Combining CMIP data with a regional convection-permitting model and observations to project extreme rainfall under climate change

Following the methodology in Klein et al (2021), and drawing on previous West African studies (Arnaud et al 1992, Laing et al 1999, Mathon et al 2002), we use thermal-infrared imagery from the Meteosat series, which we combine with microwave rainfall estimates, and ERA5 reanalysis data over the West African monsoon season May–October 2004–2018 to relate MCS rainfall intensities to atmospheric drivers. Together, these datasets capture a broad range of MCS-driver variability under current climate conditions.

Based on 10.8 µm-band brightness temperatures of the Meteosat Second Generation (MSG; Schmetz et al 2002, EUMETSAT 2021), we identify MCSs as contiguous cloud $\leqslant$−50 ^∘C regions larger than 5000 km² between 16 and 1900UTC, the time when the frequency of MCSs reaches a maximum (e.g. Mathon and Laurent 2001), for a Sahelian domain (9^∘–19^∘ N, 10^∘ W–15^∘ E). Maximum MCS rainfall (P$_\mathrm{max}$) is sampled from matched-up 'high-quality precipitation' (HQprecipitation, merged microwave-only precipitation estimate) fields of the half-hourly Final Run V06B Integrated Multi-satellitE Retrievals for Global Precipitation Measurement (IMERG-HQ; Huffman et al 2019) dataset at ∼15 km resolution. MCS snapshots with P$_\mathrm{max} \leqslant$ 1 mm h⁻¹ are removed to exclude non-precipitating cloud shields. We thus obtain conditional maximum rain rates from 22 368 MCS snapshots, for which we identify pre-storm driver conditions.

Environmental TCW and wind shear, defined here as the 925–600 hPa zonal wind difference in m s⁻¹, are sampled from ERA5 reanalysis hourly data (Hersbach et al 2020, CDS 2021) coarsened to 0.7^∘ resolution at 1200UTC, preceding afternoon MCSs, and at the location of minimum MCS temperature. The 0.7^∘ resolution for atmospheric drivers reflects the scale of smallest considered MCSs while ensuring better consistency with the coarse spatial resolution of CMIP model data. For brevity, we refer to the combination of MSG, IMERG-HQ and ERA5 data for analyses as observation-based (OBS).

This study uses data from the CP4 simulation; a 4.4 km pan-African CP climate simulation based on the Met Office Unified Model and created within the Future Climate for Africa (FCFA) Improving Model Processes for African cLimAte (IMPALA) project (Stratton et al 2018, Kendon et al 2019). The modelled historical period (CP4_H ) encompasses 1997–2006 with atmospheric boundary conditions provided by a prototype of the latest atmosphere-only UM global model GA7/GL7 at 25 km with sea-surface temperatures (SST) prescribed from observations (Reynolds et al 2007). The CP4 future projection (CP4_F ) covers ten years representative of 2100 climate conditions. It uses GA7/GL7 atmospheric boundary conditions under increased greenhouse gas concentrations in line with the Representative Concentration pathway (RCP) 8.5 at the end of this century. Driving SSTs are adjusted from CP4_H to reflect end-of-century SSTs by adding a climatological annual cycle of ΔT derived from a HadGEM2-ES climate projection (Jones et al 2011), while ozone and aerosol concentrations in CP4_F remain the same as for CP4_H .

At ∼4.4 km resolution, CP4 operates within the 'grey zone' for resolving convection (Field et al 2017), but has been confirmed to improve the intensity and distribution of precipitation across the Sahel (Berthou et al 2019). It also correctly captures climatological MCS distributions, albeit with underestimations in maximum MCS size and speeds together with an overestimation in MCS frequencies (Crook et al 2019). We extract simulated afternoon MCSs following the same approach as for OBS, converting outgoing longwave radiation into brightness temperatures and applying the $\leqslant$−50 ^∘C temperature threshold. Filtering for $\geqslant$5000 km² rainy clouds gives 45 977 MCS snapshots for CP4_H and 35 975 snapshots for CP4_F with co-located atmospheric conditions sampled at 1200UTC and at 0.7^∘ resolution. The smaller number of future MCSs is consistent with previous CP4 analyses, which found fewer but more intense rain events paired with longer dry spells across the Sahel for CP4_F (Berthou et al 2019, Kendon et al 2019). In line with IMERG-HQ rainfall, the modelled rainfall is coarsened to 15 km resolution before sampling MCS maximum rainfall. This averaging also reduces the overestimation of high-intensity rainfall CP4 shows at native resolution (cf supplementary figure S1, Berthou et al 2019).

