The current study defines Equatorial East Africa (EEA) as the area bounded by Kenya, Uganda, Tanzania, Rwanda, and Burundi (28.5^oE-42^oE, 12^oS-5.5^oN) as shown in Figure 1. Rainfall in most of EEA shows a bimodal seasonal distribution pattern influenced by the seasonal movement of the Inter-Tropical Convergence Zone (ITCZ; Mutai and Ward 2000; Camberlin and Philippon 2002). This movement of ITCZ is accompanied by a change in wind direction that brings in moisture from the adjacent Indian Ocean. The rainfall peaks occur in March-May and October-December seasons and are separated by short dry periods in June and July and January and February. The mean annual precipitation varies from about 400 mm in the eastern and northern parts of Kenya to more than 2800 mm over the Lake Victoria basin. The region is dotted with lakes, key among them L. Victoria which drives a local land/lake breeze that controls the convective activities around the lake basin (Anyah et al., 2006; Onyango et al., 2020). On the other hand, the Kenyan highland deflects the cross-equatorial flow strengthening it during boreal summer (Peagle and Geisler, 1986; Chakraborty et al., 2009).

While in situ rainfall data remains the most accurate, the gauge network density in EA is sparse, thus inadequate to describe drought’s spatial distribution accurately. In addition, the limited length and inconsistency in the observational data (Schreck and Semazzi, 2004; Shilenje and Ongoma, 2016) limit our ability to assess the long-time observed drought changes over the region. In this study, the monthly rainfall and potential evapotranspiration data from CRU TS 4.05 (Harris et al., 2014) were used to calculate the drought index.

One of the reasons for using CRU data is because it includes all the climate variables that allow for the SPEI index computation. Secondly, all grid points have complete data over a 120-year long period during 1900–2020 with relatively higher resolution of 0.5° x 0.5°. Meanwhile, the PET data from CRU uses the FAO-recommended Penman-Monteith method that incorporates daily temperature, radiation, humidity and wind speed. Moreover, the CRU dataset performs relatively well in hydrometeorological studies over East Africa. Ongoma and Chen (2017) found that CRU data outperforms the GPCC dataset in reproducing the East Africa rainfall cycle. The CRU dataset has also been widely used for climate variability studies over East Africa, including the rainfall decadal variation and the model performance evaluation (e.g., Yang et al., 2014; Ayugi et al.,2020; Mbigi et al., 2022). It is noted that, the spring season in this study refers to March, April and May (MAM). Monthly Nino3.4 and PDO indices data were obtained from the NOAA Climate Prediction Center (https://psl.noaa.gov/gcos_wgsp/Timeseries/).

In order to clarify the sub-regional difference in drought characteristics over EEA, this study adopted the maximum loading rule (Vincent Serrano, 2006) to identify the sub-regional boundaries for drought analysis. Figure 2 presents the spatial pattern of the top six dominant REOF modes that were well-demarcated and shows clear spatially disjunctive patterns that did not overlap when a loading threshold of ±0.05 was applied. The six modes explained approximately 77.9% of the total variance in the original SPEI series. The leading mode of variability explains 18.5% of the total variance and has a maximum magnitude in northeastern Kenya. The second mode represents the south of Lake Victoria basin, central and western Tanzania and explains 18.3% of the total variance. The third mode represents Kenya and Tanzania’s coastal strip and explains 12.9% of the total variance. The fourth mode explains 10.1% of the total variance and has its center in western Kenya, Central and northern Uganda. The fifth mode has its maximum magnitude in southern Tanzania and explains 9.9% of the total variance. Lastly, the sixth mode has the largest magnitude in south Uganda, Rwanda and Burundi.

