Wetland inundation dynamics in a model of land surface climate: Evaluation in the Niger inland delta region



[1] Observed river gauging data show significant evaporative losses from the land and water surface in the Niger inland delta. These losses indicate an important potential feedback between the land surface and atmosphere. Moreover, the reduction in river flow downstream of the wetland has clear implications for water management in the region and beyond. Here we have modeled the evaporative losses that occur over the Niger inland delta by adding an overbank flow parameterization to the Joint UK Land-Environment Simulator (JULES) land surface model. The hydrological component of this model comprises a probability-distributed model of soil moisture and runoff production coupled with a discrete approximation to the one-dimensional kinematic wave equation to route river water downslope. We use subgrid-resolution topographic data to derive a two-parameter frequency distribution of inundated areas for each grid box which we then employ to represent overbank inundation in the model. The model was driven using data from the ALMIP experiment (ALMIP stands for AMMA Land surface Model Intercomparison Project, wherein AMMA stands for African Monsoon Multidisciplinary Analyses). The model reproduces the salient features of the observed river flow and inundation patterns; these include significant evaporative losses from the inundated region accounting for doubling of the total land-atmosphere water flux during periods of greatest flooding. Our predictions of inundated area are in good agreement with observed estimates of the extent of inundation obtained using satellite infrared and microwave remote sensing.

1. Introduction

1.1. Background

[2] Soil moisture influences atmospheric processes through its impact on heat and water fluxes between the land surface and the atmosphere. Using an ensemble of global climate models (GCMs), Koster et al. [2004] have estimated the strength of land-atmosphere coupling globally. They identified several regional hot spots where soil moisture exerted a strong control on forecast skill. The hot spots included the central Great Plains of North America, the Sahel, equatorial Africa, and India. These are regions in transition zones between wet and dry climates where moist convection is sensitive to surface fluxes and where evaporation is sensitive to soil moisture.

[3] Modeling the availability of moisture at the land-atmosphere boundary is a key challenge in understanding land-atmosphere feedbacks, and in improving the predictability of summer climate over continental regions [Koster et al., 2004]. One potential improvement is the better resolution of the extent of seasonal wetlands in climate models, because these regions represent key sources of moisture [Gedney et al., 2004]. Global climate simulations in which surface waters are included are often in greater agreement with observations [Coe, 1997]. Furthermore, a recent study by Taylor [2010] identified from satellite observations a significant increase in the number of storms triggered when a seasonal wetland floods in Mali. Indeed, in many applications it is vital that the surface waters are treated as interactive components of climate models and not simple static parameterizations [Coe, 1997; Papa et al., 2006; d'Orgeval et al., 2008].

1.2. Aims

[4] The present paper has three aims: (1) to introduce a computationally efficient method for estimating the extent of fluvial inundation using subgrid-scale elevation data within a land surface model; (2) to evaluate the model's performance over the Niger inland delta; and (3) to quantify the effects of wetland inundation on land-atmosphere heat and moisture fluxes in the study region. To undertake this work, we have used the Joint UK Land-Environment Simulator (JULES) land surface model [Cox et al., 1999] together with the CEH Grid-to-Grid (G2G) 1D kinematic wave river routing model [Bell et al., 2007; S. J. Dadson et al., Evaluation of a grid-based river flow model configured for use in a regional climate model, submitted to Journal of Hydrology, 2010]. We have added a parameterization of subgrid-scale topography to represent overbank inundation, which is described below. The model was driven using meteorological data from the ALMIP experiment (ALMIP stands for AMMA Land surface Model Intercomparison Project, wherein AMMA stands for African Monsoon Multidisciplinary Analyses) [Boone et al., 2010]. We use this model to estimate losses to evaporation from the inundated part of the river channel and floodplain, and to evaluate the effects of evaporation on latent heat flux and surface temperature.

1.3. Study Region

[5] This paper focuses on the Niger inland delta because the Sahel belt is an area of strong land-atmosphere coupling [Koster et al., 2004; Taylor and Ellis, 2006], and recent observations suggest that the seasonal flooding of the wetland induces a significant increase in storms. From the Guinea Highlands (250 km east of the Atlantic coastline) the Niger River flows toward the northeast, passing through an inland delta near (5°W, 15°N), after which its course turns southeast through Niger into Nigeria (Figure 1) [Grove, 1985]. The Niger River is characterized by strongly seasonal flows, with high flows causing floods late in the wet season. Within the 15,000 km2 inundated area, annual evaporation exceeds 2200 mm [Gourcy et al., 2000, p. 702]. The isotopic composition of river waters in the delta region indicates that evaporation is responsible for a reduction in water volume of 25 percent across the inland delta region [Gourcy et al., 2000, p. 706]. By comparing the stable isotope composition of river water and groundwater, Fontes et al. [1991, p. 199] have demonstrated that while aquifer recharge may have occurred in the Niger inland delta region during extensive flooding in wetter periods of the Holocene, modern groundwater recharge rates are extremely low relative to rainfall and evaporation.

