Modelling Soil Erosion at European Scale: Towards Harmonization and Reproducibility. Bosco, C.; de Rigo, D.; Dewitte, O.; Poesen, J.; and Panagos, P. 15(2):225–245.
Modelling Soil Erosion at European Scale: Towards Harmonization and Reproducibility [link]Paper  doi  abstract   bibtex   
Soil erosion by water is one of the most widespread forms of soil degradation. The loss of soil as a result of erosion can lead to decline in organic matter and nutrient contents, breakdown of soil structure and reduction of the water-holding capacity. Measuring soil loss across the whole landscape is impractical and thus research is needed to improve methods of estimating soil erosion with computational modelling, upon which integrated assessment and mitigation strategies may be based. Despite the efforts, the prediction value of existing models is still limited, especially at regional and continental scale, because a systematic knowledge of local climatological and soil parameters is often unavailable. A new approach for modelling soil erosion at regional scale is here proposed. It is based on the joint use of low-data-demanding models and innovative techniques for better estimating model inputs. The proposed modelling architecture has at its basis the semantic array programming paradigm and a strong effort towards computational reproducibility. An extended version of the Revised Universal Soil Loss Equation (RUSLE) has been implemented merging different empirical rainfall-erosivity equations within a climatic ensemble model and adding a new factor for a better consideration of soil stoniness within the model. Pan-European soil erosion rates by water have been estimated through the use of publicly available data sets and locally reliable empirical relationships. The accuracy of the results is corroborated by a visual plausibility check (63\,% of a random sample of grid cells are accurate, 83\,% at least moderately accurate, bootstrap p ≤ 0.05). A comparison with country-level statistics of pre-existing European soil erosion maps is also provided. [Excerpt: Physically based and empirical models] Distinct modelling approaches can lead to significantly different soil erosion rates even when the same model is applied within the same region (Shen et al., 2009). The way the model is implemented (i.e. with the selection of different algorithms when available), the use of data sets with different resolution or accuracy (Merritt et al., 2003) and the provenance of a given data set (Buneman et al., 2000; Simmhan et al., 2005) can play a key role on the output. When incomplete or missing, these pieces of information may affect the assessment of the actual accuracy of data to be used as input, therefore weakening - or in some circumstances even compromising - the application of theoretically accurate models (Saltelli et al., 2010). [\n] While physically based models can in principle offer scientifically sound methods for deriving soil erosion rates from a plethora of detailed input data, their practical suitability for regional- or continental-scale assessment is controversial (Bras et al., 2003). The enormous gap between the type and accuracy of the needed input parameters and the actual availability of harmonized, verifiable large-scale data sets limits the applicability of such models (Stroosnijder, 2005). In theory, when working with physically based models, possibly all the requested parameters are measurable and then could be considered as ” known”. In practice, often the parameters have to be calibrated against observed data (Beck et al., 1995; Wheater et al., 1993). This calibration adds nonnegligible uncertainty in the parameters' values. The heterogeneity, variability and uncertainty associated with input parameter values and their interpolation in spatial or temporal domains outside the observed ones should be considered as key factors (Saltelli et al., 2010; Jetten et al., 2003) which may partially explain why lumped regression-based models can perform better than more complex physically based models (Bosco et al., 2013; de Vente et al., 2013). [\n] If at watershed scale a trend is observed (Daniel et al., 2011) to complement or replace physically based models with machine-learning techniques (which are advanced empirical modelling techniques), at regional scale the adaptation of widely adopted empirical models and their improvement with the same techniques could play a meaningful role. Regional-scale approximations with robust empirical modelling could provide useful - even if necessarily less accurate - support for risk assessors involved in decision-making processes over wide spatial extents. The main limit of such approach is that empirical models do not necessarily model the right process and should only be used for the range of conditions they were developed for (Hessel, 2002; de Vente et al., 2013). [\n] Computational science is emerging as one of the central topics within environmental modelling (Casagrandi and Guariso, 2009). To overcome the above problems, reproducible computational methods based on free software and data are increasingly needed (Stallman, 2005, 2009; Peng, 2011). This may also help to reuse in a controlled way empirical equations for compensating the lack of detailed data. [\n] [...] [Results and discussion] [...] The well-known role of natural vegetation in mitigating soil erosion (Cerdan et al., 2010; de Rigo and Bosco, 2011; Maetens et al., 2012) may be observed by comparing the presented map with pan-European forest maps (e.g. Kempeneers et al., 2012) and vegetation maps (e.g. Martin et al., 2010, derived from CLC 2006). Brittany, northern Portugal and western Norway show high soil erosion rates