The Agoufou watershed (Fig. 1) is located in the Gourma, a region of northern Mali delimited by the Niger river to the north and the border with Burkina Faso to the south.This region has been extensively monitored by the AMMA-CATCH observatory (Analyse Multidisciplinaire de la Mousson Africaine – Couplage de l'Atmosphère Tropicale et du Cycle Hydrologique) and before by ILCA (International Livestock Centre for Africa) and IER (Institut d'Economie Rurale in Mali) providing historical data (Hiernaux et al., 2009b; Lebel et al., 2009; Mougin et al., 2009).

As elsewhere in the Sahel, the climate is tropical semiarid with a unimodal precipitation regime. The rainy season extends from late June to September, and is followed by a long dry season. Precipitation comes from tropical convective events, 25 to 50 per year, brought by the West-African monsoon (Frappart et al., 2009; Vischel and Lebel, 2007). Its long-term evolution has been characterized by a wet period between 1950 and 1970 followed by a long dry period with extreme droughts in 1972–1973 and again in 1983–1984. The last 15 years have shown a partial recovery of rainfall, with large events seemingly occurring more often (Frappart et al., 2009; Panthou et al., 2012). Average rainfall was 345 mm yr-1 over the 2000–2015 period and 382 mm yr-1 over the 1960–1975 period. Potential evaporation was much higher than precipitation, with averages of 3235 and 2930 mm yr-1 for the two periods, respectively.

The Agoufou watershed extends over 245 km2 and ranges between latitudes of 15.3 and 15.4∘ N and longitudes of 1.4 and 1.6∘ W. The Gourma region is endorheic, which means that it is a mosaic of closed drainage watersheds that does not provide outflow to the Niger river and thus to the Atlantic Ocean. The Agoufou lake is the outlet of the watershed. As with the majority of lakes and ponds in the region, it showed an important surface area increase over the last 50–60 years (Gal et al., 2016) and nowadays, typically reaches about 3 km2 at the end of the rainy season.

The geology of the watershed is characterized by Upper Precambrian schists and sandstones partially covered by staggered ferricrete surfaces (Grimaud et al., 2014), silt depositions and sand dunes. The study site has been extensively described in Gal et al. (2016). The northern part of the watershed (Fig. 1) consists of outcrops and shallow soils lying on sandstone, schist or iron pans. Some of these soils are fine textured soils (silt flats) which favor the frequent generation of runoff. The southern part is dominated by deep sandy soils with high infiltration capacity. The altitude range is 92 m, and the average slope of the main reach is equal to 0.22 %.

The vegetation is typical Sahelian vegetation with an herbaceous layer almost exclusively composed of annual plants, among which grasses dominate, plus scattered bushes, shrubs and low trees (Boudet, 1972; Hiernaux et al., 2009a, b). Almost continuous on sandy soils, except for a few deflation patches and bare dune crests, the herbaceous layer is highly discontinuous on shallow soils and clay plains, leaving large bare areas prone to runoff. The density and crown cover of woody plants are low in average, usually between 0 and 5 % (Hiernaux et al., 2009b). Woody plants concentrate along drainage lines, around ponds, in the interdune depressions and also sometimes on shallow soils, with a regular pattern of narrow linear thickets set perpendicular to the slope known as “tiger bush” (Hiernaux and Gérard, 1999; Leprun, 1992). These thickets live on the water and nutrients harvested on the impluvium made by the bare soil upstream, and their development efficiently limits runoff further downstream (D'Herbès and Valentin, 1997).

Casenave and Valentin (1989), among others, have demonstrated that the Sahelian hydrological processes are largely dependent on land surface conditions: soil properties, crusting, topography and vegetation cover (Albergel, 1987; Collinet, 1988; Dunne et al., 1991; Hernandez et al., 2000). Low soil infiltrability associated with the convective nature of the precipitation favors runoff generation by infiltration excess (Descroix et al., 2009, 2013; Leblanc et al., 2008; Peugeot et al., 2003) commonly known as Hortonian runoff.

