin et al., 2016).

A network of observational sites was assembled within 100 kilometers (km) of Manaus, Brazil, to examine the background atmospheric state and to investigate the influence of urban pollution (the Manaus plume) on the pristine Amazonian ecosystem (Martin et al., 2016). Two intensive observation periods (IOPs) took place during 2014 in the wet season (IOP1; 1 February to 31 March) and dry season (IOP2; 15 August to 15 October). These data serve as an independent verification of our methodology, which is applied over a larger spatiotemporal scale using meteorological reanalyses.

Radiosondes were launched from the T3 site (−60.60°W, −3.21°S,) located in a pasture 70 km downwind from Manaus (Figure 1). Launches occurred approximately every 6 hr (01:30, 07:30, 13:30, and 19:30 local solar time, Vaisala RS-92 radiosondes; ARM, 1993) providing a high-resolution picture of the local thermodynamic state. High-frequency surface meteorological variables were also collected at T3 including precipitation, wind components, temperature, and atmospheric pressure.

Details are in the caption following the image
The location of the analysis domain (black grid), the GOAmazon analysis domain (red), the coastal domain (blue), and the interior domain (dark green). Elevation contours (meters) are labeled and colored. The lower plot shows a closer look at the GOAmazon domain with major rivers and relevant sites labeled.

2.1.3 TRMM

The Tropical Rainfall Measuring Mission (TRMM)-adjusted merged-infrared 3B42 Version 7 data set (Huffman et al., 2007) provides precipitation rates (Precip) at 0.25° × 0.25° spatial resolution and 3-hourly temporal resolution extending from 50°S to 50°N. The TRMM data are interpolated to 1° × 1° only when matching regimes for consistency with reanalysis Precip; otherwise, the native resolution is used.

The 3B42 precipitation algorithm consists of two steps. First, Visible and Infrared Scanner and TRMM Microwave Imager orbit data (TRMM products 1B01 and 2A12) are used to produce monthly infrared (IR) calibration parameters. Second, the monthly calibration parameters are used to adjust the merged-IR precipitation data retrieved from geostationary satellites (Kummerow et al., 1998).

Oliveira et al. (2016) compare the diurnal cycle of precipitation from several satellite rainfall products to the S-band weather radar located in Manaus, Brazil, during GOAmazon. They find seasonally dependent biases in satellite rainfall estimates, including an overestimation (underestimation) of heavy precipitation events in excess of 10 mm/hr during IOP1 (IOP2). Diurnally, these biases arise due to the overestimation of occurrence frequency and volume of heavy rainfall primarily between 20:00–00:00 and 11:00–14:00 local time (LT) during IOP1 and due to an underestimation of occurrence and frequency of heavy rainfall events between 09:00 and 17:00 LT during IOP2. These diurnal biases are attributed to the spatiotemporal features of precipitation systems, including the inability of satellites to detect local convective cells in drier months and the systematic wet bias in the satellite rainfall retrieval algorithms for larger-scale systems more common in the wet season (Oliveira et al., 2016).

2.2 Reanalysis Models

2.2.1 MERRA-2

The Modern-Era Retrospective Analysis for Research and Applications-2 (MERRA-2) is the most recent version of the reanalysis produced by NASA to replace and extend MERRA, ending February 2016. MERRA-2 uses version 5.12.4 of the Goddard Earth Observing System (GEOS) atmospheric model and data assimilation system with horizontal resolution of 0.5° latitude × 0.625° longitude and vertical resolution of 72 hybrid-eta levels at hourly and 3-hourly temporal resolution for single-level variables and vertical profiles, respectively (Bosilovich et al., 2015). The GEOS system has undergone several important updates since MERRA (GEOS-5.2.0; Rienecker et al., 2011). A full description of the various updates to the GEOS model is available in Bosilovich et al. (2015), Molod et al. (2015), and references therein. McCarty et al. (2015) provide a full list of the new input observations in MERRA-2.

MERRA-2 includes a Tokioka-type trigger on deep convection as part of the Relaxed Arakawa-Schubert (Molod et al., 2015; Moorthi & Suarez, 1992; Rienecker et al., 2011) convective parameterization scheme, which uses a priori power law probability density functions defined for various horizontal resolutions to govern the lower limit on the allowable entrainment plumes (Arakawa & Wu, 2013; Bacmeister & Stephens, 2011; Molod et al., 2015).

The land surface hydrology has been improved in MERRA-2 due to the application of observation-corrected precipitation to the land surface forcing and due to improved assimilation of rain gauges over land (Bosilovich et al., 2015; Molod et al., 2015; Reichle et al., 2017). Globally, the diurnal cycle of the model-generated precipitation shows little improvement from MERRA to MERRA-2 (Bosilovich et al., 2015). In the current study, we use the model-generated precipitation and instantaneous assimilated profiles of temperature, specific humidity, wind vectors, and vertical velocity in pressure coordinates (omega) obtained from the surface to 100 hectopascals (hPa). Profiles are interpolated to 1° × 1° spatial resolution and are available at 3-hourly snapshots beginning at 00:00 Greenwich Mean Time (GMT).

2.2.2 ERA-Interim

ERA-Interim is a reanalysis produced by the ECMWF recently replaced by ERA5. The data assimilation system uses a four-dimensional variational scheme with 12-hourly analysis cycles at 00:00 and 12:00 GMT. A diurnally complete record is obtained by concatenating 6–18-hr forecast steps from both analysis times to allow the forecast model to spin-up (Balsamo et al., 2015). The native spatial grid (0.75° × 0.75°) is interpolated to 1° × 1°. Forecasts are produced by the ECMWF Integrated Forecast System (IFS) release Cy31r2 using a spectral model with T255 (79-km) horizontal resolution with 60 vertical model levels (Dee et al., 2011).