For brevity, the zones were named as follows; Northern Kenya (NK; REOF1), Central Region (CR; REOF2), Coastal Strip (CS; REOF3), Northern Uganda (NU; REOF4), Southern Tanzania (ST; REOF5) and Western Sector (WS; REOF6). We further performed a Spearman correlation between the rotated principal component (RPC) for each dominant REOF mode and the averaged SPEI values corresponding to each region. The correlation coefficients were generally higher than 0.70, except for the NU Region, with a coefficient value of 0.46. The high correlation coefficients between any two corresponding regions imply that the demarcation is appropriate for studying regionally homogeneous drought characteristics.

The estimated SPEI values were used to characterize drought based on frequency, intensity and duration for the analysis period. The spatial distribution of drought characteristics over EA is shown in Figure 3, with lower drought frequency (<20%) found over the region. More specifically, the frequency is relatively lower in the WS sub-region, with a value below 14%. Conversely, CR, NU, ST, and NK sub-regions have a relatively higher frequency (15–20%) of drought occurrence (Figure 3A). The frequency has a maximum magnitude of ∼20% in the CR region at the border of Kenya and Tanzania. On the other hand, Figure 3B reveals that drought in the MAM season lasts between 2.2 and 2.8 months. Drought rarely occurs in all months of the season as it coincides with the long rainfall season when the region receives its maximum rainfall. Higher drought duration occurs mostly in NU, with a large portion of the sub-region having an average duration of 2.7 months. Parts of the western CR and southern WS sub-regions also experience a high average drought duration of 2.6 months. The CS region shows the lowest drought frequency relative to the other sub-region, with the highest duration of approximately 2.2 months. Figure 3C shows that CR and NK sub-regions exhibit relatively lower drought intensity (∼1.44/month) than the rest of EEA, while the WS shows the largest intensity (1.6/month). Generally, the higher frequency and lower intensity of drought in the CR and NK sub-regions indicates that the zones experience more droughts that are less intense, probably due to the reduced rainfall received in the season. In contrast, the NU sub-region showed higher frequency that lasted a relatively longer duration hence the large drought intensity over the region during the MAM season.

The spatial distributions of the drought trend for the entire study period are mapped in Figure 4. The linear trend of MAM SPEI3 majorly returned drying over EA. The trend is insignificant at a 95% confidence level in most parts of EA except southwestern Tanzania, northern Tanzania and parts of northern Kenya and Uganda, where the magnitude is about −0.008/yr (Figure 4B). The result shows that most regions had a mild to moderate increase in drought. Conversely, a small region north of Lake Victoria Basin and southeastern Tanzania shows a weak wetting trend of about 0.002 years⁻¹. Lake Victoria is known to control the diurnal variation of rainfall in the surrounding area, making the basin wetter than the surrounding. Skliris et al. (2016) found that the Lake makes its basin wetter relative to other regions through the “amplification of the water cycle” that occurs at a slower rate due to global warming. Figure 4B reveals a mixture of an increasing and a decreasing trend in drought duration that is largely insignificant over EA. CS sub-region mainly shows an increase in drought duration with an annual trend reaching 0.008/yr for some areas that exhibit a significant trend. Similarly, the NK sub-region shows a positive trend in most parts but appears insignificant, with an annual trend value of about 0.004/yr. Figure 4C shows that drought intensity also shows a combination of both downward and upward trends. The reduction in intensity is mainly located in northwestern Kenya (∼0.006/yr) and parts of western CR sub-regions (∼0.003/yr), where the trend is significant. The decreasing trend is mainly located in WS and western parts of NU sub-regions, where the trend is as high as 0.007/yr.

