Identifying changing precipitation extremes in Sub-Saharan Africa with gauge and satellite products

Some of the satellite products overlap REGEN by 31 years (CHIRPS, ARC2, and PERSIANN), so we use that overlap to determine if these products exhibit similar trends in the quality data areas. For this analysis we compute spatial means for each dataset and examine trends in these area-averaged time series, similar to section 4.1. We also evaluate the gauge-based CPC and GPCC datasets. TAMSAT v3 was omitted from this comparison due to gaps in daily data during the 1980s. We focus our analysis on the two most significant 1983–2013 REGEN trends: Wet season SDII and annual R×1day. Other indices did not show clear trends during this period.

Figure 2 shows the Theil-Sen estimated slope for WS SDII and annual R×1day area-averaged time series for each product, first for the all-SSA quality pixels and then for dry, intermediate, and wet area subsets of those data. There is general agreement between the gauge-based REGEN and GPCC and satellite-gauge CHIRPS and ARC2 as to increases in wet season precipitation intensity (figure 2(a)). This is seen for the all-SSA time series and for the dry and intermediate subsets, though only REGEN, ARC2, and CHIRPS trends are significant. All of the products show a positive trend in the dry area subset, though trend magnitudes vary greatly. There are large differences between product trends for the wet area subset. The REGEN trend is negligible, GPCC shows a large increase in WS SDII, and CHIRPS and PERSIANN show significant positive and negative trends, respectively.

Zoom In Zoom Out Reset image size

Products show comparatively less agreement for annual R×1day trends (figure 2(b)). There is some agreement for the dry area subset, but for other subsets the gauge and satellite-gauge products mainly show opposite sign trends. REGEN and several products show positive trends for dry area annual R×1day (GPCC, ARC2, and PERSIANN) while others have trends that are near zero (CPC and CHIRPS). Similar to WS SDII, products disagree most for the wet area subset. We have low confidence in the GPCC wet area annual R×1day and WS SDII trends. These time series look odd, with very high values in 1983–1986 followed by values ∼25% lower and then a steep increasing trend from the early 1990s to 2013. We suspect this is related to low-quality GPCC data in western Ethiopia and northern Zambia, as GPCC and REGEN per-pixel time series are poorly correlated in these areas even for the less challenging index, WS PRCPTOT (${\rho }_{1983-2013}$ < 0.4). The GPCC intermediate area annual R × 1 d and WS SDII time series also begin with suspiciously high values but data from 1985 to 2013 look reasonable and are similar to the REGEN time series (e.g. WS SDII ρ_{1983–2013 =} 0.6).

Results thus far have shown that with respect to time trends in area-averaged indices, in quality data areas there is decent agreement between gauge and satellite products for increasing WS SDII. For positive annual R×1day time trends, however, gauge and satellite products show agreement only in the dry areas. Disagreement in wetter areas resulted in trends of opposing sign for the SSA average. These variations, differences between the gauge-based products, and the limitations of poor-quality data coverage, highlight wet areas of SSA as being the areas we are seemingly least likely to find convergence with regard to climate change impacts on high precipitation extremes.

Here we examine two questions. First, which products most consistently agree with gridded in situ measurements? Second, which indices are associated with the most interproduct agreement? To address this we compare products and indices to REGEN data using data from all quality area pixels rather than the area-averaged time series. We make this comparison with the five longer record products over 1983–2013 period and with all the products over the 1998–2013 period. For the first comparison we are mainly interested in whether or not the product can recreate the temporal variability in REGEN data. To reduce the influence of spatial biases for each product, time series at each grid cell were first converted into standardized anomalies using the local 1983–2013 mean and standard deviation. Figure 3 shows the per-product correlation with REGEN based on these data. Time series from all pixels in the area of interest were first stacked, for each dataset, then Spearman's Rho correlation was calculated using these data.

Zoom In Zoom Out Reset image size

As figure 3 demonstrates, satellite products struggle with correctly identifying extreme precipitation amounts on a daily time scale. The products do most poorly for annual R×1day and WS CDD. This problem appears related to both accuracy and precision, based on the improved but still relatively weak correlation R×1day for annual R×1day totals. Agreement for 1 d extremes would be surprising, however, given that the observation-based products also show weak correlations R×1day (p < 0.5). For the observation-based products, comparatively higher correlations for annual R×1 d (GPCC ρ ∼ 0.6) are somewhat promising and lead us to infer that, for SSA, more confidence might be placed in climate change assessments of extreme storm totals rather than extreme 1 d precipitation. On the other hand, given the similarities in inputs and methods in the REGEN and GPCC datasets, the modest 0.6 correlation for annual R×5 day time-series underscores the greater uncertainties in estimates of extreme statistics.

