ABSTRACT
ANGUS, MICHAEL PETER. Rainfall in East Africa: Uncertainties in Data Availability and the Recent Drying Trend. (Under the direction of Dr. Fredrick Semazzi.)
Reliable observation networks are an essential component in predicting and understanding
the mechanisms underlying seasonal rainfall in East Africa. However between 1980 and 2010,
the number of station observations in the region declined substantially, reducing the ability of
these networks to reliably inform our understanding of the seasonal rainfall. Simultaneously, the
networks recorded a significant decline in the total rainfall amount of the Long Rains. Several
mechanisms have been proposed for the decline of the Long Rains overall precipitation, but
no attempt has been made to assess the impact of station data loss on the observed seasonal
rainfall change. This study aims to address this to increase confidence in future projections of
rainfall variability for this vulnerable region.
The characteristics of the recent drying trend are first defined through a comprehensive
review of the previous literature and by directly comparing multiple precipitation data. The
Long Rains trend is observed to be independent across the three months of the season and
as such conclusions are drawn for each separately. The relationship between precipitation and
station decline is clearly observed in May, where a significant drying trend is only observed in
gauge networks which retain local station information. No significant trends are observed during
March in any available dataset, motivating a move away from the standard "MAM" Long Rains
season.
A two pronged approach is taken to quantify the uncertainty surrounding the Long Rains
decline; a simulation of station loss in East Africa between 1951 and 1980 is developed to isolate
the impact of varying gauge number, and alternative measures of drying are analyzed to isolate
the impact of rainfall decline. Regional trends driven by station data loss are of equal magnitude
to the observed drying trend, particularly in southern Tanzania where station decrease has been
most extensive, casting doubt on the observed rainfall trend during April in this region. Proxy
trend during May, necessitating increased station coverage in these countries. Based on these
results it is proposed that the search for causes of the recent drying trend shift focus from April
Rainfall in East Africa: Uncertainties in Data Availability and the Recent Drying Trend
by
Michael Peter Angus
A dissertation submitted to the Graduate Faculty of North Carolina State University
in partial fulfillment of the requirements for the Degree of
Doctor of Philosophy
Marine, Earth and Atmospheric Sciences
Raleigh, North Carolina
2018
APPROVED BY:
Dr. Walter Robinson Dr. Anantha Aiyyer
Dr. Nagiza Samatova Dr. Fredrick Semazzi
BIOGRAPHY
Michael Angus was born in the North East of England, where he attended high school in
the market town of Chester-le-Street. After finishing school, Michael ventured into the outside
world to begin his B.Sc in Meteorology and Oceanography at the University of East Anglia in
2006. During his studies, he completed a year long work experience program at the National
Oceanography Centre in Liverpool, focusing on coastal flooding in the Thames Valley region.
Upon graduating, Michael moved further afield to join North Carolina State University in 2010,
where he began a masters in Internal Wave Physics under the tuition of Dr Ping-Tung Shaw,
gaining his M.S. degree in 2012. He went on to study for his PhD, switching topics to climate
variability in East Africa. While in the Marine, Earth and Atmospheric Sciences Department,
Michael became President of the Graduate Student Association, helped to set up the graduate
student symposium, and led a number of intramural sports teams to miserable but enjoyable
failure. Taking the opportunity to explore the US during his studies, Michael made it to 25 out
of 50 states, which he thinks is pretty okay.
During this project, Michael found the uncertain future of the people of East Africa a
constant source of motivation. In this spirit, he would like to include the following lines from
"Equatorial Rainstorm" written by American poet Tom Heidlebaugh while living in Kenya:
The rain said, receding:
I only consider myself complete
When every wet chilled man prays
ACKNOWLEDGEMENTS
I would like to thank my adviser Dr Fredrick Semazzi for his helpful suggestions and professional
instruction throughout this PhD project as well as the other members of my advisory committee
for agreeing to serve and advise: Dr Walt Robinson, Dr Anantha Aiyyer and Dr Nagiza Samatova.
All of my committee members were vital to my success, but I would particularly like to thank Dr
Robinson for providing the wider perspective I at times lost somewhere in my code. My colleagues
in the computer sciences department of NC State were instrumental in developing some of the
techniques applied here, and I would especially like to thank Mandar Chaudhary and Gonzalo
Bello. I would also like to acknowledge my lab mate Masilin Gudoshava for helpful suggestions
throughout, and Dr Stefan Liess of the University of Minnesota for some methodology pointers.
Multiple data sets were provided by the NOAA/OAR/ESRL PSD, Boulder, Colorado, USA,
from their Web site at http://www.esrl.noaa.gov/psd/, and the CMIP5 data portal hosted
by Lawrence Livermore National Laboratory at https://esgf-node.llnl.gov/projects/esgf-llnl/.
This project was partially funded by NSF award grant No. 1029731. On a personal note, I am
extremely grateful for the patience and overly generous support of my partner and occasional
lab mate Rowan Argent.
TABLE OF CONTENTS
LIST OF TABLES . . . vi
LIST OF FIGURES. . . vii
Chapter 1 Introduction . . . 1
1.1 Background and Motivation . . . 1
1.2 An Overview of the East African rainfall paradigm . . . 2
1.3 The recent Long Rains precipitation decline . . . 5
1.4 "The East African Climate Paradox" . . . 8
1.5 Objectives . . . 10
1.6 Figures . . . 12
Chapter 2 Recent decline in consistency amongst gridded precipitation datasets in East Africa . . . 15
2.1 Introduction . . . 15
2.2 Data and Methods . . . 17
2.3 Climatology and Trends in the Long Rains . . . 19
2.4 East African station data loss . . . 22
2.5 Discussion and Conclusion . . . 24
2.6 Tables . . . 27
2.7 Figures . . . 28
Chapter 3 The impact of simulated station data loss on the Long Rains decline 45 3.1 Introduction . . . 45
3.2 Data and Methods . . . 47
3.3 Comparison of Interpolation methods . . . 49
3.4 Gauge density variation in the interpolated GHCN data . . . 51
3.5 Trend analysis . . . 53
3.6 Spatial homogeneity . . . 56
3.7 Conclusion . . . 56
3.8 Figures . . . 59
Chapter 4 The long rains decline in precipitation dependent proxy datasets. . 76
4.1 Introduction . . . 76
4.4.1 Data and Methods . . . 87
4.4.2 Results . . . 89
4.5 Summary and Future Work . . . 91
4.6 Tables . . . 93
4.7 Figures . . . 99
Chapter 5 The Long Rains decline in Atmospheric Climate Models and Re-analyses . . . .111
5.1 Introduction . . . 111
5.2 Data and Methods . . . 112
5.3 Model Performance . . . 112
5.4 Model Trends . . . 114
5.5 Summary and Future Work . . . 116
5.6 Tables . . . 118
5.7 Figures . . . 120
Chapter 6 Conclusion . . . .129
REFERENCES . . . .132
APPENDIX . . . .148
Appendix A . . . 149
LIST OF TABLES
Table 2.1 Gridded precipitation datasets available in East Africa. TIR = Thermal Infrared, PMW = Passive Microwave, both from satellite measurements. . 27 Table 2.2 Correlation between GPCC precipitation for each of the months in the
Long Rains season, averaged over the whole East African domain for the period 1951-1980. Correlations in bold significant at the 90% confidence interval (P<0.1) . . . 27