The choice of this watershed has been motivated by three characteristics: (1) this site is instrumented by the SO-AMMA-CATCH, which provides field measurements on vegetation and soil characteristics, meteorological variables, and lake height estimates over a long period of time (starting in 1984 for the long-term ecological survey), as well as the good field knowledge of coauthors. (2) The Agoufou lake has experienced a spectacular increase in inflow over the past 60 years despite the decrease in precipitation (Gal et al. 2016), which is a very good example of the Sahelian paradox and of the evolution of surface water observed more generally in the Gourma region (see Gardelle et al., 2010) and elsewhere in the Sahel (Mauritania and Niger, Gal et al., 2016). (3) This site is a pastoral watershed where agricultural activity is almost nonexistent. It is thus different from the watersheds that were addressed by hydrological studies on the Sahelian paradox up to now. It can therefore shed a new light on the debate over land use versus land cover as possible explanation of the Sahelian paradox.

Gal (2016) carried out a literature review of 20 different hydrological models (global, distributed and semidistributed) in order to identify the most appropriate models to simulate hydrological processes in the study area and to meet the objectives of this study.

The kinematic runoff and erosion model (KINEROS-2) was considered the most suited among those analyzed. KINEROS-2 (K2; Goodrich et al., 2011; Semmens et al., 2008; Smith et al., 1995) is the second version of KINEROS (Woolhiser et al., 1990). It is an event-oriented physically based model describing the processes of infiltration, surface runoff, interception and erosion for small watersheds, and most of its applications concern arid and semiarid areas (Hernandez et al., 2005; Kepner et al., 2008; Lajili-Ghezal, 2004; Mansouri et al., 2001; Miller et al., 2002). The surface runoff simulation is based on the numerical solution of the kinematic wave equations (Wooding, 1966), solved with a finite-difference method. It assumes that runoff can be generated by exceeding the infiltration capacity (Hortonian mechanism) or by soil saturation, depending on rainfall intensity and soil properties (infiltration capacity). The infiltration process is based on the Smith and Parlange (1978) equation defined by soil and land cover parameters: soil water capacity (the difference between soil saturation capacity and initial saturation), saturated hydraulic conductivity, soil porosity, net capillary drive, pore distribution, roughness coefficient and percent of canopy cover. Evapotranspiration and groundwater flow are neglected (Mansouri et al., 2001) but K2 takes into account canopy interception and storage. Soil water is redistributed during storm intervals (Corradini et al., 2000) based on the Brooks and Corey relationship, corresponding to an unsaturated permanent flow.

The watershed is treated as a cascading network of planes and channel elements. Channels receive flow from adjacent planes and/or upslope channels. Each element is assigned homogeneous parameter values that describe geometry and hydrological parameters (slope, vegetation cover, soil properties, initial conditions, etc.) and control runoff generation (Goodrich et al., 2011).

Element definition is done with the automated geospatial watershed assessment (AGWA) tool which is the GIS-based interface (Miller et al., 2007). From the topography, AGWA discretizes the watershed into subwatersheds (or planes) according to the contributing source area (CSA) defined by the user. The CSA is the minimum area that is required for initiation of channel flow. The number of subwatershed (or planes) and the density of the channel network increase with decreasing CSA. Each plane is considered as homogenous and its hydrological parameters are derived from soil-surface-characteristic maps based on soil texture (FAO classes) and vegetation properties.

K2 needs four input datasets, the digital elevation model (DEM), the soil map and the land cover map, that are necessary to describe the watershed in term of hydrological and geometric parameters and the precipitation data that are needed at a small time step (5 min) to take into account the short and intense rainfall events, typical of the Sahelian monsoon. The input data used in this study are summarized in Table 1 and described below. Further details and analysis of in situ data can be found in Frappart et al. (2009) Guichard et al. (2009), Timouk et al. (2009) and Gal et al. (2016).

Land cover and soils maps have been derived from satellite data for the present period (2011) and from aerial photographs for the past period (1956). For each period, four major groups of landscape units have been distinguished: sandy soil units (S), outcrop units (O), pediment units or “glacis” (P) and flooded zone units (F). These are further divided into subunits with different soil and land cover types and hydrological properties (Table 2). This classification is based on a long-term ecosystem survey (Mougin et al., 2009) and studies carried out in the Sahel by Casenave and Valentin (1989), Valentin and Janeau (1988), Kergoat et al. (2015) and Diallo et al. (1999). Each landscape unit has unique vegetation and groups soil types with similar hydrodynamic properties.

