In this context, a Bayesian Network model was constructed to assess the impact of water management options on the hydrosystem of the Hérault Middle Valley.
th
7 International Conference on Hydroinformatics HIC 2006, Nice, FRANCE
WATER MANAGEMENT IMPACT ASSESSMENT USING A BAYESIAN NETWORK MODEL SANDRA LANINI Water Department, BRGM, 34000 Montpellier, France
The research project AGIRE is dedicated to support tools and models for integrated water management. In this context, a Bayesian Network model was constructed to assess the impact of water management options on the hydrosystem of the Hérault Middle Valley (France). The network includes all the variables and elements that are supposed to have an influence on the ecological quality of the catchment. Links between nodes representing causal relations are described in conditional probability tables, allowing the model to deal with uncertainties. Tables were filled thanks to calculations, evaluations and/or expert opinions. The numerical model was then developed using Microsoft MSBNX©. Simulations consist in changing the values of the input nodes according to different water management scenarios, in order to observe the subsequent variations of ecological indicators. Three of them were selected to represent the ecological quality of the catchment: landscape quality, ecological value and fishermen satisfaction. Management options are related to the socio-economic background (demography, laws on gravel pits, agricultural practices…) or directly to hydrological criteria (number of days the river discharge or the water table is below given thresholds). These hydrological inputs which in turn depend on the management options are given by an independent lumped model of the socio-hydrosystem (WASS-Hérault). The sensibility analysis performed on the model proved the relevance of every node in the network. It also showed that the final indicators are sensitive to changes in every parent node and especially in hydrological inputs (river discharge and piezometric level). Possible improvements concern the way pollutant transfers in soils and groundwater are accounting for. Then, the model will have to be presented to stakeholders in order to be fully validated.
INTRODUCTION The concept of Integrated Water Resources Management (IWRM) appeared in the early 1970s but has been only endorsed by the United Nations at the Dublin International Conference on Water and the Environment in 1992. The Dublin statement encouraged participatory approaches involving users, planners and policymakers in order to avoid conflicts over water uses. Despite the European Water Framework Directive (2000) which has also promoted participation of affected citizens to water management,
1
2 stakeholders are still insufficiently involved in the decisional process (Soncini-Sessa et al. [12]). Water users and managers generally lack tools to investigate and compare the impact of various water management scenarios. Actually, Decision Support Systems (DSS) applied to hydrology, integrated water management and more generally sustainable development have been developed since the late 1980s (Jamieson & Fedra [6]; Parker et al. [10]), but it is only recently that they have been designed and adapted for non-expert users, taking advantage of improvements in computing science and technology. For instance, a system like the TIDDD (Pereira et al. [11]) brings scientific expertise in the decision process, but also proposes a representation of the hydrosystem that is shared by all the stakeholders. Bayesian Network models, which come from a marriage between probability theory and graph theory (Jordan [7]) provide interesting features to deal both with the complexity of natural system and with the need to support decision despite scientific uncertainty and lack of knowledge (Borsuk [3]). Being graphical models, Bayesian Network structure the problem such that it is visually interpretable by stakeholders and decision-makers (Ames et al. [2]). Moreover, as results are presented in the form of probability distributions, BN give an explicit representation of uncertainty (Bromley [5]). The accuracy of the model results can thus be explained to managers enhancing the decision making process (Acreman [1]). In more details, Bayesian networks, also called Belief Networks, are directed graphical models, which consist in series of nodes, representing variables which interact with each other. An arc from node A to node B can be interpreted as “A causes B”. These interactions are expressed as probabilistic dependencies between variables, which are discretized into distinct states allowing to define the probability distributions through Conditional Probability Tables (CPTs) (Ames et al. [2]). BN can be used to infer the values of the hidden causes (bottom-up reasoning) or to predict the effects of a given set of input (top-down reasoning) (Murphy [9]). Within the BRGM research project AGIRE1 “Decision Support Tool for Integrated Water Resources Management", different approaches were tested in order to model complex systems in which physical, environmental, social and economical variables are coupled. First, a lumped model of the functioning of the socio-hydrosystem of the Hérault middle valley was developed. It is based on a comprehensive knowledge and expertise of geological, hydrogeological and hydrological features of the catchment basin, as well as on extensive consultation of the stakeholders involved in its water resource management. It has been translated in a numerical simulator called WASS-Hérault (for Water Scenario Simulator) using the Matlab/Simulink© software (Lanini et al. [8]). This tool allows to easily test some water management options and to observe their impact on the aquifer levels and the river discharges in several points. But it can only deal with determinist 1 http://agire.brgm.fr
3 relationships and quantitative data, and thus does not supply any information on the quality of the ecosystem. In order to evaluate the environmental impact of different management scenarios, a complementary model was then constructed relying on the Bayesian Network approach. It is presented in this paper after a brief description of the study area.
SHORT DESCRIPTION OF THE STUDY AREA The Hérault River Basin is a Mediterranean watershed of 2500 km² located in the Languedoc Roussillon Region (France). The middle valley extends over 100 km² and corresponds to the upstream section of the alluvial plain of the watershed. The alluvial aquifer is made of two types of Quaternary deposits: recent alluvium around the presentday river and old alluvium in superimposed terraces.