We analyse driver changes in wind shear and TCW out to 2100 in simulations from 38 CMIP5 (Taylor et al 2012) and 26 CMIP6 models (Eyring et al 2016); one realisation per model i.e. 'r1i1p1' members (see supplementary table 1; available online at stacks.iop.org/ERL/16/104023/mmedia) for which data were available for both variables. Only simulations forced by the RCP 8.5 (Shared Socioeconomic Pathway 5-8.5 for CMIP6) are analysed, consistent with CP4. Three different 30-year time slices are considered; 2030–2059 ('2040'), 2050–2079 ('2060'), and 2070–2099 ('2080'). The reference period covers 1950–1999, which was driven by historical anthropogenic and natural forcings. The models were interpolated onto a common 1.25^∘ latitude × 1.875^∘ longitude grid and averaged for respective time slices. For the rainfall reconstruction, we use a seasonal average of the changes in atmospheric drivers during the peak monsoon months July–September (JAS), during which MCS activity is at a maximum in the Sahel (Lafore et al 2011, Nicholson 2018). For a subset of models, we also evaluate changes in 3-hourly and daily rainfall extremes, depending on availability (cf supplementary table 1).

Using the identified MCSs from OBS and CP4, we define MCS extreme rainfall as the 95th percentile of the maximum MCS rainfall distribution ($P_\mathrm{max^{95}}$). We assume a linear relationship between changes in $\Delta P_\mathrm{max^{95}}$ and atmospheric driver changes. The individual driver contribution is then defined as

$$\begin{equation} \Delta P_{\mathrm{max}^{95},t} = \beta_{t} \times \Delta \mathrm{TCW}, \end{equation} \tag{ 1 }$$

and

$$\begin{equation} \Delta P_\mathrm{max^{95},s} = \beta_\mathrm{s} \times \Delta \mathrm{shear}, \end{equation} \tag{ 2 }$$

where ΔTCW and Δshear denote the differences between JAS average future and historical conditions of the respective variable from either CMIP models or CP4. Reconstructed rainfall from JAS CP4-drivers will be used to evaluate our scaling approach in comparison to the CP4 modelled rainfall change. $\beta_\mathrm t$ and $\beta_\mathrm s$ represent the associated rainfall/driver relationships based on hourly atmospheric data, either as simulated by CP4 or derived from OBS. The change in driver-based $P_\mathrm{max^{95}}$ is then reconstructed as a linear combination of individual driver contributions:

$$\begin{equation} \Delta P_\mathrm{{max}^{95}, t+s} = \Delta P_\mathrm{{max}^{95}, t} + \Delta P_\mathrm{{max}^{95}, s}.\end{equation} \tag{ 3 }$$

Our use of the driver scaling factors $\beta_\mathrm t$ and $\beta_\mathrm s$ depends on the precipitation/driver relationship remaining the same in the current and future climates, and on CP4 capturing the relationship under both climates. This will be discussed in the following.

We first evaluate how the rainfall-driver scaling compares between OBS and CP4_H for the atmospheric drivers TCW and shear. In figures 1(a) and (b), we stratify the MCS sample according to driver strength and compute the 95th percentile of the $P_\mathrm{max}$ distribution to obtain the intensity of the 5% most intense storms that can be supported by any joint TCW-shear combination (i.e. $P_\mathrm{max^{95}}$). This allows us to ascertain how $P_\mathrm{max^{95}}$ changes in response to one driver while the other remains approximately constant.