The time series of SPEI in different sub-regions may help understand the temporal evolution of drought indices at a sub-domain scale and their regional differences. Figure 5 gives the temporal evolution of SPEI over the six sub-regions. The linear trend line and the local regression (LOESS) curves with 11 years are fitted to the SPEI time series to represent the linear and non-linear patterns of SPEI. From Figure 5, we can find that the SPEI3 drought index shows considerable inter-annual variation between wet and dry years. Meanwhile, the 11-year loess curve reveals that negative SPEI3 values have dominated the SPEI evolution since the 1980s in all sub-regions. This indicates that a large portion of EEA is gradually becoming drier in recent decades. The evolution of the SPEI drought index reveals the widespread drought events over EA in 1902, 1921, 1924/25, 1933, 1943, 1949, 1953, 1955, 1961, 1969, 1971/73, 1976, 1983/84, 1992/93, 2000/01, 2004/05, 2007/09, 2011/2012, 2014, and 2017. The drought events identified in the second half of the 20th century correspond well with the findings of previous studies (e.g., Nicholson, 2014; Agutu et al., 2017).

Regarding the regional differences, even more, drought years were identified in different years. For instance, the NK sub-region had drought events in 1928/29, 1938,1945,1976,1980, and 1999 (Figure 5A); CR had droughts in 1909 and 1993 (Figure 5B); CS had droughts in 1961 and 1962 (Figure 5C). For ST, more drought can be found in 1919, 1954, 1957, 1964, 1970, and 1988 (Figure 5D); for NU in 1905, 1939, 1943, and 1953 (Figure 5E) and WS in 1901, 1904, and 2020 (Figure 5F). Generally, many years identified as having sub-regional droughts had negative SPEI values in some of the sub-regions that did not satisfy our drought threshold. This indicates that while there would be a large-scale driver of drought in a particular year, each zone’s distinct interaction with the local physical mechanism may suppress or enhance the dry condition. The time evolution of SPEI shows a weak negative trend in all the sub-regions. The trend is significant in CR (0.005/yr.), CS (0.006/yr.), and ST (0.005/yr.) and insignificant in NK (0.003/yr.), NU (0.001/yr.) and WS (0.004/yr.).

The areal drought extent was computed as a ratio of grid points with SPEI values less than −1 to the total number of all the grid points in that sub-region. Figure 6 shows the time series of percentage drought areas for the six sub-regions of EEA, we can find that the area affected by drought fluctuated significantly year by year, with magnitude ranging between 0% and 95% for the six sub-regions during the study period. Meanwhile, increasing trend of drought affected areas can also be found for all six sub-regions.

The variability of the drought area suggests that the positive linear trend observed may not be monotonous during the entire study period. Moreover, the high percentages of drought areas experienced in the later stages of the study period are likely to affect changes in the slope of the linear fitting. Consequently, we performed a non-linear trend analysis using singular value decomposition (SSA) for each sub-region separately. We chose a window frame of 11 years and reconstructed the time series using the dominant mode to represent the time evolution of the series better. Figure 6 reveals a gradual rise in the area under drought for NK, CS, ST, and NU sub-regions, while the CR and WS show a flat curve, with relatively smaller increasing trends.

The percentage drought areas can further be categorized into four classes, i.e., moderate, severe and extreme drought area. For each category, the percentage area is obtained by comparing the total number of grids under that specific category to the total number of grids in each year. The result revealed that the total increase of percentage drought area is dominated by an increase in the area under moderate and severe droughts, rather than that under extreme drought (Supplementary Figure S2).

In most regions with larger increasing trend of percentage drought area extent, large spatial extension of drought can be also found after the 1980s. Previous study suggested that the decadal variation of rainfall anomalies in East Africa could be linked to the Pacific decadal oscillation (PDO; Bahaga et al., 2018). To understand the potential linkage between drought variation in EEA and PDO, the time series of the PDO index and 11-year running mean SPEI3 drought index over EEA region during 1900–2020 is presented (Supplementary Figure S3). It is found that the correlation between PDO index and SPEI drought index is weak during the whole study period, with temporal correlation coefficient of only 0.17. However, the correlation increase to around 0.88 after 1980. These suggest that the PDO signal can enhance the drought extension after 1980s, but this relationship is not significant over the whole time period.

