Anomaly in the rainfall-runoff behaviour of the Meuse catchment

F. Fenicia, H. H. G. Savenije, and Y. Avdeeva Water Resources Section, Faculty of Civil Engineering and Geosciences, Delft University of Technology, P.O. Box 5048, 2600 GA Delft, The Netherlands Received: 10 June 2008 – Accepted: 10 June 2008 – Published: 10 July 2008 Correspondence to: Fabrizio Fenicia (f.fenicia@tudelft.nl) Published by Copernicus Publications on behalf of the European Geosciences Union.

authors considered the way in which the land is used by for instance agricultural and forestry practises. This paper focuses on the possible effects of land use and land use management on the hydrological response of the Meuse catchment. In absence of detailed information on land use over the observation period, we adopted a fully "top-down" approach to the problem. The approach consists of a dynamic evaluation 10 of a conceptual hydrological model and the interpretation of the temporal trends of model parameters. It appears that land use has had a considerable impact on the hydrological behaviour of the Meuse catchment. The time lag of the catchment has reduced markedly over time, possibly related to more intensive drainage and river training works. Moreover we hypothesise that forest rotation has had a significant impact on the

Introduction
The Meuse river basin extends over an area of 33 000 km 2 which covers parts of France, Belgium, Luxembourg, Germany and the Netherlands (Fig. 1). During the past century, this catchment has faced several modifications which may have influenced its 25 hydrological behaviour. In general the catchment has undergone substantial change 1788 as a result of population development, migration and economic development. Changes in land cover and land use involved urbanization, industrialization, afforestation, deforestation, intensification of agricultural practices, construction of storage works and increase of drainage efficiency (Garcier, 2007). The drainage improved in agricultural fields, in urban areas and also in forests as a result of forest road construction. Road 5 construction in general can have a marked influence on drainage, particularly during peak events. Also the drainage efficiency of the water courses improved as a result of river rectification and building of weirs, locks and canals. An impact seldom mentioned is the effect of different land use management strategies, which primarily relates to the ways in which agricultural areas and forests are tilled, cropped and harvested. 10 After severe floods in the Meuse in 1993 and 1995, a meeting was held at the EGU in 1995 on the possible causes of the extremity of these floods (Savenije, 1995). The opinions of authors could be grouped in: (1) nothing out of the ordinary happened; it is part of the natural variability (Ulbrich and Fink, 1995); (2) there are trends visible indicating more intensive floods in recent years, indicating at climatic change (e.g., 15 Black, 1995;Bronstert, 1995;Caspary, 1995); and (3) there is a significant effect of human interference due to more intensive drainage, road construction and river training (Savenije, 1995). Since then, the effect of climate and land use change on the behaviour of the Meuse catchment has been subject of several studies. The daily record of hydrologically useful information starts in 1911, and it has been shown that the 20 rainfall-runoff relation of this catchment has experienced significant alterations (Tu Min et al., 2005). However, whether the observed changes in hydrological behaviour should be attributed to human induced change, or to climate variability, or that they are simply an effect of the large uncertainties in the data was not well understood.
Tu Min et al. (2005) applied change point analyses to investigate the temporal change 25 in the flood peaks of the Meuse River over the past century, accounting for the potential effects of climate and land use change. They determined that both the antecedent precipitation depth and the flood peaks increased significantly since the 1980s, while the land cover remained relatively stable during the past century. They concluded Introduction Tables  Figures   Back  Close Full Screen / Esc

