Implementing contour bank farming practices into the J2000 model to improve hydrological and erosion modelling in semi-arid Western Cape Province of South Africa

Due to low annual precipitation volumes and the high spatio-temporal variability in precipitation, semi-arid areas are often challenged by limited water resources. In addition population growth, increasing per capita water demand and large scale irrigation practices result in further pressure being placed on water resources. This occurs particularly in areas which experience inadequate land management practices and the combined effect of these factors often causes severe environmental problems such as desertification and droughts, soil erosion and salinisation and the loss of biodiversity (Hughes 2008; Wheater 2008). To better understand and assess such impacts, innovative tools are required, which aim to sustainably manage the limited land and water resources. Integrated hydrological models have the potential to support management decisions by providing information on system dynamics and change impacts (Abbott et al. 1986; Wheater 2008; Biondi et al. 2012). To sufficiently address land management practices and environmental conditions, the underlying model concept of such models needs to be process-based, i.e. needs to represent the processes controlling lateral as well as vertical fluxes (Hughes 2004; Arnold & Fohrer 2005).

One of the most significant environmental problems occurring within the semi-arid south-western Cape region of South Africa is the process of soil erosion (Meadows 2003). According to Meadows & Hoffman (2002) the mean annual soil erosion caused by water in South Africa ranges from 0.8 to 4 tonnes per hectare (t ha⁻¹), which may lead to severe land degradation. To prevent and reduce land degradation, contour bank farming (also called contour banking) is commonly practised as a management instrument of water and soil conservation works within the agricultural areas of the Western Cape Province. These earthen structures are constructed perpendicular to cultivated slopes, as well as at pre-determined intervals down slopes in order to reduce slope lengths, flow velocity of water and to intercept and prolong surface runoff before it causes erosion (Wakindiki et al. 2007). Contour banks also channel surface runoff into stable waterways, natural depressions or grassed areas adjacent to a paddock (Department of Environment and Resource Management (DERM) 2004; Figure 1).

Contour bank channels have a small gradient (0.1–0.4%) to prevent channel flow reaching high and erosive velocities in non-vegetated areas. The distance between contour banks is usually determined by the local slope. Other factors influencing the distance between contour banks are the soil type, the applied cropping practice and previous erosion events. According to DERM (2004) and Freebairn (2004) two different types of contour bank cross sections exist. The narrow-based type of contour bank is used on steeper cultivated slopes (5–12%) with batters which are too steep to be cultivated and are usually covered with grass. The channels of narrow-based contour banks are normally part of the contour bay and cultivated and used for crop production. Broad-based contour banks are usually built on gentle slopes (<5%), with deep soils and are often cultivated (DERM 2004).

The impact of contour bank farming on flood reduction and sediment dynamics has been studied in various semi-arid Mediterranean basins (Kingumbi et al. 2004; Nasri et al. 2004a, b; Leduc et al. 2007; Nasri 2007; Baccari et al. 2008; Lesschen et al. 2009; Ouessar et al. 2009) and in Australia (Callow & Smettem 2009; RPS 2010). Nasri et al. (2004a) published results from an experimental catchment in Tunisia showing that soil conservation techniques significantly reduce flood risks due to the reduction of surface runoff on steep slopes. At a larger scale, Nasri et al. (2004b) demonstrated that the introduction of contour ridges in a semi-arid, medium-sized (18.1 km²) basin led to a decrease of catchment outflow by 50–80%, a reduction in peak discharge by 60–90% and the virtual disappearance of observed erosion. Similar results regarding the quantification of water holding capacities and efficiencies of benches in Tunisian catchments were reported by Nasri (2007), Leduc et al. (2007) and Lacombe et al. (2008). Some studies (Nasri 2007; Baccari et al. 2008) also showed that contour banks lose their efficiency in reducing runoff as a result of sedimentation and filling of the sinks over time.

Studies from Australia indicated that contour banks have trapping efficiencies ranging from 54 to 84% (compared to areas without contour banks) and therefore reduce the direct sediment transport into the stream (RPS 2010). It was also demonstrated that 10–40% of transported material is deposited on the hillslope before reaching the contour bank channel. Furthermore, Callow & Smettem (2009) investigated the effect of farm dams and man-made earth banks on hydrologic connectivity and runoff estimation in 12 sub-catchment basins in a dryland agricultural region. Based on geomorphic and hydrologic descriptors (cumulative area distribution, hypsometric curve, simplified width function and instantaneous unit hydrograph) they have shown that the basin hydrology, in particular hillslope processes, is significantly altered by farm dams and earth banks.