Based on the SPEI definition, the driving factors considered were rainfall and PET and, thus, their contribution to the total drought area. However, the influence of PET and rainfall on SPEI depend on their independence from one another (Cook et al., 2014). To quantify the contribution of rainfall and PET on SPEI-based drought changes, we adopted the detrending method described by Hamlet and Lettenmaier (2007), which references the SPEI by incorporating changes in either variable. To isolate the impact of rainfall, we calculated the SPEI index using the original data that retained the transient changes in rainfall and detrended PET and named it SPEI-PRE. Similarly, to isolate the impact of PET, SPEI was calculated using detrended rainfall and the original values of PET and referred result as SPEI-PET.

As shown in Figure 9, the combined effects of rainfall and PET on drought either reinforced the area under drought (mid-20th century) or acted in opposition (beginning and end of the 20th century) to mitigate the effect of each other for most regions in EEA. Additionally. Figure 7 also reveals that in the first half of the 20th century, the percentage of drought area computed using detrended rainfall (PET) was higher (lower) than the observed conditions in all six regions. Consequently, the findings suggest that lower PET values reduced drought intensity. However, the result reverses in the post-1975 period, where the increase in PET had a positive effect on drought, indicated by the rise in drought area calculated using varPre. Therefore, the apparent increase in areal drought extent in the post-1980s may be attributed to the high evaporative demands that exert greater stress on an already declining rainfall, thus exacerbating the drought severity. The impact of PET was pronounced in CR, NK, WS, and NU regions (Figure 7).

To explore the impact of PET on areal drought extent further, we compared the area under drought computed using SPEI to that SPI index (Supplementary Figure S4). The result shows divergent evolution of the percentage of land area affected by drought defined by the two indices. Furthermore, the result indicates that since the 1950s, the SPEI index has progressively identified more drought extent, thus pointing to the role of PET in the increase of drought coverage in all the sub-regions. Therefore, the apparent growth in drought areal extent in the post-1980s may be attributed to the high evaporative demands that exerted significant stress on an already declining rainfall, thus exacerbating the drought severity.

REOF helps identify drought variability patterns but fails to provide information on the periodicities of the observed variabilities. Consequently, we apply the wavelet analysis to isolate dominant periodicity in the drought records and identify the coherence between drought and teleconnection indices. Figure 8 shows the result of the wavelet power spectrum of spatially averaged SPEI3 over the six sub-regions. The thick black line encloses values that have passed the red noise test at a 95% confidence level as determined by the Monte Carlo process (Jevrejeva et al., 2003). However, the Cone of Influence (COI) gives the region potentially influenced by the edge effect of data during wavelet transformation, and the area inside it provides an accurate time-frequency representation of the data. In contrast, the area outside it suffers from edge effects, thus unreliable.

For the NK sub-region (Figure 8A), a significant power region exists in the 3 years band in the 1920s, increasing to a periodicity of 6 years in the 1940s before decreasing and decaying by 1955. In addition, there is a strong amplitude between the 1970s and 1985 in the 4 years band and a weak amplitude in the 8 years band between 1940 and 2000s. Figure 8B shows that for the CR sub-region, significant power regions are distributed in the 3–6 years band between 1920 and 1950 and a 9–14 years band between the 1960s and 1990s. Figure 8C shows intermittent wave spectra between 1901 and 1980 in the CS sub-region. A weak 8–12 years band exists between 1960 and 2010, as well as a strong band outside the cone of influence in the 30–48 years band. Figure 8D shows the local wavelet spectra in the 2–7 years band for the NU sub-region. The result appears weak throughout the study period except in 1920, 1930–1950, and 1980–1985. For the ST sub-region, the local wavelet spectra are in the 2–6 years band and 8–30 years band between 1950 and 2000 (Figure 8E). The inter-annual periodicity decreases from the 6 years to 2 years band from the 1980s. The sub-region also has a significant wavelet power in the 30–48 years band outside the cone of influence. In contrast, Figure 8F shows that WS has the weakest periodicity of all the sub-regions throughout the study period, with a reduced level of activities in pre-1920 and post-2000. In general, the regular periodicity is relatively stable at 2–6 years, indicating a large inter-annual variability within the period 1920–1980, while a decadal (8–16 years) variability also exists in some sub-regions. This suggests that inter-annual periodicities (<6years) dominate drought variability across EA. Overall, CR, CS, and ST sub-regions revealed the strongest interannual variations with maximum yellow regions.

