Artificial surfaces characteristics and sediment connectivity explain muddy flood hazard in Wallonia

Highlights

• A methodology was developed to predict muddy floods with good explanatory power.

• The characteristics of urbanized land are key variables.

• Sediment connectivity indices improve the predictive power of the model.

Abstract

Over the last decades, the off-site damages caused by muddy floods have been of growing concern throughout much of Western Europe, and particularly in Wallonia (Belgium). A reliable identification of locations with a high muddy flood hazard is thus a key issue in this context. The main objective of this study was therefore to build and evaluate a muddy flood hazard prediction model in order to assess the probability of occurrence of muddy floods at any specific location. A logistic regression approach was used to explain muddy flood occurrence using a database of 442 muddy flood-affected sites and an equal number of homologous non-flooded sites. Explanatory variables were related to geomorphology, land use, sediment production and sediment connectivity in the contributing area. The prediction quality of the model was then validated using an independent dataset composed of 48 pairs of flooded and non-flooded sites. The best muddy flood hazard assessment model required a total of 5 explanatory variables as inputs: the mean slope, a sediment connectivity index, as well as the proportion, spatial aggregation and proximity to the outlet of artificial surfaces. The model resulted in a prediction quality of 76% (calibration dataset) and 81% (validation dataset). Including the characteristics of artificial surfaces substantially improved the model quality (p-values from 10^− 11 to 10^− 5). All three variables related to artificial surfaces showed negative correlations with the muddy flood hazard. The proportion of cropland was not included in the final model, but this variable was strongly inversely correlated to the proportion of artificial surfaces. Besides the characteristics of artificial surfaces, sediment connectivity also showed significant explanatory power (p-value of 10^− 12). A positive correlation between sediment connectivity and muddy flood hazard was found. Future muddy flood hazard models should therefore include both artificial surfaces characteristics and sediment connectivity-related information. Given the good prediction quality, the developed statistical model could be used as a reliable tool to prioritize sites at risk of muddy floods in order to install mitigation measures.

Introduction

Over the last two decades, muddy flood events have drawn widespread attention from the public, decision makers and scientists alike. Muddy floods are caused by water flowing from agricultural fields and carrying large quantities of suspended sediments or bedload (Boardman et al., 2006). This soil erosion-related process triggers off-site problems such as the flooding and damaging of infrastructure or private property downslope. Muddy floods thereby not only result in financial costs to restore the damaged infrastructure, but also in psychological stress to the people regularly exposed to them, in clogging of trenches, sedimentation of gravel-bedded rivers, silting up of retention ponds, or contamination of surface waters (Bielders et al., 2003, Boardman et al., 2009, Verstraeten and Poesen, 1999).

As for many other regions in Europe, the loess belt, located in central Belgium (Fig. 1), is frequently affected by muddy floods (Verstraeten and Poesen, 1999, Bielders et al., 2003, Evrard et al., 2010). Indeed, over a 10-year period, 79% of the municipalities in central Belgium were confronted with one or several muddy floods (Evrard et al., 2007). Among those, 22% experienced > 10 floods in 10 years (Evrard et al., 2007). The resulting off-site damage has been estimated to range between 12.5 and 122 10⁶ € yr^− 1 depending on the year (Evrard et al., 2007). In order to tackle this issue and to better target erosion and muddy flood mitigation measures, tools must be developed that allow assessing the muddy flood hazard for any given location at the regional scale based on a better understanding of the relative importance of various driving factors.

