Differential signal of change among multiple components of West African rainfall

Rainfall components likely differ in the magnitude and direction of their long-term changes for any given location, and some rainfall components may carry a greater regional signal of change than rainfall totals. This study evaluates the magnitude of change of multiple rainfall components relative to other components, and the greatest regions of change across all rainfall components in West Africa. Hourly rainfall data from the ERA5 reanalysis dataset was used to derive twelve rainfall components, which were evaluated for long-term means, interannual variability, and long-term changes. For rainfall totals and rainfall intensity, the central Sahel is witnessing increasing trends while the western Sahel is experiencing significant decreasing trends. In general, decreasing trends predominate in the study domain, especially in the northwestern Congo Basin, where annual rainfall is decreasing by 120 mm per decade. Importantly, rainfall frequency accounts for 49% of all significant grid-point trends for the whole domain. In contrast, rainfall totals account for 26% of all combined significant trends across the domain, while rainfall intensity (12.6%), rainy season length (9.5%), and seasonality (3.3%) account for the remaining signals of change. Most of the changes among the rainfall components are in the tropical wet and dry regions (59% of all significant trends); the Saharan and equatorial regions account for the least changes. This study finds evidence that rainfall frequency is changing more across the regions compared to rainfall totals and should be explored as rainfall inputs in climate models to potentially improve regional predictions of future rainfall.


Introduction
Studies of tropical climate variability and change have traditionally been carried out with a disproportional focus on rainfall totals (Trenberth et al. 2007;Feng et al. 2013) or other single rainfall 'components.' However, tropical rainfall can be disaggregated into a whole range of different components, e.g., rainfall frequency, intensity, timing, seasonality, zones, etc., (Olaniran 2002), in addition to just totals, all of which may be impacted as a manifestation of global-scale climatic change. Analyzing rainfall totals is important in understanding seasonal rainfall distribution, but it is often a quick way of conceptualizing rainfall as a climate variable to permit easy regional comparison. Also, rainfall totals are the major rainfall parameters in climate models. However, since tropical rainfall can be disaggregated into several different components, the narrow focus on one of them does not provide a full picture of the potential changes to the tropical climate over the past four decades in response to both external forcing and internal climate variability.
The IPCC (Stocker 2014) reported that there was no significant trend in total tropical rainfall for the 1901-2008 and 1951-2008 period because decreasing trends in the 1970s-1990s had been offset by increasing trends in the last couple of decades (Trewin 2014), especially in the Sahel. It then becomes pertinent to ask: Is this outcome peculiar to rainfall totals alone, or have other rainfall components shown similar trends? How have other rainfall components (besides rainfall totals) changed in the studied period? Without considering other rainfall components, we risk missing aspects of the changing tropical rainfall that might be more important than totals. More importantly, the responses of many systems are more critically impacted by these other rainfall components. For example, the timing of rainfall is more crucial to West Africa's rain-fed agricultural sector than rainfall totals (Obarein and Amanambu 2019;Adefolalu 1986), and the frequency and intensity of rainfall have a greater influence on floods than rainfall totals. Furthermore, there is some evidence in the literature that rainfall totals are often less sensitive to climate variability compared to other rainfall components (e.g., Alpert et al. 2002), and any two rainfall components may show different trends (in both direction and magnitude) for the same location and period (Lodoun et al. 2013). Therefore, it is likely that other rainfall components are more sensitive (or at least differ in their respective sensitivities) to climate variability and change than rainfall totals. These relative climate sensitivities have not been explicitly investigated. Detailed estimates of past tropical rainfall changes and accurate projections of future tropical rainfall changes should be cognizant of all disaggregated tropical rainfall components, without which such projections will be incomplete.
This study, therefore, sought to evaluate the magnitude of change of each disaggregated rainfall component relative to all other components, and the locations of greatest change across all the different rainfall components in West Africa. This study also explores ENSO teleconnection with the rainfall components because, while the lagged impact of ENSO on West African rainfall is well documented, this teleconnection has been mostly linked to rainfall totals.
The choice of West Africa is not arbitrary because its many rainfall regions are arguably more representative of tropics-wide conditions than most tropical sub-regions and thus ideal for this study. Also, West Africa is one of the most vulnerable regions to climate change-induced rainfall extremes (Aryeetey-Attoh 1997; Niang et al. 2014;Yaro and Hesselberg 2016). The Sahel, in particular, continues to be the focus of research in tropical rainfall climatology because of its sensitivity to substantial multi-decadal variability (Akinsanola and Zhou 2019;Mohino et al. 2011;Owusu and Waylen 2009;Mahe et al. 2001). The Sahelian drought of the 1970s and 1980s, which was largely responsible for the reduction in global land precipitation (Dore 2005), was preceded by the wet decades of the 1950s and 1960s (Bichet & Diedhiou, 2018), and followed by the recovery in the 1990s. Current estimates of future West African rainfall-mostly of rainfall totals-are uncertain (Monerie et al. 2020a, b;Sylla et al. 2016;Niang et al. 2014). Therefore, by unearthing evidence of a higher signal of rainfall change in other rainfall components (besides rainfall totals), the case can be made for including that in future estimates of West African rainfall.