Printer-friendly Version
Interactive Discussion that the increase in high flow events should be attributed primarily to climate variability rather than to land use change. Ward et al. (2008) performed a hypothetical experiment where they coupled a climate model and a hydrological model to simulate daily discharge during several millennia. They noticed that the observed reforestation over the last century should theoretically have led to decreased discharge and that the main 5 mechanism responsible for the larger flood peaks in the latest years is increased precipitation. Increases in annual and winter maximum daily discharge and precipitation were also demonstrated in the neighbouring Alzette catchment in Luxembourg by Pfister et al. (2000).
Building on the work of Booij (2005), Ashagrie et al. (2006) investigated the effect of 10 land cover change within the catchment by using a semi-distributed conceptual model (HBV). According to their analysis, the main change in land cover was a shift from deciduous to coniferous forest, while the total forest cover and agricultural land remained relatively stable. The model application indicated that while the observed discharge was well represented at the beginning and at the end of the simulation period, a con-15 sistent overestimation occurred in the central part of the observation record (between years 1933 and 1968) (see Fig. 2). This deviation in discharge volumes could not be explained by a change in forest types. To test if the anomaly was due to data errors, Ashagrie et al. (2006) performed the same analysis on the Moselle catchment, which neighbours the Meuse on the east 20 and has similar size, history and physical characteristics. Surprisingly, the analysis produced similar results, with even larger systematic deviations in the central part of the observation record. As the data of the two catchments can be considered independent, it can be hypothesized that such systematic deviation is not merely a consequence of data errors, but it is the result of processes, mechanisms or catchment structural 25 modifications that where not identified in previous studies.
Overall, these studies suggest that land use changes have had a negligible impact on the hydrological behaviour of the Meuse catchment. Climate change, in contrast, is seen as the dominant control. However, while climate change may account for the Introduction

Conclusions
References Tables  Figures   Back  Close Full Screen / Esc

Printer-friendly Version
Interactive Discussion increased flood peaks of the last decades, it cannot explain the anomaly in the rainfallrunoff relation identified by Ashagrie et al. (2006). This motivated the investigation of additional hypotheses of catchment behaviour, which is the main purpose of this work. The evaporation process significantly affects the catchment water balance. In the Meuse catchment, 41% of the annual rainfall evaporates, and the remaining part ends 5 up as runoff. Evaporation can be significantly affected by human activities such as drainage and agricultural and forestry practices. It is difficult to obtain full information on the developments that have taken place over such an extended period of time. However, it can be argued that the gradual shift from deciduous to coniferous forest was accompanied by a change in forest management policy as well. An important 10 hydrological parameter in forest management is the rotation of the stands, since forest age is linked to the actual transpiration of the trees.
In the twentieth century 90% to 99% of the forest area in Europe could be considered as functional (production) forest (FAO, 1995). Industrial development, mining and the growth of population led to an increase in European consumption of wood and wood 15 products. The total consumption of wood rose from 270 to 340 million cubic meters in the period from 1910 to 1960 (FAO, 1964). During the Second World War, similar to other resources, forests were overexploited (Pearson, 2006). After the Second World War, afforestation activities in Europe grew to a high level, which slowed down during the last two decades of the 20th century (Dirkse and Daamen, 2004). Efforts were 20 made to improve both forest productivity and resource conservation. As a result, during the last decades of the 20th century the average forest age had a tendency to increase. Dirkse and Daamen (2004) state that in the Netherlands, due to changes in forest management (less clear cut and more thinning), between years 1980 and 2001 the average age of trees in standing forests increased by 10 years (from 43.3 to 53.3 25 years).
The relation between stand age and evaporation in different types of trees and climates is not yet fully understood. However there is an increasing body of evidence that demonstrates that young forest evaporates considerably more than mature forest. Introduction