Although these studies demonstrate the impact of contour banks on runoff generation and sediment fluxes, their successful implementation in distributed and process-based model exercises lacks evidence in previously published studies. However, Ouessar et al. (2009) introduced a modified version of the SWAT (Soil and Water Assessment Tool; Arnold et al. 1998) model to assess the impact of water harvesting systems (earthen dikes) on hydrological processes in semi-arid environments. The authors conclude that the model reproduces all water balance components sufficiently, but is limited in representing the spatial variability of runoff generation mechanisms. By integrating hillslope and plot features influencing the hydrologic connectivity in basin modelling, Lesschen et al. (2009) successfully applied the LAPSUS model to predict realistic values for total runoff and sediment yield at catchment scale for one single event. All studies show that there is a clear indication that contour banks change natural flow paths and inner basin hydrological process dynamics and that they have to be integrated into distributed, process-based modelling to allow for realistic assessments of water and nutrient transport dynamics in regions affected by contour bank farming. Despite those rare, but promising, model approaches, the problem of erroneous runoff and erosion modelling in areas with contour banks due to the lack of suitable modelling tools still requires research efforts (Jetten et al. 1999; Bugan et al. 2009; Lesschen et al. 2009).

This study was conducted in the Swartland region in South Africa (Western Cape), where contour banks which total approximately 25,000 km in length were constructed between 1972 and 1995 (Meadows 2003). Little is known, however, about their quantitative impact on runoff generation and erosion control. Bugan et al. (2009) concluded in their modelling studies within the Sandspruit catchment (a dryland agricultural catchment in the Swartland region) that the contour banks have a significant influence on the catchment water balance, and also that common model approaches are not capable of representing contour bank farming.

The aim of this study is to describe how contour banks may be integrated to improve distributed, process-based hydrological and erosion modelling in semi-arid environments which are affected by contour bank farming. Thus, the J2000 hydrological model has been enhanced with the implementation of contour bank and erosion modules. The contour banks have been considered in the delineation and routing of spatial model entities. The developed model tools were tested and used to simulate runoff generation and sediment transport and to assess the topological control of the contour banks in the Sandspruit catchment.

The Sandspruit catchment (152 km²) is located in the Western Cape Province, South Africa (Figure 2) and is a tributary basin of the Berg River. The region is characterised by a Mediterranean-type climate with warm dry summers and cool moist winters (Department of Water Affairs and Forestry (DWAF) 2007; Bugan et al. 2009). The mean annual temperature is approximately 19 °C. Precipitation dominantly occurs during the winter months (April–October) in the form of long duration and low intensity frontal rain from the north-west. Summer precipitation is predominantly convective, and thus rainfall events are generally highly variable in space and time (Meadows 2003; Kruger et al. 2010). The mean annual precipitation exhibits large spatial variability as shown for four stations located in the study area (Table 1, DWAF 2007; Bugan et al. 2009). The spatial variability of precipitation is a function of both topography and the distance from the coastline. Bugan et al. (2012) presented rainfall data for stations located within the Sandspruit catchment. Analysis of these data indicates that a rainfall gradient exists across the catchment, i.e. decreasing from south to north. In 2009, the annual rainfall recorded in the upper- (south), mid- and lower-parts (north) were 494, 391 and 319 mm respectively. In addition to the higher elevation, the higher rainfall in the upper parts of the catchment is also interpreted to be a result of orographic rainfall caused by Kasteelberg. In terms of rainfall shadowing effects, Kasteelberg is not interpreted to have any effect to its west and north as the rainfall approaches from the north-west. The area to the east and south falls outside of the Sandspruit catchment and hence was not included in this study. The Quaternary catchment G10J (including the Sandspruit catchment) is reported to exhibit a potential evapotranspiration (ET) range between 1,500 and 1,700 mm/a. (Midgley et al. 1994).