The moist convection scheme is based on the mass flux approach from Tiedtke (1989), which separately simulates deep, shallow, and midlevel convection. For cloud depths less than 200 hPa, the shallow convective scheme is used. For deep convection, the mass flux removes CAPE over a given timescale. For midlevel convection (elevated moist layers), the mass flux is set according to the large-scale vertical velocity. Recent improvements include a modified entrainment formulation leading to an improved representation of tropical convective variability (Bechtold et al., 2008) and a modified CAPE closure leading to improvements in the diurnal cycle of convection (Bechtold et al., 2014).

Estimates of precipitation are produced by the IFS based on temperature and humidity information derived from the assimilated observations (Dee et al., 2011). Approximations used to represent moist processes in the IFS strongly affect the quality of the hydrological cycle, which must be considered when using ERA-Interim precipitation or fluxes (Dee et al., 2011). A detailed documentation of ERA-Interim is provided in Berrisford et al. (2009).

Instantaneous 3-hourly forecast profiles of temperature, specific humidity, u and v wind components, and omega are obtained from the surface to 100 hPa and interpolated to 1° × 1° to match MERRA-2.

3 Methodology

The study goal is to better understand the sensitivity of the CDC process to atmospheric conditions. Our approach employs several steps. First, the CDC manifests under a wide range of atmospheric conditions and involves a number of interacting processes that depend nonlinearly on atmospheric conditions. Therefore, the relationship between the CDC and atmospheric conditions varies with atmospheric conditions (e.g., Dodson & Taylor, 2016; Taylor, 2014a). To investigate this hypothesis, we control for significant differences in the large-scale atmospheric state using a convective precipitation rate threshold of 0.5 mm/hr (Song & Zhang, 2017) during any time step. Lastly, the CDC sensitivity to atmospheric conditions is investigated using composite analyses and conditional averaging.

To better compare reanalysis data and observations, a subset of matched days is used. Matched days are defined as days when both reanalyses and TRMM either produce convective precipitation rates (CON regime) or do not produce convective precipitation rates (STA regime) during any time step of the diurnal cycle process. This approach avoids comparing cases where the reanalysis produces convection that is not observed, and vice versa. Additionally, statistics are shown for all days in the time frame (ALL regime).

Similar to Itterly et al. (2016), each day and grid box is considered a realization of the CDC process, requiring a specific start and end time. The CDC process is defined to begin at sunrise and end at the next sunrise—6:00 to 6:00 local solar time (LST) to center the analysis on the solar insolation forcing. We perform temporal interpolation on all satellite and reanalysis data sets to align diurnal cycles at various longitudes and temporal resolutions in order to ensure robust comparisons. We performed sensitivity tests to compare this interpolation approach with a coarser shifting of the data to the closest hour box approach, revealing similar results and conclusions.

Key CDC metrics, including phase (local time of maximum daily value), and amplitude (half the difference between the daily maximum and the daily minimum) are considered. To refer to a diurnal statistic of a diurnal cycle variable (DCV) in a regime of convective intensity, we add subscripts to the diurnal phase (P) or amplitude (A); for example, PCON,LWCF refers to the LWCF phase in the CON regime.

The study domain covers a large area (70–45°W, 20°S–5°N) to compare the wide range of atmospheric state and convective forcing mechanisms in the Amazon region (Figure 1, black box). In order to investigate specific climatological features and convective processes, we choose three subregions (colored boxes, Figure 1). The green box is located in the western Amazon (70–68°W; 4–2°S), the red box is located in the central Amazon around Manaus (61–59°W; 4–2°S), and the blue box is located near the coast (50–48°W; 4–2°S). As shown later, each subregion falls into a different bin of climatological precipitation phase; thus, our motivation is to extrapolate physical insights learned from each subregion to the entire domain.

We calculate composite diurnal cycle variable anomalies (DCV'DC) for the entire region and each subregion stratified by convective intensity regimes. DCV'DC are computed at each day in each grid box then totaled to form composites by region and regime. For shortwave variables, we omit values when solar zenith angle exceeds 82° when computing diurnal anomalies.

To investigate the sensitivity of the CDC to atmospheric conditions, we analyze vertical diurnal composites of RH, winds, and omega over the subregions for each regime. These atmospheric state composites are then related back to the corresponding diurnal composites for each region to qualitatively investigate the relationships between atmospheric state and DCV'DC.

Additionally, we conditionally average PrecipDC by bins of atmospheric state variables (ASVs) in each subregion to further assess the model's internal sensitivity to atmospheric state. We repeat this analysis with GOAmazon IOP1 observations using the radiosonde for ASVs and the collocated Precip gauge for the DCV.

For CAPE and convective inhibition (CIN) calculations, we use the average of the lowest 100 hPa as the parcel. For MERRA-2 and ERA, we use the 2-m temperature, pressure, and humidity as the surface value and then use the next lower pressure level as the second level. For the radiosonde, we interpolate the profiles to 10-hPa bins from the surface to 100 hPa. No vertical interpolation is performed when calculating ASVs for the models; however, for the vertical contour plots, we interpolate the radiosonde and MERRA-2 to match ERA pressure levels. TDEF (Tawfik & Dirmeyer, 2014) is the temperature increment at the surface needed for convective initiation; convection is expected to initiate when TDEF equals zero. RH is computed with respect to liquid from specific humidity and temperature profiles.

