Modelling stream flow and quantifying blue water using a modified STREAM model for a heterogeneous , highly utilized and data-scarce river basin in Africa

Integrated water resources management is a combination of managing blue and green water resources. Often the main focus is on the blue water resources, as information on spatially distributed evaporative water use is not as readily available as the link to river flows. Physically based, spatially distributed models are often used to generate this kind of information. These models require enormous amounts of data, which can result in equifinality, making them less suitable for scenario analyses. Furthermore, hydrological models often focus on natural processes and fail to account for anthropogenic influences. This study presents a spatially distributed hydrological model that has been developed for a heterogeneous, highly utilized and data-scarce river basin in eastern Africa. Using an innovative approach, remote-sensingderived evapotranspiration and soil moisture variables for 3 years were incorporated as input data into the Spatial Tools for River basin Environmental Analysis and Management (STREAM) model. To cater for the extensive irrigation water application, an additional blue water component ( Qb) was incorporated in the STREAM model to quantify irrigation water use. To enhance model parameter identification and calibration, three hydrological landscapes (wetlands, hillslope and snowmelt) were identified using field data. The model was calibrated against discharge data from five gauging stations and showed good performance, especially in the simulation of low flows, where the Nash–Sutcliffe Efficiency of the natural logarithm ( Ens_ln) of discharge were greater than 0.6 in both calibration and validation periods. At the outlet, theEns_ln coefficient was even higher (0.90). During low flows, Qb consumed nearly 50 % of the river flow in the basin. TheQb model result for irrigation was comparable to the field-based net irrigation estimates, with less than 20 % difference. These results show the great potential of developing spatially distributed models that can account for supplementary water use. Such information is important for water resources planning and management in heavily utilized catchment areas. Model flexibility offers the opportunity for continuous model improvement when more data become available.


Introduction
Hydrological models are indispensable for water resource planning and management at catchment scale, as these can provide detailed information on, for example, impacts of different scenarios and trade-off analyses.Society's demand for more accountability in the management of externalities between upstream and the downstream water users has also intensified the need for more predictive and accurate models.However, complexity of hydrological processes and high levels of heterogeneity present considerable challenges in model development.Such challenges have been exacerbated over time by land use changes that have influenced the rainfall partitioning into green (soil moisture) and blue (runoff) water resources.In spite of these challenges, it is still desirable to develop a distributed hydrological model that can simulate the dominant hydrological processes and take into account the various water uses.In large catchments with high heterogeneity, key variables such as water storage (in unsaturated and saturated zones) and evaporation (including transpiration) are difficult to obtain directly from point measurements.This becomes even more difficult for ungauged or poorly gauged river basins.
In most cases those variables are derived from models using (limited) river discharge data which increases equifinality problems (Savenije, 2001;Uhlenbrook et al., 2004;Mc-Donnell et al., 2007;Immerzeel and Droogers, 2008).On the other hand, grid-based distributed models at fine spatial scales do not explicitly account for additional blue water use (Q b ), such as transpiration from supplementary irrigation or withdrawals from open water evaporation.In fact, in tropical arid regions, Q b can be a large percentage of the river discharge during low flow.Calibrating models using modified stream flow data may lead to incorrect parameterization, and may lead to high predictive uncertainty in the hydrological model outputs, especially when dealing with scenarios for water use planning.
To overcome these challenges, many researchers have opted for simple, lumped and or parsimonious models with a limited number of model parameters.The models are simplified by bounding and aggregation of some functionality in the complex system (Winsemius et al., 2008).In doing so, models may become too simplified to represent hydrological processes in a catchment (Savenije, 2010).Therefore, Savenije (2010) proposes a conceptual model mainly based on topographic characteristics to represent the dominant hydrological processes.The model maintains the observable landscape characteristics and requires a limited number of parameters.Other researchers have used secondary datae.g. from remote sensing to calibrate or infer model parameters as much as possible (Winsemius et al., 2008;Immerzeel and Droogers, 2008;Campo et al., 2006).This has been possible in the recent past because of the availability of satellite images with finer spatial resolutions.Advancement in remote sensing algorithms has also resulted in wider-range spatial data of reasonably good accuracies.Such spatial data include actual evapotranspiration (ET a ) derived from remote sensing data, such as TSEB (Norman et al., 1995), SEBAL (Bastiaanssen et al., 1998a, b), S-SEBI (Roerink et al., 2000), SEBS (Su, 2002) and METRIC (Allen et al., 2007).Spatial data on soil moisture can also be derived from satellite images, for example from ERS-1 Synthetic Aperture Radar (SAR) combined with the TOPMODEL topographic index (Scipal et al., 2005) or from Advance Very High Resolution Radiometer (AVHRR) combined with the SEBAL model (Mohamed et al., 2004).It is also evident that distributed models perform well with finer-resolution data, as demonstrated by Shrestha et al. (2007).Using different resolution data (grid precipitation and grid ET a ) and a concept of IC ratio (input grid data area to catchment area), they found that a ratio higher than 10 produces better performance in the Huaihe River basin and its sub-basin of Wangjiaba and Suiping in China (Shrestha et al., 2007).
Furthermore, remotely sensed data at finer resolutions offer great potential for incorporating blue water, in the form of (supplementary) water use (Q b ) in model conceptualization.This opportunity arises from the fact that remotely sensed ET a based on energy balance provides total evapotranspiration that already accounts for Q b .For instance, Romaguera et al. (2012) used the difference between the Meteosat Second Generation (MSG) satellites data (total ET a ) and the Global Land Data Assimilation System (GLDAS), which does not account for Q b , to quantify blue water use for croplands in Europe with a reasonable accuracy.However, the spatial scales of such data sets (GLDAS (1 km) and MSG (3 km)) limit the application.Nevertheless, Romaguera et al. (2012) recommended such an application to recently available data of wider spatial and temporal coverage -e.g.data derived from Moderate-resolution Imaging Spectroradiometer (MODIS) 250 m and 500 m.
However, the literature shows limited applications of utilizing grid data for distributed hydrological models in poorly gauged catchments.Winsemius et al. (2006) showed that the soil moisture variations from the Gravity Recovery And Climate Experiment (GRACE) could provide useful information to infer and constrain hydrological model parameters in the Zambezi River basin.Campo et al. (2006), using an algorithm developed by Nelder and Mead (1965), used remotely sensing soil moisture information to calibrate a distributed hydrological model in the Arno Basin, Italy.Immerzeel and Droogers (2008) used remotely sensed ET a derived from SE-BAL in the calibration of a Soil and Water Assessment Tool (SWAT) model of the Krishna Basin in southern India in which the model performance (r 2 ) increased from 0.40 to 0.81.Recently, Cheema et al. (2014) used satellite-derived rainfall to parameterize the SWAT model, while ET a from ETLook was used to calibrate the model to determine the contribution of groundwater use to the total blue water use in the Indus Basin.
The factors that may have limited the application of remote sensing (RS) data on hydrological modelling include (a) limited flexibility of hydrological models to utilize spatially distributed data.This is normally the case where the user has no control over the model source code.The user is therefore limited to optimizing model performance using secondary data.(b) Limited availability of RS data at the appropriate spatial and temporal scales to capture dominant hydrological processes in a catchment.(c) The lack of technical skills by most hydrologists and water resource specialists on how to transform RS data into hydro-meteorological data (Schultz, 1993).The opportunities and challenges for the wider application of remote sensing for hydrological modelling are discussed by De Troch et al. (1996) and Schultz (1993).
This paper presents a novel method of using ET a and soil moisture data derived from satellite images as input in a distributed hydrological model.The upper Pangani River basin in eastern Africa has been used as a case study.This river basin has heavily managed landscapes dominated by small and large-scale irrigated agriculture.The secondary data used in this study have been generated using MODIS satellite information and the SEBAL model on 250 m and 8-day resolutions for the period 2008-2010 (Kiptala et al., 2013b).Here the STREAM model has been modified to incorporate blue water use.The model parameters have also been confined further by the topographic characteristics and groundwater observations using the hydrological conceptualization developed by Savenije (2010).