Data and methods
The study domain covers West Africa (Fig. 1). The domain stretches from 1°N -21°N and 21°W -18°E, but with analyses performed on land-based grid-points over West Africa. Hourly rainfall from the ERA5 gridded reanalysis dataset developed by the European Centre for Medium-Range Weather Forecasts (ECMWF; Hoffmann et al. 2019) was used to derive the rainfall components/indices used in this study. The dataset has a spatial resolution of 0.25 o × 0.25° and covers 1979 to 2018. Reanalysis numerical weather prediction (NWP) models are fed by many familiar sources of observational climate data (such as ground-based stations, weather ships, radiosondes, and weather satellites). However, the ERA5 reanalysis is greatly improved through its assimilation of observations from the Advanced Scatterometer (ASCAT), the Microwave Humidity Sounder (MWHS), and many other recent sophisticated weather satellite instruments (Dee and NCAR 2020;Dee et al. 2016). This improved data assimilation capacity explains why the ERA5 has a better spatial and temporal resolution and has more parameters than the ERA-Interim (Hoffmann et al. 2019;Urraca et al. 2018; Olauson 2018).

Estimating long-term trends and variability
The method of derivation for each rainfall component is summarized in Table 1. The long-term variability of each of the indices was evaluated using the median absolute deviation (MAD) statistic. The MAD is computed by first finding the median, then calculating the absolute deviations from the median, and computing the median of these absolute deviations. The commonly used standard deviation statistic is problematic because both the standard deviation and the mean are sensitive to outliers. MAD has been shown to circumvent this problem and is ideal for estimating statistical dispersion in a climatic time series (Leys et al. 2013;Kwak and Kim 2017).
At each grid point, long-term trend detection was done using the Theil-Sen slope estimator. It is a method for fitting a line to a set of points that chooses the median slope among all lines connecting all possible pairs of two-dimensional sample points (Sen 1968;Theil 1950). It is insensitive to outliers and is more robust than the simple linear regression method of least squares, especially in datasets that are not normally distributed (Ohlson and Kim 2015). The Theil-Sen estimator is becoming more popular in detecting monotonic trends in univariate time series, such as climatic datasets.
Statistical significance of trends was estimated using 1000 bootstrap resamples of the Theil-Sen slope estimation to derive 95% confidence intervals. Confidence intervals were calculated based on MATLAB's biascorrected accelerated percentile method and converted to p-values using the technique outlined in Altman and Bland (2011). To tackle potential misinterpretation of statistical significance arising from multiple statistical tests at many spatially autocorrelated grid points, we controlled for the false detection ratio (FDR) as suggested by Wilks (2016). Since bootstrapping and Theil-Sen are quite computationally expensive, simple linear regression was used to save computational time in computing dailylevel rainfall trend estimates. To visualize decadal-scale changes in rainfall components, the Theil-Sen slope estimates have been multiplied by 9 (n-1; where n is 10, the number of years in a decade).
Since diverse definitions of rainfall timing currently exist for West Africa (see Fitzpatrick et al., 2015;Bombardi et al. 2019), this study compared the result of the trend analysis of Ilesanmi's (1972) method with the method advanced by the Famine and Early Warning System (FEWS) and described in AGRHYMET (1996). They describe onset as any one dekad with 25 mm of rainfall, followed by two dekads with at least 20 mm of rain. Cessation defined as any day after 01 September with 20 consecutive days of no rainfall (Salack et al. 2011;Hachigonta et al. 2008).

Estimating relative changes using standardized z-score values
To better quantify the magnitude of change of each rainfall component relative to other components (i.e., to examine which rainfall components are changing the most and where), we must transform the slopes (at each grid-point for all rainfall indices) into equivalent units. A modified 'robust' Z score formula (Eq. 2) was used (see Choudhury et al., 2017 andSrivastava 2017), which simply replaces the mean and standard deviation in the well-known standardization procedure (Eq. 1) with the median slope value and MAD of the set of slopes (across space) for each rainfall component, respectively. Since the aim is to quantify the spatial distribution of relative changes (i.e., relative to the other components), the sign of Z is less important (and potentially misleading) than the absolute values of Z-which were thus used in the results below.
where x, − x, and s are the sample, the sample mean, and the sample standard deviation respectively where x, M, and MAD are the slope values, median slope value, and MAD, respectively.
(1) Z = x − x S ; (2) Relative changes were also quantified by summing the count of all significant grid-point trends in each rainfall component.