4 Results

4.1 Variations of Diurnal Statistics With Convective Intensity

Figure 2 shows the percentage of convective days for each grid box during December-January-February (DJF; the wet season) and June-July-August (JJA; the dry season) from 2002 to 2016 and IOP1 and IOP2 for GOAmazon in 2014 using TRMM, MERRA-2, ERA-Interim, and only days when all agree on each regime (matched days). In the wet season, TRMM identifies 60%–80% of days as convective in the central Amazon with the lowest frequencies (0%–40%) occurring north of the equator and in mountainous terrain (Figure 2a). MERRA-2 produces fewer convective days overall except for isolated areas near the northeast coast and in the mountains of Bolivia (Figure 2b). ERA produces far too many convective days across the entire region (Figure 2c). The results indicate a spatial dependence of the matched CON days with 30%–50% frequency near the coast and in the central Amazon and between 0%–20% north of the equator and in the mountainous terrain (Figure 2d). Matched STA days rarely (0%–10%) occur in the interior Amazon (Figure 2e).

Details are in the caption following the image
Frequency of occurrence of convective days during (a–e) the wet season, (f–j) the dry season, (k–o) IOP1, and (p–t) IOP2. Regimes are shown separately for (a, f, k, and p) TRMM, (b, g, l, and q) MERRA-2, (c, h, m, and r) ERA-Interim, (d, i, n, and s) matched CON days, and (e, j, o, and t) matched STA days.

In Figure 2c, ERA overestimates convective days in JJA for nearly all grid boxes (Figure 2h) and MERRA-2 more closely matches observations except for an underestimation of convective days in the western Amazon south of the equator between 0 and 10°S (Figure 2g). Matched STA days occur between 50% and 100% of the time in the southern half of the domain in the dry season (Figure 2j).

TRMM data indicates widespread, above-average convective activity during IOP1 (Figure 2k) compared to DJF (Figure 2a), most notably along the northeast coast of Brazil owing to a more frequent influence of the coastal front during the GOAmazon study period relative to the DJF climatology. Both MERRA-2 and ERA capture the increased frequency of convection along the coast (Figures 2l and 2m). During IOP2, TRMM identifies less frequent convection in northeastern Brazil (Figure 2p) than average JJA conditions (Figure 2f) and significantly more frequent convection south of 5°S owing to the inclusion of transition months in the IOP2 dates. Matched CON days in IOP2 are generally restricted to northwestern Brazil (Figure 2s), whereas matched STA days make up the majority of matches elsewhere, especially over the ocean and mountainous terrain (Figure 2t).

To investigate atmospheric state, we plot averages of MERRA-2 and ERA Precipitable Water (PW) and CIN by regime in DJF (Figure 3) and JJA (Figure 4). On CON days in DJF, PW values range from 50 to 60 mm for all low-elevation points (pink colors, Figures 3a and 3g). CIN is lowest (dark blue, 0–10 J/kg) near the coast for MERRA-2 (Figure 3d) and ERA (Figure 3j). ERA CIN is lower than MERRA-2 for the entire interior Amazon, perhaps contributing to the higher frequency of occurrence of convection in ERA (Figure 2c).

Details are in the caption following the image
Wet season averages by regime of (a–c) MERRA-2 precipitable water, (d–f) MERRA-2 convective inhibition, (g–i) ERA precipitable water, and (j–l) ERA convective inhibition convective inhibition in for (a, d, g, and j) convective days, (b, e, h, and k) stable days, and (c, f, i, and l) all days. Black colors are greater than the largest value, white colors are less than the smallest value, and gray colors are missing data.
Details are in the caption following the image
Same as Figure 3, but for the dry season.

During STA days, PW is 5–10 mm lower, whereas CIN is significantly higher than CON days in both MERRA-2 (Figures 3b and 3e) and ERA (Figures 3h and 3k) for nearly all points. The largest CIN values occur at elevation in southeastern Brazil, exceeding 50 J/kg in both. During all DJF days, the largest differences occur near the coast and in the central Amazon, where MERRA-2 CIN and PW are higher than ERA (Figures 3c, 3f, 3i, and 3j). Similar averages of CAPE (not shown) reveal larger CAPE in ERA than MERRA-2 overall except in northwestern Brazil. Differences in CAPE between the regimes are negligible (or CAPE is slightly larger on STA days) in the interior Amazon for both models; however, in areas of less frequent convection (e.g., north of the equator, along the coast, or southern Brazil), CAPE is significantly larger on CON days than STA days (not shown).

On JJA CON days, average CIN north of the equator is lower than DJF for both MERRA-2 (Figure 4d) and ERA (Figure 4j). ERA CIN and PW (Figures 4g and 4j) remain slightly lower than MERRA-2 (Figures 4a and 4d) inland from the coast and south of the equator, similar to Figure 3. On STA days, MERRA-2 CIN is significantly lower than ERA north of the equator (Figures 4e and 4k). In this area, CIN exhibits little sensitivity to regime in MERRA-2 (Figures 4d and 4e), whereas ERA CIN is larger on STA days compared to CON days (Figures 4j and 4k). Near 5°S on STA days, CIN exceeds 30 J/kg in both models, creating a swath stretching southwestward from the subequatorial coastline to the Bolivian highlands (Figures 4e and 4k). Looking back at Figures 2f–2j, this region aligns with the area where convection is essentially nonexistent in the dry season; additionally, this is where PW values drop off considerably (Figures 4b and 4h). For all JJA days, spatial averages of CIN north of this swath are almost identical between models (Figures 4c and 4f). South of this swath, MERRA-2 CIN is higher than ERA, while PW is nearly identical between models, consistent with more convective days occurring here in ERA (Figures 2g and 2h).

Wet season domain-averaged DCV'DC stratified by convective intensity regime are shown for αCLD (Figures 5a–5c), LWCF (Figures 5d–5f), and Precip (Figures 5g–5i). The largest model biases are evident in the Precip'DC for all regimes and both reanalyses. The observed αCLD'CON, DC decreases between 7:30 and 11:30 LST followed by a symmetric afternoon increase (Figure 5a, black line). ERA (red line) and MERRA-2 (blue line) αCLD'CON, DC begin to increase 3 hr earlier than observations leading to positive afternoon αCLD'CON, DC biases (Figure 5a). The observed αCLDSTA,DC decreases between 7:30 and 10:30 LST then increases at an accelerating rate through 16:30 LST—likely influenced by a daytime increase in shallow cumuli (Figure 5b). MERRA-2 αCLDSTA,DC captures the morning decrease then increases more rapidly in the afternoon than for observations, whereas ERA αCLDSTA, fails to decrease in the morning and increases earlier than observed (Figure 5b).