For the past period, photo interpretation of stereopairs is used so that the relative elevation of the different units can be derived from the three-dimensional view, which is helpful for identifying units on panchromatic images. For the present period, the very high resolution of satellite images and the true color composites allows each unit to be discriminated rather easily. For both periods, units have been delimited independently and manually to maximize consistency. When photo interpretation is not sufficient to discriminate some landscape units, changes between present and past are considered null.

The objective of this work is to use the K2 model to analyze changes between the past and present periods and to employ the model as a diagnostic tool. For this reason, model parameters are prescribed as realistically as possible for these two periods, and calibration is kept at a minimum in order not to mask out or distort the impact of observed landscape evolution on past and present surface runoff. The only calibrated parameters are those for channels since these are the least known from literature and they are difficult to identify with precision from remote sensing. Therefore, the change attribution study is carried out for planes only, which are not calibrated. Details on the model setup and on the determination of the different parameters are given in the next subsections.

Two reference simulation cases are designed together with a suite of academic simulations to quantify and rank the effects of the landscape and meteorological changes observed over time.

The first reference simulation is the “present case”, which builds on the soil and vegetation map of 2011, with a simulation period extending from 2000 to 2015 (n=15). The present case, which has the highest number of observations available, combining the channel setup and evaluation period, is considered as the “baseline” simulation. The second reference case is the “past case”, which builds on the soil and vegetation map of 1956, with a simulation period extending from 1960 to 1975 (n=15).

From the present case, a suite of landscape changes, identified through the comparison of the Agoufou watershed in 1956 and 2011 (see Sect. 4.1), lead to simulations C, D, V and S, and a meteorological change leads to simulation P. These changes are implemented in the model first independently, then in combination. The simulation setup is summarized in Table 5 together with the associated forcing.

The impact of the different factors considered in the different simulations is expressed as a fraction of the difference between present and past mean annual discharge (Ex in percentage, %; Eq. 5), with 100 % corresponding to the past discharge and 0 % to the present. Ex=(AQpr-AQx)×100(AQpr-AQpa) Here, AQpa is the past annual discharge, averaged over 1960–1975, AQpr is the present annual discharge averaged over 2000–2015 and AQx the annual discharge of each simulation. The different factors can therefore be ranked according to their effect on runoff. Additional simulations (CD, VS and CDVS) also address the effects of factors combination.

A sensitivity analysis was carried out to assess the robustness of the model in ranking the factors responsible for the increase of surface runoff, considering the uncertainties associated with planes and channel parameters.

According to sensitivity studies previously carried out for K2 in semiarid areas, the Ks and the Man are the most important parameters affecting the simulated surface runoff (e.g., Al-Qurashi et al., 2008; Smith et al., 1999). A first sensitivity test was carried out on planes Ks and Man. The range of variability in plane parameters was based on data compiled by Casenave and Valentin (1989) for Sahelian soils, resulting in a factor of 2.5 for Ks and 1.75 for Man intervals. A second sensitivity analysis was carried out for the channel parameters, to compare the parameters giving the lowest RMSE and the lower bias during the channel setup period (2011–2015).

To represent the spatial evolution of these two sensitive hydrological parameters (Ks and Man) and surface runoff (Q) within the watershed, watershed maps were constructed for the monsoon period over the past and the present period.

The 1956 and 2011 land cover maps are presented in Fig. 3a and b, together with the corresponding drainage networks. For each landscape unit, the difference between these two periods has been computed (Fig. 3c).

Drainage network and flooded zones (F). The drainage network significantly increased between the two periods, with a total channel length of 71 km in 1956 against 104 km in 2011, corresponding to a drainage density increased by a factor of 1.5. Four zones (Z1, Z2, Z3 and Z4) underwent a particularly strong development of the drainage network (Table 6 and Fig. 4). Furthermore, a fraction of the watershed, located in the western region, has become a contributing area of the watershed in 2011, as can be seen by the active drainage network development. Floodplains (F1) have also expanded from 6.5 km2 in 1956 to 12 km2 in 2011. This change coincides with an important increase of the open water area (F2), especially marked for the watershed outlet (the Agoufou lake).