Figure 1: Localization and map of the Hérault middle valley
The study area includes a large irrigated zone where water is supplied by a gravityirrigation canal. However, agriculture (especially vineyards) increasingly competes with tourism for water as recreational activities like swimming, fishing and canoeing attract a growing number of tourists to the river. As a result, there is an increasing demand for in-
4 stream preservation of water quantity and quality along the Hérault River. Moreover, the rapid economic growth of the city of Montpellier, which lies outside the Hérault catchment, has induced an intensive urbanization of the alluvial Hérault valley. The demand for domestic water has thus increased significantly. In the framework of the GOUVERNe European project2, an ecological diagnosis of the Hérault middle valley was performed (Bouche [4]). The results indicated that dams and sills on the Hérault River, as they modify the river flow and prevent fish migration, as well as tourism and recreational activities which disturb the bank environment, seem to be the main threats for the ecosystem quality and variety. On the contrary abandoned gravel pits, which have become wetlands, may have a positive impact on environment.
MODEL DEVELOPMENT The first step to construct the BN applied to the Hérault middle valley consisted in selecting the different variables to be included in the model and then defining their possible states. Then, conditional probability tables (CPTs) were filled for each nodes. All this work was done thanks to discussions with ecologists, experts and stakeholders of the basin. The BN, constructed with the Microsoft MSBNX© software, comprises 26 nodes, divided into three classes: 9 input nodes (2 of them representing hydrological variables) which are directly related to the water management options and evolution scenarios; 14 intermediate nodes; and 3 output nodes which supply the expected results. The number of states for each nodes was limited (2 or 3) in order to keep manageable CPTs, as suggested by Bromley [5]. The input nodes, that have no parents and are described by unconditional probabilities, correspond to the variables that are more or less directly modified by political and management decisions: - land use (non cultivated / traditional agriculture / organic agriculture): % of the surface area in each category; - gravel pits regulations (high / low): the two states correspond to the extreme following cases: the law make it compulsory to ecologically restore all the abandoned pits / there is no law concerning pits restoration; - bank degradation (high / low); - sills & dams (yes / no): this variable only indicates if there is sills or dams on the river;
2
European project No. EVK1-1999-00032
5 -
tourism: the two distinct states are bounds representing the minimum and maximum numbers of persons who come to the Hérault River each year (swimmers, walkers, fishermen); - population: upper and lower bounds; - impervious surface: upper and lower bounds; - Hérault River discharge (enough / not enough): estimated according to the number of days in summer during which the discharge remains under a given threshold (see Table 1); - Piezometric level (enough / not enough): estimated according to the number of days per year with a piezometric level under a given threshold. The threshold and classes for the discharge and piezometric distributions were defined thanks to observed data. Table 1: Unconditional probabilities for the “Hérault River discharge” variable Probability Enough = 1 / Not enough = 0 Enough = 0,9 / Not enough = 0,1 Enough = 0,8 / Not enough = 0,2 Enough = 0,7 / Not enough = 0,3 Enough = 0,6 / Not enough = 0,4 Enough = 0,5 / Not enough = 0,5 Enough = 0,4 / Not enough = 0,6 Enough = 0,3 / Not enough = 0,7 Enough = 0,2 / Not enough = 0,8 Enough = 0,1 / Not enough = 0,9 Enough = 0 / Not enough = 1
Number of days in summer with a discharge under 1m3/s 0-5 5-10 10-15 15-25 25-45 45-80 80-105 105-110 110-115 115-120 120-125
Intermediate nodes are described by two basic states (high/low or yes/no or enough/not enough). They were defined based on the following main assumptions: - “River pollution” and “Soil pollution”: As there is no industrial pollution source in the study area, pollution comes mainly from agricultural practices (fertilizers and pesticides), runoff on roads (heavy metals) and wastewater (organic pollution); - “River perturbation by hydraulic structures”: the “high” state gives the % of river length disturbed by eddies (upstream of the hydraulic works); - “Ecological restoration of gravel pits”: the ‘yes” state gives the % of gravel pits that have been restored. Four environmental indicators were identified as endpoints of the model: pits quality, terrestrial environment quality, aquatic environment quality and biocenosis quality. With
6 a perspective of decision support, they are gathered in three final output nodes which synthetize the environmental criteria: - Landscape (yes/no): assess the attraction of the study area for pleasure and leisure, regarding only the quality/beauty of landscape; - Ecological value (important/low): assess the ecological wealth (species, biodiversity) of the study area - Fishermen satisfaction (yes/no): depends on the number of fish species and fish quantity in the Hérault River. Conditional probabilities tables were filled and calibrated such that, for the values of input nodes corresponding to the current situation, the BN estimates a probability to have a quality for the terrestrial and aquatic environments and for the biocenosis which is consistent with the ecological diagnosis results (Bouche [4]). As an example, the CPT for the node “River Pollution” is presented below (Table 2). It has to be read line by line. For instance, given that the agricultural pollution is low, the runoff pollution is low and that the wastewater treatment capacity is “not enough”, the probability to have a high river pollution is equal to 0.6. Table 2: CPT for the node “River Pollution” Agriculture Pollution
Runoff pollution
Wastewater treatment
high high high high low low low low
high high low low high high low Low
not enough enough not enough enough not enough enough not enough enough
River pollution High 1 0.9 0.9 0.8 0.9 0.8 0.6 0
Low 0 0.1 0.1 0.2 0.1 0.2 0.4 1
SENSITIVITY ANALYSIS A sensitivity analysis was performed on the model as a first step toward validation. The BN model was first run with input values corresponding to the current situation. Then, the model was run successively with one of the input node set to its best state, then on its worth state, and this was repeated for each of the nine input nodes. The variations induced on the value of the three output nodes were observed. The result of this analysis
7 is shown on figure 2 for the output node “Ecological value”. The dash-line corresponds to the current situation. It appeared clearly that: a) all the nodes are useful in the network (all of them have an influence on the results) b) the input nodes that have the greater impact on the final results are the two hydrological variables (river discharge and piezometric level). The sensitivity analysis also showed that the results have a large amplitude of variation according to the input data values. For example, the ecological value has a probability of 94.6% to be high if all the input nodes are set to their best states, against only 18.5% if they are all set to their worst states.