Details are in the caption following the image
Composite diurnal cycle anomalies of (a–c) cloudy-sky albedo, (d–f) LWCF, and (g–i) precipitation on (a, d, and g) convective days, (b, e, and h) stable days, and (c, f, and i) all DJF days over the entire Amazon region using observations (black), MERRA-2 (blue), and ERA-Interim (red). The standard error of the composites is shaded in corresponding color.

The observed LWCF'CON, DC decreases between 7:30 and 10:30 LST followed by a 30 W/m2 afternoon increase through 18:30 LST and a steady nocturnal decrease as convection dissipates (Figure 5d). LWCF'STA, DC decreases through 13:30 LST indicating clearing from the previous day then increases rapidly between 12:00 and 17:00 LST and again near midnight (Figure 5e). The later increase is possibly due to nonlocal or propagating convective cloudiness. MERRA-2 more closely reproduces the observed shape of LWCF'DC for all regimes (Figures 5d–5f, blue line). ERA LWCF'CON, DC begins to increase 6 hr earlier than observations and decreases too rapidly after 16:30 LST in all regimes (Figures 5d–5f, red line).

Large differences are evident between Precip'DC in observations and models (Figures 5g–5i). Models fail to represent the phase and amplitude of the Precip'CON, DC maximum between 16:30 and 19:30 LST (Figure 5g). MERRA-2 Precip'CON, DC exhibits more variability (illustrated by the wider blue shading in Figure 5g) and smaller diurnal amplitude than observed. ERA Precip'CON, DC maximizes 3 to 6 hr too early at 13:30 LST with larger diurnal amplitude than observed (Figure 5g, red line). The observation-corrected Precip'DC in MERRA-2 is similar to ERA with a systematic near-noontime bias (not shown).

For the dry season, the diurnal composites within the CON and STA regimes are similar to Figure 5 (not shown). The largest DCV'DC difference between the seasons occurs in the ALL regime due to the higher occurrence frequency of the STA regime.

Domain-averaged composite diurnal cycles are useful tools to visualize model biases; however, they do not capture regional CDC variability or inform if models simulate certain regions better than others. To investigate the spatial dependence of model diurnal cycle biases, we plot PLWCF (Figure 6) and PPrecip (Figure 7) spatially by regime in the wet season.

Details are in the caption following the image
The composite LWCF phase during the wet season on (a, d, and g) convective days, (b, e, and h) stable days, and (c, f, and i) all days using (a–c) CERES, (d–f) MERRA-2, and (g–i) ERA-Interim.
Details are in the caption following the image
The composite precipitation phase during the wet season on (a, d, and g) convective days, (b, e, and h) stable days, and (c, f, and i) all days using (a–c) TRMM, (d–f) MERRA-2, and (g–i) ERA-Interim.

CERES PCON,LWCF generally occurs in the late afternoon or overnight (Figure 6a) with nocturnal maxima occurring in western Brazil, the mountains of Bolivia, and inland from the coastal front (purple and blue shading) due to the longer lifetime of propagating large-scale squall lines and convective systems. CERES PCON,LWCF occurs shortly after local noon (yellow shading) over the ocean and near Manaus due to the variety of convective systems encountered in these regions (Figure 6a). CERES PSTA, LWCF generally occurs between 4:00 and 10:00 LST (blue shading, Figure 6c) consistent with the clearing of convective cloudiness from the previous day.

MERRA-2 captures the nocturnal PCON,LWCF in southwestern Brazil and shows limited evidence of the coastal front propagating inland overnight (Figure 6d). Elsewhere, MERRA-2 PCON,LWCF occurs several hours later than CERES. ERA PCON,LWCF captures general features of the observed CDC including late afternoon values in eastern Brazil and late evening values in western Brazil but fails to capture propagating mesoscale features such as the coastal front (Figure 6g). Both models struggle to simulate the early morning PSTA, LWCF although MERRA-2 does better overall (Figures 6e and 6h).

Similarly, Figure 7 shows spatial plots of PPrecip by regime. TRMM PCON,Precip resembles Figure 6a, highlighting the inland nocturnal propagation of the coastal squall line (pink and purple shading), early morning maxima in the high terrain of Bolivia (blue and green shading), and elsewhere occurring between 16:00 and 20:00 LST (orange and red shading) consistent with afternoon deep convection (Figure 7a). Spatially, PSTA,Precip lacks propagating features and is noisy across the central Amazon, alternating between midafternoon and early morning maxima (Figure 7b). MERRA-2 captures the nocturnal propagation of the coastal squall line and regime dependence of PPrecip well (Figures 7d–7f). In central and southeastern Brazil, MERRA-2 precipitation maximizes too late in the day similar to Figure 6d. ERA PPrecip occurs between 10:00 and 14:00 LST (light green to dark green shading) for all low-elevation land points indicating a systematic bias (Figures 7g–7i). For the ocean and high elevations, ERA PPrecip occurs in the early morning (blue and purple shading) across all regimes (Figures 7g–7i).

For the remaining plots, we focus our analysis on the coastal subregion during the wet season (Figure 8) and the GOAmazon subregion during IOP1 (Figure 9) to investigate model CDC biases in two areas of large model disagreement as seen in Figures 6 and 7. Similar to Figure 5, Figure 8 shows the DCV'DC for the coastal subregion (Figure 1, blue box). We choose this subregion to compare the model performance in an area affected by the inland propagation of the coastal front overnight. The coastal front propagation impacts the observed climatological TOA flux diurnal phase and amplitude over a thousand kilometers inland (Itterly & Taylor, 2014). Although reanalysis models generally do not attempt to simulate the propagation of synoptic/mesoscale convective features (Lee et al., 2007), it is crucial to analyze the biases to understand the downwind implications of this deficiency.