Study area
The upper Pangani River basin (13 400 km 2 ) covers approximately 30 % of the total area of the Pangani River basin and is located in eastern Africa (Fig. 1).It is the main headwaters of the basin and derives its water sources from Mt. Kilimanjaro (5880 metres above sea level (m a.s.l.)) and Mt.Meru (4565 m a.s.l.) catchments.The flows to the lower basin are regulated by a large dam (storage capacity 1.1 × 10 9 m 3 ), the Nyumba ya Mungu (NyM) reservoir.The lower Pangani River basin has three operational hydro-electric power (HEP) stations: NyM HEP, Hale HEP and the New Pangani Falls HEP station.These provide up to 91.5 MW or 17 % of Tanzania's electricity.
The catchment has a highly varied climate mainly influenced by topography.The high altitude slopes around the mountain ranges have an Afro-Alpine climate and receive nearly 2500 mm yr −1 of rainfall.The lower parts have a sub-humid to semi-arid climate and the rainfall varies between 300 to 800 mm yr −1 .The rainfall has a bimodal pattern where long rains are experienced in the months of March to May (masika season) and the short rains in the months of November and December (vuli season).It is during these two seasons when most crops are grown.Rainfed and supplementary irrigated croplands are the dominant agricultural systems.However, grasslands and shrublands are also dominant land use types (see Sect. 3.1.3)(Kiptala et al., 2013a).

Hydro-meteorological data
Daily rainfall data for 93 stations located in or near the upper Pangani River basin were obtained from the Tanzania Meteorological Agency and the Kenya Meteorological Department.The data was screened and checked for stationarity (Dahmen and Hall, 1990).Of the original group, 43 stations proved useful after data validation for the period 2008-2010.Unfortunately, there were no rainfall stations at elevations higher than 2000 m a.s.l., where the highest rainfall actually occurs.Spatially distributed rainfall can also be provided by satellite sensors to augment rainfall data from the ground stations (Huffman et al., 2001).Such satellites sensors include the Tropical Rainfall Measuring Mission (TRMM).The Famine Early Warning System (FEWS) product also provides remotely sensed rainfall data in Africa.The satellite-based rainfall has uncertainties that can be corrected using limited ground rainfall measurements (Hong et al., 2006;Cheema and Bastiaanssen, 2012).Since there were no rainfall stations at the mountainous areas, the satellite-based rainfall could not be validated (Haque, 2009).
According to PWBO/IUCN (2006), the maximum longterm mean annual precipitation (MAP) at the Pangani River basin is estimated at 3453 mm yr −1 at elevation 2453 m a.s.l.The estimates were based on a rain gauge station that is no longer operational.Therefore, a linear extrapolation method based on the concept of double mass analysis (Wilson, 1983) was used to derive the seasonal rainfall up to the mountain peaks.Double mass analysis assumes relatively consistent correlation between time series of rainfall data at nearby stations with similar hydrological conditions (Chang and Lee, 1974).In the analysis, the seasonal precipitation at the mountain peak (Y ) is assumed to have a linear relation to the seasonal precipitation of the nearby stations (X) scaled by a proportionality factor (α).The proportionality factor, α is the average slope of the long-term MAP for the two reference points.Y is therefore given as [Y = αX].The rainfall was maintained constant above this elevation to 4565 m a.s.l. for Mt.Meru and 5880 m a.s.l. for

J. K. Kiptala et al.: Modelling stream flow and quantifying blue water
Mt. Kilimanjaro.This assumption is expected to have negligible effects at the Pangani River basin because of the relatively small area above this elevation (3 %).Six dummy stations were therefore extrapolated from the existing rainfall stations to the mountain peaks.
River discharges for six gauging stations were obtained from the Pangani Basin Water Office (Moshi, Tanzania), see Fig. 1.The measurements were obtained as daily water level measurements and converted to daily discharge data using their corresponding rating curves equations for the period 2008-2010.

Evaporation and soil moisture
The actual evapotranspiration (ET a ) and soil moisture data for the upper Pangani River basin were obtained from recent and related research by Kiptala et al. (2013b).ET a and soil moisture data for 8-day and 250 m resolutions for the years 2008-2010 were derived from MODIS satellite images using the Surface Energy Balance Algorithm of Land (SEBAL) algorithm (Bastiaanssen et al., 1998a, b).Actual evapotranspiration (ET a ) is comprised of interception (I ), soil evaporation (E s ), open water evaporation (E o ) and transpiration (T ).

Land use and land cover types
In this study, we employed the LULC classification for the upper Pangani River basin as developed by Kiptala et al. (2013a).They derived the LULC types using phenological variability of vegetation for the same period of analysis, 2008 to 2010.LULC types include 16 classes dominated by rainfed maize and shrublands that constitute half of the area in the upper Pangani River basin.

Other spatial data
Elevation and soil data were also obtained for the upper Pangani River basin.A digital elevation model (DEM) with 90 m resolution was obtained from NASA's Shuttle Radar Topography Mission (SRTM) (Farr et al., 2007).The soil map was derived from the harmonized world soil database which relied on soil and terrain (SOTER) regional maps for northern and southern Africa (FAO/IIASA/ISRIC/ISS-CAS/JRC, 2012).