The topography within the catchment is undulating with elevations ranging from approximately 30 metres a.s.l. (near the confluence with the Berg River) to 950 metres at the Kasteelberg Mountain in the southern part of the catchment. The Sandspruit catchment comprises sequences of rocks of the Malmesbury Group, the Table Mountain Group and younger Cenozoic deposits. The Malmesbury Group comprises erodible, soft rocks within the mid to lower parts of the catchment which are steeply folded along a north-west striking axis. The hard and resistant quartzitic sandstones of the Table Mountain Group, which act as secondary water-bearing features, occur only in the southern parts of the catchment around the Kasteelberg. The Cenozoic sediments in the northern and north-western parts of the catchment are the result of the Tertiary sea-level transgression and regression processes and consist primarily of aeolian sands, gravel and clayey deposits (DWAF 2007). According to Bugan et al. (2009) the soils within the catchment are shallow and poorly developed with soil thicknesses between 0.5 and 1 m. Almost 70% of the catchment area is dominated by the lithic Glenrosa (Gs) soil form which is highly erodible. These soils appear predominantly on the substrates of the Malmesbury and Table Mountain Groups and are pre-eminently associated with convexity and the processes of divergent water flow (Görgens & de Clercq 2006; Fey 2010). Along the streambed of the Sandspruit, the Dundee soil form is dominant and covers approximately 10% of the catchment area. These soils consist of varying shallow to deep soil deposits (alluvial material). The stratified sandy layers, varying in texture, colour and thickness are the result of continuous deposition of soil material by floods. Within the northern parts of the Sandspruit streambed, the moderately deep and non-gravelly, wet saline duplex soils of the Estcourt soil form are dominant. This soil form has a low agricultural production potential especially during the wet season (Lambrechts 2007). The north-eastern parts of the catchment are dominated by the soils of the Swartland soil form, which developed above the weathered Malmesbury shales and have a higher agricultural potential with depths partially exceeding 50 cm. Furthermore there are minor occurrences of colluvial soils of the Oakleaf form with a high production potential in the northern and eastern parts of the catchment (Lambrechts 2007). The land use within the catchment is dominated by cultivated lands and pastures (approximately 145 km²) with dryland winter wheat and canola being the dominant vegetation types (Bugan et al. 2009). To minimise the detachment of soil particles from erosive runoff and the transport of the resulting soil material, farming utilising contour banks to prevent gully formation is applied as a land management technique within the catchment.

Hydro-meteorological data for model input and calibration parameters were available from different sources. Since no climate station is located in the Sandspruit catchment, climate data from four nearby stations (Figure 1) were obtained from the Drysal River Basin Information System (DrysalRBIS, link: leutra.geogr.uni-jena.de/drysalRBIS/), the Agricultural Research Council (ARC) and the Department of Water Affairs (DWA) in South Africa. Daily maximum, minimum and mean air temperature [°C], precipitation sums [mm], relative humidity [%], wind speed [m s⁻¹] and sunshine duration [h] were available for the De Hoek, Langgewens, Moorreesburg and Voëlvlei stations (Figure 1). All climatological time series were edited regarding homogeneity, consistency and gaps and corrected using regression analysis. As a result, corrected data sets were available for the hydrological years from 1986 to 2010. Daily time series of river discharge [m³ s⁻¹] were available for the gauging station G1H043 (Vrischewaagd) (Figure 2). Runoff records were also tested for homogeneity and consistency and corrected using regression analysis. In addition, non-continuous records of soil water and groundwater measured at several sites across the study area were available for model calibration and validation.

Geographic Information System (GIS) data are available for the study area from different sources, i.e. soils (Görgens & de Clercq 2006), land cover (Council for Scientific and Industrial Research (CSIR) and ARC 2005) and geology (Visser 1989). A Digital Elevation Model (DEM) derived from the Shuttle Radar Topography Mission (SRTM) with 90 m resolution was obtained from the US Geological Service (USGS 2003), providing slope, aspect, river network and watersheds. Comparing the extracted river network with existing data (e.g. Google Earth images (Google Inc. 2011)) revealed inconsistencies in the generated river network. Thus, the DEM was corrected using a river network available from topographic mapping and filling. In addition, all GIS data products, including the location of climate stations and gauges, were projected to Universal Transverse Mercator Zone 34 South.

During the analysis of the different spatial data sets, inconsistencies and discrepancies such as soils on bare rock areas or soils without hydromorphic characteristics in floodplain areas were identified. Those areas were extracted and reclassified using neighbourhood functions and cross-validation as well as visual comparisons to identify the respective class of highest probability.

An enhanced version of the process-oriented, spatially distributed J2000 modelling system (Krause 2001) was used for the simulation of the hydrological and erosion dynamics at catchment and model unit scales. For a better representation of spatially explicit hydrological dynamics and linkages, J2000 is following a fully distributed modelling approach. Here, the catchment is subdivided into Hydrological Response Units (HRUs) (Flügel 1996) which are delineated according to Wolf et al. (2009). Represented as individual polygons, the delineated HRUs are linked to each other in an explicit topology which is in contrast to semi-distributive modelling approaches as used, for example, by the SWAT (Arnold et al. 1998).