Zoom In Zoom Out Reset image size

There is a marked tendency for $P_\mathrm{max^{95}}$ to increase with higher TCW as well as with wind shear in OBS (figure 1(a)). This behaviour cannot be explained by any correlation between TCW and shear, which is negative for OBS ($r = -$0.21, $p\leqslant$0.01). CP4_H similarly shows higher $P_\mathrm{max^{95}}$ as TCW increases, but exhibits little sensitivity to ambient shear. Averaging across TCW-bins, figure 1(c) confirms that CP4_H $P_\mathrm{max^{95}}$-scaling with TCW shows good correspondence with a rainfall change of 0.69 mm h⁻¹ per unit increase in TCW (mm) compared to 0.71 mm h⁻¹ for OBS. At the same time, CP4_H considerably underestimates the rainfall spread introduced by wind shear per TCW-bin (blue spread) and consequently does not reproduce the observed $P_\mathrm{max^{95}}$-increase of 0.78 mm h⁻¹ per unit shear (m s⁻¹) shown in figure 1(d).

This result is in line with previous studies of CP4, which found realistic MCS rainfall sensitivity to TCW but little shear dependency (Fitzpatrick et al 2020, Senior et al 2021). Our observation-based results however highlight the importance of shear for MCS maximum rainfall intensity on synoptic time scales; a relationship which is backed by theory (Alfaro 2017) and can be captured by idealised models below 1 km spatial resolution (Bickle et al 2021).

We will therefore rely on the historical absolute $P_\mathrm{max^{95}}$-shear scaling as inferred from OBS following equation (2) with $\beta_\mathrm s$ defined as

$$\begin{equation} \beta_{s} = \left(\frac{\partial P_\mathrm{max^{95}}}{\partial \mathrm{shear}} \right)_\mathrm{OBS} = 0.78\pm0.15~\frac{\mathrm{mm~h}^{-1}}{\mathrm{m~s}^{-1}}~~(\pm \mathrm{SE}) \end{equation} \tag{ 4 }$$

representing the absolute change in rainfall per unit shear from observations (figure 1(d)) ± standard error (SE). This is applied under the assumption that scaling of rainfall with wind shear remains constant across climates, which is supported by the good fit of the layer-lifting model of convection in Bickle et al (2021), since this model depends on MCS-relative flows of moisture and hence shear. For TCW on the other hand, CP4 shows realistic behaviour, which we exploit in the next step to derive the scaling of $P_\mathrm{max^{95}}$ with increasing TCW under global warming.

Based on end-of-century projected changes by CP4, we now consider the climate change sensitivity of $P_\mathrm{max^{95}}$ to TCW relative to the historical period. Different from the effect of wind shear, the rainfall-humidity relationship cannot be assumed to remain the same across climates as similar levels of TCW do not result in similar rainfall intensities.

This is illustrated in figure 2, where in addition to CP4_H , the pre-storm atmospheric drivers are separated for CP4_F MCSs, similar to the approach used to obtain figure 1(c), but by averaging the ${P_\mathrm{max^{95}}}$ distribution across 5-percentile TCW-bins. Compared to CP4_H , the CP4_F MCS distribution shows a marked shift towards higher TCW. At the same time, when MCSs occur, similar levels of TCW in CP4_F result in higher $P_\mathrm{max^{95}}$ than in CP4_H . This is in contrast to our current understanding for likely changes in mean rainfall, for which less rainfall is expected in a warmer climate for similar TCW, as more moisture is necessary to reach similar levels of relative humidity. A possible explanation for this behaviour is that extreme MCS rainfall occurs when convection is strong. Convective updraughts are expected to widen and intensify under global warming (Prein et al 2017), with the latter similarly identified for CP4 (Jackson et al 2020). The dynamical MCS intensification may then result in higher extreme rainfall in the future for similar TCW levels.

Zoom In Zoom Out Reset image size

We follow a 'quantile-projection' approach to map the historical driver distribution and associated $P_\mathrm{max^{95}}$ onto future conditions, as represented by climate change vectors ${\beta}_\mathrm{t}$ between CP4_F and CP4_H in figure 2. The resulting ${\beta}_\mathrm{t}$ ranges from 1.07 up to 1.33 (figure 2(a) inset), with a mean ${\beta}_\mathrm{t}$ according to

$$\begin{equation} {\beta}_\mathrm{t} = \overline{\left(\frac{\Delta P_\mathrm{max^{95}}}{\Delta \mathrm{TCW}} \right)}_\mathrm{CP4} = 1.2\pm0.13~\mathrm h^{-1}~~ (\pm \mathrm{min/max}). \end{equation} \tag{ 5 }$$

The fact that hourly extreme rain scales with preceding TCW with ${\beta}_\mathrm{t}\gt1$ points towards the importance of dynamical processes that help to increase the vertical transport of moisture in MCSs and the MCS moisture supply from the surroundings. In the following, ${\beta}_\mathrm{t}$ is used to calculate $P_\mathrm{max^{95},t}$ from TCW changes in CMIP models.