Model development
The hydrological model was built to simulate stream flow for the period 2008-2010 for the upper Pangani River basin.An 8-day time step and 250 m moderate resolutions were used to correspond to availability of remotely sensed ET a data for the period of analysis.The 8-day time step is sufficiently short for the agricultural water use process, which has a timescale range of between 10 and 30 days (unsaturated zone storage over transpiration rate).In addition, this timescale is assumed to be sufficiently large to neglect travel time lag in the river basin.The other general hydrological processes in the river basin are estimated to have larger timescales (Notter et al., 2012).The spatial scale of 250 m is limited by the available MODIS satellite data.This is reasonably representative of the sizes of the small-scale irrigation schemes in the upper Pangani River basin.STREAM, a physically based conceptual model, was developed in the PcRaster modelling environment (Aerts et al., 1999).The PcRaster scripting model environment consists of a wide range of analytical functions for manipulating Raster GIS maps (Karssenberg et al., 2001).It uses a dynamic script to analyse hydrological processes in a spatial environment.The PcRaster environment allows for tailored model development and can therefore be used to develop new models, suiting the specific aims of the research, including the availability of field data.The STREAM model in the PcRaster environment allows the inclusion of spatially variable information, like ET a and soil moisture, in the model.Furthermore, STREAM model is an open source model which has been applied successfully in other data-limited river basins, especially in Africa (Gerrits, 2005;Winsemius et al., 2006;Abwoga, 2012;Bashange, 2013).
In the STREAM model, surface runoff is computed from the water balance of each individual grid cell, which is then accumulated in the local drainage direction derived from the DEM to the outlet point (the gauging station).The model structure consists of a series of reservoirs where the surface flows are routed to the rivers.We modified the STREAM model by including an additional blue water storage parameter (S b ) that regulates Q b in the unsaturated zone.Q b can be derived from the groundwater as capillary rise, C(t), or river abstraction, Q d (t).The input variables for the modified STREAM model are precipitation (P ) and interception (I ), calculated on a daily basis as a pre-processor outside the model.Evaporation (E s , E o ) and Transpiration (T ) denoted as [E+T ] were derived by subtracting I for the total evaporation (ET a ) derived from SEBAL [ET a −I ].The minimum soil moisture, S u,min , is also derived from SEBAL.The other parameters are determined through calibration.Figure 2 shows the modified STREAM model structure for upper Pangani River basin.
In the model, E + T and S u,min are the main drivers of the hydrological processes in the unsaturated zone of the model.E + T is the evaporation (soil moisture) depletion component, while S u,min is the depletion threshold.It is assumed that excess water from the upstream cells or pixels would supplement water needs of the middle or lower catchments where supplementary water is used.The upper Pangani River basin is a typical river basin, where precipitation exceeds ET a in the upper catchments and hence contributes river flow to the downstream catchments.
The rationale for accounting for Q b in the model is motivated by the incapability of the original STREAM model if applied in irrigated landscapes to simulate actual transpiration.The original STREAM model was developed specifically for natural landscapes dominated by woody savannas and wetlands with high storage capacity (Dambos) in the Zambezi River basin (Gerrits, 2005;Winsemius et al., 2006).The blue water use is therefore limited and has been accounted for by the capillary rise only.The total transpiration was therefore derived only as a function of potential evaporation and the soil moisture (from precipitation) in the unsaturated zone using the relation by Rijtema and Aboukhaled (1975).Bashange (2013), using the original STREAM model, found that simulated E + T for irrigated croplands were significantly lower compared to SE-BAL E + T for dry seasons in the Kakiwe Catchment, upper Pangani River basin.The result was attributed to lower soil moisture levels in the unsaturated zone (not replenished in the model by blue water use).

Interception (I )
When precipitation occurs over a landscape, not all of it infiltrates into the subsurface or becomes runoff.Part of it evaporates back to the atmosphere within the same day the rainfall takes place as interception.The interception consists of several components that include canopy interception, shallow soil interception or fast evaporation from temporary surface storage (Savenije, 2004).The interception is dependent on the land use and is modelled as a threshold value (D).The interception process typically has a daily timescale, although some work has been done to parameterize the interception threshold on a monthly timescale (De Groen and Savenije, 2006).
In our case, we calculate the daily interception according to Savenije (1997Savenije ( , 2004) outside of the model (see Eq. 1): where I d is the daily interception, D d is the daily interception threshold and P d is the observed precipitation on a rainy day.Since I d occurs on a daily time step during a precipitation (P d ) event, the interception at 8 days (I d( 8) ) is derived from the accumulated daily interception computed based on daily precipitation.The interception thresholds (D d ) vary per land use and have been adopted from the guidelines provided by Liu and De Smedt (2004) and Gerrits (2010).As such, an interception threshold of 2.5 mm day −1 was used for croplands and natural vegetation, and 4 mm day −1 for forest.

Net precipitation (P e )
The net precipitation (P e( 8) ) is calculated by subtracting the accumulated interception (I d( 8) ) from the accumulated precipitation (P d( 8) ) for the 8-day timescale.
P e( 8) is split through a separation coefficient (c r ) into the two storages, unsaturated and saturated (groundwater) storages.c r is a calibration parameter that is dependent on the soil type and land use types.

Evaporation depletion (E + T )
The evaporation depletion (E + T ) is derived by subtracting the interception component of the actual evapotranspiration (ET a ) at each time step.ET a from SEBAL includes I d(8) for 8-day time steps.

Unsaturated zone
The maximum soil moisture storage (S u,max ) was defined based on land use and soil types.Water available for evaporation depletion includes water infiltrated from precipitation (c r × P e ) and blue water use (Q b ), consisting of water from capillary rise (C) and river abstraction (Q d ).During the dry (non-rainy) periods, the spatial variation in soil moisture is controlled by vegetation through the uptake of blue water resources (Seyfried and Wilcox, 1995).The model assumes a minimum soil moisture level (S u,min ) which varies for managed and natural landscapes.Soil moisture status at each time step (S u ) is therefore a key variable controlling water and energy fluxes in soils (Eqs.4 and 5).
As a result, the green water use is defined as the evaporation depletion less the blue water use (Eq.6).
The value for S u,min for each land use type is assumed to be realized during the dry months and is expressed as a fraction of S u,max (soil moisture depletion fraction).S u,min is derived in the SEBAL model for dry months as an empirical function of the evaporative fraction, (the ratio of the actual to the crop evaporative demand, when the atmospheric moisture conditions are in equilibrium with the soil moisture conditions) (Ahmed and Bastiaanssen, 2003), see Eq. ( 7).
where f is the soil moisture depletion fraction expressed as a fraction of soil moisture, S u,min , to the moisture value at full saturation, S u,max , for the dry months.S u,min was realized in the month of January, which is the driest period in the river basin.Values for f are given in Fig. 3 for selected land use types for the dry month of January, averaged over 2008-2010.
The soil moisture levels agree reasonably well with previous field studies that have shown similar ranges for natural land use types in sub-humid and semi-arid areas (Fu et al., 2003;Korres et al., 2013).It is also noted that the SEBAL model has some level of uncertainty regarding soil moisture storage and water stress (Ruhoff et al., 2012).In recognizing this uncertainty, the modified SEBAL model also uses a water balance approach, where lower S u,min levels can be tolerated with respect to the available Q b during the dry season for natural land use types.