The J2000 modelling system allows the simulation of the energy and water balance of a watershed across temporal (hourly to annually) and spatial (micro to macro) levels of aggregation. Hydrological processes modelled include interception, ET, snow accumulation and ablation, horizontally differentiated soil- and groundwater balance, and distributed runoff generation by explicitly computed lateral flows and flood routing in the catchment's river network (Krause et al. 2006).

All relevant hydrological processes are implemented as individual process modules (Krause 2002). The interception module uses a simple storage approach following Dickinson (1984), which calculates a maximum interception storage capacity based on the Leaf Area Index (LAI) of the respective land use. As long as this maximum capacity is not exceeded, precipitation is stored in the actual interception storage, which is depleted by evaporation. When the maximum storage capacity is exceeded, any surplus of rainfall is treated as throughfall and passed to the next module. Potential ET is calculated using the Penman–Monteith equation. Parameters required to solve the Penman–Monteith formula are provided by input data derived from the given climate stations. In the case of missing variables, J2000 offers algorithms to calculate variables such as net radiation balance, real and saturated vapour pressure, heat fluxes and absolute humidity (Krause 2001).

The J2000 soil-water balance module (Figure 3) represents the soil zone as a regulation and distribution system and interacts with nearly all other J2000 process modules. Based on empirical equations, the module separates precipitation into Runoff Direct 1 (RD1, surface runoff), water stored in depression storages (once the maximum infiltration rate is exceeded) and infiltrating water. This separation, takes the soil infiltration capacity, maximum infiltration rate, surface conditions (structure and slope) and calibration parameters into account. To simulate water movement in the unsaturated soil zone, the module implements a unique water storage concept based on two different compartments. It divides between the middle pore storage (MPS) in which stored water is held against gravity and only drained by an active tension, and the large pore storage (LPS), which is not able to hold water against gravity. The total amount of outflow from the LPS is calculated by a nonlinear relationship taking the relative saturation of the storage into account. It is split into (1) the horizontal component Runoff Direct 2 (RD2), which is interflow from the unsaturated soil zone, and (2) the vertical components (percolation, ET) with respect to topographic (e.g. slope) as well as soil (e.g. soil depths) parameters.

The groundwater module incorporates two storages representing the shallower weathering zone and a deeper aquifer (Figure 3) for each spatial model unit. The water input received from percolation into the groundwater module is distributed among the two storages based on slope and a calibration parameter. The outflow from the two storages occurs either as Runoff Groundwater 1 (RG1) which could be seen as slow interflow, i.e. water movement in the shallower weathering zone, or Runoff Groundwater 2 (RG2) which is baseflow. RG2 represents the water movement in the deeper aquifer and/or in fractures and is computed based on the actual storage content, a recession coefficient, and a global calibration parameter (Krause 2001; ILMSWIKI 2012).

Runoff is simulated at HRU level, with subsequent calculation of runoff concentration processes (through a lateral routing scheme) and flood routing in the stream channel network (Figure 2). All runoff components (surface runoff (RD1), interflow from the unsaturated soil zone (RD2), interflow from the saturated weathering layer of the underlying hydrogeological unit (RG1), and saturated baseflow (RG2)) are simulated for each HRU and for the total watershed (Figure 2). After calculation of the runoff generation processes, runoff concentration is calculated based on topological routing of the HRUs (Pfennig et al. 2009). The routing of water between adjacent HRUs is realised by using a multi-dimensional approach developed by Pfennig and Wolf (2007), and is an extension of the one-dimensional approach introduced by Staudenrausch (2001). Here, each runoff component generated on a single HRU is laterally connected either to a receiving HRU or to a receiving stream reach. The routing inside the stream network is simulated by connecting the reach storages, receiving the water from the topologically connected HRUs by a hierarchical storage cascade and calculating the flow velocity inside the channel using the Manning–Strickler equation. Thus, the travel time is primarily controlled by only a roughness coefficient. The outflow of each stream segment is transferred as inflow to the downstream segment. All runoff is accumulated as the basin runoff at the reference gauge at the watershed outlet.

J2000 also offers routines for the regionalisation and correction of climate data used in the process modules, based on statistical tests (regression analysis) and geostatistical approaches (Inverse Distance Weighting) (Krause 2001). Tools for sensitivity and uncertainty analysis (Fischer et al. 2009), accuracy assessment and output capabilities to export or visualise model results are also incorporated (Krause et al. 2006).