A muddy flood risk assessment starts with the characterization of its hazard, i.e., the probability for a muddy flood to happen in a specific place and period of time (Varnes, 1984). Various ways to quantify the muddy flood hazard have been developed. Muddy flood susceptibility of a given site is derived either directly from field measurements or indirectly from intrinsic properties, causal factors or processes believed to influence the occurrence of a muddy flood (Boardman et al., 2003). Whereas field-based methods are highly appropriate for small-scale studies, they are time consuming and costly to implement over large areas. Hence, indirect methods are the most appropriate for regional approaches, despite frequent limitations in the availability of relevant data. The various parameters believed to govern muddy flood susceptibility may be used in heuristic approaches by assigning different weights depending on their relative importance but this kind of approach may entail subjectivity. Statistical techniques such as stepwise modeling via univariate or multivariate models are more objective quantitative approaches. In these approaches, the functional relationship between a set of factors (explanatory variables) and muddy flood susceptibility is investigated. The most commonly used models are discriminant analysis, neural networks and regressions (Shu-Quiang and Unwin, 1992, Atkinson et al., 1998, Chung and Fabbri, 1999, Guzzetti et al., 1999, Boardman et al., 2003, Akgün and Türk, 2011). This approach comes with the advantages of conceptual simplicity and statistical objectivity, but requires continuous and sufficient data (Boardman et al., 2003). Moreover, the results cannot be extrapolated to different contexts as they are often dependent on the dataset used (Evrard et al., 2010). Another possibility is to model processes believed to govern the muddy flood occurrence with the use of physically-based equations (Stankoviansky et al., 2010). This kind of technique is potentially the most accurate in terms of characterization of the processes but requires complex inputs, often unavailable at large spatial scales. Regardless of the method used to quantify the hazard, the factors and processes believed to govern the muddy flood occurrence need to be known.

Numerous factors influencing muddy flood susceptibility have been identified (Table 1). Some of these factors are invariant or almost invariant over time periods of a few years, whereas others are variable at annual or even sub-annual time scales. Topography (slope length and steepness, slope configuration) and the intrinsic soil properties are mostly invariant. The proportion of cropland (without distinguishing among crop types), pasture, forest, urbanized areas, roads, and features affecting surface drainage (e.g. ditches, banks), are also reasonably perennial in the context of western agriculture. On the other hand, weather conditions (rainfall intensity), the spatial organization of crops, agricultural practices and the resulting surface conditions (e.g., roughness, vegetation cover, structure of the topsoil) are highly dynamic parameters at the intra-annual scale. Although time varying factors have proven useful in explaining muddy floods a posteriori (Bielders et al., 2003, Evrard et al., 2010), it is difficult to use such factors in predictive models since it is generally not known how they will evolve, even one year in advance. For instance, the assignment of various crops to the parcels in a watershed is dependent on non-concerted decisions by individual farmers, who themselves are influenced by internal constraints (e.g., crop rotation requirements) and external driving forces (e.g., markets).

None of the above-listed factors explicitly take into account sediment production and routing. The development of sediment connectivity indices to characterize sediment transport in the landscape has been at the center of an active field of research for several years (Hooke, 2003, Croke et al., 2005, de Vente and Poesen, 2005, Borselli et al., 2008, Sougnez et al., 2011, Vigiak et al., 2012, Cavalli et al., 2013, Jamshidi et al., 2014, Gay et al., 2016). Sediment connectivity is defined as “the degree of linkage which controls sediment fluxes through landscapes, and, in particular, between sediment sources and downstream areas” (Cavalli et al., 2013). Previous studies have demonstrated the importance of sediment connectivity-related features in determining the muddy flood hazard (Boardman et al., 2009, Alder et al., 2015). Evrard et al. (2007) for instance, emphasized the importance of landscape features such as roads, sunken lanes, ditches, and gullies that enhance the transfer of water and sediment towards the flooded areas. Verstraeten and Poesen (1999) noticed that in numerous cases, muddy floods are related to roads situated in a thalweg position. Describing sediment connectivity in a landscape allows to consider the spatial organization of various physiographic units and their contribution to sediment production, transport and deposition (Borselli et al., 2008, Boardman and Vandaele, 2016). The recent development of sediment connectivity indices now makes it possible to include this concept explicitly in predictive models of muddy floods at large spatial scales.

The main objective of this study was therefore to quantify the contribution of various intrinsic watershed characteristics to the muddy flood hazard in Wallonia, and to develop a reliable statistically-based model capable of assessing the muddy flood hazard at any given location in Wallonia. A secondary objective was to evaluate whether the inclusion of recently developed sediment connectivity indices would improve the predictive power of such a model. The resulting model is to be used for land use planning purposes and for targeting priority areas with anti-erosion measures.