Details are in the caption following the image
Composite diurnal cycle anomalies of (a–c) cloudy-sky albedo, (d–f) LWCF, and (g–i) precipitation on (a, d, and g) convective days, (b, e, and h) stable days, and (c, f, and i) all DJF days over the coastal subregion using observations (black), MERRA-2 (blue), and ERA-Interim (red). The standard error of the composites is shaded in corresponding color.
Details are in the caption following the image
Same as Figure 5, but during IOP1 for the central Amazon subregion.

On convective days, the observed DCV'DC for all three variables decreases between 06:00 and 13:30 LST before rapidly increasing through the afternoon and evening, maximizing near 22:30 LST for Precip'CON, DC (Figure 8g) and 23:30 LST for LWCF'CON, DC (Figure 8d) with large diurnal amplitudes of 15 mm/day and 25 W/m2, respectively. ERA αCLDCON, DC and LWCF'CON, DC begin to increase too early in the day. Both models lack diurnal amplitude although MERRA-2 more closely reproduces the observed CDC shape. Interestingly, MERRA-2 Precip'CON, DC shows a 12 mm/day decrease between 06:00 and 13:00 LST followed by a late afternoon increase maximizing near 02:00 LST (Figure 8g, blue line). Although still slightly out of phase with the observed Precip'CON, DC, MERRA-2 clearly captures the influence of the coastal front. These improvements, however, are uncoupled to the representation of cloud radiative characteristics (Figure 8d). Overall, MERRA-2 matches the observed diurnal cycle better than ERA for the coastal and western Amazon subregions (not shown) in the wet season.

In Figure 9, composite DCV'DC are shown during IOP1 for the central Amazon subregion (Figure 1, red box) to evaluate the model performance during the GOAmazon study period. On convective days, the observed LWCF (black line) maximizes around 16:00 LST, 2 hr earlier than Figure 3 and with slightly decreased amplitude (Figure 9d). ERA (red line) captures the PCON,LWCF better than MERRA-2 (blue line), which maximizes around 21:00 LST. TRMM Precip'CON, DC maximizes near 3:30 LST (Figure 9g) indicating frequent morning rainfall due to the influence of larger-scale systems (Tang et al., 2016). MERRA-2 misses the positive morning Precip'CON, DC, while ERA maximizes near noon with exaggerated amplitude. Additionally, MERRA-2 simulates the observed morning decrease in αCLD'STA,DC, while ERA shows an increase through midafternoon (Figure 9b). For all IOP1 days, ERA simulates the LWCF'ALL,DC and Precip'ALL,DC better than MERRA-2 (Figures 9f and 9i). Interestingly, TRMM PALL,Precip shifts to a midmorning maximum when including all IOP1 days (Figure 9i).

Extending the temporal range of Figure 9 to all DJF days between 2002 and 2016, the DCV'DC hardly changes for either model (not shown). For observations, LWCF'CON, DC increases in amplitude by 5 W/m2 and TRMM Precip'CON, DC shifts to negative anomalies in the morning followed by a larger afternoon amplitude maximizing ~16:30 LST (not shown). Overall, the persistent early convective initiation bias in ERA leads to better agreement with the observed CDC in the central Amazon subregion compared to MERRA-2.

4.2 Vertical Diurnal Composites of Atmospheric State by Regime

In the following section, we analyze the sensitivity of DCV'DC to atmospheric state using MERRA-2 and ERA-Interim vertical profiles composited over the subregions in Figure 1 to investigate relationships between atmospheric conditions and the DCV'DC composites discussed in section 4.1. Results for the dry season are similar and are not shown.

Figure 10 shows vertical diurnal composites of RH (filled contours), omega (black contour lines), and winds (barbs) in MERRA-2 and ERA computed over the central Amazon subregion (Figure 1, red box) stratified by convective regime. Additionally, we compute the corresponding composite diurnal cycles of TDEF, CAPE, CIN, latent heat flux (LHF), sensible heat flux (SHF), 2-m temperature (T2m) and humidity (RH2m), omega at 300 hPa (w300), OLR, and Precip stratified by regime in Figure 11.

Details are in the caption following the image
Vertical diurnal cycle composites of relative humidity with respect to liquid (colored), wind vectors (shown as barbs), and omega (contour lines) over central Brazil in the wet season. Composites are computed on (a, d) convective days, (b, e) stable days, and (c, f) the full wet season climatology using (a–c) MERRA-2 and (d–f) ERA-Interim. The number of valid days contained in each composite is shown in red below the x axis. Units are in % for RH, kPa/d for omega, and knots for wind speed. Dashed lines denote negative omega, dotted lines denote positive omega, and solid lines denote zero omega.
Details are in the caption following the image
Wet season diurnal cycle composites over the central Amazon subregion of (a, d) CIN, CAPE, and Precip, (b, e) OLR, T2 m, and LHF, and (c, f) RH2m, SHF, w300, and TDEF using (a–c) MERRA-2 and (d–f) ERA-Interim. Solid lines are used for convective days, and dashed lines are used for stable days.

Above 700 hPa, MERRA-2 upper tropospheric humidity (UTH; above 400 hPa) and middle tropospheric humidity (MTH; between 700 and 400 hPa) is 10%–20% higher compared to ERA before 18:00 LST (Figures 10a and 10d). ERA lower tropospheric humidity (LTH; between 1,000 and 700 hPa) is higher than MERRA-2, especially in the morning. This top-heavy structure in the MERRA-2 humidity profile acts to raise the lifted condensation level (LCL), increasing CIN and TDEF in the early morning (Figure 11a), possibly due to a slower dissipation of midlevel cloudiness from previous convection. The bottom-heavy structure in the ERA humidity profile leads to lower morning LCLs, reducing morning CIN and TDEF (Figures 11d and 11f) contributing to earlier convective initiation (Figure 9).