Saturated zone
Apart from the net precipitation component ((1 − c r ) × P e ), the saturated zone receives water from the unsaturated zone when the soil moisture S u reaches field capacity (S u,max ).Excess overflow (Q u ) is routed to the groundwater using a recession factor, K u .The saturated zone consists of three linear outlets which are separated by S s,min to represent the minimum storage level, S s,q to represent quickflow threshold and S s,max to represent rapid subsurface overflow.The flows are routed using K o , K q and K s calibration coefficients, respectively.
When the groundwater storage (S s ) exceeds the S s,max , then saturation overland flow (Q o,s ) occurs: where K o is the overland flow recession constant.
The second groundwater flow component is the quick groundwater flow (Q q ).It is assumed to be linearly dependent on the S s and a quick flow threshold S s,q determined through calibration (Eq.9).
where K q is the quick flow recession constant.The third component is the slow groundwater flow (Q s,g ) which is dependent on the S s levels: where K s is the slow flow recession constant.K o , K q , K s equal 1, 2 and 28, respectively, and were determined from recession curve analysis (where 1 unit is equal to the 8-day time step).
The maximum saturation storage (S s,max ) is a key variable that determines the dominant hydrological processes in the saturated zone.Three hydrological zones can be delineated from S s,max -wetland, hillslope and snow/ice zone.When S s,max is low, the saturation excess overland flow is dominant.This is characteristic for wetland systems described in detail by Savenije (2010).It occurs in the low-lying areas of the Pangani River basin where slopes are modest, or with shallow groundwater levels.During a rainfall event, there is no adequate storage of groundwater, leading to saturation excess overland flow.The wetland system is therefore dominated by Q o and as such, the S s,max is set very low or at zero (fully saturated areas) and c r at 1.
As the elevation and slope increases, the groundwater depth as well as the S s,max increase gradually.This is characteristic of the hillslope system where storage excess subsurface flow is the dominant runoff mechanism.Topographic indicators can be used to identify and separate this zone from the wetland system (where S s,max is near zero).Recently developed indices that can be used include the elevation above the nearest open water (H ) (Savenije, 2010), or the height above the nearest drainage (HAND) (Nobre et al., 2011;Cuartas et al., 2012)  near zero.This was estimated from 92 groundwater observation levels located in the lower catchments of the river basin (Fig. 4).
Figure 4 shows the delineation of the dominant hydrological processes in the upper Pangani River basin, including the wetland and hillslope (includes snowmelts at the peak of the mountains).
S s,max is not completely available for groundwater storage due to the soil texture (porosity and soil compression).According to Gerrits (2005), the maximum groundwater storage, S s,max [mm] for hillslope can be estimated using the natural log function of water storage depth, H s (Eq.11).
where H s [m] is the normalized DEM above H (where active groundwater storage is assumed zero).It is noteworthy that the wetland system can still exist along the drainage network of a river system beyond H .This is possible since the H s would still ensure a low groundwater storage (S s,max ) which makes the wetland system the dominant hydrological process.As observed in Fig. 4, the middle catchment forms the transition from the wetlands to the hillslope.It is noted that the hydrological landscape plateau (dominated by deep percolation and hortonian overland flow) described in detail by Savenije (2010), is not existent on the slopes of Kilimanjaro and Meru, the higher elevations are forested and active in the rainfall-runoff process.It is therefore modelled as forested hillslope.
The third zone delineated is the snowmelt.The amount of snow in the river basin is limited to the small portion of the mountain peaks of Mt.Kilimanjaro and Mt.Meru.The snowmelt occurs at elevation ranges of 4070 to 5880 m a.s.l. and is derived from the land use map (Kiptala et al., 2013a).During rainfall seasons, the snow is formed and stored in the land surface.During the dry season, the snow gradually melts into the soil moisture and groundwater.This is unlike the temperate climate, where a lot of snow cover is generated during the winter seasons which may result in heavy or excess overland discharge during the summer sea-sons.Furthermore, Mt.Kilimanjaro has lost most of its snow cover in the recent past due to climate variability/change, with significant snow visible only on the Kibo Peak (Misana et al., 2012).According to Grossmann (2008) the snowmelt contribution to groundwater recharge is insignificant for the Kilimanjaro aquifer.Simple representation of snowmelt can therefore be made using the hillslope parameters, where the precipitation is stored in the unsaturated zone (c r = 1 for snow) as excess unsaturated storage.The snowmelt is thereafter routed by K u (unsaturated flow recession constant) to the groundwater over the season.This model conceptualization enables the hydrological model to maintain a limited number of parameters.

Interaction between the two zones
Capillary rise only occurs when groundwater storage is above a certain level, S c,min .S c,min can be a fixed or a variable threshold value of the groundwater storage (S s ).Winsemius et al. (2006) adopted a fixed value of 25 mm as the S c,min for the Zambezi River basin.Since S s,max (from Eq. 11) is a function of H s , a fixed threshold is not possible in this study.S c,min is made a function of groundwater storage S s to provide a spatially variable threshold through calibration over the river basin.Capillary rise above this threshold is estimated on the basis of the balance between water use needs at the unsaturated zone, and water availability in the saturated zone.Actual capillary rise is determined implicitly using the maximum capillary rise C max (calibration parameter for each land use type), evaporation depletion (E + T ) and the available groundwater storage S s .Below S c,min , a minimal capillary rise C min is possible and is assumed to be zero for this study (timescale of 8 days is assumed low for substantial C min to be realized).
where the active groundwater storage for capillary rise, S = S s − S c,min.However, since the capillary flow is low compared to water use for some land use types, supplementary blue water from river abstractions (Q d ) is required in the system.The third blue water storage term, S b , is introduced to regulate blue water availability from capillary rise, C, and river abstractions, Q d .River abstractions include water demands from supplementary irrigation, wetlands and open water evaporation for lakes or rivers derived directly from the river systems.
where Q b is the blue water required to fill the evaporation gap that cannot be supplied from the soil storage.For irrigated croplands, Q d is assumed to represent the net irrigation abstractions in the river basin.The assumption is based on the 8-day time step that is considered sufficient for the return flows to get back to the river systems, that is the flow is at equilibrium.Q d is therefore modelled as net water use in the river system.Since river abstractions mainly occur in the middle to lower catchments and the accumulated flow would have a resultant net effect equivalent to the total simulated discharge, Q s , at a downstream outlet point or gauge station (Eqs.15 and 16). Q

