Uncertainty in the modelling of large scale flood events in the - PowerPoint Presentation

Slide1

Uncertainty in the modelling of large scale flood events in the

Barotse

floodplain, Zambia

Tom Willis

1, Mark Smith1, Donall Cross2, Andrew Hardy3, Georgina Ettritch3, Happiness Malawo4, Mweemba Sinkombo4, Cosmas Chalo4, Elizabeth Mroz1 and Chris Thomas51. School of Geography, University of Leeds, Leeds, UK2. Institute of Biological, Environmental and Rural Sciences, Aberystwyth University, Aberystwyth, UK3. Department of Geography and Earth Sciences, Aberystwyth University, Aberystwyth, UK4. Water Resource Management Authority WARMA, Mongu, Zambia5. College of Science, University of Lincoln

Zambia

Ministry

of Health

Introduction to the

Barotse Floodplain and the FLOODMAL Project

The modelling work presented here is part of a wider project - FLOODMAL. This study investigates the development of process based models to predict the formation of water bodies and vector critical to the malaria cycle in the Barotseland Floodplain, Western Province, Zambia.Introduction to the hydrodynamic modelling

An summary of the approach to the modelling, the location of the model domain, datasets used and the model setup.

Calibration, Uncertainty and Sensitivity analysisOverview of the modelling results, and the implications for modelling large scale hydrological features. Overview of the Display

FLOODMAL

1. FLOODMALProject Overview

FLOODMAL is an interdisciplinary approach to understanding key processes of the relationship between flooding in major river systems and malaria, using the Zambezi river as a test case.

By coupling existing techniques and models in hydrology, remote sensing and spatial ecology the aim is to provide predictive power to the availability of water bodies crucial to mosquitos that are the main vector for malaria

1. The

Bartose Floodplain

The

Barotse

Floodplain is located in the Western Province, Zambia. The Zambezi river is the main source of flow in the region running from north to south. A number of other rivers, including the

Luena

and the

Luanginga

also enter the valley. The valley is a major flood attenuation feature for flooding along the Zambezi river.

The flooding of the region is an annual event. A rainy season for the Central African region lasts from November to January, caused by the movement of the ITCZ from North to South leads to a

floodpeak

in March to April. The valley will slowly drain from this time to a minimum value in October

© Contains Bing Satellite data

1. The Bartose Floodplain

The floodplain is approximately 250km long, and 50km wide and a low gradient across the floodplain with an elevation change of 40m from north to south. The floodplain is covered with

Salharian

sands, which are dominated by grass lands. With the exception of the causeway, linking the west and east sides of the floodplain, no significant topographical features occur that impede flow.

The region is home to approximately 100,000 people, the

Lozi people, who migrate between the lower parts of the floodplain during the minimum river level phase, and higher parts during the flood peak. Fishing and subsistence farming are the dominant activities in the floodplain, with Malaria rates in the floodplain have continued to rise, in comparison to the rest of the Zambia where rates have steadily fallen in the last 10 years.

As flooding of the region is independent of rainfall, the hydrodynamic modelling of the floodplain is focused on representing floodplain flow paths and channel conveyance. A 1D-2D LISFLOOD-ACC formulation provides an ideal compromise between speed, channel representation and floodplain dynamics.

The model domain and main river network to be included in the model were determined from analysis of the daily river flow gauges located throughout the model domain (blue dots)

The model was analysed against gauge data from the

Senanga

guage (at the southern end of the domain) and LandSAT imagery of flooding throughout the year2. Hydrodynamic Modelling

2. Hydrodynamic Modelling

The model required parameters including surface roughness, evaporation rates, infiltrations rates and bathymetry data to effectively capture the main processes in the hydrological dynamics of the river

The terrain data for the model is 12m TanDEM-X1 data, resampled to 900m for calibration runs. Surface roughness values for the floodplain (Manning’s n) were determined though analysis of

LandSAT

imagery Infiltration rates were determined from surface sampling in the floodplain, and estimating appropriate rates based on established values. Evaporation rates from Global CLIM datasets were used as inputs to the model

2. Hydrodynamic Modelling

Bathymetry data was collected from 80 Cross sections and Thalweg

surveys of the River Zambezi, and 25 from minor channels and canal, using a georeferenced depth range sonar

. The data were used to create a model of the channel based on location and width, to determine channel depth and profile at further locations

© Contains Bing Satellite data

3. Calibration and Uncertainty analysis

The model was developed and run for the 2009 flood event which represented the largest event in the gauge record to which corresponding

LandSAT

imagery was also available. This allows the model to be calibrated with a multi-objective approach, and allows the model to be calibrated to both peak and minimum