Relationship with ENSO
The time series of the different rainfall components were analyzed for associations with ENSO signals. To evaluate possible ENSO signals on each tropical rainfall component, first, NOAA's Oceanic Niño Index (ONI) (which is NOAA's primary indicator for monitoring ENSO (Dahlman 2016) is used to quantify ENSO activity. The ONI is a 3-month moving average SST anomaly for the Niño 3.4 region. An annual ONI was computed that classifies each season into one of three groups: Neutral event, La Nina event, or El Niño event (Supplementary material, Table S1). Based on NOAA's operational definition of the ONI, a year was designated an El Niño year if it has five consecutive 3-month moving average SST anomalies at or above + 0.5 °C. A La Nina year would be one with five consecutive 3-month moving average SST anomalies at or below − 0.5 °C. A neutral year does not satisfy these conditions. The annual ONI was computed by averaging all 3-month moving average SST anomalies in each ENSO season, starting in the boreal winter.
Typically, ENSO-rainfall teleconnection is investigated by making the ENSO indices lag rainfall indices by 3-9 months (e.g., Xu et al. 2004;Manatsa et al. 2017;Lee et al. 2020, among others). Wang et al. (2020) and Camargo and Sobel (2005) found that a lag time of six months The annual count of qualified rain-days (defined as those exceeding 0.25 mm) with daily rainfall amounts at or below the 10 th percentile of all qualified rain-days Moderate rain-days (days/year): The annual count of qualified rain-days (defined as those exceeding 0.25 mm) has daily rainfall amounts between the 10 th and 90 th percentile of all qualified rain-days Heavy rain-days (days/year): The annual count of qualified rain-days (defined as those exceeding 0.25 mm) with daily rainfall amounts at or above the 90 th percentile of all qualified rain-days Mean daily hours of rainfall (MDHR) per year (hr/yr): Daily rainfall hours summed for each year and averaged Rainfall intensity Rainfall Intensity (mm/year): The ratio of the total annual rainfall to the annual number of days of rainfall Rainfall timing Defined using the mean cumulative rainfall method by Ilesanmi (1972): the mean annual rainfall ) that occurs at each 5-day interval (a pentad) and expressing each pentad rainfall as a percentage of the total annual rainfall. This percentage is cumulated for each year and across grid-points Onset (Julian days): The 8% cumulated rainfall Cessation (Julian days): The 90% cumulated rainfall Rainy season length (RSL) (Julian days): the intervening days between the onset and cessation dates Rainfall seasonality Seasonality Index (SI): The sum of the absolute deviations of mean monthly rainfalls from the overall monthly mean, divided by the mean annual rainfall (Walsh & Lawler 1981) produced the strongest correlation between ENSO and SSTAs in the Bohai Sea and the Yellow Sea, and between ENSO and tropical cyclone activities, respectively. However, because the situation may be different for West Africa, this research correlated rainfall indices with the annual ONI of each ENSO season, using the Pearson's moment correlation coefficient, r, effectively minimizing any significant difference in correlation arising from using different lag times.
p-values were calculated from the correlation coefficients to test the hypothesis of no correlation against the alternative hypothesis of a non-zero correlation (with α = 0.05).

Rainfall regionalization
Theoretically, rainfall zones-regions of homogenous rainfall characteristics-can be defined using any one of the tropical rainfall components. However, only rainfall totals are used to determine rainfall zones herein to stay within the convention and avoid unnecessary complexity. The main rainfall regions were defined using k-means based regionalization of the entire 40-year dataset. This method begins by randomly placing a predetermined number of k centroids in space based on the number of clusters required. The next step involves computing the distance between each grid-point and a centroid and assigning a grid-point to the closest centroid. This first iteration creates k number of clusters, but on the next iteration, the centroids shift to occupy a new position in space based on the average of all the members in the cluster. Some grid points may change clusters if it becomes closest to a new centroid as a cluster centroid shift. Iterations are performed until no grid-point changes cluster membership (Hannachi 2004;Carvalho et al. 2016). K-means was applied to the 12 averaged (over 40 years) monthly mean rainfall amounts for each of the 9066 grid points, and each grid point was classified into one of 5 clusters/regions. No consistent standard zonation of West African rainfall exists in the literature. Cluster evaluation metrics yielded varying results, so five clusters were chosen, as this solution yielded five rainfall zones, with distinctive mean annual rainfall totals, mean monthly totals, and mean annual regimes (Fig. 2). Clusters 1 and 5, representing the Sahara Desert zone and the rainy equatorial southern coastal zone, respectively, have the most contrasting rainfall characteristics and are two of the most easily recognizable rainfall zones. The Sahelian zone (cluster 4) has received considerable attention because of its well-documented multi-decadal rainfall oscillations and fragile environment (Dong and Sutton 2015;Giannini et al., 2013;Bichet and Diedhiou 2018). As cluster 4 shows, most of the rainfall in this zone occurs in the boreal summer, with a peak in August (Nicholson et al. 2018). Clusters 2 and 3 have also featured prominently in previous literature as the other three regions described above, with the terms 'Guinean' and 'Sudanian', used to describe them, respectively (Akumaga and Tarhule 2018;Camberlin et al. 2020), possibly due to the vegetation class in these regions. Cluster 2 is closer to the equatorial zone and so receives more total rainfall than cluster 3, whose northern boundary is the Sahelian zone. However, the most contrasting feature of these two regions is their annual cycle, clearly depicted in Fig. 2a. The Tropical Wet region (cluster 2) has a bimodal annual rainfall distribution because the ITCZ makes two overhead migrations as it moves northward and southward during the West African Monsoon (Diallo et al. 2016). By contrast, the tropical dry region (cluster 3) has a unimodal annual rainfall distribution (Nkrumah et al. 2019). In all, these five clusters capture most of the known spatial variability in total rainfall characteristics across West Africa.

Validating the ERA5 precipitation dataset
For all the benefits of reanalyses datasets, legitimate concern exists in the climate science community about the accuracy with which reanalysis datasets reproduce atmospheric variables, especially precipitation, on local to regional scales.
There are also concerns about the use of reanalysis data for trend analysis. However, reanalysis data are the only source of complete long-term data at regular temporal and spatial intervals in West Africa and other regions that suffer from a sparse network of ground-based observations. Reanalyses are a necessity for many climate studies in these areas. In all the evaluations of the ERA5 reanalysis reviewed for this research, ERA5 outperformed previous reanalysis products (see Zhang et al. 2019;Delhasse et al. 2019, andMahto andMishra 2019). And although validations of ERA5 precipitation data for Africa relative to the Climate Hazards Group InfraRed Precipitation with Stations version 2 (CHIRPS v2) yielded significant improvements over ERA-Interim (Gleixner et al. 2020), no such validations exist for the West African region specifically. In this section, the ERA5 precipitation data is validated against a diverse set of precipitation products: ground observation weather stations (the full list of stations is in Supplementary Table S2), Tropical Rainfall Measuring Mission (TRMM) data, and Tropical Applications of Meteorology using Satellite and ground-based observations (TAMSAT) data. Information on the time span, spatial resolution, and data repository are summarized in Table 6. All validation data are monthly rainfall totals. The ERA5 data, which is monthly averaged precipitation, was resampled to the grid size of each validation data to facilitate grid-by-grid validation of raw data. Mean monthly precipitation and precipitation variability were also evaluated over the entire study area. The MAD was used to measure variability, and all validation was done by correlating both sets of data with the Pearson product-moment correlation coefficient.