On STA days in the wet season, descending motion is present in both models in the morning between 400 and 800 hPa (albeit twice as strong in ERA) followed by weak ascending motion in the lower levels as the land surface is heated (Figures 10b and 10e). Both models show large morning CIN and diurnal minimum CIN occurring around 12:00 LST as turbulent mixing moistens the lower troposphere (Figures 11a and 11d). However, MERRA-2 CIN (CAPE) is lower (higher) than ERA by late afternoon (Figures 11a and 11d), resulting in a more moistened middle and upper troposphere (Figures 10b and 10e).

Using the full wet season climatology, ERA produces slightly stronger rising motion than MERRA-2 during midday in the middle and upper troposphere (Figures 10c and 10f) likely due to the increased frequency of convective days (Figures 2b and 2c). However, ERA also produces stronger descending motion overnight between 400 and 700 hPa, acting to reduce the nocturnal MTH (Figure 10f), and dissipate midlevel clouds, allowing for solar radiation to penetrate deeper into the column (increasing surface fluxes) and also enhancing the entrainment of dry air during the following daytime.

To further investigate the sensitivity of the model CDC to atmospheric state, Figure 12 shows composites of PrecipDC conditionally averaged by bins of 09:00 LST TDEF using MERRA-2 and ERA. Daytime PrecipDC shows a strong sensitivity to 09:00 LST TDEF in MERRA-2 (Figure 12a). The lowest TDEF bins (red and orange lines) lead to larger morning precipitation rates and earlier PPrecip, CON (Figure 12a). A similar relationship (higher TDEF bins shift PPrecip, CON later) is seen in ERA (orange, green, and blue lines) albeit obscured by the less variable PrecipDC (Figure 12d). TDEF in ERA is biased toward lower values due to the more humid lower troposphere (Figure 11f). During ALL days, the mean state and amplitude of ERA Precip, DC is more affected by 09:00 LST TDEF than the phase (Figure 12f). For example, the lowest TDEF bin (red line) shows approximately 4 times higher Precip rates than the highest TDEF bin (blue line) around noon (Figure 12f). For MERRA-2, lower values of 09:00 LST TDEF shift the phase to the early morning, increasing both the diurnal amplitude and the mean state of PrecipDC (Figure 12c).

Details are in the caption following the image
Precipitation diurnal cycle conditionally averaged by bins of 09:00 LST TDEF over the central Amazon subregion using (a–c) MERRA-2 and (d–f) ERA-Interim on (a, d) convective days, (b, e) stable days, and (c, f) all days.

Vertical diurnal composites of atmospheric state over the coastal subregion (Figure 1, blue box) are shown in Figure 13. Due to the drastically different composite DCV'DC characteristics in this subregion (Figure 8) compared to the central Amazon (Figure 9), we expect to see a corresponding shift in the atmospheric state, especially for MERRA-2, which captures the observed nighttime PCON,Precip, whereas ERA does not (Figures 7 and 8).

Details are in the caption following the image
Same as Figure 10, but for the coastal subregion.

Similar to Figure 10, MERRA-2 MTH is 10%–20% higher compared to ERA on CON days (Figures 13a and 13d). Consistent with Precip'CON, DC from Figure 8g, the upward vertical velocity is strongest overnight in MERRA-2 (Figure 13a) and weaker overall in ERA, maximizing between 12:00 and 18:00 LST (Figure 13d). Additionally, MERRA-2 CIN is twice as large as ERA in the early morning, allowing the atmosphere to generate higher values of CAPE and to produce stronger updrafts (not shown).

On STA days, RH is 20%–30% lower than CON days in the middle and upper troposphere for both models (Figures 13b and 13e), revealing a larger difference between regimes than the central Amazon (Figure 10). Weak descending motion is present above 800 hPa for all times of day in ERA (Figure 13e). Below 800 hPa, upward vertical velocity maximizes near 15:00 LST at 900 hPa as buoyant thermals produce shallow cumuli and moisten the layer above the boundary layer (Figure 13e). Descending motion is present in the morning during STA days in MERRA-2; however, intense ascending motion appears near midnight—likely due to the coastal front—acting to moisten the middle troposphere (Figure 13b).

4.3 GOAmazon Diurnal Cycle

Composite diurnal profiles of RH and wind barbs at the T3 site are stratified by convective regime during IOP1 and IOP2 in Figures 14 and 15, respectively. We use the 6-hourly GOAmazon radiosonde data set and the nearest profiles from MERRA-2 and ERA to compare the diurnal cycle of tropospheric RH and winds to elucidate the model DCV'DC biases (Figure 9).

Details are in the caption following the image
Vertical diurnal cycle composites of relative humidity (shaded) and wind components (shown as barbs) at Manaus, Brazil, during IOP1 on (a–c) convective days, (d–f) stable days, and (g–i) all days using (a, d, and g) radiosonde observations, (b, e, and h) ERA-Interim, and (c, f, and i) MERRA-2. The number of valid days contained in each composite is shown in red below the x axis.
Details are in the caption following the image
Same as Figure 14, but for IOP2.