Tables Figures
Back Close

Full Screen / Esc
Printer-friendly Version Interactive Discussion Kuczera (1987) developed an idealized curve for mountain ash forest that relates annual water yield to stand age (Fig. 3). This curve is based on real data of different catchments characterized by a mean annual rainfall of 1800 mm/a, and shows that mean annual discharge after harvest and regeneration of Eucalyptus regnans forest reaches approximately half its original value about 25-30 years after harvest, and then 5 recovers. Similar curves were obtained by Shiklomanov and Krestovsky (1988) for 33 types of forest of the northern-east part of Russia. Vertessy et al. (2001) analysed the influence of stand age on the catchment water balance separating different sources of evaporation including overstory, understory, soil and litter evaporation and rainfall interception. They estimated that annual overstory transpiration declines by 66% when tree age increases from 15 to 240 years. Scott and Lesch (1997) determined that the afforestation of a catchment with an average runoff of 236 mm/a caused the stream to dry up completely after 12 years. Several paired catchment experiments were used to assess the impact of afforestation on water yield (see Brown et al., 2005 for a review). Using data from South African catchments, Scott 15 and Smith (1997) developed an empirical model that predicts the reduction in water yield as a function of plantation age and species planted. According to their model the reduction in stream flow assuming a full forest cover can be as much as 100%.
The results of these studies are difficult to generalise as the relation between stand age and catchment water balance depends on climate, tree species, and catchment 20 characteristics. Farley et al. (2005) compiled data from different parts of the world and concluded that runoff reduced significantly during the first 20 years after planting, while these reductions where percentage-wise larger in dryer catchments, but in absolute terms larger in wetter catchments. Watson (1999) investigated the various processes that determine tree transpiration 25 and how they modify with tree age. He concluded that the flow of water through the leaves (leaf conductance) declines markedly with age. This is a consequence of the observation that, while the mean sapwood velocity is constant with age, the stand sapwood area per unit leaf area decreases significantly. The decrease in leaf conductance HESSD 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment F. Fenicia et al. in old forests may be caused by slower leaf replacement and hence a greater abundance of older gummed-up leaves (Watson, 1999). Previous modelling studies of the Meuse catchment considered the effect of land use change accounting for the different evaporation and interception characteristics of different types of vegetation. However, they did not consider the forest age distribution, 5 which varied during the period of simulation. This motivated the main objective of this paper, which consists of evaluating whether the observed dynamics of catchment response can be explained by variations in the evaporative demand of growing forest.
Even if the ecology of growing forests were well understood, there are no ready to use models that are applicable to different forest types under different climatic condi-10 tions. Moreover, there is a lack of knowledge on the exact history of the forest management in the Meuse basin, in terms of tree species, tree distribution and average age of the trees. Hence, we used a top-down approach to investigate the hypothesis that a difference in leaf conductance, as a result of stand age, can explain the flow anomaly observed in previous studies. 15 Although the motivation of this work is specific, our purpose is more general. It aims at assessing the possible impacts of human interference (not only of forest management) on the rainfall-runoff behaviour of the Meuse basin. The approach adopted consists of using a conceptual model which is dynamically evaluated over time in order to detect temporal trends or change points in model parameters. The model is con-20 structed in a way that it includes parameters that explicitly refer to leave conductance, parameters which were considered constant in previous studies, as is commonly done in hydrological modelling applications. The model evaluation framework is based on the dynamic identifiability analysis proposed by Wagener et al. (2003). Subsequently, we try to relate trends in parameter values to changes that may have occurred in the 25 catchment. The paper is organized as follows. We first present a description of the catchment and data, then we describe the model structure and set up, then we introduce the model evaluation framework and finally we present and discuss the results. 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment F. Fenicia et al.