Sandy soils (S). Sandy dunes (S2) and deep sandy soils (S3) exhibit limited changes. The total surface of these two units is 54 % of the total watershed in 2011 against 60 % in 1956. The conversion of S2 into agriculture enclosure (S4) explains most of this change, since enclosures occupy 10 km2 in 2011 against 2.5 km2 in 1956. Isolated dunes (S1) are found at the same location for both periods, but have been eroded and partially encrusted. Today, approximately 30 % of their surface is covered by crusts (i.e., 30 % of S1 in 1956 correspond to S1c in 2011). Overall, the sandy soils represented 63 % of the total watershed in 1956 and 60 % in 2011. Their hydrological properties are similar for the present and the past periods, except for crusted isolated dunes which represent 0.36 % of the total watershed in 2011 and were not detected in 1956.

Outcrops (O). Conversely, outcrops markedly developed in the northern part of the watershed. For instance, large areas in the northeastern part changed from pediment to O2 outcrops. Overall, the surface of the outcrop classes has increased from 18 km2 in 1956 to 27 km2 in 2011.

Pediment (P). Although the overall proportion of the pediment class (P) on the watershed has not really changed between 1956 and 2011 (24 and 26 %, respectively), this unit underwent great changes within the pediment class. Indeed, all tiger bush units (P3) have completely disappeared, leaving impervious denudated soils (P4), sometimes with some rare trees or bushes, witnesses of the old tiger bush. In addition, the silt layer (P2) has increased from 7 to 11 km2. The watershed map of 1956 also shows a large central area occupied by shallow sandy soils (P1v) that has largely disappeared and has been replaced by mostly impervious rocky pediments (P1). This last landscape unit occupied 10 % of the total watershed in 1956 against 19 % in 2011.

These changes strongly impact the watershed hydrological properties, since P1 and P4 favor surface runoff compared to P1v and P3 (Table 3). In addition, the silt layer (P2) has increased from 7 to 11 km2, mainly in areas where the drainage network became highly developed, reflecting the transition from sheet runoff to concentrated runoff and/or the increase of overland flow.

The best agreement between observed and simulated annual discharges for the channel setup period is obtained with a channel Man of 0.03 s m-1/3 and a channel Ks of 30 mm h-1 (Table 7), which is close to the best guess for these parameters. The corresponding bias is 2.7 % of the averaged discharge and the RMSE is equal to 6.4×105 m3 (corresponding to 2.6 mm yr-1 over the total watershed, n=5). As expected, several combinations of Man and Ks give close results, with higher Man compensating lower Ks. The optimized values of Man and Ks correspond to rather impervious channels, which infiltrate much less than what is found in the literature for Sahelian watersheds, which reports a Man close to 0.03 s m-1/3, but a Ks commonly reaching 150 to 250 mm h-1. These studies, however, concern particularly sandy areas where channels are several meters deep and are sometimes preferential infiltration sites (Chow, 1959; Estèves, 1995; Peugeot et al., 2007; Séguis et al., 2004). Conversely, the Agoufou watershed is characterized by very shallow silted soils or outcrops (northern part of the watershed) and silted channels, which is consistent with lower values of Ks.

For the validation period (6 years with available observation during the period 2000 to 2010), the bias on the annual discharge is -1.2 % and the RMSE is 1.1×106 m3 (4.5 mm yr-1; n=6), showing that the model performs reasonably well. For both the channel setup and validation periods, the interannual variability of the simulations is slightly greater than the observed one, with an underestimation for 2000–2001 (mid- to low-discharge years) and an overestimation for 2010 and 2013 (high-discharge year, Fig. 5). The intraannual variability of the simulated discharge is also reasonably close to the observations, considering the significant scatter of the simulated ensembles due the statistical rainfall disaggregation.

Overall, the annual cumulated discharge are close to the observations, with simulated mean annual discharges of 3.72×106 m3 (15.2 mm yr-1; n=5) and 3.85×106 m3 (15.7 mm yr-1; n=6) for the channel setup and validation period, respectively, against 3.42×106 m3 (14 mm yr-1; n=5) and 3.47×106 m3 (14.1 mm yr-1; n=6) for the observations. The mean relative bias between observed and simulated discharge during the whole period is 0.5 % (n=11) with a RMSE of 9.35×105 m3 (3.8 mm yr-1; n=11).