Table 4.1 2012 Landcover classification from MODIS data set MCD12Q1 . . . 93 Table 4.2 East Africa wide changes in Cassava Yield, for 1982-2008. Bold Trends
significant at the 95% confidence level, based on the Mann-Kendall test. Decadal averages provided for reference . . . 94 Table 4.3 East Africa wide changes in Maize Yield, for 1982-2008. Bold Trends
signif-icant at the 95% confidence level, based on the Mann-Kendall test. Decadal averages provided for reference . . . 95 Table 4.4 East Africa wide changes in Sorghum Yield, for 1982-2008. Bold Trends
significant at the 95% confidence level, based on the Mann-Kendall test. Decadal averages provided for reference . . . 96 Table 4.5 Climate Indices chosen for community guided feature selection . . . 97 Table 4.6 Percentage of experiments (out of 30) during which a climate index was
selected as a response to the given rainfall timeseries, using Community Guided Feature Selection for 1951-1980. . . 98
Table 5.1 Members of the CMIP5 AMIP Family of Models selected for this study. Each model includes data from at least 1982-2008. . . 118 Table 5.2 Reanalysis data sets, for which precipitation estimates were calculated from
LIST OF FIGURES
Figure 1.1 East Africa domain, with topographical height shown in background. Inset shows location globally (red box). . . 12 Figure 1.2 a) Monthly observed precipitation for 1982-2008, spatially averaged over
the East Africa domain. CenTrends has been interpolated to the same grid as the GPCC rainfall data set. b) Monthly rainfall maximum over the same period in the CenTrends rainfall dataset. . . 13 Figure 1.3 Climatological mean precipitation (mm/month), averaged between 27.5◦E
and 52◦E, for gridded gauge data (CenTrends) over two year periods 1955-1956 to 2007-2008. . . 14
Figure 2.1 Percentage of rainfall for each month as a total of the annual rainfall in East Africa. Climatology based on the GPCC data set, 1951-1980. . . 28 Figure 2.2 Climatological mean precipitation (mm/month) during the month of March,
1982-2008 for gridded gauge (GPCC, CRU, PRECL, CenTrends) and satellite incorporating (CHIRPS, CMAP, TARCAT) data sets. . . 29 Figure 2.3 Climatological mean precipitation (mm/month) during the month of April,
1982-2008 for gridded gauge (GPCC, CRU, PRECL, CenTrends) and satellite incorporating (CHIRPS, CMAP, TARCAT) data sets. . . 30 Figure 2.4 Climatological mean precipitation (mm/month) during the month of May,
1982-2008 for gridded gauge (GPCC, CRU, PRECL, CenTrends) and satellite incorporating (CHIRPS, CMAP, TARCAT) data sets. . . 31 Figure 2.5 Climatological mean precipitation (mm/month), averaged between 27.5◦E
and 52◦E, 1982-2008 for gridded gauge (GPCC, CRU, PRECL, Cen-Trends) and satellite incorporating (CHIRPS, CMAP, TARCAT) data sets. . . 32 Figure 2.6 Linear trend in precipitation (mm/month/year) during the month of March,
1982-2008 for gridded gauge (GPCC, CRU, PRECL, CenTrends) and satellite incorporating (CHIRPS, CMAP, TARCAT) data sets. Grid cells which pass the Mann-Kendall test at the 95% significance level indicated by black crosses. . . 33 Figure 2.7 Linear trend in precipitation (mm/month/year) during the month of April,
1982-2008 for gridded gauge (GPCC, CRU, PRECL, CenTrends) and satellite incorporating (CHIRPS, CMAP, TARCAT) data sets. Grid cells which pass the Mann-Kendall test at the 95% significance level indicated by black crosses. . . 34 Figure 2.8 Linear trend in precipitation (mm/month/year) during the month of May,
1982-2008 for gridded gauge (GPCC, CRU, PRECL, CenTrends) and satellite incorporating (CHIRPS, CMAP, TARCAT) data sets. Grid cells which pass the Mann-Kendall test at the 95% significance level indicated by black crosses. . . 35
Figure 2.9 Precipitation trend (mm/month/year), averaged between 27.5◦E and 52◦E, 1982-2008 for gridded gauge (GPCC, CRU, PRECL, CenTrends) and satellite incorporating (CHIRPS, CMAP, TARCAT) data sets. . . 36 Figure 2.10 Precipitation trend (mm/month/year), averaged between 27.5◦E and 52◦E,
1955-1981 for gridded gauge (GPCC, CRU, PRECL, CenTrends) data sets. 37 Figure 2.11 a) to c): Linear precipitation trend (mm/month/year) for the total East
African domain, with error bar representing 95% confidence intervals. d) to f) Median linear precipitation trend (mm/month/year), as calculated for each grid cell within the total East African domain, with error bar representing 75% and 25% quantiles. Trends calculated over 1982-2008 for gridded gauge (GPCC, CRU, PRECL, CenTrends) and satellite incor-porating (CHIRPS, CMAP, TARCAT) data sets during March (a and d), April (b and e) and May (c and f). . . 38 Figure 2.12 a) to c): Linear precipitation trend (mm/month/year) for anomalously
high monthly rainfall regions (>15% of annual rainfall), with error bar representing 95% confidence intervals. d) to f) Median linear precipita-tion trend (mm/month/year), as calculated for each grid cell within the anomalously high monthly rainfall region, with error bar representing 75% and 25% quantiles. Trends calculated over 1982-2008 for gridded gauge (GPCC, CRU, PRECL, CenTrends) and satellite incorporating (CHIRPS, CMAP, TARCAT) data sets during March (a and d), April (b and e) and May (c and f). Y-axis values adjusted in c) and f) . . . 39 Figure 2.13 Station Decline shown on left side of image represented by difference
be-tween number of stations per grid cells bebe-tween 1980 and 2010. CenTrends only as an approximation, based on number of stations per country (Funk et al. 2015). Light blue at 0◦N 33◦E represents Lake Victoria, for which no data is provided. . . 40 Figure 2.14 Running 20-year (1901-1920,1902-1921,etc.) pearson correlation between
GPCC and labelled data set for total rainfall in anomalously high monthly rainfall regions (>15% of annual rainfall). Correlations below 0.45 are insignificant at 95% confidence level. Y-axis values adjusted during May. . 41 Figure 2.15 Difference between GPCC and CRU precipitation, standardized at each
grid cell during a) March b) April and c) May over the period 1955-1981. d-f: same as a-c, but over the period 1982-2008. . . 42 Figure 2.16 Difference between GPCC and PRECL precipitation, standardized at each
Figure 3.1 Pink markers designate the location of stations in GHCN which report rainfall for all of the years 1951-1980, during March (a), April (b) and May (c). Background contours indicate elevation from 0 to 2400 at an interval of 100m. For each station, the adjusted R-squared value between distance in meters is also reported for the same months (b,d,f). . . 59 Figure 3.2 Correlation Decay Distance calculated for two separate grid cells in East
Africa. Each black cross represents one grid cell in CenTrends, following interpolation to the same grid as GPCC. Correlation for March rainfall only, over period 1951-1980. Orange line indicates exponential decay fol-lowing equation 3.1. Blue dotted line indicates CDD value at each location, for threshold r=0.36. . . 60 Figure 3.3 Interannual variability of peak (>15% of annual total) rainfall during
March (a,b), April (c,d) and May (e,f). On the right hand side, the central mark indicates the median, and the bottom and top edges of the box indi-cate the 25th and 75th percentiles, respectively. Whiskers extend a further 1.5 times 25th and 75th percentile, with outliers marked by red crosses. . 61 Figure 3.4 Root mean square difference of peak (>15% of annual total) rainfall
be-tween GPCC and gridded data sets (blue) and bebe-tween GPCC and calcu-lated interpolations (black), expressed as a fraction of the climatological mean. Calculated over the period 1951-1980. Central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. Whiskers extend a further 1.5 times 25th and 75th percentile, with outliers marked by red crosses. . . 62 Figure 3.5 Black crosses indicate GHCN network stations available from 1951 to 1980.