Sensitivity and uncertainty analysis
Since a number of assumptions were introduced to simulate the hydrological processes in the basin, a sensitivity analysis was performed to assess the influence of model input parameters to the variation of model performance.The parameter adjustments were done during the calibration process manually by trial and error.Some parameter values were manually altered within parameter ranges while others were calibrated freely.According to Lenhart et al. (2002), the parameter sensitivity can be achieved by varying one parameter at a time within the parameter range or using a fixed percentage change of the base value while holding the others fixed.Three parameter values -interception threshold (D), separation coefficient of net precipitation between the unsaturated and saturated zones (c r ) and the quick flow components (q c ) -were varied within the parameter ranges.Three parameter values for maximum storage in the unsaturated zone (S u,max ), maximum storage in the saturated zone (S s,max ) and maximum potential capillary rise (C max ) that were calibrated freely were varied by a fixed change of the base value.The other three parameter values representing runoff timescales (K o , K q , K s ) were also varied by a fixed value from the estimates determined from the recession curve.
A sensitivity coefficient was computed to represent the change in the response variable that is caused by a unit change of an input variable, while holding the other parameters constant (Gu and Li, 2002).The sensitivity coefficient (SC) was normalized by reference values representing the range of each output and input variable to give the sensitivity index (SI) represented by Eq. ( 17).
where x 0 and y 0 are the base input parameter value and model output from the final model calibration, respectively; x i and y i are the varied input parameter and the corresponding model output, respectively.SI makes it feasible to compare the results of different input parameters independent of the chosen variation range (Lenhart et al., 2002;Bastiaanssen et al., 2012).The SI can be positive or negative depending on the co-directional response of the model performance to the input parameter change.The absolute higher SI values indicate higher sensitivity.

Model performance
The modified STREAM model was calibrated and validated against measured daily discharge data from five gauging stations in the basin (see Fig. 1).One discharge gauge station, 1dd55, had a lot of missing data.Nevertheless, the limited information from this station, most upstream and the only one in the upper Mt.Meru, was useful in the calibration process of the downstream gauge stations.The daily discharge data were aggregated to an 8-day timescale for the period 2008-2010.Since the secondary data from remote sensing (ET a and f ) were available for only 3 years, 1 year of data was used for calibration while the remainder of 2 years' data was used for the validation.An initial 1 year (46 simulations) was used as a warm-up period to stabilize the model parameters using the mean input values.In total, the model was simulated for 184 time steps (4-year period).
The following goodness of fit statistics were used to evaluate the model performance: the Nash-Sutcliffe model efficiency coefficient (E ns ) (Nash and Sutcliffe, 1970), mean absolute error (MAE) and the relative mean square error (RMSE) in Eqs. ( 18), ( 19) and ( 20), respectively.
where Q s and Q o are simulated discharge and observed discharge, respectively, Q o is the mean of the observed discharge and n is the discharge data sets (n = 46 calibration; n = 92 validation).
Since the model priority objective is to simulate low flows, the E ns_ln was also evaluated using the natural logarithm of the variables in Eq. ( 18).The E ns values range within [−∞, 1], with 1 being the optimum (Ehret and Zehe, 2011).The range of MAE and RMSE is [0, ∞], with zero being the optimum (Murphy, 1995).The model is optimized using these parameters to achieve a balance between the correlation, the bias, and the relative variability in the simulated and observed discharge (Gupta et al., 2009).The model estimates for irrigation water use (Q b(I) ), defined as Q b for all the irrigation land use classes, were also compared with the field data on net irrigation water use from the river basin agency, Pangani Basin Water Office.

Scenario development
In the Pangani River basin, blue water use is currently over-exploited (Kiptala et al., 2013b).The implication for additional water allocation on stream flow to the nationally important hydropower stations needs to be known.This may also result in water-savings or tradeoffs with other interventions or water uses.The crop yields for rainfed and supplementary irrigated lands are also low, leading to low crop water productivity (Makurira et al., 2010).A few water management scenarios targeted on water-savings and improved crop water productivity is explored using the modified STREAM model.They include (i) water-saving through increased irrigation water efficiency, (ii) increased crop productivity for rainfed lands, and (iii) modifying the landscape for increased agricultural production.
To meet the first objective, the non-beneficial component of evaporation (soil evaporation) for irrigated landscapes is targeted for reduction.Soil evaporation (E s ) can account for up to 40 % of evaporation depletion (E +T ) in irrigated landscapes (Bastiaanssen et al., 2012;Burt et al., 2001).In the Pangani River basin, located in a tropical climate, the irrigation system used by smallholder farmers that conveys water using small earthen furrow canals may have high levels of E s .It is noteworthy that interception (I ) also includes shallow (fast) soil evaporation that is implicitly derived only from precipitation.For demonstrative purposes, a reduction of 5 % in E + T for supplementary irrigated mixed crops is targeted (Scenario 1).The reduction represents about 15 % of E s if we assume a conservative E s of 30 % of E + T in the supplementary irrigation systems.There are several methods for reducing E s .They may include the lining of the main canals or using more efficient micro-irrigation systems.Further reduction can also be achieved by either straw or mechanical mulching (Prathapar and Qureshi, 1999;Zhang et al., 2003).
To meet the second objective, productive transpiration for rainfed maize (highland) is increased by 30 % (Scenario 2a).According to Makurira et al. (2010), the crop water productivity for smallholder rainfed farms can be improved by using systems innovations (SIs).The study was done in Makanya catchments within the Pangani River basin.The SIs used combined runoff harvesting with in-field trenches and soil bunds which resulted in an increase of transpiration of 47 %.The SIs also aimed at preventing soil and nutrient loss.An increase in T would result in an increase in biomass production and thus crop yields (Steduto et al., 2009).The rainfed maize in the highland areas was targeted due to the relative high precipitation during the rainy seasons.In-field trenches and soil bunds (fanya juus) is normally associated with high infiltration levels and higher soil moisture retention (Kosgei et al., 2007;Makurira et al., 2010).An additional increase in S u,max of 30 % is also investigated in addition to the increased transpiration for highland rainfed maize and coffee (Scenario 2b).
For the third objective, the area for irrigated sugarcane is doubled to its potential (Scenario 3).Currently, TPC irrigation scheme covers an area of 8000 ha, for which 7400 ha is under sugarcane cultivation, with the reminder providing the irrigation services.The potential irrigation area is esti- mated at 16 000 ha constrained by limited water resources.The expansion of the irrigation system is of great interest in the basin due to the high sugar demand and increasing potential for bio-fuels.