A detailed description of all process modules of J2000 is given in Krause (2001) and Krause (2002). Figure 3 gives an overview of the concept of J2000.

The delineation of spatially distributed model entities, which integrate hydrological relevant characteristics of landscape features such as topography, land use, soil and geology, is based on the modification of the HRU approach (Flügel 1996). To better represent topography in the hydrological modelling, the HRU approach was modified by incorporating a clustering of selected topographic indices derived from the SRTM-DEM as shown by Wolf et al. (2009). The topographic indices included are the Topographical Wetness Index (TWI) (Kolberg 1997; Böhner et al. 2002; Wolf et al. 2009), the Annual Solar Radiation Index (ASR) (McCune & Keon 2002) and Mass Balance Index (MBI) (Möller et al. 2008; Wolf et al. 2009).

The TWI is used to describe the spatial pattern of depth of the water table and spatial distribution of soil moisture where a high TWI corresponds to a moist area with rapid hydrological response (Kolberg 1997; Wolf et al. 2009). The ASR estimates the potential direct incident radiation, which represents an improved relation to potential evaporation by integrating the slope, exposition and latitude (McCune & Keon 2002; Wolf et al. 2009). The MBI is a relative index for the determination of erosion-, transit- and accumulation zones. Furthermore it represents the spatial distribution of soil water (Wolf et al. 2009). The clustering was performed by a complete linkage cluster analysis using a program for iterative combining of homogeneous areas (IVHG) (Friedrich 1996; Metzler 2002). Here, agglomerative distance procedures are linked with a spatial neighbourhood analysis. With this approach, only grid cells having a very small displacement to each other in the multivariate space and being spatial neighbours within the data matrix are aggregated. In order to derive homogeneous relief units the distance grouping method, using the smallest Euclidian displacement, is applied (Friedrich 1996, 1998). More details on the process of clustering are documented in Pfennig & Wolf (2007). To represent the basin heterogeneity by process-oriented HRUs as described by Flügel (1996), the spatial model entities derived from clustering were merged with spatial information relating to soil, geology and land use. Here, re-classified layers of soil (Classes: Glenrosa, Oakleaf, Swartland, Estcourt and Dundee), land cover (Meadows and Pastures, Deciduous Forest, Mixed Forest, Farmland and Shrubs) and geology (Fine Sediment, Granite, Greenstone, Quartzite, Shale, Silcrete-Ferricrete), which are considered as relevant for hydrological process dynamics in the Sandspruit catchment (Flügel 1995; Bugan et al. 2009), were overlaid with the output of the clustering. To address relationships between runoff generation mechanisms and storage dynamics, a topological routing was performed for lateral water transport modelling between HRUs, determining the watershed spatial connections between two HRUs based on a n:m topology, or between HRUs and stream reaches, or between stream segments (Pfennig et al. 2009).

The study also revealed that contour bank farming is widely restricted to slopes ranging between 6 and 18%. Using these criteria the contours were calculated for all HRUs. Since the mean slope length for each HRU is required for erosion modelling, the mean HRU slope length was then divided by the number of segments of contour lines crossing HRUs. Finally, the HRUs with assigned contour banks had to be routed to the adjacent stream segment by following the direction towards the steepest slope (O'Callaghan & Mark 1984). For this purpose, the stream network was extended allowing the direct linking of contour banks in areas without direct linkage to the stream network.

J2000 parameterisation requires the generation of input parameter files and the parameterisation of the process modules. With respect to the classes used for the HRU delineation, physical parameters for each class of land cover (e.g. albedo, monthly surface resistance for water saturation, LAI, efficient vegetation height, rooting depth, sealed grade), geology (e.g. maximum storage capacities, storage coefficient) and soil (e.g. soil depth, permeability coefficient, air capacity, usable field capacity) were taken from Bugan et al. (2009). HRU related parameters such as position parameters, elevation, slope and aspect, class IDs (soil, geology, land cover) and routing parameters were derived from GIS analysis. Model parameter values for the process modules were also taken from Bugan et al. (2009), but adjusted following assessment of parameter sensitivity during the calibration process.