Red boxes indicate points contributing to the GHCN2010 interpolation experiment. Turquoise squares indicate the location of stations reporting in the GPCC data set for each month of 2010. . . 63 Figure 3.6 Stations providing data during each year of the interpolation for March
(a,b), April (c,d) and May (e,f). Colors on the left hand side indicate whether the station was removed during the interpolation (red) or constant throughout (black). The total number of stations included per year for East Africa is shown on the right hand side. . . 64 Figure 3.7 Stations providing data during each year of the GHCN gain interpolation
for March (a,b), April (c,d) and May (e,f). Colors on the left hand side indicate the year at which the individual station data began providing data. The total number of stations included per year for East Africa is shown on the right hand side. . . 65 Figure 3.8 Root mean square difference of peak (>15% of annual total) rainfall
be-tween GPCC and GHCN interpolation experiments, expressed as a frac-tion of the climatological mean. Calculated over the period 1951-1980. Central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. Whiskers extend a further 1.5 times 25th and 75th percentile, with outliers marked by red crosses. GHCN loss represented by ensemble mean station removal. . . 66
Figure 3.9 Linear trend in precipitation (mm/month/year) 1951-1980 for gridded gauge (GPCC) and GHCN kriging interpolated data sets. Grid cells which pass the Mann-Kendall test at the 95% significance level indicated by black crosses. Final column expresses absolute difference between the first two. . 67 Figure 3.10 Linear trend in precipitation (mm/month/year) 1951-1980 for gridded
gauge (GPCC) and GHCN kriging 2010 interpolated data sets. Grid cells which pass the Mann-Kendall test at the 95% significance level indicated by black crosses. Final column expresses absolute difference between the first two. . . 68 Figure 3.11 Linear trend in precipitation (mm/month/year) 1951-1980 for gridded
gauge (GPCC) and GHCN kriging gain interpolated data sets. Grid cells which pass the Mann-Kendall test at the 95% significance level indicated by black crosses. Final column expresses absolute difference between the first two. . . 69 Figure 3.12 Distribution of total East Africa Trend in station loss ensemble, with
indi-vidual trends from each other experiment also marked for March (a), April (b) and May (c). (d) to (f) as above, but for anomalously high monthly rainfall regions (>15% of annual rainfall). . . 70 Figure 3.13 Number of ensemble members, from a total of 1000, at each grid cell which
pass the Mann-Kendall test at the 95% significance level. Blue (Red) colors indicate positive (negative) rainfall trends. . . 71 Figure 3.14 Linear trend in precipitation (mm/month/year) 1951-1980 for gridded
gauge (GPCC) and ensemble mean of the GHCN kriging loss interpolated data sets. Grid cells which pass the Mann-Kendall test at the 95% signif-icance level indicated by black crosses. Final column expresses absolute difference between the first two. . . 72 Figure 3.15 a) to c): Linear precipitation trend (mm/month/year) for the total East
African domain, with error bar representing 95% confidence intervals. Lin-ear precipitation trend (mm/month/yLin-ear) for anomalously high monthly rainfall regions (>15% of annual rainfall), with error bar representing 95% confidence intervals.Trends calculated over 1951-1980 for gridded gauge (GPCC), and GHCN interpolations (GHCN kriging, GHCN 2010, GHCN gain, GHCN loss) during March (a and d), April (b and e) and May (c and f). Error bars of GHCN loss represent 10th and 90th member of ranked ensemble trend. . . 73 Figure 3.16 Correlation Decay Distance (km) in the GPCC data set. Final column
Figure 4.2 NOAA OLR Climatology (W/m2), over the period 1982-2008. Bottom right panel average calculated between 27.25◦E and 52◦E . . . 100 Figure 4.3 Top: NDVI from GIMMS3g (color axis). Red shading represents areas for
which burn data exceeds 10% of the grid cell area during any month of the year, yellow dashes outline protected areas from the World Database of Protected Areas. Bottom: MODIS Landcover class during 2012, see table 4.1 for classifications. . . 101 Figure 4.4 Linear trend in OLR (W/m2/year) over the period 1982-2008. . . 102 Figure 4.5 Linear trend in OLR (W/m2/year) over the period 1982-2008. Average
calculated between 27.25◦E and 52◦E . . . 103 Figure 4.6 Linear OLR trend (W/m2/year) for the total East African domain, after
subtracting tropics wide (20◦N to 20◦S) trend. Error bars represent 95% confidence intervals. Calculated over period 1982-2008 . . . 104 Figure 4.7 March, April, May and June NDVI trend over 1982-2008 as a percentage
change of the 1982-2011 climatological mean. Only grid points that pass the Mann-Kendall test at the 95% significance level are shown. . . 105 Figure 4.8 Median Percentage change with 25th and 75th quartile for: a-c, NDVI,
Protected areas of NDVI (PA), No Burn areas of NDVI (NB) and pro-tected areas with no burn areas for NDVI (PA NB), April âĂŞ June. d-f) April-June trends in NDVI for Land classes that remained consistent from 2001-2012, evergreen broadleaf forest (ebf), Open Shrubland (OS), Woody Savannas (WSA) Savannas (SA), Grasslands (GL), Cropland/natural veg-etation (CNV), Barren or sparsely vegetated (BSV). g-i) as in d-f, but for no burn areas. All Trends calculated over 1982-2008 . . . 106 Figure 4.9 Linear percentage change of agricultural yield (Kg/ha/year) for the total
East African domain, with error bar representing 95% confidence intervals. Calculated over period 1982-2008 . . . 107 Figure 4.10 Agricultural Yield (kg/ha) for Cassava, Maize and Sorghum across four
distinct regions in northern East Africa from 1982 to 2008. Data first normalized by dividing the period mean . . . 108 Figure 4.11 Agricultural Yield (kg/ha) for Cassava, Maize and Sorghum across three
distinct regions in southern East Africa from 1982 to 2008. Bottom left represents the East Africa total. Data first normalized by dividing the period mean . . . 109 Figure 4.12 Second Principal component mode of CenTrends Precipitation in East
Africa from 1951-1980. Non-Dimensional Timeseries for each mode pro-vided in bottom right panel. May loading pattern has been multiplied by -1 for direct comparison with others. . . 110
Figure 5.1 a) Climatological mean precipitation (mm/month), averaged between 27.5◦E and 52◦E and over 1982-2008 for satellite observations (CHIRPS). b-g) Re-peated for AMIP models A-H (see Table 5.1 for details). Where applicable, the ensemble average is calculated. . . 120
Figure 5.2 Climatological mean precipitation (mm/month), averaged between 27.5◦E and 52◦E and over 1982-2008 for AMIP models I-Q (see Table 5.1 for details). Where applicable, the ensemble average is calculated . . . 121 Figure 5.3 a-b) Climatological mean precipitation (mm/month), averaged between
27.5◦E and 52◦E and over 1982-2008 for AMIP models R-S (see Table 5.1 for details). Where applicable, the ensemble average is calculated. c-g) As above, but for Reanalysis data sets 1-7 (see Table 5.2 for details). . . 122 Figure 5.4 Spatial Taylor Diagram for a) a latitude-monthly HovmÃűller and b) a
spatial map of East Africa. Red letters indicate AMIP family members, while brown numbers indicate Reanalysis (See Tables 5.1 and 5.2 for de-tails). Each numeral represents the ensemble mean. Reference data calcu-lated from CenTrends precipitation gauge data set. . . 123 Figure 5.5 a) Linear Trend in precipitation (mm/month/year), averaged between
27.5◦E and 52◦E and over 1982-2008 for satellite observations (CHIRPS). b-g) Repeated for AMIP models A-H (see Table 5.1 for details). Where applicable, the ensemble average is calculated. . . 124 Figure 5.6 Linear Trend in precipitation (mm/month/year), averaged between 27.5◦E
and 52◦E and over 1982-2008 for AMIP models R-S (see Table 5.1 for details). Where applicable, the ensemble average is calculated. c-g) As above, but for Reanalysis data sets 1-7 (see Table 5.2 for details). . . 125 Figure 5.7 Linear Trend in precipitation (mm/month/year), averaged between 27.5◦E
and 52◦E and over 1982-2008 for AMIP models R-S (see Table 5.1 for details). Where applicable, the ensemble average is calculated. c-g) As above, but for Reanalysis data sets 1-7 (see Table 5.2 for details). . . 126 Figure 5.8 Each grid square represents an ensemble member with the x-label