The sensitivity analysis results are assessed by a comparison between the Ex values obtained by the attribution simulations for a reference case and the values obtained by the two sensitivity tests (Table 8) on planes and channel parameters. The reference case and the sensitivity tests are run using only one precipitation member, namely the closest to the ensemble average. This indeed leads to Ex values (Table 8, Column 3) very close to those reported for the attribution simulations on the 10-member ensemble (Sect. 4.3.2). Regarding the first test, multiplying the Ks of all planes by 2.5 and Man by 1.75 decreases runoff by a factor of 3 (3.3 mm yr-1 against initially 13.4 mm yr-1 for the “PRES Ref” test), but it does not change the ranking of the different factors (Table 8, Columns 3, 6 and 9), which is the primary goal of these attribution simulations. The main difference in the Ex values is found for simulation “C”, with a more important contribution of dune crusting to the runoff increase, for planes with higher Ks and Man. The second sensitivity test uses Ks equal to 40 mm h-1 (which gives the lowest RMSE for the channel setup period) instead of 30 mm h-1 (which gives lowest bias). This results in a small decrease in total runoff in all simulations but it does not change the ranking of the different factors. The impact of the drainage network development is, however, sensibly lower than in the reference simulations, which is consistent with channels with higher infiltration capacity. Overall, these two sensitivity tests show the robustness of our results concerning the ranking of the different factors contributing to runoff changes between the past and the present. In particular, the impact of vegetation changes and the evolution of soil properties, which alone are sufficient to simulate the past–present difference, is a robust feature.

The maps of landscape units were derived from different data (aerial photography and satellite images) of various spatial and spectral resolutions. Delimitation and identification of the landscape units proved easier for the present than for the past. Panchromatic aerial photographs give limited information, and on many occasions the three-dimensional visualization is necessary to clearly identify the units. Photo interpretation for the past aimed to be conservative, meaning that obvious changes only were retained while ambiguous cases were considered as “no change”. Overall, despite the uncertainties related to the photo interpretation and mapping of landscape units, that are not easily estimated, the land surface changes of the Agoufou watershed are important and clearly observable.

The results reported in Fig. 3 highlight the increase in drainage density, which reaches a factor of 1.5 over the whole watershed and is accompanied by an expansion of open water surface (F2) and alluvial plains (F1). Similar changes were also observed by Massuel (2005) and by Leblanc et al. (2008) in southwestern Niger, where the drainage density increases by a factor of more than 2.5 between 1950 and 1992, as well as in a small watershed in northern Mali by Kergoat et al. (2015), who reported a factor of 2.8 between 1956 and 2008. Considering that few changes are observed on the southern sandy part of the basin, the evolution of the drainage network for Agoufou is consistent with the values found in the literature.

Woody vegetation, and especially the thickets of the tiger bush unit, and some of the shallow sandy soils have completely disappeared and have been replaced by hardpan outcrops and silt surfaces, which is consistent with several studies (Hiernaux and Gérard, 1999; Leblanc et al., 2007; Abdourhamane Touré et al., 2011; Trichon et al., 2012). Hiernaux et al. (2009b) have observed that the woody vegetation of the Gourma region has declined since the 1950s and particularly from 1975 to 1992 over shallow soils. Given that tiger bush thickets grow perpendicularly to the water flow and therefore protect the soil against erosion, capture runoff and favor infiltration (Valentin et al., 1999 among others), disruption of thickets leads to runoff concentration and changes the overall hydrological properties. Sighomnou et al. (2013) in Niger and Kergoat et al. (2015) in Mali have also noted a significant decrease of vegetation over shallow soils and a corresponding increase in denudated surfaces. In a watershed in Niger, Abdourhamane Touré et al. (2011) estimated that the tiger bush occupied 69 % in the 1970s and has disappeared in the 2000s. Man-driven deforestation has been put forward as a cause for thickets clearing in southwestern Niger. For Agoufou, such activity is not reported and remains of dead trees can be observed, testifying natural death of vegetation.