Results and discussion
The results and discussion section will focus on the unique rainfall components and indices out of the 13 that were evaluated. The long-term means, variability, and trends of all indices are shown in Supplementary Figures S1, S2, S3, and S4, respectively. Supplementary Table S3 and S4 presents a breakdown of regional trends for all 13 indices as decadal changes and average percentages. Finally, Supplementary Tables S5 and S6 show the regional percentage of grid-point trends for all rainfall indices and ENSO association with all rainfall indices.
Daily, annual, and JJAS rainfall totals show very similar long-term means and trends. All rainfall frequency indices have similar long-term means, with moderate rainfall and total rain-days having similar long-term patterns. The rainy season length (RSL) is a function of the onset and cessation of rainfall. Therefore, this study will focus primarily on five rainfall components/indices: annual rainfall, total rainfall days, rainfall intensity, rainfall seasonality, and rainy season length.

Long-term means of rainfall components
There is an observable latitudinal variation in long-term means across all rainfall components in the study domain (Fig. 3). Total rain-days and rainfall seasonality ( Fig. 3b and d) surprisingly show more glaring latitudinal rainfall distribution than rainfall totals and rainy-season length (RSL) (Fig. 3a and e), which have been traditionally used to describe the latitudinal distribution of the West African rainfall regime (e.g., Hagos and Cook 2007;Trewin 2014). The strength of the latitudinal distribution in rainfall frequency and seasonality is seen in the continuous east-west band of rainfall characteristics. Unlike rainfall totals, rainfall frequency and seasonality appear insensitive to localized increases in rainfall in high elevation areas. The latitudinal discontinuity in the distribution of total rainfall and RSL is seen in highland areas and primarily at the coast where relative dryness occurs-which breaks the coastal band of Fig. 3 Long-term means of a annual rainfall, b annual number of rain days, c annual rainfall intensity, d seasonality, e RSL (1979RSL ( -2018. Note the different scale for each plot heavy rainfall between the coast of Cameroon through the Niger Delta to the coast of Liberia and Sierra Leone, where mean annual rainfall is at least 3000 mm. Still, the latitudinal distribution of rainfall totals is a prominent feature of West African climate and tropical climates found, to varying degrees, elsewhere, such as central Africa and the Indonesian/Australian region (Trewin 2014). This feature involves a transition from the heaviest rainfall zone, outside the equator, to a vast zone of relatively lesser rainfall and then to a dry zone-often a desert. The rainy coastal zone of West Africa occurs near 4-5°N latitude, and from this zone, rainfall decreases with distance away from the coast, northward up to the Saharan region.
Unlike rainfall totals and RSL, there is little reference to rainfall frequency and seasonality as rainfall components that exhibit a precise latitudinal distribution. The earliest attempts at regionalizing West African rainfall only considered rainfall totals and RSL (Nicholson 1979(Nicholson , 1980(Nicholson , 1993 This contrast in capturing localized rainfall disparities gives an insight into the climatological difference and similarity between rainfall totals on the one hand ( Fig. 3a) and rainfall frequency and seasonality ( Fig. 3b and d), on the other. It is possible that failing to capture localized orographic rainfall, rainfall frequency, and seasonality heavily reflects midtropospheric rainfall mechanisms that drive the development of mesoscale convective systems (MCSs), and thus of rainfall (Akinsanola and Zhou 2019;Nicholson 2009;Cretat et al. 2015). Statistically and visually, total rain-days appear slightly independent of rainfall totals because more rainfall occurrences do not always translate into more rainfall totals: the correlation between spatially averaged mean annual rainfall and rain-days is 0.688.
North of 15°N, there is an almost total spatial uniformity in mean distribution among the rainfall components because of the extensively dry expanse of the Sahara Desert. With a mean annual rainfall total of less than 3 mm, parts of the Sahara are the driest places on Earth (Kelley 2014;Ghoneim and El-Baz 2020). RSL, rainfall frequency, and rainfall intensity are also the least in this region, and the rainfall here is very seasonal because it is limited to 1-2 boreal summer months.