rep-resenting a specific model and the y-axis the ensemble member. Color axis represents the median trend in East Africa, for the period 1982-2008. Numbers 1 to 19 represent models A-S from previous figure, while 20-26 describe the reanalysis data listed as 1-7. . . 127 Figure 5.9 Expressed relationship between model rank (x-axis) and ensemble mean
observed median trend in East Africa from 1982-2008 (y-axis). Rank de-termined by Taylor diagrams (Figure 5.4). . . 128
Figure A.4 Interannual variability of total rainfall in the East Africa domain during March (a,b), April (c,d) and May (e,f). On the right hand side, the central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. Whiskers extend a further 1.5 times 25th and 75th percentile, with outliers marked by red crosses. . . 151 Figure A.5 Root mean square difference of total domain rainfall between GPCC and
gridded data sets (blue) and between GPCC and calculated interpolations (black), expressed as a fraction of the climatological mean. Calculated over the period 1951-1980. Central mark indicates the median, and the bottom and top edges of the box indicate the 25th and 75th percentiles, respectively. Whiskers extend a further 1.5 times 25th and 75th percentile, with outliers marked by red crosses. . . 152 Figure A.6 Climatological mean precipitation (mm/month) during the month of March,
1951-1980 for: (a) GPCC (b) GHCN kriging (c), GHCN 2010 (d), GHCN gain (e) and GHCN loss ensemble mean. Differences between GPCC ex-periment and GHCN reconstructions (b) to (e) shown in (f) to (i). . . 152 Figure A.7 Climatological mean precipitation (mm/month) during the month of April,
1951-1980 for: (a) GPCC (b) GHCN kriging (c), GHCN 2010 (d), GHCN gain (e) and GHCN loss ensemble mean. Differences between GPCC ex-periment and GHCN reconstructions (b) to (e) shown in (f) to (i). . . 153 Figure A.8 Climatological mean precipitation (mm/month) during the month of May,
1951-1980 for: (a) GPCC (b) GHCN kriging (c), GHCN 2010 (d), GHCN gain (e) and GHCN loss ensemble mean. Differences between GPCC ex-periment and GHCN reconstructions (b) to (e) shown in (f) to (i). . . 153
1.
Introduction
1.1
Background and Motivation
Regions of the world in which the local population rely on seasonal rainfall for their livelihood
are particularly sensitive to climate change. Natural fluctuations in the interannual variability
can often lead to both flooding and drought, with prevalence towards these extremes dictated
by the relationship between seasonal rainfall and the global climate system. Long term trends in
seasonal rainfall are therefore the subject of intense scientific interest, to help prepare vulnerable
communities facing an uncertain future climate. This study aims to address some of the key
uncertainties of the current projected trend for one such community: the 300 million people
dependent on seasonal rainfall in East Africa.
Although "East Africa" as a region has no strict boundaries, it is typically defined (Funk
et al. 2016, Nicholson 2017) as the countries of the East African Community: Burundi, Kenya,
Rwanda, South Sudan, Tanzania and Uganda and those of the Greater Horn of Africa: Djibouti,
Ethiopia, Eritrea, and Somalia. The spatial dimensions chosen here (Figure 1.1) are purposefully
broad to encompass all possible definitions of East Africa, including the countries listed above
and sections of South Sudan and the Democratic Republic of the Congo. Geographically, this
large domain includes an extreme variety of environments, such as the second largest freshwater
lake in the world (Lake Victoria), the two mountainous branches of the Great Rift Valley, the
Failure of one or both of the two rainy seasons frequently leads to drought (Hastenrath et
al. 2007, Lott et al. 2013, Lyon et al. 2014). Repeated drought conditions throughout the past
quarter of a century have made the region highly food insecure (Funk et al. 2008), with the
Food and Agriculture Organization of the United Nations (FAO) currently estimating that 124
million people are undernourished in East Africa (FAO 2015) due to the strong dependence of
East African countries on locally grown crops for food (Funk and Brown 2009). In addition,
high population and urban area growth (Seto et al. 2013) continues to increase demand on both
agricultural production and foreign aid (Funk and Brown 2009). The impacts of drought may
not be limited to famine; Raleigh and Kniveton (2012) hypothesized that the scant resources
available during drought years may lead to an increase in local conflicts, although there is little
direct evidence to support this (Ayana et al. 2016). Seasonal rainfall predictions, such as those
made by the Famine Early Warning Network (FEWS NET) are therefore vital in preparing the
region for drought events when coupled with effective information dissemination (Washington
and Downing 1999, Hillbruner and Moloney 2012). An accurate understanding of the climate
factors which control rainfall in the region is essential for these predictions to be accurate (Verdin
et al. 2005), motivating this study.
1.2
An Overview of the East African rainfall paradigm
The two principle rainfall seasons in East Africa are the "Long Rains" from March-May and the
"Short Rains" from October to November. This bimodal distribution is asymmetrical, with a
clear maximum during the Long Rains (Figure 1.2). The importance of the Short Rains should
not be discounted though as they can be vitally important for the agriculture and livelihood of
the local population, particularly during years in which the Long Rains are below average (Hoell
and Funk 2014). The interannual variability of each season is independent from the other and
as such they are typically analyzed separately (e.g. Liebmann et al. 2014, Yang et al. 2015).
Conventionally, the Intertropical Convergence Zone (ITCZ) has been understood to be the
primary driver of East Africa’s bimodal rainfall distribution (Riddle and Cook 2008; Riddle
and Wilks 2013) as the timing of the rainy seasons are coincident with the motion of the
overhead sun. This can perhaps best be observed by looking at a longitude average hovmöller
plot averaged over two years of the annual cycle (Figure 1.3). During the boreal winter and
summer months precipitation is constrained to south and north of the equator respectively,
while during spring and autumn rainfall extends throughout East Africa, producing the observed
bimodal distribution in maximum rainfall. The use of the term ITCZ is problematic however, as
the seasonal rainfall observed in equatorial Africa bares no relation to the meeting of the trades
winds usually defined as the ITCZ (Nicholson 2018). Instead, it is more accurate to refer to a
seasonal rainfall belt, the timing and location of which is driven by a number of diverse factors
including the regional circulation, geography and remote forcing from large scale teleconnections
(see Nicholson 2017 for summary). Much of this variability is driven by large scale ocean and
climate modes (Indeje et al. 2000, Mutai and Ward 2000, Diro et al. 2011) such as the El Nin˜o
Southern Oscillation (ENSO). These teleconnections are common in the tropics (Bjerknes 1969),
and create a degree of predictability through statistical relationships between rainfall and ocean
and atmospheric indices (Mutai et al. 1998, Chen and Georgakakos 2014).
The Long Rains season can perhaps best be understood by considering the onset and
cessa-tion of the rains separately. As Camberlin and Philippon (2002) and Eden et al. (2014) show,
the temporal coherence of the rainfall among the months March, April and May is relatively low
with different climate factors controlling the interannual variability of each. During March, the
total amount of rainfall is strongly dependent on the timing of the equatorial rain belt moving
North (Camberlin et al. 2009), with early onset highly correlated with higher total rainfall.
Low level moisture flux from the Indian Ocean also plays a role, with strong westerly
the lower troposphere can disrupt this process. Additionally, strong heating in the Indo-Pacific
ocean basin produces a Walker circulation during the Long Rains season (Williams and Funk
2011), with the strength and location of the descending branch dependent on temperatures in
the Indian and West Pacific oceans (Hoell and Funk 2014). These factors are not independent,
as anomalously warm Indian Ocean temperatures can both increase the occurrence of westerly
flow across East Africa through a surface anticyclone and zonally displace subsidence over East
Africa.