Calibration and validation results
Figures 5 and 6 show the comparison of the observed and simulated hydrographs and the average precipitation for five outlets (gauge stations) in the upper Pangani River basin.The figures provide a visual inspection of the goodness of fit of the data with an additional scatter plot for the most downstream outlet (1dd1).The model simulates the base flows very well both during the calibration and validation periods.The peak flows for the larger streams (1dd54, 1dd1) were better simulated than for the smaller streams (1dc8a, 1dc5b,  1dc11a).It is to be noted that the observed discharge data is also subject to uncertainty which is more pronounced for the smaller streams.The remotely sensed data, ET a and f , also have a higher uncertainty during the rainy season (peak flow season).This is the period when most clouded satellite images exist and the cloud removal process is subject to uncertainty (Kiptala et al., 2013b).
Table 1 shows the performance model results for the validation and calibration of the modified STREAM model in the upper Pangani River basin.The Nash-Sutcliffe Efficiency, E ns , for the calibration period scored > 0.5 (except for 1dd11a, where E ns = 0.46) which is indicative of good model performance.In the validation period, two outlet points had scores < 0.5 (1dd11a -0.33 and 1dd54 -0.42) which indicates a moderate performance.The Nash-Sutcliffe Efficiency for a natural logarithm, E ns_ln , which emphasizes the base flow, resulted in better results, with all outlet points scoring ≥ 0.6.There was a slight reduction in E ns_ln in outlet points 1dd54 (calibration) and 1dd8a, 1d5b (validation) but overall the model performance on the low flows was good.
MAE ranged between 0.62 and 2.08 m 3 s −1 for the larger streams in the calibration period.A big difference is observed between the RMSE and MAE (up to four times) for the downstream stations 1dd54 and 1dd1 during the calibration period.The result is indicative of large (noisy) variations between the simulated and observed discharges.Figure 5 also shows that the large deviations arise during the rainy periods (masika and vuli seasons).This may be attributed to the uncertainties of the remote sensing data in the clouded periods (rainy days).Such uncertainties can be avoided by using passive microwave imagery (Bastiaanssen et al., 2012).Furthermore, the river gauging stations are poorly maintained in the river basin.The discharge rating curves are also not regularly updated despite the changes in the river regime.Model conceptualization assumptions such as irrigation water use and return flows may also not coincide in space and time with the actual processes in the river basin.Errors in boundary conditions on the representation of groundwater may also occur if they do not coincide with the river systems.

Sensitivity analysis
The sensitivity analysis of the input parameters is given in Table 2.The sensitivity index (SI) was analysed using the RMSE and MAE model performance indicators for the entire simulation period using the discharge measurements at outlet point (1dd1).The base input values (x 0 ) were the final calibrated values that were varied by a fixed or percentage change (x 1 or x 2 ).Decrease in S u,max by 25 % resulted in the highest SI of −1.97 for RMSE.However, a similar increase of 25 % did not have any significant change in model output.The sensitivity is mainly attributed to the overland flow that is influenced by the water storage in the unsaturated zone.Similar changes in S s,max also resulted in moderately high sensitivity for both RMSE and MAE.This is mainly because the saturated zone controls all the runoff components.Separation coefficient c r that separates the net precipitation between unsaturated and saturated zones, and the quick flow coefficient, q c had high sensitivity.The values used (c r = 0.75 and q c = 0.75 -aggregated averages) for various land use types were generally derived from previous modelling experiences and were based on the soil type and land use.
The soil moisture depletion fractions (f ) were derived from the SEBAL model for various land use types.An aggregated average f value of 0.33 was adopted from the mean values for the land use types that ranged between 0.2 for natural land use types to over 0.6 for irrigated agriculture (also see Fig. 3).These parameters resulted in minimum sensitivity  since the ranges used (±25 % of the base values) were reasonable within the derived estimates from remote sensing.The runoff timescales parameters K o and K q also had low sensitivity because the flow times were short and within the estimates derived from the recession curves.The timescale K s for slow groundwater flow, with higher flow times, had a moderate sensitivity.A lower timescale for K o of 1 time step (8 days) may introduce some uncertainty if the model was used to simulate flood events that are critical at shorter timescales of 1-2 days.However, for hydrological processes that characterize agricultural water use, such as irrigation scheduling or dry river flows, the uncertainty is minimal.The maximum capillary rise (C max ) was calibrated through a water balance process to maintain the evaporation depletion (E + T ).An aggregated average value of 2 mm day −1 was achieved, and ranged from 1.1 mm day −1 for woodland landscape in semi-arid areas to a maximum of 2.8 mm day −1 in the natural dense forest in humid climate.The calibrated values were within the ranges for natural vegetation reported in literature (Shah et al., 2011).In natural and rainfed systems, only C max was calibrated to maintain the evaporative capacity of the unsaturated zone.The actual capillary rise (C) would not change with an increase in C max .However, a decrease in C max would constrain C, thus resulting in lower soil moisture conditions in the unsaturated zone.For irrigated land use types, the evaporative capacity (E +T ) is maintained by both C and irrigation (Q d ).The changes in C due to high or lower C max threshold will correspond to a similar change in Q d .C max was therefore a less influential parameter with low sensitivity in natural vegetation.Interception threshold, D, also showed low sensitivity to changes within the parameter range.D was computed on a daily basis using the interception threshold for various land use types derived from literature.However, the actual interception is more dependent on the daily variability of rainfall than the total interception threshold.Similar findings were observed by De Groen and Savenije (2006).While the interception thresh-  old is not an influential parameter, actual interception (I ) is still an important water balance component, as the water for the other processes is dependent on the net precipitation after interception (Makurira et al., 2010).

Interception and transpiration
There is general consensus that actual interception (I ) is a key component in hydrology and water management.I influences the net precipitation and therefore the amount of water available for evaporation (E + T ).Evaporation depletion (E + T ) influences the stream flow dynamics and is the manageable component of ET a in biomass production.Therefore, there is a need to distinguish E + T from the calculated I as a deficit of total ET a (SEBAL), Fig. 7.
The mean annual I ranged between 8 and 24 % of the total evapotranspiration.The land use types in the upper catchments, for example forest, rainfed coffee and bananas, had higher I .Irrigated sugarcane and natural shrublands located in the lower catchments had lower I .The variation is mainly influenced by the maximum threshold (D) and the rainfall (intensity and frequency) which are relatively higher for land use types in the upper catchments.The forest interception average estimate of 24 % of the total evapotranspiration (or 22 % of the total rainfall) is comparable with field measurements from previous studies that found forest canopy interception of about 25 % of the total rainfall in a savannah ecosystem in Africa (Tsiko et al., 2012).Q b contributions (e.g.irrigation) also enhanced the evaporation depletion (E + T ) component of ET a , resulting in relatively lower I for irrigated croplands.Any intervention to change I would influence antecedent soil moisture conditions, especially during small rainfall events (Zhang and Savenije, 2005).This may influence the productivity of E+T and/or the stream flow generation in the river basin.However, more research is required to estimate explicitly the changes in I from certain field-based interventions.The outcome of such studies maybe incorporated in the STREAM model.Rainfed crops and natural vegetation including the forests had a lower Q b , mainly stemming from groundwater (and snowmelts).Sparse vegetation, bushlands, grasslands, and natural shrublands had Q b contributions of less than 1 % of total ET a , while rainfed maize (middle catchments) and rainfed coffee (upper catchments) had Q b contributions of 2 and 7 % of ET a , respectively.Dense forest and afro-alpine forest had slightly higher Q b contributions (ranging between 7 and 9 %) attributed mainly to the availability of groundwater from snowmelts in the upper mountains.