Interannual Variability in Rainfall Components
Rainfall in dry tropical areas is marked by high interannual variability (relative to the mean) compared to wetter parts of the tropics where the higher rainfall tends to be more reliable (Nieuwolt 1982;Ayoade, 2004). Figure 4 displays the spatial pattern of normalized interannual variabilities among West African rainfall components. To calculate this, the MADs are divided by the averages of the rainfall component for each grid point. The MADs are normalized because the spatial distribution of MADs is skewed by those grid points with higher values of rainfall characteristics.
The most variability is in the dry Sahara Desert, particularly the eastern portion, the driest in West Africa. The interannual variation in rainfall totals, frequency, intensity, and seasonality from the Sahara decreases, almost latitudinally, southwards to the coast. Rainfall intensity (Fig. 4c) displays the most latitudinal, north-south decrease in rainfall variation, while rainfall seasonality (Fig. 4d) displays the least latitudinal variation. In contrast to long-term means, the interannual variability in rainfall frequency and intensity south of 15 o N appears constant, compared to the dry north, where there is more considerable variation.
In general, interannual variability in West Africa is higher than interannual trends and even decadal trends for all rainfall components (Supplementary Figure S3). Globally, twentieth-century decadal trends in rainfall totals also tend to be lower than interannual variability (New et al. 2001). Figure 2b shows the result of the K-means cluster analysis of rainfall totals over the study domain for the 40 years of the study. Five distinct rainfall regions have been produced, based on the mean monthly rainfall of each grid point. These five regions capture almost all the spatial variability observed over the decades in this region. The regionalization also captured the whole spectrum of the mean annual rainfall cycle (Fig. 2a) characteristic of West Africa. For all regions, the peak of the WAM is in the boreal summer which lasts from June to September.

Regionalization of rainfall totals in West Africa
In addition to the latitudinal distribution of the rainfall regions, the regionalization also depicts the necessary zonal discontinuity that captures the actual spatial variation of the rainfall regime in West Africa. Only the Saharan, Sahelian, and tropical dry regions stretch unbroken from east to west. Still, longitudinal bands of more homogenous rainfall characteristics have been identified in these regions, as seen in previous research (Nicholson 1993;Klutse et al. 2016). Overall, the drier, northern rainfall regions are more spatially contiguous than the wetter, southern regions. A tabulation of some rainfall characteristics in each rainfall region is presented in Table 2.