The cessation of the Long rains during May is primarily determined by the strength, timing
and location of the Somali Jet (Findlater 1969). This northward cross-equatorial low to
mid-level jet occurs at the Western boundary of the Indian Ocean, and is forced by the land-sea
temperature contrast and orography of the east African coast (Rodwell and Hoskins 1995). By
blocking the flow of moisture from the Indian Ocean into East Africa, variations in the Somali
Jet impact the cessation date of the Long Rains, and subsequently the amount of May rainfall
(Riddle and Wilks 2013). As the Somali Jet acts to transport moisture from the South Indian
Ocean northwards, the cessation of the Long Rains has also been linked to the onset of the East
Asian Monsoon (Zorita and Tilya 2002, Camberlin et al. 2010).
The Madden-Julian Oscillation (MJO) can also influence interannual variability of the Long
Rains during both the onset and cessation (Pohl and Camberlin 2006a). The sign and timing of
this relationship depends on the location of the convective envelope in the Tropics (Berhane and
Zaitchik 2014) which can act to induce positive or negative Sea-level pressure (SLP) anomalies in
the West Equatorial Indian Ocean. Similar to other interannual sources of variability this can act
to weaken or strengthen the easterly moisture flux. Pohl and Camberlin (2006b) hypothesize that
the impact of the MJO on intraseasonal timescales interferes with the Long Rains connection
to other teleconnections, which may help to explain why the Long Rains are the most difficult
seasonal rainfall to predict in Africa (Jury and Mpeta 2009).
Conversely, the Short Rains are more homogenous (Camberlin et al. 2009), and predictions
for this season have longer lead times than the Long Rains (Nicholson 2014). The most important
source of variability for the Short Rains is the Indian Ocean Zonal Mode, or Indian Ocean Dipole
(IOD). During the extreme positive phase of the IOD, anomalous equatorial easterlies develop,
opposing the mean flow and weakening the flow of moisture away from East Africa (Black et al.
2003), leading to increased rainfall. However, this response is nonlinear, with only strong events
affecting the Short Rains. Black et al. (2003) also draw attention to the forcing of the IOD
by ENSO, which explains the relationship between ENSO and the Short Rains (Omondi et al.
2012). ENSO is not the only trigger for strong IOD events though and the IOD also influences
the Short Rains independently.
In addition to the Pacific and Indian Ocean influences Omondi et al. (2012) also showed that
a dipole pattern in the Tropical Atlantic is well correlated with a negative rainfall anomaly over
much of East Africa during the Short Rains, which varies on decadal timescales. The mechanism
for this relationship was not discussed however. In general, most of the variability of the Short
Rains is interannual with little to no long term trend, although drought events have increased
in recent years (fig 4, Hoell and Funk 2014).
While the primary focus of the literature discussed here is the seasonal variability of East
African rainfall, it is important to provide context on the nature of the precipitation. As
cat-egorized by Dezfuli et al. (2017), the majority of rainfall events in East Africa are localized
convection, driven by surface convergence and/or surface heating. Short (less than forty
min-utes), low intensity (less than 10 mm/h) events are the most commonly occurring, but the
majority of precipitation is observed during high intensity convection events or longer mesoscale
convective systems (MCSs). The Ethiopian Highlands (approx. 8◦ N, 40◦E) and Lake Victoria
are local maxima of MCSs (Jackson et al. 2009) driven by orography and land-lake breeze
sur-face convection respectively. Accurately categorizing rainfall in these two regions is therefore a
al. 2015). Specifically, gauge based precipitation reconstructions such as the Global
Precipita-tion Climatology Center V6 (GPCC; Schneider et al. 2014), and CenTrends V1.0 (Funk et al.
2015) as well as combined satellite gauge products such as the Global Precipitation Climatology
Project (GPCP; Huffman et al. 2009) and the Climate Prediction Center Merged Analysis of
Precipitation dataset (CMAP; Xie and Arkin 1997) show good agreement in decreasing rainfall
(Yang et al. 2014, Funk et al. 2015). The Climate Research Unit TSv3.23 (CRU; Harris et al.
2014) also records a pronounced decline although the magnitude is comparable to the negative
anomalies observed at the beginning and middle of the 20th century (Yang et al. 2014), unlike
the unprecedented decline observed in GPCC or CenTrends. These observations have led to the
consensus that the Long Rains will experience more drought in the short-term future (Funk et
al. 2014, Lyon 2014).
The cause of the current decline is a source of significant debate in the literature. Williams
and Funk (2011) found that an a extension of the warm pool in the Indian Ocean driven by
the anthropogenic warming of sea surface temperatures (SSTs) was responsible for extending
the Walker circulation westward, which caused subsidence and subsequent drying in the region.
Hoell and Funk (2014) and Funk et al. (2015) provided support for this drying hypothesis
by showing the long term SST trend had more influence on the rainfall than the interannual
variation, implying an anthropogenic source. In this hypothesis subsidence over East Africa is
specifically driven by the gradient between the tropical Pacific and Indo-Pacific, rather than
natural variations in Pacific ocean temperatures. Conversely, Lyon and DeWitt (2012) and
Lyon et al. (2014) focus on the variability of the Pacific Decadal Oscillation (PDO), which cause
anomalous westerlies across the Indian Ocean driven by forced stationary Rossby waves. In this
hypothesis the decline in East African rainfall is driven by a homogeneously warm northern
Pacific ocean rather than an anthropogenic signal. Hoell et al. (2017) seek to reconcile these two
theories by separately simulating model response of East African rainfall to a uniform warming
of the Pacific, and to a temperature gradient response. Declining rainfall is observed in the
models in both cases implying a combination of these factors as the cause.
In contrast to that debate, there is little discussion the veracity of the rainfall decline in
the observational datasets. In summary of the current understanding of the Long Rains, Rowell
et al. (2015) comment that there is already compelling evidence for a Long Rains decline over
the past two or three decades based on unanimous agreement amongst observations, with the
caveat that the weaker decline in CRU requires further study. While there is strong agreement
overall that a decline in the Long Rains has occurred over the previous three decades, the
extent of that decline remains uncertain, with the range estimated at -14 to -65 mm/yr per
decade by Maidment et al. (2015). Although the agreement in sign among the observations does
inspire confidence there are reasons to be wary. Firstly, these precipitation datasets are not
independent, sharing gauge data from the global telecommunication system (GTS) and Global
Historical Climate Network (GHCN, Peterson and Vose 1997). Strong agreement is therefore
to be expected over this period, particularly as the relative percentage of gauges in the GPCC
dataset from GTS has increased (Schneider et al. 2014). The gauge density throughout the East
Africa region has simultaneously decreased during the Long Rains decline (Dinku et al. 2007)
and the extent of the uncertainty introduced by this decline in available station data has yet
to be examined. Yin and Gruber (2010) found that in the Congo River Basin spurious trends
were introduced to the GPCP record by changes in the number of reporting stations between
1987 and 1991, supporting the hypothesis that trends in station number can influence trends in
recorded rainfall.
In addition to the uncertainties in the commonly used precipitation datasets, alternative
measures of precipitation frequently do not agree on the extent of the decline. Higher resolution
satellite based precipitation reconstructions show less severe and spatially inconsistent downward
trends in East Africa from 1982-2010 (Maidment et al. 2015). In particular, the strong rainfall
inconsistent with the extent of the observed GPCP decline and show negligible change over the
past 30 years (Conway et al. 2004, Seleshi and Zanke 2004, Seleshi and Camberlin 2006, Viste
et al. 2013). Some alternative indicators of a move towards drought like conditions do support a
significant decline in rainfall however such as a decline in the total water storage in East Africa
from 2002-2010 (Omondi et al. 2014) and a significant decline in vegetation between 1981 and
2003 for pastoral regions in Kenya (Stige et al. 2006). In a wider context, proxy data from a
marine sediment core also indicate that the rate of 20th century drying in East Africa is unusual
in the context of the last 2000 years (Tierney and Ummenhofer 2015, figure 1).