The satellite products perform better for the indices based on longer accumulation periods. Two of the products (CHIRPS and PERSIANN) stand out as best performers for annual and wet season precipitation totals, coming close to the agreement seen between REGEN and GPCC (ρ ∼ 0.8). This is promising, given that they may also perform reasonably well in less-gauged regions. CPC and ARC2 have lower correlations but are still around R ∼ 0.65. Compared to the precipitation total indices, correlations for wet season rain days (WS R1mm) and average intensity (WS SDII) are lower for all products and are between ρ ∼ 0.45 and ρ ∼ 0.6. Most of the satellite products perform similarly to GPCC for WS R1mm, as do CHIRPS and PERSIANN for WS SDII. This finding may suggest that satellite-gauge products are good tools for exploring changes in R1mm and WS SDII.

When based on a shorter time period but more products, satellite product performance remains highest for wet season precipitation totals. This larger set of products shows consistently higher accuracy for number of rain days than for dry spell length or precipitation intensity. These aspects are shown in figure 4, in which Taylor Diagrams are used to compare 1998–2013 data from 14 products (GPCC, CPC, and 12 satellite products) compare to REGEN wet season indices, again for quality data pixels only. Products were evaluated based on years with complete data. Products with fewer than 16 years evaluated were TAMSAT (2000–2005 and 2007–2013), RFE2 (2001–2004, 2006–2013), and GSMAP and TRMM 3B42 (both 2001–2013). A number of products do well at WS PRCPTOT and WS R1 mm, including TRMM 3B42, CHIRPS, GPCC, TAMSAT, PERSIANN, GPCP, and CMORPH. This again suggests that changes in R1 mm may be a quite tractable target for which the satellites provide valuable information. Correlations (indicated by azimuth angle), are highest for WS PRCPTOT and WS R1 mm for most of the products (R ∼ 0.9), lower for WS CDD (R ∼ 0.7), and lowest for WS SDII (R ∼ 0.6). Correlations shown in figure 4 are higher than in figure 3 because these are based on non-standardized values that contain more spatial covariance. Correlations for all the standardized products were similar to those shown in figure 2, with correlations for WS R1 mm ranging from R ∼ 0.3 to R ∼ 0.7 and correlations for WS PCRPTOT ranging from R ~ 0.5 to R ∼ 0.8.

Zoom In Zoom Out Reset image size

It is interesting to see that a number of satellite products perform as well as or better than the gauge-only data. As expected, gauge-only GPCC (orange dot) ranks high for most indices. However, GPCC overestimates REGEN variability in WS SDII and underestimates variability for WS R1mm. Over or underestimation of REGEN variability is shown by the ratio of product standard deviation to REGEN standard deviation (distance from the standard deviation = 1 curve). Mean square error is proportional to the linear distance between product symbols and the REGEN symbol (black circle) on the x axis. The other gauge-only product, CPC Unified (gray dot), is among the least skillful products for WS PRCPTOT and WS R1mm. CPC does about as well for WS SDII as CHIRP (purple box), which is the satellite-only version of CHIRPS.

No satellite product scores highest across all four indices, but TRMM 3B42 (light green triangle) and CHIRPS (purple triangle) show the best performance overall. PERSIANN (green asterisk) also ranks consistently high but does not do as well recreating the observed variability. For WS PRCPTOT and WS R1mm, TRMM 3B42 and CHIRPS correlations are similar to GPCC and these products closely mimic REGEN variability and means, outperforming GPCC in these measures. For dry spell length (CDD), the best performing products are TRMM 3B42, GSMAP-gauges, ARC2, GPCC, SM2RAIN, CMORPH, and RFE2. CHIRPS, PERSIANN, and TAMSAT all struggle with under or overestimating CDD variance. For SDII, CHIRPS, PERSIANN, and TRMM 3B42 do somewhat better than the other products, including CPC and GPCC.

Blending stations usually results in higher performance, as can be seen from lower bias and higher correlations in the CHIRPS compared to CHIRP and TRMM 3B42 compared to TRMM 3B42RT (triangle versus square symbols). An exception is for WS R1mm—both versions of TRMM 3B42 perform well. For CHIRPS, the blended stations are what makes it one of the top performing products in WS PRCPTOT and WS R1mm. The satellite-only CHIRP overestimates the WS R1mm variance and underestimates the WS CDD variance.

We complete our evaluation by comparing spatial patterns of SSA trends for indices well represented by the satellites in quality data areas: precipitation totals, rain days, and SDII for the local wettest season. Figure 5 shows the Theil-Sen slope calculated at each pixel using 1983–2013 data. Slope values that did not pass significance with p < 0.10 were set to white. WS SDII (bottom row) is calculated from the ratio of total precipitation (top row) to the number of rain days (middle row), thus it is these first order variables that are behind trends in SDII. Earlier we showed agreement from multiple gauge-based and satellite-based products as to positive WS SDII trends during 1983–2013 for quality data area-averaged time series (except for in wet areas, figure 2). For all the products, trend maps for annual PRCPTOT (not shown) are similar to WS PRCPTOT maps but are larger in magnitude.