Long-term changes in rainfall components
The long-term changes in rainfall components are mapped as decadal changes in Fig. 5. Regionally-averaged decadal changes for each rainfall component are summarized in Table 3. Perhaps one of the most consistent and clearest signals of long-term changes among West African rainfall components is the contrasting trend between the central and western Sahel for annual rainfall totals (Fig. 5a) and rainfall intensity (Fig. 5c). Previous studies have reported this zonal contrast (e.g., Lebel and Ali 2009;Nicholson et al., 2018;among others). On average, annual rainfall totals are increasing by 40 mm per decade in the central Sahel and decreasing by 120 mm per decade in the western Sahel-these decreases are statistically significant.
Very little research has sought to proffer an explanation for these contrasting trends in western and central Sahelian   . Accounting for only internal climate variability, they found that the zonal contrast is associated with northern African surface warming and sea-surface temperature anomalies in the Pacific and tropical Atlantic oceans. In Getani et al.
(2017), forcing the CMIP5 with higher CO 2 concentrations produces wetness in the Sahel, but with a stronger response in the central Sahel, whereas SST warming produces dryness in the Sahel, with a stronger response in the western Sahel. These two competing and contradictory responses may collectively (rather than individually) account for the observed zonal contrast in Sahelian rainfall trends. The literature provides some evidence about the likely time when the central/western Sahelian rainfall trends contrast began to occur. The year 1994 was reported as the  wettest year in the Sahel since the drought began in the 1970s (Nicholson et al. 1996;LeCompte et al. 1994), and Nicholson et al. (1996) found that most of those positive rainfall anomalies occurred in the central Sahel, with little increases in the western Sahel. In addition, these spatial anomalies were limited to the boreal summer and continue to drive the increases in central Sahelian rainfall, in agreement with Nicholson et al. (1996), Sanogo et al. (2015), and Biasutti (2013). Vegetative data analysis by Anyamba and Tucker (2005) also explicitly puts the start of the recovery of Sahelian rainfall in 1994, given the uptick in vegetation in that year. The contrasting trends between the western and central Sahel for annual rainfall totals and summer rainfall have likely grown stronger over time, and some studies now predict that it would persist to the end of the twenty-first century (Monerie et al. 2017). The increasing trend in the Guinea coast and other coastal regions (for rainfall totals, intensity, and seasonality; Fig. 5a, c, and d) is surprisingly not experienced in other highland areas that make up the equatorial zone. This difference suggests that increasing trends in these coastal areas maybe feedbacks from a warming adjacent ocean, as precipitation is sensitive to local SST changes (He et al. 2018;Balas et al. 2007). Nicholson et al. (2018) and Jin and Huo (2018) have found that the Guinea coast of West Africa showed a very high correlation between rainfall and local SSTs in the JAS season for the 1969-2014 period compared to other coastal areas.
Furthermore, widespread decreasing trends exist across all rainfall components. One area of concern is the northwestern Congo Basin, experiencing significant decreases in rainfall totals, intensity, and seasonality (Fig. 5a, c, and d). Annual rainfall in the northwestern Congo Basin is decreasing at 120 mm per decade, and total rain-days are decreasing by up to 2% (relative to their mean) per decade. The rainfall decline in the world's second-largest rainforest is well documented from empirical data on long-term rainfall totals (Malhi and Wright 2004;Asefi-Najafabady and Saatchi 2013;Hua et al. 2016;Jiang et al. 2019;and others), and observations of rainforest greenness (Zhou et al. 2014). Warm air masses in tropical regions have become more frequent, in lockstep with increases in air temperatures (Lee 2020), but the increased absolute water vapor content of humid air has not led to increased rainfall in the Congo Basin and in many tropical rainforests (Malhi and Wright 2004). Other studies have also noted decreasing trends in cloud cover in the Congo Basin (Lee 2020), which may also contribute to declines in precipitation. Jiang et al. (2019) and Zhou et al. (2014) have drawn attention to the observed massive intensification of the dry season, which has potentially created a positive feedback loop of vegetation drying and drought. Long-term drying can increase the land evaporative demand, reduce cloud cover, and increase aridity; and the deforestation in the Congo Basin only helps amplify the drying (Nogherotto et al., 2013). This study finds evidence of increased dry season duration in the shrinking of the RSL (by 10 days per decade [ Fig. 5e]).
The widespread decreasing trend in rainfall totals and rainfall frequency in West Africa may be part of a larger drying signal in tropical land areas, particularly in the northern hemisphere tropics where this study domain lies (Neelin et al. 2006;Zhang et al. 2007). Malhi and Wright (2004) have reported precipitation declines in tropical rainforest regions, mainly driven by anthropogenic forcings-mostly logging and agricultural clearing. Attribution science in tropical rainfall changes is still somewhat fuzzy, and there is some consensus about the cause of the drying in the tropics; however, it is possible that the role of natural forcings or internal climate variability-such as SST anomalies and ENSO-have become relatively less prominent, especially in the last few decades as anthropogenic influences have become stronger (IPCC, 2021). Still, the oceans play a crucial role in modulating tropical rainfall and even causing sustained summer drying over land. For example, Hua et al. (2016) reported that Indo-pacific SST teleconnection with central African rainfall was the likely cause of the drying in the Congo Basin. The Congo Basin, regardless of the absence of any consensus on its drying, remains a major convection hotspot that influences regional climate by modulating the energy balance and hydrological cycle, and its future change will have significant regional climatic implications.
Importantly, the large-scale drying in the tropics seems to contradict the widely held paradigm that dry areas will get drier and wet areas will get wetter. Chadwick et al. (2013), Greve et al. (2014;[1948-2005), and (Feng and Zhang 2015[1979 found that a robust "dry gets drier and wet gets wetter" pattern occurred in only 10.8% and 15.12% of global land areas respectively: the vast portion of humid tropical Africa, especially the Congo Basin and the majority of the West African domain, are getting drier. The "dry gets drier and wet gets wetter" pattern is more applicable to patterns of precipitation trends over the ocean (Hu et al. 2019;Greve et al. 2014); it is unlikely that land surfaces will follow a simplistic climate change pattern given the complex landvegetation-climate feedback system, some of which have been mentioned above.
As mentioned above, RSL is shrinking, driven by delayed onset and earlier-than-normal cessation ( Supplementary Figure S4). The delayed onset of the rainfall is consistent with African-wide trends reported in Dunning et al. (2018). The RSL is generally more dependent on the onset than on the cessation (Obarein and Amanambu, 2019), and a strong correlation between the former and the latter has been reported in some studies in Africa (Kanemasu et al. 1990;Mubvuma 2013;Mugalavai 2008). This interdependency explains why a delayed onset is observed with a shorter rainy season in West Africa.
Averaged by rainfall regions (Table 3), most of the trends are negative and statistically significant, an unsurprising result given the broad swaths of statistically significant decreasing trends across the study domain in most rainfall indices. Areas of increasing trends are too localized to sway the general regional trend away from more widespread decreases in rainfall components. It is noteworthy that the Saharan region has no statistically significant trend for any rainfall components. Oppositely, the tropical dry and wet regions are seeing the most significant trends.
Overall, the rainy south has witnessed more changes than the dry north. The trends in rainfall totals and rainfall frequency are broadly corroborated by the IPCC's (Niang et al. 2014) estimates: statistically significant decreasing trends in most of the domain and non-statistically significant increasing trends mainly in the central Sahel and the Guinea Coast. The IPCC projects that extreme rainfall days will increase by the end of the century, but the projections are low to medium confidence.
The results of the long-term means, interannual variability, and long-term trends of rainfall timing, using the Aghrymet-FEWS method, are mapped in Supplementary Figure S6, S7, and S8. Although many grid points and years in the dry Sahara and wet equatorial regions did not meet the required threshold (as is the case with many agrometeorological definitions [Marteau et al. 2009]), there was broad agreement between both methods. For example, rainfall onset and RSL in both methods are trending in the same direction, with their magnitude of changes also similar.