1.4
"The East African Climate Paradox"
The two most recent generations of the Climate Model Intercomparsion Project (CMIP3 and
CMIP5) project that the precipitation total of the Long Rains will increase during the 21st
century (Hulme et al. 2001, Shongwe et al. 2011, Otieno and Anyah 2013a, Rowell et al. 2015).
This is consistent with the "wet get wetter" pattern of future precipitation projected throughout
much of the tropics (Seager et al. 2010, Balan Sarojini et al. 2012) but inconsistent with the
current understanding of the Long Rains. This contradiction between the observed and projected
rainfall trend in the region has been referred to as the "East Africa Climate Paradox". To assess
how much confidence should be ascribed to the future projections and understand the possible
causes of this paradox it is vital to first ascertain how accurately climate models represent the
historical Long Rains. Dynamically downscaled regional models, atmospheric models forced by
ocean observations, and global climate models have all been assessed within this context.
Historical simulations of the Long Rains using an ensemble of global climate models from
CMIP5 show no skill in recreating the observed trend (Yang et al. 2015, Rowell et al. 2015). The
annual cycle of East African rainfall is also poorly represented with the Short Rains anomalously
wetter than observed and the Long Rains anomalously drier (Yang et al. 2015). The model
un-derestimation of the SST during March-May in the western Indian Ocean leading to a decrease
of moisture flux into East Africa is a possible explanation for this bias (Yang et al. 2015).
porting this hypothesis, the Atmospheric Model Intercomparison Project experiments (AMIP5)
capture the annual cycle successfully, albeit with large variations in skill among the model
en-semble members (Otieno and Anyah 2013b). Certain members of the AMIP5 enen-semble also
capture the observed rainfall decline with some degree of success, although the trend is within
model variability (Maidment et al. 2015) and significantly weaker than the GPCC and GPCP
observations, particularly to the northeast of the region (Liebmann et al. 2014). Despite these
limitations the capture of the main facets of the East African climate system is encouraging.
Regional models within the Co-ordinated Regional Downscaling Experiment for Africa
(CORDEX) framework also simulate the current drying bias in the Long Rains successfully
(Nikulin et al. 2012, Dosio and Panitz 2016). Similarly to the CMIP5 ensemble, future projections
using the CORDEX experiment domain primarily indicate a wetting trend throughout much
of East Africa with some drying in South Sudan (Souverijns et al. 2016). Alternative regional
climate model projections such as Cook and Vizy (2013) and Buontempo et al. (2015) do project
a decline in the Long Rains during the 21st century, stressing the importance of local processes.
For example Cook and Vizy (2013) attribute the simulated drying to a secondary circulation
from an increase in convection over the Congo Basin, driven by increased moisture flux from
the Guinea Coast. This regional complexity makes East Africa particularly difficult for global
climate models to correctly simulate, particularly when combined with the uncertainty in tropical
precipitation projections as a whole (Kent et al. 2015). Disagreement between regional and
global climate model projections of the Long Rains decline therefore cast further doubt on the
ability of the historical CMIP5 simulations to correctly predict trends for precipitation in the
21st century. Conversely, the strong agreement between regional and SST observation driven
1.5
Objectives
Given the uncertainties discussed above, the overarching objective of this study is to assess the
confidence with which the Long Rains can be said to have declined since 1980. To achieve this,
the following objectives and hypotheses are defined:
1. Clearly define the East African Long Rains decline in the rainfall observations, by directly
comparing multiple precipitation representations across the whole domain and temporally
across all three months (March, April, May). Presented in Chapter 2, this analysis provides
the necessary context of the observed decline as understood by the current state of the
art literature.
2. Examine the impact of station data loss in East Africa on the Long Rains decline. The
hypothesis that spurious trends can be introduced to gridded rainfall datasets by varying
yearly station number is tested in Chapter 3, by creating a control dataset of gridded
observations over the period 1951-1980. Experiments in which yearly station count is
varied in the interpolated grid are then compared to those in which the station count
remains the same.
3. Assess the late 20th century/early 21st century trend in a range of outcomes that would
be consistent with the observed decline in the Long Rains. Chapter 4 comprises of
ob-servations available in East Africa which are strongly associated with declining rainfall
such as normalized difference vegetation index (NDVI), crop yield and large scale climate
indices such as the PDO, which have been associated with the long rains variability on
multidecadal timescales. In Chapter 5, the trend of simulated rainfall in climate model
and reanalysis data is calculated. The extent to which these separate metrics of declining
rainfall agree or disagree across the region will be used to build a comprehensive overview
of the extent of the current rainfall decline.
4. In chapter 6, the results of chapters 3 to 5 are synthesized with reference to the observed
trend described in chapter 2. As well as validating the observational datasets, this will
also provide stakeholders with an assessment of how confident we are in representing the
current Long Rains climate variability and with a measure of how that confidence varies
1.6
Figures
Rwanda Tanzania Kenya Uganda Zambia Mozambique Lake Malawi Lake Tanganyika Lake Victoria Lake Tana D. Rep. Congo Somalia Djibouti Ethiopia South Sudan Sudan BurundiFigure 1.1 East Africa domain, with topographical height shown in background. Inset shows location globally (red box).
a)
Figure 1.3 Climatological mean precipitation (mm/month), averaged between 27.5◦E and 52◦E, for gridded gauge data (CenTrends) over two year periods 1955-1956 to 2007-2008.
2.
Recent decline in consistency amongst gridded
precipitation datasets in East Africa
2.1
Introduction
Given the variety of definitions in the geographical boundaries of East Africa it is perhaps
un-surprising that there is no clear agreement on the exact magnitude of the recent drying trend.
Restricting the region to just the Greater Horn of Africa (GHA) Rowell et al. (2015) estimated
the trend to be between -3 and -15 mm/month/decade depending on month, with strong
agree-ment between gridded observations and low resolution satellite approximations (2.5◦x2.5◦) that
the trend is most significant during April. The weakest trends (<9 mm/month/decade) occurred
during March, while the largest disagreement occurred during May, spanning the entire range.
The study period was limited to 1986-2007. Using the same data sets, but with the rainfall
decline defined as occurring over 1983-2012, Yang et al. (2014) found trends ranging from
ap-proximately -7 to -32 mm/decade for the whole March-May season. In addition to the expanded
time series the East Africa domain was also more all-encompassing, similar to Figure 1.1 of this
study. Using a similar spatial domain Maidment et al. (2015) estimated the trend to be between
-14 and -65 mm/decade between 1983 and 2008 for the entire Long Rains. As the climatological
average total rainfall for the Long Rains is approximately 300mm (Figure 1.2), this represents
of political conflict and lack of investment (Dinku et al. 2016). Although this data loss is
ac-knowledged in the decline estimates of Yang et al. (2014) and Rowell (2015), no attempt has
been made to account for its impact. The impacts of station data loss globally were studied by
Wan et al. (2013), who examined the variability of precipitation globally by masking available
stations and comparing the resulting trends and mean rainfall to a control experiment. They
found more uncertainty in the positive trend in rainfall throughout the tropics in the dataset
without masking, caused by the deliberate sampling errors. This implies that a decrease in data
availability not only decreases confidence in gridded observations but also impacts our
under-standing of long term climate variability. Therefore a similar study for East Africa is necessary
to place the observed rainfall trends in their proper context.