During IOP1, 25 matched CON days occurred and only five matched STA days occurred out of the 58 possible days. The radiosonde reveals primarily northeasterly winds from the surface to 500 hPa and primarily southerly/southeasterly winds above 400 hPa (Figure 14a). The models agree on the general behavior of the winds (Figures 14b and 14c). Much larger disagreements are seen in the diurnal cycle of MTH and UTH; both are fundamentally tied to the behavior of convective parameterizations and therefore to the CDC. Consistent with CERES LWCF'CON, DC in Figure 9, the radiosonde UTH maximizes near 18:00 LST (Figure 14a). ERA MTH is 10%–20% too dry in the afternoon, and ERA UTH shows an enhanced diurnal amplitude peaking too late in the evening (Figure 5b). Consistent with Figure 9, the diurnal cycle of MERRA-2 UTH exhibits a moist bias overnight and the diurnal maximum occurs later in the evening than observed (Figure 14c). MERRA-2 LTH is lower than both ERA and the radiosonde. The radiosonde RH diurnal cycle (Figure 14a) shows an upper tropospheric maximum above 200 hPa directly following peak of convection possibly indicating that convection is responsible for injecting this moisture, a feature missing from both reanalysis data sets. Alternatively, this feature could be an artifact of radiosonde-measured humidity in a near-saturated column. Others have found evening values of PW from radiosondes near Manaus to be biased high relative to Global Position System radio occultation measurements (David Adams, personal communication).

On IOP1 STA days, the radiosonde RH is 5%–20% lower than CON days for much of the lower and upper troposphere (Figure 14d). The UTH DC reaches diurnal minimum at 18:00 LST and builds to its diurnal maximum near midnight. ERA captures this diurnal behavior and mean state (Figure 14e). MERRA-2 UTH is biased moist, and LTH is biased dry (Figure 14f). Combining all available IOP1 days highlights similar biases (Figures 14g–14i). Both observations and models show a westerly wind component between 400 and 500 hPa, perhaps revealing the source of drier air acting to stifle convection during STA days (Figures 14d–14f).

During IOP2, only five matched CON days and 23 STA days occurred out of 61 possible days (Figure 15). On convective days, wind directions below 500 hPa are primarily easterly and wind speeds are 5–10 knots higher than during IOP1 (Figures 15a–15c). The radiosonde MTH is 20%–30% lower overnight and in the early morning than IOP1 CON days followed by an intense increase in the afternoon as deep convection initiates and rapidly dissipates in early evening (Figure 15a). This finding is qualitatively consistent with other GOAmazon results (e.g., Tang et al., 2016), which show that dry season convection is more intense and less driven by large-scale moisture flux convergence in the middle troposphere compared to the wet season. For ERA, LTH (MTH and UTH) is 5%–15% too moist (dry) and the UTH maximum occurs 3 to 6 hr later than observed (Figure 15b). For MERRA-2, both MTH and UTH are 10%–20% higher than the radiosonde overnight and in the early morning, maximizing around 18:00 LST then remaining 10%–20% too moist overnight instead of the observed rapid drying (Figure 15c).

On IOP2 STA days, the radiosonde MTH and UTH are below 50% for all times of day, while LTH maximizes in the afternoon near 850 hPa as turbulent mixing forms shallow cumuli (Figure 15d). Except for slight moist biases in the upper troposphere, both models capture this behavior (Figures 15e and 15f). During all days in IOP2, MERRA-2 shows better agreement with the radiosonde in the lower troposphere and ERA shows better agreement in the upper troposphere (Figures 15g–15i).

Using GOAmazon observations during IOP1, Figure 16 shows the composite diurnal cycles of CAPE, CIN, TDEF, and PW stratified by regime. For the remaining plots, regimes are defined by applying the 0.5 mm/hr threshold to 3-hourly averaged precipitation rates from a collocated rain gauge at the T3 site (Figure 1). The most significant differences between the regimes occur in the morning hours for all ASVs. At 1:30 LST, the average CAPE is 1,100 J/kg on CON days compared to 600 J/kg on STA days. Percentage differences between regimes are even more significant for morning CIN (Figure 16b) and TDEF (Figure 16c), while PW is only a couple of millimeters higher during CON days compared to STA days (Figure 16d). Statistically significant differences in the ASV DCs persist through the afternoon. For example, CAPE decreases (increases) from 1:30 to 13:30 LST on CON (STA) days (Figure 16a). This opposite diurnal behavior between the two regimes reveals the importance of large-scale versus local-scale forcing mechanisms during CON and STA days, respectively.

Details are in the caption following the image
Composite radiosonde-measured diurnal cycles of (a) CAPE, (b) CIN, (c) TDEF, and (d) PW during IOP1 on convective days (blue), stable days (red), and all days (black). The standard error is shown as error bars.

Figure 17 shows conditional averages of the observed PrecipDC from the T3 rain gauge to the collocated 7:30 LST radiosonde measurements of CIN (Figures 17a–17c) and TDEF (Figures 17d–17f). Convective mornings with large CIN reveal a later PCON,Precip and a larger ACON, Precip because the atmosphere has more time to generate CAPE (green and blue lines, Figure 17a). Before noon, most of the Precip occurs when CIN is close to zero (red line, Figure 17a). On stable days, precipitation is generally suppressed on days with large CIN (blue line, Figure 17b).

Details are in the caption following the image
GOAmazon rain gauge diurnal cycle conditionally averaged by bins of (a–c) 07:30 LST CIN and (d–f) 7:30 LST TDEF measured by the collocated radiosonde during IOP1 on (a, d) convective days, (b, e) stable days, and (c, f) all days.

Using identical TDEF bins as Figure 12, the observed PrecipDC exhibits a qualitatively similar response to bins of morning TDEF compared to MERRA-2 from Figure 12 (Figures 17d–17f). Specifically, lower morning TDEF bins (red and orange lines) shift the PCON,Precip earlier in the day and increase ACON,Precip, while the higher TDEF bins (blue and green lines) lead to lower than average Precip during the morning, acting to shift PCON, Precip to the afternoon or evening (Figure 17d). The more significant PrecipDC differences conditionally averaged by bins of CIN compared to bins of TDEF potentially reveal that morning CIN is the more robust predictor of the CDC during IOP1.