Study area and data description
The study area is the Meuse basin upstream of the gauging station of Borgharen, which covers an area of approximately 21 000 km 2 . The Meuse basin is characterized by a rainfall-evaporation regime, which produces low flows during summer and high flows during winter. Mean annual precipitation over the basin is 950 mm/a, and is reasonably uniformly distributed throughout the year, while potential evaporation has a strong seasonal trend, with higher values during summer and lower values during winter (de Wit et al., 2007). The lithology of the catchment is characterized by three main zones. The southern part of the basin is dominated by sedimentary consolidated rocks. The central part be-10 longs to the Ardennes Massif, and consists of methamorphic rocks which are relatively impermeable. The northern part of the basin consists of lowlands characterized by sedimentary, unconsolidated rocks. Hydrologically useful information is available starting from year 1911 and consists of daily values of discharge measured at Borgharen, meteorological variables measured at De Bilt in the Netherlands, and precipitation amounts 15 form several raingauges in the catchment (Ashagrie et al., 2006).
The discharge at Borgharen is affected by extraction of water which is diverted outside the basin. The observed time series were corrected for these canal extractions (Ashagrie et al., 2006). However, detailed information was available only after 1990. Prior to this date the water extraction has been estimated on an annual basis, and has 20 been set at a fixed rate. Moreover, the discharge of the Meuse river is influenced by the operation of weirs and reservoirs, which particularly affect discharge during low flows. In fact, the Meuse river is navigable over a substantial part of its total length, connecting the Rotterdam-Amsterdam-Antwerp port areas to the industrial areas upstream. During low flow periods the weirs are operated to maintain a minimum stream 25 level for shipping.
Meteorological data required to determine potential evaporation were obtained from the station of De Bilt, located 180 km north of Borgharen. They include relative sunshine HESSD 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment Interactive Discussion duration, wind speed, air temperature, and relative humidity. These data were used to determine the potential evaporation with the Penman-Monteith equation. Observations of meteorological variables within the catchment area were not available for the entire observation period. However, analyzing the latest period, we could determine that the data from De Bilt provide an acceptable agreement with observations within the 5 catchment area. Precipitation is available at seven stations within the catchments. Information about the location and the quality of the data can be found in Ashagrie et al. (2006). For this study the rainfall series have been averaged over the entire basin using the Thiessen polygon method. 10 According to the CORINE data base (http://reports.eea.europa.eu/COR0-landcover/ en), land use in the Meuse catchment at present consists of 34% agricultural land, 20% pasture, 35% forest, 9% built-up area, and 2% wetlands. Historical land use change is not readily available at the basin scale, but can be reconstructed combining information from various areas within or in the proximity of the catchment. 15 In the Walloon Region of Belgium, which includes part of the Meuse catchment, it has been estimated that the total forest cover did not vary considerably during the past century, consisting of about 30% of the total area of the region (DGRNE, 2000). However, the forest composition changed significantly. The extension of artificial coniferous forest increased from 30% to 50% of the total forest cover between the years 1929-20 1984, and decreased slightly afterwards. Production forest, after a period of very strong expansion, has stabilized and is now probably in decline.
Farmland covers about 45% of total area of the Walloon Region, and did not vary significantly over time (DGRNE, 2000). According to Tu Min et al. (2005), in the Lorraine-Meuse the portion of farmland used for agriculture showed a decreasing trend until the 25 1970s, after which it started to increase. An opposite trend is seen in the remaining part, primarily used for pasture.
The urban expansion triggered by the industrialization process has been very rapid during the 19th century, and progressed at a slower rate during the first part of the HESSD 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment Interactive Discussion 20th century. It can be considered that the region was already highly urbanized by the 1950s, with a proportion of urban land of about 10-13% of the total catchment area (Tu Min et al., 2005). The construction of sewer systems and road networks progressed together with urban expansion, however, until the 1950s waste water was discharged into the natural environment without treatment. The construction of large treatment 5 plants began only after the 1960s (Garcier, 2007).

Methodology
The assessment of the effect of land use change on rainfall-runoff behaviour has been the subject of several studies and remains an area of active research (DeFries and Eshleman, 2004). Different approaches have been employed to investigate this prob-10 lem. The most widespread is based on observations and consists of paired catchment experiments. Two neighbouring catchments are monitored simultaneously, and following a calibration period one of the catchments is subjected to treatment while the other remains as control (Brown et al., 2005). This approach has several advantages, as it allows filtering out climate variability and it enhances the understanding on hydrologi-15 cal processes and their interaction with land use. However, its application is limited to relatively small catchments, and the results are difficult to generalize. Another common approach concerns the use of hydrological models. Model parameters are designed to represent specific catchment characteristics, and changes in land use are then interpreted based on parameter variation. Following this approach, Eck- In contrast to previous modelling studies which required prior knowledge of the type HESSD 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment Interactive Discussion of land use change which occurred during the time considered, we here adopt an approach that is fully "top-down", and may be useful for catchments where little information is available on structural modifications. The approach is based on the model identification framework proposed by Wagener et al. (2003), which consists of a dynamic model evaluation in a moving time window. Our purpose is to extract significant 5 trends in model parameter values which may then be related to land use change. The advantage of this approach is that it does not require prior information on land use change. Moreover, it provides continuous analysis of the time series and allows the identification of change points in hydrological behaviour.