Over EEA region, the dominant inter-annual (2–6 years) oscillations observed in the periodicity are mainly modulated by the El-Niño southern oscillation (ENSO; Ogallo, 1989; Kiladis and Diaz, 1989; Nicholson and Kim, 1997; Indeje et al., 2000). Consequently, we explore the interdependency between SPEI3 and the Niño3.4 index using cross-wavelet transformations. In case of any association between the two indices, a slowly varying phase lag is expected, and the phenomenon would be phase-locked, i.e., the phase arrows point only in one direction for a given wavelength (Grinsted et al., 2004). The use of arrows expresses the relationship between the two factors. The arrows pointing east reflect the in-phase relationship, whereas the arrow pointing west indicates an out-of-phase relationship. Conversely, arrows pointing upward (downward) indicate drought lag (lead) the index by a quarter cycle and reveals a non-linear relationship. Higher wavelength transforms coefficients correspond to a stronger correlation between drought and the Niño3.4 index.

Figure 9 shows the result of cross wavelet transformation between Niño3.4 and SPEI with high wavelet coherence larger than 0.6 shown with arrows. As shown in Figure 9A, NK showed high energies between SPEI and Niño3.4 in the 8–12 years frequency band in 1920–1960 years; and 4–6 years in the 1930s. The observed coherence between Niño3.4 and SPEI was in-phase during the periods as arrows points to the east. However, from 1970 to 1990, the two indices were out of phase within the 3–5 years period band. Nevertheless, the coherence in the later period failed the significance test. In the CR sub-region, a significant correlation between the indices appeared intermittently from the 1920s to 1990s in the 2–6 years band (Figure 9B). In the beginning, the two indices were in-phase, followed by Niño3.4 leading SPEI and later an antiphase relationship. Figure 9C shows that the CS sub-region exhibited a statistically significant relationship between SPEI and Niño3.4 between 1940 and 1970; however, the former index led the latter by about 3–4 months. The anti-phase association infers no or weak influence on drought by the ENSO. Another high power appears in the 16–24 years band where the relationship is in-phase. NU sub-region shows a scattered high coherence between the two indices within a 2–6 years band with in-phase coherence between 1920–1940 and Niño3.4 leading SPEI by 3–4 months in the 1950s and 1980s (Figure 9D). Figure 9E reveals that SPEI were out of phase with Niño3.4 in the 1920s over the ST sub-region with high energies lasting 2–6 years. Another statistically significant high power is evident in the 1970s and 1980s, where SPEI lags Niño3.4 by 3–4 months lasting between 9 and 15 years. Over the WS sub-region, significant high power was concentrated in the 2–6 years band and was in-phase between 1920–1940, while the variability of SPEI lagged that of Niño3.4 in 1960 (Figure 9F).

In general, the 2–6 years oscillation is more robust, suggesting a strong inter-annual coherence between drought and ENSO, and this is consistent with findings by previous studies (Rodhe and Virji, 1976; Nicholson and Entekhabi, 1986). It is suggested that ENSO can produce anomalous diabatic heating/cooling over the western Pacific Ocean, creating a dynamically forced anomalous ascent/descent in the region, thus modifying the walker-type zonal circulation in the equatorial Indian Ocean (Goddard and Graham 1999; Mutai and Ward, 2000). Pohl and Camberlin (2011) also reported that the interannual variability of rainfall in all season except the boreal summer over EEA are strongly controlled by zonal wind shear between 850 mb and 150 mb level over the equatorial Indian Ocean, which was closely correlated to ENSO in all seasons.