Zoom In Zoom Out Reset image size

Products show agreement as to increasing 1983–2013 trends in boreal summer wet season rainfall totals, wet days, and SDII in West Africa Sahel (figure 5). In southwestern West Africa there is some agreement as to declines in boreal summer rain days. Agreement is also seen in Southern Africa for increasing totals, rain days, and SDII, though these trends are spatially fragmented. The Sahel wetting trends are strongly influenced by the severe 1980s drought (Dai et al 2004, Nicholson 2013, Nicholson et al 2018), however, several regional studies have reported increases in rainfall intensity that may also be related to a more extreme climate (Giannini et al 2013, Lodoun et al 2013, Panthou et al 2014). In Southern Africa, increases in seasonal and annual totals are also reported in other studies and are thought to be related to intensification of the Walker Circulation (Maidment et al 2015). Products show some agreement as to drying in central equatorial Africa. Based on figure 5 there is more agreement between these products as to decreases in wet season rain days than rainfall totals (these correspond to boreal fall and winter seasons), however, spatial extent and magnitude of trends vary widely. Drying trends, potential drivers, and dataset discrepancies in this region have been a topic of numerous studies (Malhi and Wright 2004, Yin and Gruber 2010, Zhou et al 2014, Diem et al 2015, Hua et al 2016). According to Hua et al (2016) downward GPCC rainfall trends have been confirmed by some gauge records and are most prominent in boreal spring.

The differences between the gauge products are interesting and potentially concerning. REGEN and GPCC trends are broadly similar, though not identical, while CPC has highly pronounced negative trends across large areas of western East Africa and Central Africa. The CPC differs from REGEN and GPCC in that it mainly uses gauges from the Global Telecommunication System station (GTS) network in Africa. We suspect that changes in the GTS network, fewer stations in the post-2005 'real time' product, and interactions with its interpolation algorithm could be responsible for CPC's spurious trends. The GPCC and REGEN use interpolation algorithms based on distance-weighted sums of neighboring station anomalies. The CPC product uses a surface fitting procedure (optimal interpolation) that may be sensitive to changes in the station distributions over time. The CPC trends also remind us why it is important to thoroughly examine data inputs when working with second order climate change indices. CPC WS SDII trends in western East Africa and Central Africa are not dissimilar from the rest of the products, but trends in the first order variables are.

It is interesting to note large differences between the satellite-gauge products. CHIRPS looks similar to the GPCC and the REGEN, while the ARC2 and PERSIANN show much greater increases in precipitation and the number of rain days in West and Southern Africa. The similarity in the ARC2 and PERSIANN increases in WS PRCTOT and WS R1 mm is quite striking, given the large differences in how the products are created and constrained by station data. The ARC2 uses GTS data to correct for bias. The PERSIANN is bias corrected to match the GPCP. Both the GPCP and ARC merging procedures, combined with changing gauge networks, have previously produced spurious trends (Yin and Gruber 2010, Maidment et al 2015). Potentially spurious trends in the ARC2 dataset were recently examined by Maidment et al (2015). Given the similarities between figures 5(e)–(f) and (k)–(l), we hypothesize that the common element driving similarities between the trends in these products may reside in the underlying thermal infrared radiation datasets, and the inter-calibration of these archives between successive geostationary satellites.

Evaluation of trends in WS PRCPTOT, R1mm, and WS SDII in the satellite-only CHIRP also indicate large increases in precipitation (not shown), although the location of these increases is different than those found in the ARC2 and PERSIANN datasets. It is interesting that the station-adjusted CHIRPS seems to remove these potential discrepancies. The following explains what may be happening. The CHIRPS product is built around a high resolution climatology which makes it much more resilient to changes in network density. When there are large discrepancies between satellite-only estimates and station observations, changes in the station network over time can produce large time-varying systematic errors. Consider, for example, a location with a wet season precipitation average of 300 mm. Large biases in satellite-only estimates are common. If these averaged 200 mm for a location, and stations were only available at that location for the beginning period of record, a large spurious decline would be produced in the blended product. On the other hand, if the satellite estimates have low mean bias, shifts in the frequency of observations would have smaller impacts. The CHIRP climatology has low bias when compared to stations (Dinku et al 2018). This low bias, combined with the large number of stations used in the CHIRPS, appears to produce trends similar to the REGEN and GPCC products. However, this bias correction may reduce the rainfall variance in the CHIRPS, especially in drier locations, and this may explain why the CHIRPS trends are more muted than in the other products. The CHIRP is calculated as a percent anomaly multiplied by the climatological mean. When this mean is low the CHIRPS will almost always be low. So, while the CHIRPS seems to be preferable for analyzing long-term trends, our study has also suggested limitations in some other categories.