Blue and green water use
Notably higher Q b was experienced in the dry year of 2009 (as shown by the error bars in Fig. 9).This is attributed to higher potential evaporation from relatively drier weather conditions.The lower precipitation during this period also resulted in increased groundwater use for the afro-alpine and dense forest land uses in the upper catchments.For instance, the Q b contribution to ET a for dense forest increased from 5 % in 2008 (a relatively wet year) to 10 % in 2009.The enhanced Q b for the irrigated croplands during 2009 is also attributable to the higher potential evaporation and limited precipitation that increased the irrigation water requirement.This is illustrated by irrigated sugarcane, where Q b increased from 35 % in 2008 to 55 % in 2009.Q b for supplementary irrigation also increased from 14 to 29 % during the dry year.The Q b for year 2010 was in general average for all land use types which is indicative of the average weather conditions that prevailed during the year.

Irrigation water use
This section presents the model results for supplementary irrigation water use (Q b(I) ) as estimated at various outlet points (gauging stations) in the river basin.The annual irrigation abstractions, predominant during dry seasons, were accumulated and the average mean for the period 2008-2010 is presented in Fig. 10.Six gauge stations and three additional points (accumulation points for Kikuletwa, Ruvu and Lake Jipe) were also considered.The annual net irrigation (in million cubic metres) was converted m 3 s −1 to allow easier comparison with the discharge data in Sect.4.1.
The Q b(I) ranges from 0.06 m 3 s −1 on the smaller streams to a total of 3.4 and 4.2 m 3 s −1 in the outlets of the Ruvu and Kikuletwa river systems, respectively.A significant irrigation abstraction of 1.5 m 3 s −1 was observed by the TPC sugarcane irrigation system, the largest single irrigation scheme in the river basin.The total Q b(I) upstream of NyM reservoir was estimated at 7.6 m 3 s −1 , which represents approximately 50 % of the low flows in the upper Pangani River basin.
Open canal irrigation is the commonly used irrigation technique in the upper Pangani River basin.There are an estimated 2000 small-scale traditional furrow systems, some 200-300 years old (Komakech et al., 2012).According to records at the Pangani Basin Water Office, approximately 1200 of these abstractions have formal water rights.PWBO estimates that the total gross irrigation abstraction is approximately 40 m 3 s −1 .The irrigation efficiencies of the irrigation systems range between 12 and 15 % (Zonal Irrigation office, Moshi).Here, we adopted a higher irrigation efficiency limit of 15 % to compensate for any uncertainties that may arise from the higher irrigation efficiencies in the larger irrigation schemes.The field estimates provide net irrigation consumptions of approximately 6 m 3 s −1 (using 15 % efficiency) and about 79 % of the Q b(I) model estimates (19 % efficiency).The water leaks in the traditional furrow canals flow back to the river system.The capacity and ability of the river basin authority to monitor actual water abstraction is limited.However, considering these uncertainties, the modelled net irrigation abstraction was reasonably close to field net irrigation estimates for the upper Pangani Basin.

Open water evaporation
The blue water use by the water bodies (Q b(w) ) upstream of NyM reservoir was also estimated using the modified STREAM model.Q b(w) is the net open water evaporation from blue water which would otherwise flow into the NyM reservoir.The water bodies considered include wetlands (98 km 2 ), Lake Jipe (25 km 2 ) and Lake Chala (4 km 2 ).The total mean Q b(w) was estimated to be 53.6 × 10 6 m 3 yr −1 (1.7 m 3 s −1 ) and 22.1 × 10 6 m 3 yr −1 (0.7 m 3 s −1 ) in the Ruvu and Kikuletwa river systems, respectively.The total Q b(w) (12 % of low flows) may also provide an insight into ecosystem services or benefits provided by the natural water bodies compared with the alternative uses, such as irrigation or hydropower in the downstream part of the river basin.

Future water management scenario using a modified STREAM model
The previous sections illustrate how the modified STREAM model provides spatial information on the water use (green and blue) under the current situation.The information is useful, especially in monitoring unregulated irrigation water use.
The model also provided useful information on the implication of future water use management scenarios in the river basin.Table 3 shows the real impact of the interventions on the water resources under the scenarios defined in Sect.3.6.The changes in surface runoff were evaluated from the outlet points (1dc & 1dd) upstream of NyM reservoir, upper Pangani River basin (Fig. 10).If soil evaporation is reduced in irrigation systems (Scenario 1), real water-savings of 37.8 × 10 6 m 3 yr −1 can be achieved.The additional water saved (4 % of total river flow), mostly groundwater flow, can be utilized in the expansion of the irrigated sugarcane (scenario 3).Scenario 1, alternatively, could also release additional base flow that may be required for other water uses that include environmental and/or downstream hydropower demands.Financing of the required interventions can also form a basis for basin-wide trade-off negotiations between downstream and upstream water users.
Scenario 2(a) investigates the implications of upscaling system innovations (SIs) for the rainfed maize cultivated in the highlands.In the area targeted, mixed farming of maize and coffee is practised, and covers an area of 72 300 ha (Kiptala et al., 2013a).Half of this area is under maize cultivation.This intervention would result in additional water use of 84 × 10 6 m 3 yr −1 , which is about 10 % of the total river flows.The model simulation shows that the water use will be derived from base flow.However, small-scale runoff harvesting devices can be used to store overland flow to supplement blue water needs during the dry seasons.Scenario 2(b) shows that an increase in both T and S u,max would result in slightly higher overland flow water use.This will not only increase water availability in the unsaturated zone for transpiration, but also reduce the soil and nutrient losses normally associated with higher overland flows.
In scenario 3, the increase in the sugarcane irrigated area by 7400 ha required an additional 53.9 Mm 3 yr −1 in average of base flow.The volume required for each year: 45.6 (2008), 68.6 (2009) and 47.4 Mm 3 yr −1 (2010) varied with the climate conditions.This is about 4, 11 and 6 % of the total river flows in 2008, 2009 and 2010, respectively.Any additional conveyance and drainage losses may increase the net water use.It was also observed that the total additional blue water required in scenario 3 is consistent with the current irrigation water use (Q b ) by the existing irrigation system.