5 Discussion and Conclusion

Previous investigators have evaluated the CDC within reanalysis and large-scale atmospheric models. These evaluations consistently revealed several systematic biases in the representations of the CDC, namely, the precipitation diurnal cycle peaks too early in the day, the CDC amplitude is weaker than observations, propagating features of the CDC are missing, and the shallow cumulus diurnal cycle is not properly represented (e.g., Betts & Jakob, 2002; Derbyshire et al., 2004; Itterly & Taylor, 2014; Lee et al., 2007; Taylor, 2012; Yang & Slingo, 2001). In these studies, seasonal or annual mean diurnal cycle composites served as the primary analysis technique. The present work differs by (1) using daily data, (2) constraining days used in the comparison, and (3) stratifying by convective intensity regimes. While the results reveal new insights on the complex relationships between convective intensity and CDC behavior in the Amazon, they also reaffirm previously reported systematic biases.

The weak dependence of the precipitation CDC phase to convective regime in ERA suggests that the root cause of this bias is likely in the parameterized physics and not due to misrepresentations of atmospheric conditions. The in-phase relationship between maximum solar insolation and precipitation in models suggests that they are too sensitive to boundary layer forcing (i.e., CAPE) and large-scale controls such as moisture flux convergence are too weak (Lee et al., 2007). MERRA-2 largely reproduces the domain-averaged and spatial variations in the precipitation CDC phase across the Amazon representing a marked improvement over ERA. This improvement in the model-generated precipitation in MERRA-2 is likely a result of improved data assimilation. Despite the improvements, the propagating CDC features, most prevalent in the CON regime, remain largely absent in the radiative flux composites. Further, all reanalysis products exhibit systematically higher daily minimum OLR values than CERES on convective days (not shown). Moreover, the results indicate that reanalyses struggle to capture the diurnal evolution of radiative fluxes in the stable regime and the convective regime.

This study builds upon a framework to quantify the Amazonian CDC sensitivity to atmospheric conditions. Moreover, given the expectation that under anthropogenic climate change, the intensity and frequency of tropical convective precipitation is likely to increase (Chou et al., 2012), it will be even more important to analyze short timescale feedbacks between convection and atmospheric state in order to constrain the global energy balance and water cycle.

Comparing the sensitivity of diurnal cycle variables to atmospheric state uncovers process contributions to the model biases in the CDC. Our work reaffirms the importance of the vertical distribution of humidity to the evolution of convection; however, we additionally show that the relationships between convection and humidity are represented very differently across reanalysis data products. Although largely driven by the convective parameterizations, model CDC biases are sustained and amplified by biases in atmospheric state from the previous day through cloud and radiative feedbacks.

The largest model differences in atmospheric state occur overnight in the midtroposphere during the dissipating stage of convection. For example, the early convective initiation bias in ERA leads to enhanced nocturnal clearing relative to MERRA-2 as remnant convective cloudiness has more time to be heated, leading to dissipation via entrainment and subsidence. Ultimately, this leads to a drier upper troposphere and more humid lower troposphere in the ERA model climate relative to MERRA-2, thus lowering the lifted condensation level and reducing (increasing) morning CIN (CAPE). MERRA-2 shows enhanced upward motion overnight, enhancing cloudiness and reducing morning surface fluxes, thus delaying the onset of convection.

Wet season climatological model differences in column vertical velocity and humidity are larger for the coastal region than the central Amazon region. Compared to CERES, the model LWCF'DC is also more biased in the coastal region due to complex interactions with the coastal front. The precipitation CDC phase is shown to be more sensitive to local atmospheric state variations in MERRA-2 than ERA. The observed sensitivity of GOAmazon gauge precipitation to radiosonde measured parameters reveals qualitatively similar relationships as the models; however, MERRA-2 precipitation behaves more realistically when conditionally averaged by identical TDEF bins.

The convective intensity regime-based evaluation illustrates significant changes in the CDC shape by regime and is shown to be useful for large-scale model evaluation. Figure 4 represents a straightforward analysis stratifying the CDC by convective intensity that can be directly compared to models for evaluation. In addition, Figures 10-17 represent a more detailed model evaluation technique with the ability to identify biases in the model-simulated variability of CDC variables with respect to atmospheric conditions and to aid in model selection for process studies.

In closing, we identify diurnal cycle biases in the Amazon after stratifying by convective intensity. Moreover, we quantify the Amazonian CDC sensitivity to atmospheric conditions, revealing information on the behavior of model parameterizations compared to GOAmazon and satellite observations. Treating the CDC as a holistic process has potential to serve as a fruitful approach for evaluating and constraining large-scale atmospheric model simulations. The diurnal cycle of convective clouds and precipitation has implications on the energy budget and water cycle and therefore on our understanding of Earth's climate. Understanding the sensitivity of convection and its diurnal cycle to atmospheric state and evaluating its representation in models is critical for understanding the role of convection in the climate system.

Acknowledgments

We thank David Adams and two anonymous reviewers for their helpful comments, which have significantly improved the manuscript. This work was supported by the NASA Energy and Water Cycle Studies program through grant NNH10ZDA001N. CERES data used in this study are stored at the Atmospheric Science Data Center (ASDC) at NASA Langley Research Center. TRMM 3B42 precipitation data and MERRA-2 data were downloaded online (http://mirador.gsfc.nasa.gov). ERA-Interim data were downloaded online (http://apps.ecmwf.int/datasets). Field campaign data were obtained online from the Atmospheric Radiation Measurement (ARM) Program sponsored by the U.S. Department of Energy, Office of Science, Office of Biological and Environmental Research, Climate and Environmental Sciences Division (https://www.arm.gov/research/campaigns/amf2014goamazon).