Figure 1.

Location map showing the catchment boundary (solid black line), rivers (blue lines), and gauging stations (open circles). Colored shading indicates topography; mottled region indicates seasonal wetland.

[6] The sensitivity of West African land surface hydrology to climate variability is highlighted by Li et al. [2005]. West Africa has seen large interdecadal shifts in climatic variability over the past 50 years. The 1950s and 1960s were relatively wet, while the 1970s and 1980s were much drier. Much of this variability is driven by global and regional ocean conditions, modified by land-atmosphere interactions [Giannini et al., 2003]. Better resolution of the land-atmosphere coupling in this region is likely an essential part of improving understanding and prediction of the seasonal to interannual variations in West African climate [Boone et al., 2010].

2. Model

2.1. JULES-G2G

[7] The JULES model diagnoses the hydrological state of the surface and soil given time-varying inputs of temperature, wind speed, humidity, short-wave and long-wave radiation, and precipitation [Cox et al., 1999]. Within JULES there are four horizontal soil layers, each with an associated temperature and soil moisture content. Water and heat are assumed to move in the vertical direction only. Estimates of surface and subsurface runoff are calculated as the amount of liquid water leaving a grid square on the land and below ground, respectively (Figure 2a). The influence of stomatal resistance of vegetation is modeled explicitly in order to estimate transpiration, and account is taken of the effect of spatially varying soil properties and land cover. Estimates of liquid water leaving a grid square both on the land surface and below ground (runoff) are obtained using the probability distributed principle, where the spatial distribution of soil moisture stores follows a power law with exponent 0.5 [Moore, 2007; Blyth, 2002].

Figure 2.

Model structure. (a) Schematic diagram showing a JULES soil column highlighting stores and fluxes of water and Grid-to-Grid (G2G) model structure. (b) Hypsometric integral showing the fraction of grid box which is higher than a given elevation above the river. Lines indicate the model fit to subgrid elevation data from upland (solid line), lowland (dashed line), and wetland (dotted line) grid boxes.

[8] These runoff estimates are used by the G2G model, which is based on a discrete approximation to the 1-D kinematic wave equation with lateral inflow [Bell et al., 2007]. Kinematic routing is applied separately to subsurface and surface runoff; the model also allows for different formulations over land and river pathways. A return flow term allows for flow transfers between the subsurface and surface pathways representing interactions between surface and subsurface flow on hillslopes and in channels. Flow paths were calculated from the Hydro1k digital elevation model using the method of Paz et al. [2006]. In calibration, the hydrological routing parameters were adjusted manually in order to obtain optimal model performance over the time period for which data were available. Optimal flow routing parameters for the upper Niger were identified by altering the wave speed and return flow parameters by trial and improvement to achieve the best possible correspondence between modeled and observed monthly hydrographs. The possibility of overfitting the model to particular observed data is minimized because the data span a wide range flow conditions.

2.2. Overbank Parameterization

[9] The extent of inundated water surface was calculated using a parameterization of the overbank flow process which was developed to take advantage of information on floodplain topography that is available at spatial resolutions finer than that of the land surface model grid. Within each grid box of the land surface model, elevation data from the Hydro1k digital terrain model were used to construct a probability density of height above the grid box minimum. This measure was used to represent the height above the river channel in each grid box. Making the assumption that flow depth scales with flood discharge, which is confirmed in the region of interest by Zwarts et al. [2005], we calculated the depth of flow from the river flow output from JULES-G2G and from that have computed the fraction of the grid box that may be inundated using a lognormal cumulative distribution of elevation (Figure 2b). A Kolmogorov-Smirnoff test indicates that in the majority of grid boxes the distribution was indistinguishable from lognormal with 80 percent probability. Using this approach, the land surface model can account for the subgrid topographic properties of the region by using only two parameters: the mean and the standard deviation of the logarithm of elevation. When inundation occurs, a prorated fraction of the preexisting surface types is converted to have an “open water” land cover type. Water and energy are conserved during this procedure, and the newly formed area of open water is then used to calculate rates of evaporation and heat fluxes. The model presented here considers inundation by water that has been transported laterally along river channels. Once in the river channel, this water is accounted for separately from soil moisture and there is no further interaction between the river water and soil moisture. Vegetated zones can and do become flooded, in proportion to their original surface cover fractions. At present we have insufficient information on the detailed biophysical response of vegetation to flooding to model this process explicitly, but we note that potential improvement for future work. When floodwaters recede, preexisting land cover types are restored, in proportion to their original surface cover fractions.