Relative magnitude and spatial patterns of changes among rainfall components
The relative changes in all rainfall components/indices, estimated by taking the absolute z-scores of the trends, are shown in Fig. 6. Another measure of the relative magnitude of changes among the rainfall components is the percentage count of total significant trends (on a grid-point basis) in each rainfall component, subdivided by region (Table 4 and Fig. 7). Both analyses show that the rainfall frequency (measured by total rain-days) has witnessed the greatest significant changes, accounting for almost 50% of all combined significant trends, relative to the other four components. Annual rainfall total is the second greatest category of changing rainfall components (Fig. 6b), accounting for 26% of total significant trends. In contrast, RSL (9.5%), rainfall intensity (12.6%), and rainfall seasonality (3.3%) account for the least changes among the rainfall components in the last 40 years (Fig. 6l, h, and I; Table 4). Changes in moderate rainfall, total rain-days, and MDHR per year account for the greatest regional rainfall change signal in the study domain. These changes are mostly in the tropical wet and tropical dry regions and parts of the western Sahel (Fig. 7).
This finding is well supported by studies that have compared changes in rainfall frequency and rainfall intensity. Camberlin et al. (2009) and Moron et al. (2006), in their studies of East African and West African rainfall, respectively, found that rainfall frequency (total rain-days in particular) has a greater influence over regional-scale signals of interannual rainfall variability than rainfall intensity. Figure 8 maps the total number of statistically significant trends across all rainfall components for each grid point. In general, the south has undergone a bigger magnitude of change in the last 40 years compared to the north. In the south, two areas with the highest numbers of significant trends across all rainfall components are the northwestern Congo Basin and the broad western hinterlands that span the Tropical Wet and Tropical Dry regions. These two areas have changed significantly in almost all rainfall components, but this is not surprising because they correspond to the massive areas of statistically significant decreasing decadal trends across most rainfall components (Fig. 5). In contrast to the dominant areas of significant decreasing trend, the areas of an increasing trend across all the rainfall components (the central Sahel, central-eastern Saharan region, the Niger Delta, and coast of Liberia and Sierra Leone) are experiencing only mild changes overall, possibly because many of these increasing trends are not statistically significant.
Regionally, the tropical wet and dry regions account for 59% of all significant trends across all rainfall components. The equatorial (7%) and Saharan regions (12%) have the least number of significant trends across all rainfall components, while the Sahelian region accounts for 22% of all significant trends across all rainfall components (Table 4).
These analyses of the relative changes among the rainfall components support our hypothesis that disaggregated rainfall components are changing in different ways (with varying spatial distribution) in response to background warming because of differential sensitivity to climate dynamics (and possibly physiographic changes, e.g., deforestation). It also provides evidence that changes in rainfall totals are not necessarily the largest signals of rainfall change in West Africa.

Relationship with ENSO
Since different rainfall components possess varying spatio-temporal trends, it is expected that each rainfall component will interact differently with ENSO, and this is explored here. The correlation analysis result to measure the association between ENSO and rainfall components in the study domain is shown in Fig. 9. The correlation coefficients are generally high, ranging on average from − 0.57 to + 0.57. The dominant spatial correlation pattern occurs in moderate rain-days, total rain-days, MDHR per year, and RSL. This pattern depicts a broad southwestward diagonal band of positive correlation coefficients, with clusters of significant grid points in the coastal and the Sahelian regions. For the rainfall components mentioned above, negative correlation coefficients are found in the western Sahel, the Niger Delta, and the vast Tropical Wet region of the northwestern Congo Basin. In all, there is a Fig. 6 Absolute z-scores of the trends in a daily rainfall, b annual rainfall, c light rainfall, d heavy rainfall, e moderate rainfall, f annual number of rain-days, g MDHR per year, h annual rainfall intensity, i mean seasonality, j rainfall onset dates, k rainfall cessation dates, l RSL. 1979-2018 Table 4 Percentage of the total significant grid-points (trends) for rainfall indices in each region. The two highest percentages and the two lowest percentages are in red, boldface font and blue, boldface font, respectively considerable difference in the correlation between ENSO and these rainfall components for the western and eastern/central Sahel. Indeed, for all the rainfall components studied, the most consistent spatial correlation pattern is a dichotomy of ENSO teleconnection between the western Sahel and the eastern/central Sahel. Furthermore, the correlation coefficient pattern for annual rainfall totals is similar to that for annual rainfall intensity and heavy rain days. This pattern features a relatively strong positive correlation with ENSO in the eastern/central Sahel and weak positive to negative correlation coefficients in the western Sahel, with clusters of significant grid points in both regions. Because the height of the WAM is in the boreal summer, the correlation between ENSO and JJAS rainfall was analyzed (Fig. 9a), and they show similar results to that in annual rainfall.
ENSO anomalies generally correlate well with global temperate changes, whereby El Niño leads to global positive temperature anomalies while La Nina leads to negative global temperature anomalies. However, ENSO correlation with rainfall is much more complex and nonlinear. Increases in global precipitation with El Niño and decreases with La Nina have been observed (Adler et al., 2017) over local to regional scales, but ENSO signals over larger areas, such as the tropics or globally, are very weak (Wang et al. 2018;Gu and Adler 2011;Su and Neelin, 2003). The correlation of ENSO with spatially averaged annual rainfall for the whole domain is very weak (− 0.068 [not shown]) because of the combination of positive and negative correlations in the study domain. Even for all five regions, ENSO teleconnection is weak across all rainfall components-except for rainfall onset and cessation in the Tropical Wet region ( Fig. 9; Table 5). There are, however, sub-regional strong positive and negative ENSO correlations, especially in the central and western Sahel, parts of the Guinea Coast, and parts of the northwestern Congo Basin. The central Sahel has the most consistently strong and positive correlation with ENSO across all rainfall components, except rainfall onset and cessation. This result is corroborated by studies that have linked summer Fig. 9 Association between ENSO and a JJAS rainfall, b annual rainfall, c light rainfall, d heavy rainfall, e moderate rainfall, f annual number of rain-days, g annual rainfall intensity, h MDHR per year, i mean seasonality, j rainfall onset dates, k rainfall cessation dates, l RSL (1979RSL ( -2018. Black stippling marks significant (p < 0.05) gridpoint trends Sahelian rainfall (which is increasing in the central Sahel) to ENSO (Janicot et al. 1996;.
The strength of central Sahelian teleconnection with ENSO is part of two domain-wide patterns of correlation: (1) for rainfall totals (annual and JJAS rainfall) and rainfall intensity-which have both been found to show a similar pattern of long-term trends, and (2) for rainfall frequency (moderate rainfall, total rain-days and MDHR per year). The similarity in the correlation patterns among the rainfall frequency indices is interesting because all three have the highest significant grid points and are driving most of the changes happening in West Africa.