Several attempts have been made to address the root problem by either increasing the
number of rainfall gauges in East Africa or retrieving data not currently available in the gridded
precipitation datasets. The Enhancing National Climate Services (ENACTS, Dinku et al. 2016)
initiative, led by the International Research Institute for Climate and Society (IRI) aims to
improve availability, access to, and use of climate information in Africa. By working directly with
national meteorological services, ENACTS are able to include many more hydrological stations
in countries like Ethiopia that are currently unavailable in gridded precipitation data. Currently
however, ENACTS has only been developed in three East African countries (Tanzania, Rwanda,
Ethiopia), limiting regional analysis. A similar approach by Funk et al. (2015) to incorporate
more regional stations than in either GPCC or CRU resulted in the CenTrends V1.0 product,
which provides gridded precipitation data throughout East Africa at 0.1◦x0.1◦ grid spacing.
While certainly a welcome development, CenTrends is not a panacea for the problem of East
African data loss. For example, as with GPCC and CRU, there are no stations reporting between
2000 and 2014 for either Somalia or Eritrea, a combined area of 750,000km2. There are however
clear improvements in data coverage in Ethiopia and Tanzania (fig 1, Funk et al. 2015). The
Trans-African Hydro-Meteorological Observatory (Van de Giesen et al. 2014) are also hoping to
address this issue by adding small and reliable meteorological stations throughout sub-Saharan
Africa at 1000km2 grid spacing. While this will hopefully increase confidence in future climate
studies, the questions around the current decline will remain.
This Chapter is organized as follows: data and methods are discussed in section 2.2, section
2.3 presents an overview of the East African Long Rains and resulting trends in multiple data
sets and Section 2.4 examines the variability introduced to the gridded station data sets by
station data loss. Section 2.5 concludes the Chapter with a discussion of the results.
2.2
Data and Methods
Two broad categories of precipitation data are utilized: station based in situ gridded networks
and satellite derived gridded networks. A total of 7 datasets were included for analysis: four
station based and three satellite inclusive gridded datasets (Table 2.1). Each of the in situ
datasets extend from at least 1948-2011, and the satellite derived networks from at least
1983-2015.
The four station based precipitation datasets are the Global Precpitation Climatology Centre
V7 (GPCC; Schneider et al. 2015), the National Oceanic and Atmospheric Administration’s
(NOAA’s) Precipitation Reconstruction over Land (PRECL; Chen et al. 2002), The Climate
Research Unit TSv3.23 (CRU; Harris et al. 2014) and CenTrends v1.0 (Funk et al. 2015).
GPCC and PRECL are provided by NOAA Earth System Research Laboratory (ESRL) from
their Web site at http://www.esrl.noaa.gov/psd/, while CRU and CenTrends were obtained
from the Climate Research Unit and the Climate Hazards Group directly. GPCC, PRECL and
CRU also provide the number of stations per grid cell, and although CenTrends does not provide
the number of stations per grid cell, Funk et al. (2015) include information on the amount of
CHIRPS are available from 1981, while TARCAT is available from 1983, but with several years
of missing data. TARCAT v2.0 is available from TAMSAT directly, the Climate Hazards Group
provide CHIRPS and CMAP is made available by ESRL. To provide rainfall estimates which
are as accurate as possible, remote sensing data sets such as TARCAT, CMAP and CHIRPS
calibrate remote sensing data from passive microwave and thermal infrared sensors with gauge
data. Of the available gridded precipitation data sets available for 1981-2010 in East Africa, only
NOAA’s African Rainfall Climatology (ARC, Novella and Thiaw 2013) and GPCP are excluded;
ARC due to the strong disagreement with other observations over Central and Eastern Africa
(Maidment et al. 2015), and GPCP due to the similarity to CMAP and the confirmed issues
with bias in the Congo basin (Yin and Gruber 2010).
For direct comparison, the higher resolution networks CenTrends, CHIRPS and TARCAT
have been interpolated to the same 0.5x0.5 degree grid spacing as GPCC, CRU and PRECL
using a simple bilinear interpolation. All of the datasets except for GPCC include estimates for
the rainfall over Lake Victoria. To ensure this does not unfairly bias the results, and because of
the complexity and difficulty in representing the rainfall climatology over the lake (Argent et al.
2015), this region has been masked from all of the above data sets. Due to the lower resolution
of CMAP neither the interpolation or lake masking has been applied to this data set.
In order to make a direct comparison between the Long Rains decline in the gridded
pre-cipitation observations and any other data, it is first important to define how the Long Rains
decline is to be calculated. Here rainfall is expressed in terms of mm/month, with trends
ex-pressed in mm/month/year. To facilitate easier comparison with alternate data sets this may
also be expressed as a percentage change of the climatology (Liebmann et al. 2014). To assess
trend significance, the non-parametric Mann-Kendall test statistic (Hamed and Rao 1998) is
used which tests the null hypothesis that the data are independent and randomly ordered. For
each data set described subsequently, trends will be assessed using these metrics over the period
1982-2008. This time period is chosen for direct comparison with the trends in alternative data
(see Chapter 4).
As important as defining how East Africa rainfall decline should be measured temporally
are the spatial constraints. Any calculation of overall decline must necessarily define what the
boundaries of East Africa are, or alternatively, chose grid points based on conditions such as the
ratio of rainfall during March-May to the annual rainfall (Yang et al. 2014). Spatial variability in
rainfall trends are also important to maintain. Where possible therefore, trends are calculated on
a point by point basis across a wide ranging East Africa domain and compared two dimensionally.
For the region as a whole, trends are also calculated for those points where monthly rainfall
exceeds 15% of the annual total in GPCC (Figure 2.1). This represents 22% of the available grid
cells in East Africa during both March and May, and 43% during April. The season is split into
individual months to determine the influence of the separate climate regimes, and to verify if
the seasonal trend is consistent across the season. Correlations between the total East African
rainfall for each month of the season are relatively low (Table 2.2), supporting a monthly rather
than seasonal approach.
2.3
Climatology and Trends in the Long Rains
The regional climatology of each month is shown in Figures 2.2 to 2.4. During March, the
majority of the rainfall in East Africa occurs south of 8◦S in Tanzania to the East and Zambia to
the West. While this pattern is mostly consistent across all seven gridded precipitation datasets
the magnitude of the observed rainfall varies considerably. Both CMAP and TARCAT display
a dry bias in comparison to the observational data sets, a bias consistent with other satellite
representations of African precipitation (Maidment et al. 2013, Young et al. 2014). TARCAT is
the only data set which records rainfall of over 300mm/month in the Congo rainforest to the West
the two key areas of disagreement are the Ethiopian highlands (approx. 8◦N, 40◦E) and the
Kenyan coast (approx. 2◦S, 40◦E). In both regions, stronger rainfall is observed in the GPCC,
CRU and PRECL gauge observations than in either CenTrends or the satellite interpolations.
Across all three months, GPCC, CRU, PRECL, CenTrends and CHIRPS show good agreement
in the climatology, while CMAP and TARCAT replicate the same spatial patterns at a lower
intensity.
To place the importance of the Long Rains in context, the observed rainfall throughout
the year is shown in Figure 2.5. In all the gridded precipitation data sets, the movement of
the equatorial rain belt can be observed in widespread rainfall south of the equator until May
and after October, separated by the summer in which the average rainfall does not exceed
30mm/month. Conversely, peak rainfall north of the equator occurs during July and August.
The two rainfall seasons, March-May and October to December, are clearly observable as the
only time in which rainfall exceeding 90mm/month extends throughout the majority of the
month. The agreement among all seven gridded data sets on the seasonality of rainfall is excellent
although as with the spatial climatology, TARCAT and CMAP are observably drier, particularly
south of 8◦S during January to April and at the equator between June and September. In contrast
the spatial pattern of trends throughout the region varies noticeably among the data sets. During
March, the overall trend is mostly positive (Figure 2.6) but there is little agreement about where
this positive trend is significant. In GPCC, the strongest trends of +4mm/month/year occur in
the south west of the domain, with weak but significant negative trends in the north. While CRU
broadly shows agreement with GPCC to the south of the equator, the trends in the semi-arid
region of Somalia are of the opposite sign displaying a weak wetting. This result is supported
by PRECL, but not by CenTrends and CHIRPS which instead record weak drying trends. The
magnitudes of the southern positive trends are also weaker in these data sets than in GPCC
and CRU. Finally, there are no significant trends in either CMAP or TARCAT anywhere in the
region. As a result no clear conclusions can be drawn regarding the direction of the observed
rainfall trend for the month of March.