Model Description
The model used in this study is an adaptation of the FLEX model (Fenicia et al., 2007) which is represented in Fig. 4. The main differences concern the interception component, which is parameterized according to different land uses, and the evaporation formula, which accounts for variable leaf conductance of growing forest. Similarly to Fenicia et al. (2007), also in this application we tested different model concept con- 15 figurations. We determined that the inclusion of an interception component improved significantly model performance, which justified the presence this process in the model concept.
Five land use types have been considered: agricultural land, pasture, deciduous forest, coniferous forest, and built-up area. To each of these land use types an inter-20 ception threshold I was assigned (Savenije, 2004). All thresholds were then multiplied by a calibration parameter C i , in order to allow them to vary proportionally. Interception thresholds of coniferous forest and built up areas were given a constant value while those of pasture, agriculture and deciduous forest were allowed to vary between a minimum in winter and a maximum in summer increasing or decreasing continuously 25 in autumn and spring respectively (Table 1).
Evaporation has been calculated with the Penman-Monteith (PM) equation using measured wind speed, sunshine duration, relative humidity, and air temperature ( Differences between evaporation characteristics of the various land types have been introduced by using different values of aerodynamic resistance (r a ) and canopy resistance (r c ). The energy available for evaporation from interception has been calculated assuming zero r c (Asdak et al., 1998). Transpiration from different types of land cover has been determined assuming different values of r c which for the forest area were 5 multiplied by a calibration parameter α. The parameter α accounts for the fact that forest transpiration may vary with stand age. According to our parameterization, an increase in α from 0.5 to 2 corresponds to a decrease of forest potential evaporation of about 50%. The canopy resistance has been calculated for different types of crops as a ratio 10 between minimal stomatal resistance (r l ) and leaf area index (ILA). In our analyses r l has been set at 100 s/m for grass, 350 s/m for forest, and 50 s/m for crop (e.g., Allen et al., 1998;Rana et al., 1994;Mascart et al., 1991;Pinty et al., 1989). The ILA for different kinds of land cover has been assumed to vary seasonally, as described in Gurtz et al. (1999). The values of r l mainly depend on the plant type. However, a great variability is reported in the literature. (Noilhan and Mahfouf, 1996) report that r l can be of the order of 40 s/m for a closed live crop canopy, and reach values up to 500 when plants experience maturation or senescence. The aerodynamic resistance has been estimated as a function of wind speed (u z ) and roughness lengths governing momentum transfer (z om ) and transfer of heat and 20 vapour (z oh ). For the forest the wind speed measured at meteorological stations has been transformed into the wind speed above the forest. Roughness lengths for different types of vegetation have been estimated from crop height (Allen et al., 1998) and from the plant area index as described by Shaw and Pereira (1982).
The energy available for transpiration has been calculated subtracting from the po-25 tential evaporation the energy consumed in the interception process. The potential transpiration for the entire catchment (T p ) was calculated averaging the contributions of different land uses according to their proportion within the catchment. Model parameters are reported in Table 4, other model fluxes are calculated as HESSD 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment  Table 3, while the fluxes definition is described in Fig. 4. Further details on the FLEX model can be found in Fenicia et al. (2007).

Model Evaluation
The model evaluation framework is based on the Dynamic Identification Analysis (DY-NIA) introduced by Wagener et al. (2003). This approach is based on the GLUE (Gen-5 eralized Likelihood Uncertainty Estimation) framework of Beven and Binley (1992), and it differs from it in the fact that the model, instead of being evaluated simultaneously over the entire observation record, is evaluated on a moving time window. The approach consists of the following steps. First a uniform exploration of the parameter space is performed. For this purpose we used the Latin hypercube sampling 10 technique on a predefined parameter range. Each parameter set corresponds to a model generation for which a measure of performance can be calculated. In this case we used the Nash and Sutcliffe efficiency measure (Nash and Sutcliffe, 1970). The performance is calculated on a moving time window. We divided the observation record in periods of four years, and calculated the model performance in each of them. For each 15 time window the performance of the best performing parameter sets is rescaled to produce a cumulative sum of one, while the other parameter sets are given a performance of zero. In the present case we retained the top 20% parameter sets. The rescaled performance can be used to build a cumulative distribution function of each parameter in each time window. This function can be used to determine various statistics, such 20 as median and quantiles of model parameters.
The original purpose of the approach is to identify periods of identifiability for individual parameters and gain understanding on the information content of the data. In this case, we used this approach to determine trends in the model parameters, and possibly relate them to structural catchment modifications, such as land use change. 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment

Results
The FLEX model in its present configuration is characterized by 10 parameters. Feasible ranges for these parameters are reported in Table 4, and have been determined through previous exploration of a wider parameter space. Initially, the model has been evaluated by varying all parameters within their feasible range, subsequently the num-5 ber of free parameters that were allowed to compensate for the modelling error has been reduced stepwise. At first the parameter space was explored with 40 000 parameter sets which have been generated through the Latin hypercube technique. All 10 model parameters were allowed to vary. Results of the model evaluation approach are shown in Fig. 5 and 6. 10 Figure 5 shows the trend in model parameters as calculated by the DYNIA approach. The continuous lines represent the median of the parameters, while the dashed lines represent the 20% and 80% quantiles. Although the model has a large number of parameters which can compensate for each other due to parameter correlation, it is possible to recognize a trend in some of the model parameters. The clearest trend is displayed by the parameter N lagf , which represents the time to peak of the basin. It is possible to see that this parameter shows a continuously decreasing trend, which indicates that the catchment response has become progressively faster with time. This may be due to the progressive urbanization in the catchment, and in particular to the construction of roads and sewer systems which cause the flow to propagate at a much 20 faster rate. Other causes may lie in the improved drainage from agricultural fields and river training works. Overall, the time to peak of the basin appears to have reduced by about one day, decreasing from an average of 3.5 to an average of 2.5 days.
Due to parameter correlation, the temporal trends of model parameters may be difficult to identify. Moreover, it may not be realistic to allow the variation of all model param-25 eters. Hence we performed a parameter search constraining some model parameters to vary over a much smaller range (Fig. 5). This also allows evaluating whether the modelling error can be corrected through a smaller number of free parameters. At first HESSD 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment F. Fenicia et al. we allowed 5 parameters to vary, namely C i , α, F c , β, and N lagf (sampling the parameter space with 20.000 parameter sets). Finally only the two parameters α and N lagf were left to vary (10 000 parameter samples).
The result of the model simulations are represented in Fig. 6. The dark line represents the observations, the gray line represents the optimal model performance over 5 the entire observation period. The grey line shows a similar trend as reported by Ashagrie et al. (2006). The coloured lines represent the optimal performance when parameter sets are allowed to vary over time. We can see that time varying parameters can correct most of the modelling error. Moreover, the number of free parameters does not affect model performance. In fact, similar performance can be obtained by varying 10, 5 or 2 parameters.
This demonstrates that there is large equifinality between parameter sets, meaning that many combinations of parameters can compensate for the same effect. In particular, the parameters associated with the evaporation routine try to compensate for the error in the water balance (e.g. L p and F c , which respectively determine the effect of 15 moisture stress on transpiration and the maximum amount of soil moisture). However, it is not very likely that these parameters changed over time. Fixing these parameters forces the modelling error to be compensated by the parameter α alone, which affects the stomatal resistance. If we do that, we can observe that the trend on the parameter α becomes clearer, and the uncertainty bounds become narrower, as less parameters 20 are allowed to vary. The trend on the parameter N lagf remains the same under this operation, which shows that this parameter is not correlated to this process. This is not surprising when looking at the model structure configuration. It is in fact the only model parameter that directly controls the lag time of the system. The outlier that can be observed in most parameters (e.g. α, K s , C i ) occurs in the mid 1970s. This outlier 25 is due to the year 1976 which was extremely dry, with a total precipitation of less than 400 mm/a.
The choice of parameters to fix depends on the hypothesis used and hence on an a-priory decision on which catchment characteristics are allowed to vary over time or to HESSD 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment have varied during the period of observation, which supports the hypothesis that α should be allowed to vary over time. Assuming variability of other parameters is more difficult to justify. However, it is possible that other events outside our knowledge may have taken place. Figure 7 represents the actual evaporation simulated by the model with fixed pa-10 rameter values and by the model where parameters α and N lagf were allowed to vary. The difference between the two models demonstrates the variability in evaporation of growing forest.