The initial model calibration was realised for J2000 with the contour bank module and a multi-dimensional flow routing. The model was calibrated for the period of 1989–90, in which 1989 was used for model initialisation. So as to test overall model performance with an uncalibrated model run and to allow a better assessment of model results, validation was performed for 1991. However, the year 1990 is considered as a representative year with a total precipitation of 502 mm (mean 1987–2008: 500 mm), and a mean daily runoff of 0.36 m³ s⁻¹ caused by potentially erosive, low frequency high runoff events with local peaks between 7–9 m³ s⁻¹ (e.g. 8.96 m³ s⁻¹ on 10th July, 7.92 m³ s⁻¹ on 29th June and 8.37 m³ on 12th July; see Figure 5). Thus, the year 1990 was selected to assess the functioning of the modules rather than using the full period which would entail a change of focus from the introduction of the contour bank and erosion modules as model features to a more general discussion of the long-term hydrological dynamics. With regard to Janssen & Heuberger (1995), the calibration process was conducted combining manual and automated techniques. Initially, calibration was performed by a trial and error procedure, i.e. various model parameters ensembles were tested and evaluated regarding the representation of hydrological dynamics. A sensitivity analysis was done following the approach of Wheater (2008) and Beven (2008). Here, 10,000 Monte Carlo iterations were performed for 25 model parameters with the aim of identifying the most sensitive J2000 parameters as well as the specific parameter optimum within the manually predefined sampling interval. Based on the visual inspection of the hydrograph and objective functions, the parameters with a high sensitivity were finally re-calibrated.

Four accuracy criteria were considered for model accuracy assessments and optimisation during the calibration and validation process. In accordance with Nash & Sutcliffe (1970), Janssen & Heuberger (1995), Krause et al. (2005), Willmott (1981) and Wagener et al. (2003), the Absolute Volume Error (AvE), the Nash-Sutcliffe Efficiency (E2), Index of Agreement (IOA), and the Double Mass Gradient (DsGrad) were used to evaluate overall model performance and to iteratively optimise model parameters driving dominant hydrological processes.

The model results derived from the calibrated model run for the year 1990 were used for the assessment of the impact of contour banks on model performance. The hydrographs representing daily simulated and observed runoff for the year 1990 (Figure 5) show a good overall representation of temporal runoff dynamics and a reasonably good correspondence between measured and simulated runoff. It is demonstrated that the model is capable of representing the ephemeral characteristics of the stream during summer months with low baseflow, high ET rates and occasional runoff response resulting from convective summer rainfall events. In contrast, high runoff peaks show a good temporal agreement with rainfall events during the rainy season, even though slight over predictions occur at the beginning of the rainy season whereas the peaks in the late rainy seasons are under predicted. In terms of precipitation distribution, the west–east rainfall gradient, a result of the distance from the coastline, was accounted for by utilising data from Langgewens and Moorreesburg (west), as well as data from Voëlvlei Dam and De Hoek (east) into the hydrological modelling process. As meteorological time series data, measured within the Sandspruit catchment were not available for use in this study, the south–north rainfall gradient could not be accounted for. This was considered when evaluating the model results.

The accuracy assessment using objective functions confirms the successful modelling. An E2 of 0.65 indicates a fairly good representation of runoff response by the model. The IOA, which is proposed to resolve insensitivities of E2 to additive and proportional differences of the observed and predicted means and variances (Krause et al. 2005), is 0.80, which confirms the sensitivity of the model to extremes. However, the absolute volume error of 0.72 mm, showing that the annual water balance is nearly cleared, also demonstrates the good model performance. With respect to various guidelines and hydrological model assessment studies (Legates & McCabe 1999; Krause et al. 2005; Moriasi et al. 2007), all model efficiencies as well as visual interpretations and information on model-based processes (e.g. ET) met the author's criteria for a successful calibration. Thus, model results were used for the subsequent assessments of basin hydrology and the impact of contour bank farming on erosion dynamics.

The good model results suggest that the integrated analysis of simulated runoff components, rainfall pattern, topography and soils, provides information on spatio-temporal dynamics of runoff generation and erosion. The separation of individual runoff components further shows that subsurface runoff, i.e. fast interflow (RD2) is the dominant runoff component throughout the year. In 1990, interflow contributed 68.58% to the annual basin runoff, whereas RG1 and RG2 contributed 13.4 and 11.45% respectively and the fraction of surface runoff (RD1) is 6.57%.