The decay of shallow soil vegetation is not at odds with the Sahelian regreening that is observed all over the Sahel and in the Gourma since the 1980s (Anyamba et al., 2003; Dardel et al., 2014a, b; Heumann et al., 2007; Olsson et al., 2005). Dardel et al. (2014b) suggest that the resilience of herbaceous vegetation allows rapid regrowth over most soils in response to rainfall recovery, but that a fraction of the shallow soils may undergo long-term vegetation decay, in a way that impacts runoff but not region-average greening. In the Agoufou watershed, the vegetation changes affecting the northern part of the watershed would not be easily detected by coarse-resolution satellite datasets, as opposed to the herbaceous vegetation growing over the sandy soils of the southern part of the watershed. In addition, the greening trend is obvious since the 1980s, because of the maximal drought of 1983 and 1984, but the longer term trend is likely to be different (e.g., Pierre et al., 2016).

As far as land use change is concerned, a few additional enclosures (made of dead-wood fences whose prime function is to delineate land rights) are present nowadays in the Agoufou watershed, as a result of easier access to water year-round since the lake became permanent. Located on deep sandy soils not contributing sensibly to runoff, these enclosures do not impact the overall characteristics of the watershed. In that respect, Agoufou differs from most of the watersheds studied in the Sahel so far (Niger, Burkina Faso).

If simulations and observations are in good agreement for the present period, simulated discharge for the past period is overestimated (0.51×106 and 0.02×106 m3 for simulation and observation, respectively, or 2.1 and 0.08 mm yr-1). Different reasons could explain this. First, the error bars in Fig. 8, which represent the standard deviation of the 10 members used for the ensemble simulations, illustrate the high sensitivity of the model to precipitation intensity. Moreover, the intraannual dynamics (Fig. 6) also reflect the sensitivity of the model to a limited number of rainfall events each year. By assumption, the 5M rainfall intensity is supposed to be the same (i.e., to have the same distribution) for the past and the present periods. Lower 5M intensities in the past than in the present could then lead to lower simulated discharge values which could come closer to observations. However, there is no evidence of changes in precipitation intensity at this short timescale (Panthou et al., 2014). Second, the model was evaluated with all available observation data over the 2000–2015 period, when data are the most accurate and numerous. During the past period, only few data are available (4 years) and the estimation of the annual discharge is less precise (see Gal et al., 2016), which could also account for part of the discrepancies between simulated and observed mean discharge in the past.

As highlighted in Sect. 3.2.1, a possible time lag in precipitation falling within the watershed would have an impact on the instantaneous amount of water ending up in channels. In our simulation setup, channels are calibrated, so a possible time lag in precipitation would consequently have an overestimation of channel conductivity during calibration. However, this would not sensibly impact our results, which have been shown to be robust to changes in channel characteristics.

Moreover, all channels were considered to have the same characteristics over time and space, but field observations suggest that channel properties vary according to their geographical position, channels being possibly more permeable in the southern part than in the northern part, where they are also shallower. In addition, increasing surface runoff over time could contribute to erosion of the soil surface and an increase in sediment transport along channels downstream. If the sediment texture were mostly clay and silt, channels may have become more impervious, thus increasing runoff at the outlet. Less impervious channels in the past may therefore explain model overestimation. However, literature reports the reverse situation in the Sahel (Séguis et al., 2002), with material particularly rich in sand being transported. As for gully depth and soils, it is not clear whether the different Sahelian watersheds studied so far are comparable, given the importance of shallow soils and silt in the northern part of the Agoufou watershed.

Despite the possible sources of uncertainty previously identified, the difference between observations and simulations (0.14×106 and 0.49×106 or 0.5 and 2 mm yr-1) is largely below the difference between the past and the present period (3.41×106 and 2.78×106 for the observation and simulation discharge, respectively, or 13.9 and 11.3 mm yr-1, see Fig. 7). Simulated mean discharges for the present and the past periods are significantly different (t test for means equality) as it is the case for observations (Gal et al., 2016).