The 2009 event has 3 overpasses which provide images of sufficient quality to analyse the model, including the peak of the flood event in March, and a near minimum value in July

Calibration was based on optimising model parameters including floodplain and channel friction, infiltration and evaporation, and channel depth determined from the bathymetry model, based on a range of values (upper table) The model was calibrated using the Nash Sutcliffe coefficient and binary comparison measure. ParameterRangeDistribution

Floodplain n

0.01 – 0.06

Uniform

Channel n

0.01 – 0.03

Uniform

Infiltration (mm/h)

0.2 - 20

Uniform

Evaporation (mm/d)

3.4 – 6.6

Normal

Channel Depth (m)

4 – 5.5

Normal, based on bathymetry data

3. Calibration and Uncertainty analysis

Parameter

F² March 2009

F² April 2009

F² June 2009

Nash SutcliffeFloodplain n

93.1

89.975.1

95.0

Channel n

0.0

0.0

0.3

0.0

Infiltration (mm/h)

0.3

2.4

6.6

1.2

Evaporation (mm/d)

0.3

2.1

0.6

0.2

Channel Depth (m)

6.3

5.6

17.4

3.7

Calibration of the model is summarised in the opposite figure. The figure shows the model score for each model run across the total parameter set. Highlighted are single model runs that performed best for each individual score. The overall pattern is that single parameter set will perform better for either the peak of the flood, or the minimum, but not for both points. Further, modelling the peak of the event tends to be slightly easier, with higher binary comparison values.

The optimal parameter set used for the final model configuration is determined through evaluating the pareto front for multiple evaluation methods. The optimal set therefore compromises top score in any single evaluation, but performs suitably well across all objective functions, and is represented as the red line

Sensitivity analysis was undertaken to quantify the contribution of a parameter to the variance determined in the outputs from the calibration process. This provides insight not only into model uncertainty, but also dominant processes in the floodplain. The results, summarised in the lower table, and show that floodplain friction is the dominant control on model results.

Model with highest score for July analysis

Model with highest score for March analysis

Optimal model set

Sensitivity analysis results, describing the contribution of the parameter to the variance of the overall range of results seen in the calibration model results, in terms of percentage of contribution

3. Calibration and Uncertainty analysis

Analysis of the calibrated model against the observed data demonstrates the complexity in modelling the region. The above figure compare observed

LandSAT

data and modelled results.

For the March time slice the blue regions indicate the model performs well across the region. For the July time slice, there is a notable increase in modelled but unobserved data. The reasons behind this include the ability to correctly model the drawdown at this point in the year, and the ability of the

LandSAT data to model increasingly small water bodiesComparison of the model against the observed LandSAT imagery, March 26th 2009Comparison of the model against the observed LandSAT imagery, July 17th 2009

© Contains Bing Satellite data

© Contains Bing Satellite data

4. Conclusions

A 1D-2D LISFLOOD model has been developed to simulate the large

Barotse

Floodplain in Zambia

Model calibration shows that the model can be optimised for either the peak of the event or the minimum of the event, put the parameter sets are rarely optimised for both. This highlights that hydrodynamic model development that has focused on correctly modelling the peak may not be suitable for a wide range of conditions and that further work is required to provide more holistic models with a focus on the complete peak to

inbank

flow conditions

Sensitivity analysis demonstrates that the model is highly sensitive to the floodplain friction parameter, above channel parameters or water loss mechanisms. This is true across the year, and against the expectation that evaporation may play a significant effect in influencing the drying phase of the model

4. Way Forward

The model will be further refined by increasing the resolution of the terrain data, and testing with further events. The output of this refined model will feed into the spatial ecology model, and will define water bodies that support mosquito formation. This in turn can be used to identify a link between malaria hotspots and potential mosquito habitats.

Since the 2009 event, a major causeway has been built across the floodplain, and a number of access canals, built and maintained by local villages, have been restored. These features will modify the previous hydraulic processes, and determining the influence of this on both flooding and the formation of the mosquito water bodies will be critical in understanding future trends.

In addition to identifying the malaria hotspots, the outputs of the model will feed into other areas. A key problem for the health workers in the region is accessing and supplying health centres during the peak of the flood. A project is currently underway to model the road network and access routes, and using simulated flood depths over a year, will determine the length of time a health centre cannot be accessed by floods. By combining this with further research into identifying and rainfall patterns in the upper catchment to develop an early warning access system for health centres.

Digitised access network of the

Barotse

Floodplain

Overlay of road network with model output

© Contains Bing Satellite data

© Contains Bing Satellite data