The analysis of the spatio-temporal runoff pattern and individual events also indicates that seasonal runoff generating processes are controlled by the frequency and intensity of rainfall, as well as by soil properties and topographic features. Figure 5 reflects the typical seasonal flow pattern of the Sandspruit catchment. The majority of runoff (96%) is generated during the rainy season (April–August). In the dry season, runoff is (basically) very low. During convective summer rainfall events, surface runoff (RD1) becomes a dominant runoff component in areas without contour banks when rainfall intensities exceed maximum soil infiltration capacities. Surface runoff dominated areas were identified along footslopes and floodplains, in particular in the north-eastern parts of the catchment. This is in agreement with areas characterised by Swartland and Estcourt soil forms. The tendency of these clayey-loamy soils to waterlogging due to low air capacities and low pore volume during wet conditions, as well as their tendency to crust under dry conditions (Moberly & Meyer 1984) supports the generation of surface runoff. To reduce the effect of possible surface crusting on overland flow during convective precipitation events, a parameter controlling the maximum seasonal infiltration amount was implemented. It was also possible to take into account special infiltration conditions for short duration and high intensity convective precipitation events. During dry periods, runoff is only controlled by subsurface components which are fairly balanced.

The basin runoff significantly increases at the beginning of the rainfall season in late April (Figure 5). At this time infiltration and the corresponding subsurface flow processes become the dominant controls of runoff generation. As shown in Figure 6, the fast interflow (RD2) is the major source of catchment discharge during the winter rainfall season. It correlates strongly with the observed basin runoff (Figure 5) and the storage dynamics (Figure 7). As visible in Figure 7, a higher frequency of precipitation events and significantly higher quantities of precipitation (up to 35 mm/day) results in the rapid filling of the middle pore storage (MPS) of the generally sandy-loamy and loamy sand soils with moderate available moisture capacities (Moberly & Meyer 1984), reaching saturation in late April. As a consequence of the saturation of the MPS, water is transferred to the LPS, which is the source of the lateral subsurface flow (RD2), as well as percolation into the groundwater.

It also appears that flow components generated in the saturated zone, play only a minor role due to high ET rates, infiltration excess flow, and filling of the large pore storage (LPS) during rain events that exceed 25 mm. As such, it can be stated that the relative contribution of slow interflow (RG1) and baseflow (RG2) to catchment runoff, will increase following an increase in precipitation and percolation. This increase in interflow and baseflow is therefore caused by an increased saturation of the soil-water storage during the winter rainfall season and the decrease of interflow and baseflow by the recession periods afterwards.

Applying the modified soil erosion module, a total soil loss of 5.89 t was predicted in the Sandspruit basin for 1990. The predicted sediment loss was limited to about 7% (10.6 km²) of the basin area, in particular for those HRUs not managed by contour bank farming. The spatial assessment of the modelling results indicates that the range of soil loss varies between the HRUs depending on topography and soil characteristics (coarse fragment factor). The mean soil loss in areas without contour bank farming is 0.06 t ha⁻¹ and the maximum soil loss is 0.41 t ha⁻¹. In terms of spatial distribution, the majority of erosion affected areas are located under areas of loamy sands of the Glenrosa soil form, which is dominant within the catchment (∼70%), and on gentle slopes under agricultural use. The highest rates of soil loss are predicted on single model entities characterised by the Swartland and Estcourt soil form in the north-western parts of the catchment. Due to their texture, the erodibility ratings (Moberly & Meyer 1984) of these soil forms are characterised as either medium (Swartland soil form) or high (Estcourt and Glenrosa soil forms).

The seasonal assessment of erosion dynamics shows that sediment transport within the stream is strongly correlated with the runoff generation dynamics. The simulated catchment sediment yield at the outlet is 2.35 tons for the year. The highest sediment yields are predicted in May and July when surface runoff is generated due to excess infiltration flow.

To assess the influence of the implemented contour bank module on hydrological modelling and soil loss prediction, the calibrated model was applied with the same parameter set and an inactive contour bank module.

Application of the model with inactive contour bank module to the year 1990 yields a significant decrease in the model performance. The Nash–Sutcliffe efficiency (0.47) and the IOA (0.77) are significantly lower than with an active contour bank module, and the absolute volume error also increased by up to 46.16 mm (see Hydrological modelling results and accuracy assessment section).