3. Model Configuration and Driving Data

3.1. Model Configuration

[10] The model was run on a 0.5 degree regular latitude-longitude grid over the region outlined in Figure 1. JULES is a tiled model, in which each grid box contains a variable fraction of a range of distinct surface types, each of which can be set to have different properties relating to heat and water transport, and vegetation. In the present experiment, JULES was configured with nine surface types: broadleaf tree, needleleaf tree, C3 grass, C4 grass, shrub, urban, open water, bare soil, and ice, each with different heat and water transport properties [Essery et al., 2003]. The time-varying fractional coverage in each category was estimated using the ECOCLIMAP database [Masson et al., 2003], along with leaf area index and soil thermal and hydraulic properties. Soil albedo was derived from MODIS satellite observations [Houldcroft et al., 2009]. The open water fraction was adjusted dynamically through the course of the model run to account for floodplain inundation. In order to ensure that the sum of surface fractions remained one, the areas of the other surface types were reduced proportionally. A radiative canopy with explicitly modeled heat capacity was used. For vegetated surface types, canopy height and leaf area index were estimated from MODIS satellite observations and updated every 10 days. Nonvegetated surface types were modeled using standard JULES parameters [Cox et al., 1999], although the heat capacity for the open water tile was increased to 2.1 × 107 JK−1 to represent water with a mean depth of 5 m. A 5 year spin-up period was used to ensure that the soil moisture and river routing stores had reached a steady state.

3.2. ALMIP Driving Data

[11] We have driven the model using meteorological outputs from the ALMIP project [Boone et al., 2010]. That experiment consisted of offline simulations using eleven different land surface models forced with a combination of driving data from numerical weather prediction models and satellite observations. The data were provided every 3 h between 2002 and 2006. The driving data were obtained from the European Centre for Medium-Range Weather Forecasting (ECMWF) model, which is known to contain a systematic dry bias as a result of limited penetration inland of the West African Monsoon, and so the precipitation component was corrected with reference to TRMM-3B42 satellite observations [Boone et al., 2010].

4. Results

4.1. Modeled River Flow

[12] Model performance was assessed by comparing modeled monthly mean discharge at Tilembeya, in the center of the wetland region, with observed river flows obtained from the Global Runoff Data Centre. The location of the gauge is given in Figure 1. Observed and predicted monthly discharges are shown in Figure 3a. The timing of observed flows is accurately reproduced by the model, however, the model was found to overestimate river flow by 41 percent. Mean observed river flow is 885 m3 s−1, whereas the model predicted 1250 m3 s−1. The model efficiency (defined as the ratio of the variance of the model's errors to total observed variance, subtracted from unity [Nash and Sutcliffe, 1970]) was 0.70. Figure 3a also shows a 2 month lag between peak precipitation in the catchment and peak discharge at the gauge, a result which is in agreement with the observations of Zwarts et al. [2005]. Error bars in Figure 3a indicate the year-to-year spread in observed and modeled river flow. Interannual variability is greatest in the summer flood season and is higher in the modeled flow than in the observed.

Figure 3.

Comparison of modeled and observed land surface attributes. For all plotted quantities, error bars indicate 2 standard errors around the monthly mean. (a) Observed river flow (solid circles) compared with modeled flow (open circles) driven with TRMM-corrected ECMWF precipitation (dashed line). Grey shaded bars show catchment-averaged precipitation from TRMM-corrected ECMWF data set. (b) Satellite-observed inundation fraction (solid circles) from Prigent et al. [2007] and modeled inundation fraction driven with TRMM-corrected ECMWF precipitation (dashed line). (c) Comparison of land-atmosphere water vapor flux with (open circles, dashed line) and without (solid circles, solid line) inundation scheme operating; dotted line with squares indicates change in diurnal temperature range at Tilembeya.