Changes in vegetation and soil properties alone are sufficient to simulate the observed increase in watershed runoff over time. Previous studies of Sahelian hydrology agree on the major role of surface conditions in erosion and runoff generation (Casenave and Valentin, 1990; D'Herbès and Valentin, 1997; HilleRisLambers et al., 2001; Peugeot et al., 1997; Rietkerk et al., 2002). For the Agoufou watershed, the comparison of past and present land cover maps indicates that vegetation, mainly the dense thickets but more generally vegetation growing on shallow soils, has decayed after the severe droughts in 1972–1973 and again in 1983–1984 (Dardel et al., 2014b; Hiernaux et al., 2009b). The lack of vegetation recovery during the long drought period combined with erosion of shallow soils and runoff shift from sheet runoff to concentrated runoff is in agreement with findings by Séguis et al. (2004) who estimated, using hydrological modeling, that changes in land cover on the Wankama watershed had multiplied the mean annual runoff by a factor close to 3 for the 1950–1992 period. Valentin et al. (2004) have also shown that a general decrease in vegetation cover modified the hydraulic properties of the topsoil and led to an increase in Hortonian runoff collected in numerous gullies and ponds. Our study highlights the predominant role of land cover changes in a pastoral area as opposed to several studies conducted in cropland dominated areas, which pointed to the leading role of the land use changes in surface runoff changes (Albergel, 1987; Favreau et al., 2009; Leblanc et al., 2008; Mahé et al., 2005).

Drainage network development is a key marker of ecosystem degradation (Descroix and Diedhiou, 2012; San Emeterio et al., 2013). Studies of the direct impact of this phenomenon on surface runoff are scarce. The changes in drainage density shown in this study could be explained by the acceleration and/or concentration of surface runoff (due to vegetation decay) allowing gully erosion to develop in susceptible areas like on silty or sandy soil, for example (Leblanc et al., 2008; Marzolff et al., 2011; Poesen et al., 2003; Valentin et al., 2005). An increase in both the number and the length of channels reduces the travel time for water to reach the drainage network (which increases the total water flow when channels are more permeable than to planes, as in our case). An increase of drainage density was also reported by Leblanc et al. (2008) in Niger. In semiarid regions, gullies tend to enhance drainage and to decrease the water supply for the vegetation growing on planes and slopes (as opposed as along gullies or downstream), providing therefore a positive feedback to the vegetation decay and/or enhanced runoff system (Leblanc et al., 2008; Valentin et al., 2005).

Crusts are frequently cited as a possible explanation of the Sahelian paradox (Favreau et al., 2009; Leblanc et al., 2008; Mahé and Paturel, 2009). Our results suggest that the impact of crusted sandy dune on the surface runoff is limited. This is not necessarily the case further south such as in southwestern Niger, where some soils have a higher percentage of clay. Moreover, soil crusting in the Agoufou landscape may be slightly underestimated given the low resolution of aerial photographs in 1956. Trampling by livestock, not considered here, has an unclear impact on soil crusting: according to the work by Hiernaux et al. (1999) on sandy soils in Niger, the soil infiltration capacity slightly increases with moderate grazing, but decreases at higher stocking rates. Moreover, the evolution of the stocking rates is poorly known over 1956–2011, although an increase cannot be excluded. Besides, it should be noted that in the literature, vegetation degradation is sometimes classified as “increase in surface crusting”, while in this study changes from tiger bush vegetation into impervious soil, which are crusted, are simulated as part of vegetation and soil changes (“VS” simulation).

Finally, the increase in the occurrence of extreme rainy events in daily precipitation suggested by Frappart et al. (2009) and demonstrated by Panthou et al. (2012, 2014) is intrinsically taken into account by the use of daily precipitation series used to force the model in our study. The results suggest that changes in daily precipitation regime do not explain runoff changes between the past and the present. If this variable is only taken into account (simulation “P”), simulated surface runoff decreases rather than increasing over time. This is in line with Descroix et al. (2013), Casse et al. (2016) and Aich et al. (2015), who found that the modest increase in large rainfall amount (events > 40 mm) observed during the 2000s cannot alone explain the Sahelian paradox. However, this should be taken with caution because changes in precipitation not statistically detectable here may have occurred elsewhere, due to the high natural variability, and further studies are required to address this question in more detail.