So far, presented results did not consider sub-regional driver variability. Across the Sahel domain, TCW changes show marked spatial variability linked to a pronounced zonal asymmetry in projected moisture changes as well as to generally strong meridional gradients in this wet-dry transition region (supplementary figures S2c, S3, S4). Strongest wind shear changes tend to follow zonal bands for CP4 and CMIP, for which most models suggest a peak in the eastern Sahel (supplementary figures S2d, S5, S6).

Assuming our domain-wide scaling factors remain valid locally, we calculate a pixel-based $\Delta P_\mathrm{max^{95},t+s}$ for the 2080 period in figures 4(a) and (b), showing the CMIP ensembles' 90th and 10th percentile, respectively. In line with model agreements on peaks of TCW and shear change in the eastern Sahel, the reconstructed $\Delta P_\mathrm{max^{95},t+s}$ shows highest values over Niger and northern Nigeria. Shear contributions to $\Delta P_\mathrm{max^{95},t+s}$ $\geqslant$8% up to a maximum of 17% (hatching) cover most of the northern Sahel for both extreme ends of the CMIP range, commensurate with large CMIP uncertainties in modelling changes in southern Saharan lower tropospheric warming (Rowell et al 2021).

Zoom In Zoom Out Reset image size

Finally, we compare relative changes in reconstructed $\Delta P_\mathrm{max^{95}t+s}$ with modelled rainfall changes in CP4 and CMIP (figure 4(c)). The scaling is compared for the Sahel domain and boxes centred on the cities of Bamako (1), Timbuktu (2), and Niamey (3) (figure 4(a)), corresponding to regions along the previously discussed zonal gradient of TCW change projected by many CMIP models. For CP4, this step provides an indication of the skill of the simple linear driver-scaling presented here to reproduce regional intensities of modelled MCS rainfall. Figure 4(c) illustrates that for all regions except Bamako, the CP4-modelled $\Delta P_\mathrm{max^{95}}$ (black cross) lies within the uncertainty range of $\Delta P_\mathrm{max^{95}t+s}$ reconstructed from JAS CP4 driver changes (green cross and shading). The range reflects the uncertainty associated with the TCW and shear scaling factors and reaches a maximum of ±7.7% (Niamey). This good correspondence in spite of CP4's weak shear response may be linked to the scaling not considering additional rainfall intensification factors like instability, suggesting that our results may still be a conservative estimate.

The CMIP5/6-based $\Delta P_\mathrm{max^{95}t+s}$ (green box) is shown in comparison to the modelled change in 95th percentile daily rainfall (for wet days $\geqslant$0.1 mm) of a range of raw CMIP5/6 models (grey box) as well as to a bias-corrected version of CMIP5 (blue box). We consider the latter to be the best available dataset regarding CMIP5-projected rainfall changes in West Africa (Famien et al 2018). Across evaluated regions, the ensemble mean change in reconstructed extreme rainfall lies between +40% and 58% (+20%–31% for the 2040 period), which is markedly higher than for any ensemble-mean modelled CMIP rainfall in the same region. Furthermore, $\Delta P_\mathrm{max^\mathrm{95}t+s}$ shows a clear signal of intensification across the entire CMIP uncertainty range, linked to exclusively positive driver changes. Modelled daily extremes on the other hand include negative changes within the 10–90th percentile CMIP range for all regions, with region-dependent ensemble means between 14% and 35% for raw CMIP and 3%–48% for bias-corrected CMIP5. While we acknowledge that the change in $\Delta P_\mathrm{max^{95}t+s}$, which is based on sub-daily extremes, may behave differently from CMIP daily rainfall metrics, it allows us to compare the reconstructed results to a more commonly-used and directly-inferred extreme rainfall metric. In addition, 3 out of 4 currently available CMIP6 models that provide sub-daily rainfall did not show a stronger signal for 3-hourly compared to daily extremes (supplementary figure S7), suggesting there is no systematic intensification at sub-daily scale that applies to all CMIP models.