By ignoring the contour bank management practice, the fast surface runoff constitutes nearly the only runoff component during the first two months of the year. Comparing both models, it is apparent that the inclusion of contour banks significantly reduces the simulated peak flow caused by events with high precipitation input. A significant overprediction of the measured peak flow events is observed when applying the model without contour banks, in particular during the late autumn and early winter months of April and May. As a consequence, a higher simulated catchment runoff volume is predicted for this period, caused primarily by missing contour bank storages which would otherwise have retained some of the additional water. As is visible in Figure 8, both models are significantly under-predicting the high runoff event of June 29, 1990. One explanation for this is the fact that flood events in the Berg River lead to backwater effects at the gauging station (flooding the gauging station), which influence the depth of measured runoff.

During the wet winter and early spring months, the absolute value of the simulated runoff volume during peak flow events, is higher when simulated by the model version incorporating the inactive contour bank module. However by comparing the models between May and October, the differences within the composition of the simulated catchment runoff are negligible.

Application of the J2000 model with the inactive contour bank module shows a slight shift in the relative contributions of each runoff component to the annual discharge volume, compared to the results from the version with the active contour bank module. In particular, a slight decrease in the contribution of groundwater runoff components (RG1 and RG2) is observed in favour of the overland and subsurface flow components, when comparing the values of the model with the active contour bank module and that with the inactive contour bank module (see Table 2).

The predicted catchment sediment yield is significantly increased by applying the model with the inactive contour bank module. The simulated sediment yield at the gauging station is 22 tonnes in total, a total increase of over 19 tonnes. Overall, for an area of 66.63 km² (10.67 km² with the active contour bank module) of the Sandspruit catchment, soil loss amounting to 33.98 tonnes (5.89 tonnes with the active contour bank module) is predicted. Figure 9 is an overview of the spatial (top) and statistical distribution (bottom) of areas with sediment yield according to the model version applied for the year 1990. It becomes clear and visible, that the contour bank module and the subsequent shortening of slope length have a significant impact on erosion prediction. Indeed, it is visible from the histograms in Figure 9 (bottom) that the total number of HRUs with simulated sediment yield, as well as the sum of simulated sediment yield, decreases through application of the J2000 model with the implemented contour bank module. These decreases are caused by shorter slopes and less simulated, erosive surface runoff (RD1) into adjacent HRUs. Similarly, by applying the J2000 model without the contour bank module, a significantly higher sediment yield is simulated on a greater number of HRUs.

This research demonstrated that the effect of contour banks on runoff generation and erosion dynamics needs to be considered in semi-arid areas characterised by contour bank farming. Thus, the implementation of a contour bank module is a requirement for distributed hydrological and erosion modelling approaches.

The model results from the hydrological modelling under consideration of contour bank farming showed that the measured catchment runoff at the basin outlet could be predicted quite well. This is confirmed by the objective functions used for model evaluation. Besides this, the implementation of the contour bank module into the J2000 model framework leads to a more realistic representation of hydrologic relevant processes and their spatial distribution within the catchment. Accordingly, it could be shown, that the integration of the contour bank module led to a significant reduction in surface runoff along steep slopes, and consequently implies a significant reduction of soil loss from the affected areas, particularly during high precipitation events. For previous model versions without the active contour bank recognition, no surface runoff was simulated during high rainfall events for the contour bank areas. For the comparison of different approaches of topological linkage under contour bank areas, there was an equalisation of the approaches. The reason can be found in the underlying concept of the contour bank module, and in particular in the rerouting of water to the river system through the contour channels and the water transfer to downhill localised neighbouring entities. The runoff from an entity under contour banking depended on the total number of contour channels per entity. The higher the amount of water that can be stored behind contour banks per entity, the smaller the impact of topological linkage between the modelling entities. The impact of topological linkage became clearer, when comparing the approaches using a model version with an inactive contour bank module and analysing the spatial distribution of sediment yield prediction. It could be shown that the one-dimensional routing model simulated (spatially distributed) more erosion on less model entities. The reason is that the approach leads to swift runoff by just following the flow path over the point of highest flow accumulation. Therefore, the advantage of using a multi-dimensional approach for flow routing can be seen in a more accurate and realistic representation of the natural distribution of erosion dynamics or any other material fluxes within a catchment.

The authors would like to thank the German Federal Ministry for Education and Research (BMBF), the South African National Research Foundation (NRF) and the Water Research Commission of South Africa (WRC) for financial support. Acknowledgement is also given to Peter Krause, Christian Fischer, Markus Wolf and Daniel Butchart-Kuhlmann from Jena University for technical and scientific support. This research started on the Goedertrou farm in South Africa and the willingness of the farmer, Mr Stefaanús du Toit to participate, is greatly appreciated. We are grateful for the comments of the anonymous reviewers which helped to improve the quality of this paper.