The Long Rains decline is instead predominantly during April and May, where widespread
rainfall decline occurs in the majority of the data sets. During April (Figure 2.7), the central
region of the domain and the Ethiopian highlands are the most consistent areas of decline,
with the north west the only exception. Despite this spatial homogeneity, the regions of East
Africa for which the rainfall decline is significant differs among the data sets, particularly east
of Lake Victoria. There is also little agreement in the sign of the trend in Somalia and the
GHA. Similarly, during May (Figure 2.8), there is good agreement that both in the coastal
region to the south east and in the Ethiopian Highlands that drying has occurred, but the
magnitude varies from 7mm/month/year (CRU) to 2-3mm/month/year (CenTrends). CMAP
and TARCAT represent the two extreme cases, with significant rainfall decline throughout the
region in the former and no significant decline in the latter.
The negative precipitation trends in April and May are the strongest observed throughout
the year in every data set (Figure 2.9). However, the extent of the decline to the north and south
is highly variable among the data sets. In CMAP for example, it is noticeable that the decline is
not constrained to April and May but extends into June and July north of the equator as well,
while in TARCAT the decline is constrained to the south of the equator. A weaker decline is also
present during the October to December Short Rains in the majority of the data sets south of the
equator, with the exception of TARCAT. The Long Rains decline was not observed during the
earlier period of 1955-1981 (Figure 2.10), confirming that this is a recent phenomenon, although
there was a strong negative trend during January and February in the south of the region. In
six of the seven observational data sets, a positive rainfall trend is observed just before the Long
Rains season, extending into March. This shift implies the equatorial rain belt is arriving earlier
in the season, a possible cause of the observed decline. To support this observation, further
data sets, the confidence intervals again indicate that the sign of the domain total trend is
between positive and negative, with a range of -1.5 to 0.6 mm/month/year. It is only during
May (Figure 2.11c) that the uncertainty range lies entirely below zero for any of the datasets,
as observed during GPCC, CenTrends and CMAP. TARCAT is a clear outlier during May as
the only dataset with a mean positive trend; the uncertainty ranges of TARCAT and CMAP do
not overlap, indicating a significant difference between the two. The median of all available grid
cells shows good agreement with the mean (Figure 2.11d-f), with trends close to zero during
March and most significant during May.
The magnitude and uncertainty range increases when focusing only on the regions of East
Africa for which the monthly Long Rains exceed 15% of the annual precipitation (Figure 2.12).
For April this equates to over 40% of the region and differences among the significant trends
are minimal (Figure 2.11 and 2.12 c and e). Trends during March, constrained to the South
of the domain, are more significantly positive in CRU and CenTrends (Figure 2.12d), while
CMAP trends are significantly negative, unlike for the whole domain. Differences among the
data sets are also exaggerated during May when constrained to the Greater Horn. In CRU and
PRECL, the median trend is close to zero, whereas all of the grid cells between the 25th and
75th percentile trend are negative in GPCC, CenTrends, CHIRPS and CMAP (figure 2.12f).
For all data sets, the total May trend has an uncertainty range of -3.8 to 0.5 mm/month/year,
the largest of the three months.
2.4
East African station data loss
Concurrent with the precipitation decline, the number of rainfall gauges in East Africa decreased
markedly between 1980 and 2010 (Figure 2.13). In all four interpolated data sets the number
of gauges increased steadily until 1960, and remained stable through to 1980. In CenTrends,
the total gauge count decreased slightly in 1980 before declining from approximately 600 to 300
overall stations in the 2000s. Dramatic station loss occurs earlier in the other data sets with
total number of stations declining from 800 to 116 stations in GPCC, 375 to 73 in CRU and
361 to 21 stations in PRECL by 2000. Although the number of stations at the peak of GPCC
doubled that of PRECL and CRU, many of these stations occupied the same grid cells in the
data, and so did not increase the spatial resolution. The number of grid cells which contain
at least one station reached a maximum of around 400 in GPCC, 300 in PRECL and 230 in
CRU, declining at a similar rate to the overall station loss. The spatial resolution declined
from a maximum of approximately one observation site per 20,000km2 to one per 80,000km2
in GPCC on average, and to one per 170,000 km2 in CRU. The number of available gauges in
global gridded precipitation data sets at the start of the 21st century in East Africa is therefore
roughly equivalent to the beginning of the 20th century.
Gauge loss has occurred throughout East Africa with decline mostly consistent among the
gridded data sets (Figure 2.13). Somalia is an important exception, with gauge density increasing
during 1980 to 2010 in GPCC from 11 to 25 grid cells containing at least one rain gauge. In CRU
and PRECL this number dropped from 8 to 0 and 13 to 0 in Somalia respectively. Although
decreased by an order of magnitude, gauge density remains highest in the central region around
Lake Victoria and Ethiopia for all three gridded data sets. It is also noteworthy that by 2010
there are no station measurements available in the south west of the domain in either CRU or
PRECL, where March rainfall is the most intense.
The impact of the station data availability decline is evident in the agreement among the
recorded rainfall values. Figure 2.14 shows the correlation of twenty year periods between GPCC
and the other gridded rainfall gauge data sets, for 1901-1920 to 1996-2015. During March, the
correlation between yearly rainfall in GPCC and PRECL fell from 0.95 during 1978-1999 to 0.75
for 1981-2000 and remained below 0.75 subsequently. During April, the weakest correlation
sets has increased in all three months over time. This is most notable comparing GPCC to
PRECL and GPCC to CenTrends. Comparing GPCC to CRU (Figure 2.15), rainfall differences
are comparable between the higher and lower station density periods, with clear decline in
agreement occurring only in the Democratic Republic of the Congo during March (2.15a-b) and
Uganda during May (2.15e-f). Between GPCC and PREC, weakening agreement is prevalent
throughout the entire domain during the Long Rains with the exception of Kenya (Figure 2.16).
During April and May (2.16 c-d, e-f) regions of both Tanzania and Somalia reach differences of
50% of the climatology, an increase from less than 20% in the early period. In the CenTrends
data set, the most consistent increase is in the north east of the domain during April and May,
where station density increased in GPCC (Figure 2.13).
2.5
Discussion and Conclusion
The intensity, timing and location of rainfall are highly variable across the three months that
constitute the Long Rains season (Camberlin and Philippon 2002, Riddle and Cook 2008).
Despite this, the recent Long Rains decline is frequently expressed as the average trend of
all three months combined (Yang et al. 2014, Lyon 2014, Maidment et al. 2015). Here, the
climatology and trends for each individual month have been explored in a range of gauge-only
and satellite-gauge interpolated precipitation data sets.
A decline in the Long Rains between 1982-2008 is observed during April in all data sets,
and for six out of the seven during May, with varying degrees of uncertainty. Only CMAP is
significantly negative during March, with strong agreement otherwise among the gauge-only
data sets that the trend is either weakly positive or negligible. In light of this result, it is
advisable that future studies of the decline focus on the driving mechanisms of rainfall in the
latter months of the Long Rains season to avoid conflating these separate signals. By averaging
together trends of opposite sign between March and April-May, the true extent of the decline may
be underestimated. Equally, the difference between the stronger decline trend in low resolution
satellite datasets and gauge only (Yang et al. 2015) may be exacerbated by the opposite signal