Discussion
The effect of land use change has been extensively investigated in the literature. Var- 15 ious approaches have been introduced, including paired catchment experiments and modelling studies. When dealing with large catchments the most widespread approach is based on the interpretation of model results. In contrast to existing modelling studies which require information on the development of land use practices over time, in this study we used a fully top-down approach that does not require preliminary information 20 on land use change. Our purpose is to identify temporal trends in model parameters which can be related to hypotheses on land use change.
As a result of our approach, we have determined a significant decreasing trend in the time of concentration of the catchment. We hypothesize that this is due to urbanization, improved field drainage and river training. The construction of roads, sewer systems, the extension of impervious areas, and the improved drainage from agricultural fields are some of the causes that may have reduced the lag time of the catchment. 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment A second and more complex outcome of this study relates to the closure of the water balance of the catchment. In contrast to the lag time of the system, which is influenced by a single model parameter, the catchment water balance is influenced by a large number of parameters, which result in equifinality. A reasonable hypothesis is to assume that the average age of the forest stands has varied over time. This 5 hypothesis is supported by the fact that the forest was heavily exploited during the middle part of the previous century (primarily by mining and heavy industry), while more recently the new forest policy has led to an increase in the average stand age. This indicates towards time variability of the parameter α which scales the canopy resistance to mature forest. 10 The temporal trend of the parameter α shows that it is consistently higher than 1 in the initial period of the century, lower than 1 in the years 1930-1965, and around 1 in the last part of the observation period (corresponding to mature forest). Based on the model hypotheses, this is interpreted as a forest that evaporates more intensively during the middle part of the century than at the beginning or at the end of the observation 15 period.

HESSD
It is important to notice that the values of the parameter α correspond to plausible ranges according to the literature. Depending on the range of this parameter, the forest potential evaporation varies between + or −20%, which is consistent with field experiments at several locations (e.g. Vertessy et al., 2001). 20 Although several studies have shown that stand age has a considerable effect on forest evaporation, this knowledge is rarely implemented in model studies that attempt to quantify the effect of land use change. With this study we do not claim to prove that evaporation of growing forest is the missing process to fully explain the anomaly in the Meuse time series. Instead, we want to show that this is a plausible hypothesis to be 25 studied further.
Most previous studies trying to find explanations for the observed anomaly in the rainfall-runoff behaviour of the Meuse concluded that this is likely due to climate change or data errors, rather than land use change (Tu Min et al., 2005;Ward et al., 2008; Asha-HESSD 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment  , 2006). With this study we have demonstrated that a more likely explanation may be found in the change in land use and land management. A concern that motivated some of the previous studies is the apparent increase in flood occurrences in this catchment. Previous studies demonstrated that the main reason is an increase in rainfall amounts. However, the catchment modifications that 5 we have investigated may also have contributed to such increase. Forests have a large impact on the magnitude of floods. If a forest evaporates more, then there may be more storage available in the soil and the groundwater, which can mitigate the impact of floods. Also, a shorter time of concentration of the catchment may increase flood peaks which are less attenuated. 10 These considerations deserve further research to confirm of reject the hypothesis on the evaporative demand of growing forest. A possibility is to collect and interpret data from historical archives, which can provide information on wood production and wood consumption during the study period. This would serve as an independent source of information to test our hypothesis.

Conclusions
The purpose of this study is to analyze the effect of land use change on the Meuse basin during the past century. As little quantitative information is available over this extended period, we assumed a fully top-down approach which does not require prior information on land use. The approach is based on a dynamic evaluation of a hydro-20 logical model and on the consequent interpretation of the temporal trends obtained in model parameters.
Our results show that the time of concentration of the catchment has significantly reduced during the observation period. We think that this may be due to urbanization, improved field drainage and river training. A second result relates to the hypothesis that 25 variable forest age may have significantly affected transpiration and consequently the water balance of the catchment. We show that this is a plausible hypothesis according HESSD 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment F. Fenicia et al. to the literature and to our knowledge of the study area. The Meuse basin has been subject to several studies that investigated the reasons of the observed change in rainfall-runoff behaviour. Whereas previous studies concluded that this change is due to climate change rather than to land use change, our investigation indicate that land use and land management changes are likely causes 5 for observed changes in the hydrological behaviour of the Meuse river basin. 5,2008 Anomaly in the rainfall-runoff behaviour of the Meuse catchment F. Fenicia et al.   Table 3. Schematic description of model equations.