Socio-economic implication of findings for West Africa
In all, West Africa remains one of the more vulnerable regions to climate change-induced rainfall extremes. West Africa is still primarily a rain-fed agrarian economy and is thus more susceptible to the interannual fluctuations in rainfall frequency and rainfall totals. Drought and food insecurity in parts of the Sahel have led to civil and social unrest, like in the Horn of Africa. Yaro and Hesselberg (2016), drawing from IPCC's regional climate change report for West Africa and Africa (Niang et al. 2014;Riede et al. 2016), have detailed several mitigation and adaptation options available to West African governmental agencies and other stakeholders in combating the widespread drying and potential localized flooding in the region. But while policymakers should focus on rainfall frequency and rainfall totals, attention must be drawn to the practical role that other rainfall components such as rainfall timing and rainfall seasonality play in the livelihood of West African people. It has been shown that the timing of the rainy season is more important for a productive growing season than the rainfall amount. A year with normal (or above normal) rainfall totals but marked by delayed onset and/or premature cessation of the rainy season is worse for plants than one with a substantial shortfall in total amount but with normal dates of onset and cessation.

ERA5 West Africa precipitation data validation
Supplementary Table S2 provides the correlation between monthly ground observation station rainfall data and monthly ERA5 data. In all stations, the correlation coefficient is very high, and in 95% of the stations, the Spearman rank correlation coefficient exceeds 0.8. Furthermore, the correlation of mean precipitation and variability between ERA5 and the average for all stations are 0.95 and 0.75, respectively. TAMSAT monthly rainfall contains significant missing data, nonetheless, monthly ERA5 precipitation is very high and positively correlated with monthly TRMM and TAMSAT rainfall across most of West Africa (Figs. 10 and 11). For both validation rainfall data, the correlation with ERA5 is more than 0.8 in more than 80% of the study area ( Table 6). The very sparse network of ground stations may explain the low to negative correlations in the Sahara Desert.

4. Summary and conclusions
The overarching aim of this research is to examine the multidecadal trends in various components of West African rainfall. Several studies have been conducted that held implications that rainfall totals alone may not carry the largest signal of rainfall changes from climate change, but most of these have not been explicitly detailed. Some prior research has either reported two oppositely trending rainfall components or in the case of IPCC (2014), no significant trend in tropical rainfall totals-all of which imply that different rainfall components may vary considerably in terms of the magnitude of their contribution to the total signal of rainfall change in a region. Among other things, this study has sought to quantify these differences in the magnitude of change among the disaggregated rainfall components. Also, the sub-regional or spatial distribution of these signals of change has been examined, with the goal of not only providing information about the magnitude of change among the rainfall components, but also shedding light on the locations of greatest (and least) change in each rainfall component, and then across all the rainfall components collectively. Individually and collectively, our analyses yielded the following results: 1. Rainfall frequency accounts for the greatest changes relative to all other rainfall components. 2. Rainfall totals and rainfall intensity are experiencing the second and third most changes, respectively. 3. Rainy season length (RSL) and rainfall seasonality, on the other hand, are seeing the least amount of change. 4. The tropical dry and tropical wet regions are experiencing the greatest changes in the entire study domain largely due to the changes in rainfall frequency and rainfall totals being concentrated in these two regions. 5. By contrast, the Saharan and equatorial regions are the least-changing regions; changes in rainfall totals and rainfall frequency are minimal in these areas.
The signs of these changes are mostly negative and may be part of a broader trend of drying in tropical land areas and a consensus from previous research that anthropogenic changes drive these trends. Anthropogenic activities like land use/land cover changes are the main driver of the two most prominent local signals of change found in this study (that are mostly consistent with a growing number of studies): 1. The contrasting trends in rainfall totals and heavy rainfall between the central Sahel and the western Sahel. 2. The massive decreasing trend (in all rainfall components except light rainfall and rainfall onset) in the northwestern Congo Basin. Fig. 11 Pearson productmoment correlation between monthly TAMSAT data (mm/ month) and monthly averaged ERA5 precipitation data (mm/ month). White spaces are missing data in the TAMSAT dataset The results of this study challenge the widespread acceptance of rainfall totals as the only representation of rainfall, especially on regional to large scales. Components of rainfall frequency, such as rain-days, which show daily occurrences rather than accumulated amount, are perhaps more robust representations of the regional rainfall climatology than rainfall totals. Rainfall frequency appears to accurately depict the migration of the tropical rain belt, compared to rainfall totals that show local variations in rainfall due to elevation (Fig. 3). Rainfall frequency likely captures MCSs or large-scale rainfall-producing systems better, which may have important implications for future rainfall predictions in the domain. CMIP5 estimates of future rainfall changes in West Africa have a low to medium confidence, with considerable inter-model variation (Niang et al. 2014), partially attributable to the wellknown inability of GCMs to resolve convective rainfall. The use of rainfall frequency in place of rainfall totals should be explored in future research to see how well it improves future estimates of rainfall changes in the West African region. Additionally, it would be interesting to explore the relative magnitude and direction of rainfall component changes in the broader tropics and temperate climates where precipitation plays a less prominent role. Interdependencies among the rainfall components can also be assessed at various spatial scales.