River discharge Piezometric level Pit reglementation Land use Covered area Population River sides Dam Tourism 0,3
0,4
0,5
0,6
0,7
0,8
High ecological value probability
Figure 2: Results of the sensitivity analysis for the “Ecological value” node
CONCLUSION A Bayesian Network was constructed with the objective to assess the impact on the environment of different water management scenarios. The sensitivity analysis has shown that the model is sensitive enough to simulate different water management scenario and supply a relevant range of results. The next step is now to define the scenarios, then to run the WASS-Hérault model to obtain simulated piezometric and discharge series. These results will then be used to set the values of the two hydrological input nodes of the BN, whereas other parent nodes values will be imposed by the scenario. The work presented in this paper corresponds to a “feasibility study”, and it still has to be completed. Dedicated studies would be necessary to collect more data or develop simple models to rely on to fill some CPTs (especially those related to water pollution). Then, the Bayesian Network model will have to be validated, and the best way to achieve this is
8 probably to present the model to a group of stakeholders of the catchment and to compare the results to their opinions.
REFERENCES [1] Acreman M., “Linking science and decision-making: features and experience from environmental river flow setting”. Environmental Modelling & Software, 20(2), (2005), pp 99-109. [2] Ames D.P., Neilson B.T., Stevens D.K. and Lall, U., “Using Bayesian networks to model watershed management decisions: an East Canyon Creek case study”. Journal of Hydroinformatics, 07(4), (2005), pp 267-282. [3] Borsuk M., Clemen R., Maguire L., Reckhow K. “Stakeholder Values and Scientific Modeling in the Neuse River Watershed”. Group Decision and Negociation, 10, (2001), pp 355-373. [4] Bouche S. « Mise au point d’une méthode de diagnostique écologique – Application aux milieux naturels de la moyenne vallée de l’Hérault liés à l’hydrosystème ». Document provisoire, projet GOUVERNe, (2002) [5] Bromley J., KJackson N.A., Clymer O.J., Giacomello A.M. and Jensen F.V., « The use of Hugin to develop Bayesian networks as an aid to integrated water resource planning”. Environmental Modelling & Software, 20(2), (2005), pp 231-242. [6] Jamieson D.G., Fedra K., « The “WaterWare” decision-support system for riverbasin planning”. Journal of Hydrology, 177, (1996), pp 163-213. [7] Jordan M. I. “Learning in Graphical Models”. MIT Press, (1999) [8] Lanini S., Courtois N., Giraud F., Petit V., Rinaudo JD. « Socio-hydrosystem modelling for integrated water-resources management - The Hérault catchment case study, southern France”. Environmental Modelling & Software, Vol 19/11, (2004), pp 1011-1019 [9] Murphy, K.P.” An Introduction to graphical models”, http://www.cs.ubc.ca/~murphyk/Papers/intro_gm.pdf, (2001), 19 p. [10] Parker P., Letche, R., Jakeman A., Beck M.B., Harris G., Argen, R.M., Har, M., Pahl-Wostl C., Voinov A., Janssen M. et al., “Progress in integrated assessment and modelling”. Environmental Modelling & Software, 17, (2002), pp 209-217. [11] Pereira .A.G., Rinaudo J-D., Jeffrey P., Blasques J., Corral Quintana S., Courtois N., Funtowicz S., Petit V. “ICT tools to support public participation in water resources governance & planning: experiences from the design and testing of a Multi-Media platform”. Journal of Environmental Assessment Policy and Management, Vol. 5, n°3, (2003), pp. 395-420 [12] Soncini-Sessa R., Castelletti A., Weber E., “Participatory decision making in reservoir planning”, Proc., 1st biennial meeting of the IEMSS, Lugano, Switzerland, 3, (2002), pp 34-44.