[13] We suggest that the model's overprediction of discharge may be attributed to one or several of the following causes: (1) losses due to abstraction of water from the river by humans for irrigation (estimated at approximately 10–15 percent by Zwarts et al. [2005]); (2) groundwater recharge in the delta region (although available measurements indicate that groundwater recharge is insignificant [Fontes et al., 1991]); (3) uncertainty in the observed record, resulting from a combination of gauging uncertainty at high flows on a wide floodplain; (4) an underestimate of evaporation from the land surface or the river channel by the model; (5) uncertainty in precipitation amounts in the driving TRMM-corrected ECMWF data set; and (6) uncertainty in the calculation of flow paths at 0.5° spatial resolution (see Davies and Bell [2009] for a detailed discussion).

4.2. Inundated Fraction

[14] Satellite observations of the fraction of the land surface that is inundated are available from Prigent et al. [2007], who have used passive microwave land surface emissivities, active microwave scatterometer responses, and visible and near-infrared reflectances to estimate the fraction of the land surface that is inundated on a 25 km grid each month between 2000 and 2006. The patterns of modeled and observed inundation extents clearly reflect those of river discharge (Figures 3b and 4a). The model accurately reproduces the peak inundation extent (0.45, which is only 18 percent lower than observed). The efficiency statistic calculated for inundation extent is 0.52, and is diminished slightly by the early arrival of the modeled inundation peak, and by the fact that the period of simulated inundation longer than observed. The inundation model is clearly sensitive to the topographic parameters used to calculate the inundated fraction. For a typical grid box in the Niger catchment, a 10 percent change in the log-mean parameter would lead to an 11 percent change in the inundated area. The sensitivity is higher in the wetland region, owing to the lower topographic variability. For a typical wetland grid box a 10 percent change in the mean of the natural logarithm of elevation would lead to a 15 percent change in inundation. Auxiliary material Figure S1 shows the effect of the topographic parameters on inundated area in more detail.

Figure 4.

Spatial pattern of model predictions. (a) Inundated fraction. (b) Land-atmosphere water vapor flux. (c) Temperature anomaly in inundated regions, measured as difference between inundated open-water tile and grid box mean over all tiles.

4.3. Moisture Flux and Temperature Anomaly

[15] The role of the Niger inland delta as a source of water vapor and a sink of latent heat is well documented, although few quantitative constraints have been offered [Zwarts et al., 2005]. Our model indicates that the evaporative moisture flux from the delta region (Figures 3c and 4b) is doubled (from 0.47 mm d−1 to 1.1 mm d−1, averaged over the whole year) when the inundation model is employed. The effect is greatest between September and November, when the peak extent of the flooded region coincides with high temperatures. Moreover, the latent heat flux associated with this evaporation is sufficient to reduce the diurnal surface temperature range in the wetland region by an average of 10 K during the flood season (Figure 3c) and to depress the temperature over the entire inundated region by up to 5 K (Figure 4c). This model result is consistent with airborne observations over the region during June 2006 (C. M. Taylor, personal communication, 2008).

5. Discussion and Summary

[16] The model presented in this paper reproduces the salient features of the observed river flow and inundation patterns in the Niger inland delta; these include significant evaporative losses from the inundated region accounting for a doubling of the total land-atmosphere water flux during the period of greatest inundation. Taken together, these findings have two major consequences: (1) the seasonal expansion and contraction of the inundated region exerts a significant control on the regional water budget and may therefore determine the properties of the regional precipitation field [see, e.g., Taylor, 2010] and (2) climate and numerical weather prediction models operating without this dynamical land surface feedback may seriously underestimate the temporal and spatial variability of water vapor and latent heat fluxes in key regions where evaporation rates are controlled by seasonally inundated open water bodies of variable extent. The modeling scheme that we have described provides a computationally efficient means of accurately representing wetland inundation in a land surface model and has applications both in regional inundation assessment and in providing improved predictions of land-atmosphere energy and water vapor fluxes.


[17] We thank Aaron Boone and the ALMIP participants for making their meteorological data available; the Global Runoff Data Centre, Koblenz, for providing access to river gauging data; and C. Prigent for sharing remotely sensed inundation data. This research was undertaken as part of the European Union (FP6) funded Integrated Project called WATCH (contract 036946). Development of the overbank inundation model was supported by a NERC grant (NE/E011969/1) under the Ecology and Hydrology Funding Initiative. We thank two anonymous reviewers for comments that substantially improved the manuscript.