Conclusions
This paper presents a novel method of developing a spatially distributed hydrological model using blue and green water use at pixel scale.This methodology allows for unprecedented insights into hydrological processes at smaller scales of land use classes.The hydrological model was developed for a heterogeneous, highly utilized and data-scarce landscape with a sub-humid and arid tropical climate.The blue water use was quantified by employing a time series of remotely sensed evapotranspiration data as input in the STREAM model.The model was also constrained by satellite-based soil moisture estimates that provided spatially (and temporally) realistic depletion levels during the dry season.To further enhance model parameter identification and calibration, three hydrological landscapes -wetlands, hillslope and snowmelt -were identified using topographic data and field observations.Unrealistic parameter estimates, found for example in natural vegetation either through overestimation of satellite-based data or model structure, were corrected in the model conceptualization through the water balance (at pixel scale).The modified STREAM model provided a considerably good representation of supplementary blue water use, which is dominant in the upper Pangani River basin.
The model performed well on discharge, especially in the simulation of low flows.The Nash-Sutcliffe coefficient (E ns_ln ) ranged between 0.6 and 0.9 for all outlet points in both calibration and validation periods.At the outlet, the model performance was best (E ns_ln = 0.90).The large difference between MAE and RMSE was indicative of large errors or noisy fluctuations (see Figs. 5 and 6) between actual and simulated discharges during the rainy seasons.This was mainly attributed to the uncertainties of the remote sensing data during cloudy periods.The uncertainties may also have been exacerbated by possible errors in the hydrometeorological data and model conceptualization.Model parameters that were freely calibrated for different land use, such as maximum unsaturated and saturated storages (S u,max , S s,max ), separation coefficient (c r ), and quick flow coefficient Hydrol.Earth Syst.Sci., 18, 2287-2303, 2014 www.hydrol-earth-syst-sci.net/18/2287/2014/ (q c ) resulted in high sensitivity.The model calibration of these parameters can be improved in the future by field measurements or by analytical relationships.The simulated net irrigation abstractions were estimated at 7.6 m 3 s −1 which represents approximately 50 % of low flows.Model results compared reasonably well with field estimates with less than 20 % difference.In addition, the model yields spatially distributed data on net blue water use that provides insights into water use patterns for different landscapes under different climate conditions.Blue water use (Q b ) contribution to ET a during a dry year (2009) increased from 5 to 10 % for dense forest, 35 to 50 % for the wetlands and irrigated sugarcane, and 14 to 28 % for supplementary irrigation compared to the wet year (2008).Three water management scenarios on water-saving and increased crop productivity were also explored using the STREAM model.Reduced soil evaporation of 15 % in supplementary irrigation systems would result in real blue water-savings of 37.8 × 10 6 3 yr −1 (4 % of total river flows).The watersaving could alternatively be used to expand the existing sugarcane irrigation scheme (7400 ha on sugarcane) that required 6 % of total river flows if its area is doubled.Upscaling of systems innovation for highland rainfed crops to achieve a 30 % increase in productive T resulted in an additional blue water requirement of 84 × 10 6 m 3 yr −1 .The additional water requirement can be generated from runoff harvesting and storage to save on the already over-exploited blue water resources.This information may form a basis for socioeconomic trade-off analysis, on the basis of which various basin strategies and financial mechanisms can be formulated for efficient, equitable and sustainable water resources management at the river basin.
The development of advanced methods of generating more accurate, remotely sensed data should go hand in hand with ways to improve distributed hydrological models.Such methods may include the use of passive microwave imagery to generate cloud-free ET a estimates (Bastiaanssen et al., 2012).Future modelling improvements should also aim at simulating the model for longer time series, using long-term rainfall and RS data (evapotranspiration and soil moisture).The data could be based on stochastic or probabilistic techniques (Salas et al., 2003).In so doing, data can be interpreted in a way that is useful for management and decisionmaking.

Figure 1 .
Figure 1.Overview of entire Pangani River basin and the upper Pangani River basin.

Figure 2 .
Figure 2. Modified STREAM conceptual model for upper Pangani River basin.

Figure 3 .
Figure 3. Soil moisture depletion fraction (defined using average values of the dry month of January of 2008, 2009 and 2010) in the upper Pangani River basin for selected land use types.

Figure 4 .
Figure 4. (a) Wetland-hillslope (snowmelt) hydrological system.(b) Shallow groundwater observation wells with mean surface water levels (0.3-40 m) in the lower catchments of the upper Pangani River basin for the period 2008-2010.

Figure 5 .
Figure 5. (a) Comparison of observed (Q o ) and the simulated discharge (Q s ) and precipitation at the outlet points for calibration period 2008 (8-day periods 1-46) and validation 2009 and 2010 (8-day periods 47-138) in the upper Pangani River basin; (b) the corresponding scatter plots of Q o and Q s for four upstream gauge stations.

Figure 6 .
Figure 6.(a) Comparison of observed (Q o ) and simulated discharge (Q s ) and precipitation for calibration period 2008 (8-day periods 1-46) and validation 2009 and 2010 (8-day periods 47-138) in the upper Pangani River basin; (b) the corresponding scatter plot of Q o and Q s for the most downstream gauge station.

MeanFigure 7 .
Figure 7. Mean interception, I , and evaporation depletion, E + T for different land use classes in upper Pangani River basin for the period 2008-2010.

Figure 9 .
Figure 9. ET a and the corresponding Q g and Q b water use for selected land use types averaged per year over 2008-2010 in the upper Pangani River basin (error bar indicates the upper and lower bounds for mean Q b for dry year 2009 and wet year 2008, respectively).

Figures
Figures 8c and 9 show the resultant blue water use (Q b ) and the direct contribution of precipitation (Q g ) to the ET a (actual evapotranspiration) for various land use types.Q b is closely related to the land use and the ET a as observed in Fig. 8a and b.Water bodies (lakes and reservoir) and the wetlands have the highest Q b , contributed by the high open water evaporation.The averageQ b for water bodies is approx.56 % of the ET a with a maximum of 74 % (1642 mm yr −1 ) observed at the lower end of the NyM reservoir.The Q b is high in the NyM reservoir because of the high potential evaporation attributed to hotter climatic conditions and lower precipitation levels in the lower catchments.Wetlands and swamps located in the lower catchments also resulted in high Q b of approximately 42 % of ET a .In irrigated croplands, the Q b was also moderately high, with a range of between 20 % for irrigated mixed crops and bananas in the upper catchments, and 44 % for irrigated sugarcane in the lower catchment.Rainfed crops and natural vegetation including the forests had a lower Q b , mainly stemming from groundwater (and snowmelts).Sparse vegetation, bushlands, grasslands, and natural shrublands had Q b contributions of less than 1 % of total ET a , while rainfed maize (middle catchments) and rainfed coffee (upper catchments) had Q b contributions of 2 and 7 % of ET a , respectively.Dense forest and afro-alpine forest had slightly higher Q b contributions (ranging between 7 and 9 %) attributed mainly to the availability of groundwater from snowmelts in the upper mountains.

Figure 10 .
Figure 10.Total net irrigation water use (Q b(I) ) estimated upstream of the gauge stations using a modified STREAM model in the upper Pangani River basin (averaged over the years 2008-2010).
. The first topographic indicator, H (elevation above the nearest open water) is used in this study.H is derived from the level where groundwater storage is low or

Table 2 .
Sensitivity of model performance due to change in model input parameters.

Table 3 .
Impact of three water management scenarios on the surface runoff.