that seem to be related to the pattern of the interpolated rainfalls. Especially in northern Portugal and Norway, the positive relationship between erosion rates and slope length (Cerdan et al., 2010), that increases the runoff rates, appears to be enhanced by the intense precipitation pattern. [...] [\n] The map shows that approximately 14 % of the European territory is characterized by a significant soil erosion rate (moderate/high level), which is in line with previous estimations that 15-16 % of Europe's land area is affected by soil erosion (Cerdan et al., 2010; EEA, 2003). [\n] It is clear from the map that soil erosion by water is a major problem in many parts of Europe. The average rate of soil erosion by water across the EU-28 is 2.76 t ha-1 yr-1, excluding Cyprus (CY), Greece (GR) and Malta (MT). Just over 7 % of cultivated land (arable and permanent cropland) in the EU-25 (excluding GR, CY and MT) is estimated to suffer from moderate to severe erosion (i.e. OECD definition of $>$ 11 t ha-1 yr-1 ). This corresponds approximately to the entire area of Bulgaria. In comparison, only 2 % of permanent grasslands and pasture in the EU-25 is estimated to suffer from moderate to severe erosion. This demonstrates the importance of maintaining permanent vegetation cover as a mechanism to combat soil erosion. [\n] Several countries appear as not affected by notable soil erosion. Others, mainly in the southern part of Europe, are particularly susceptible to erosion, showing a soil erosion rate much higher than the European average. However, such values can be misleading: erosion rates in many areas can be considerably higher, even in those countries having a low average. The opposite is also true for countries with higher values. [\n] Considering the European ecozones (based on the FAO ecological zoning FAO, 2001, 2012), the mountain system shows a mean soil erosion rate 2-3 times higher than the average (from 4.06 t ha-1 yr-1 of the subtropical mountain system to 7.8 t ha-1 yr-1 of the boreal mountain system). [\n] As already mentioned, there is a high probability for some of the model results to be overestimated. The R factor uncertainty, the coarse resolution of the layers used for calculating the K factor, the CLC 2006 misclassification and the presence of areas having a stoniness value much higher than the value indicated by the underlying soil database (e.g. northern Scotland) can be at the basis of many of the uncertain estimations. Further investigations are suggested on the key role of land cover changes and misclassifications (CORINE Land Cover 2006 is found currently accurate in no more than 69 % of the sampled cells, bootstrap p ≤ 0.05) and of forests and vegetation, especially in mountainous areas with intense precipitation. [...]
@article{boscoModellingSoilErosion2015,
  title = {Modelling Soil Erosion at {{European}} Scale: Towards Harmonization and Reproducibility},
  author = {Bosco, Claudio and de Rigo, Daniele and Dewitte, Olivier and Poesen, Jean and Panagos, Panos},
  date = {2015-02},
  journaltitle = {Natural Hazards and Earth System Science},
  volume = {15},
  pages = {225--245},
  issn = {1684-9981},
  doi = {10.5194/nhess-15-225-2015},
  url = {https://doi.org/10.5194/nhess-15-225-2015},
  abstract = {Soil erosion by water is one of the most widespread forms of soil degradation. The loss of soil as a result of erosion can lead to decline in organic matter and nutrient contents, breakdown of soil structure and reduction of the water-holding capacity. Measuring soil loss across the whole landscape is impractical and thus research is needed to improve methods of estimating soil erosion with computational modelling, upon which integrated assessment and mitigation strategies may be based. Despite the efforts, the prediction value of existing models is still limited, especially at regional and continental scale, because a systematic knowledge of local climatological and soil parameters is often unavailable. A new approach for modelling soil erosion at regional scale is here proposed. It is based on the joint use of low-data-demanding models and innovative techniques for better estimating model inputs. The proposed modelling architecture has at its basis the semantic array programming paradigm and a strong effort towards computational reproducibility. An extended version of the Revised Universal Soil Loss Equation (RUSLE) has been implemented merging different empirical rainfall-erosivity equations within a climatic ensemble model and adding a new factor for a better consideration of soil stoniness within the model. Pan-European soil erosion rates by water have been estimated through the use of publicly available data sets and locally reliable empirical relationships. The accuracy of the results is corroborated by a visual plausibility check (63\,\% of a random sample of grid cells are accurate, 83\,\% at least moderately accurate, bootstrap p ≤ 0.05). A comparison with country-level statistics of pre-existing European soil erosion maps is also provided.

[Excerpt: Physically based and empirical models]

Distinct modelling approaches can lead to significantly different soil erosion rates even when the same model is applied within the same region (Shen et al., 2009). The way the model is implemented (i.e. with the selection of different algorithms when available), the use of data sets with different resolution or accuracy (Merritt et al., 2003) and the provenance of a given data set (Buneman et al., 2000; Simmhan et al., 2005) can play a key role on the output. When incomplete or missing, these pieces of information may affect the assessment of the actual accuracy of data to be used as input, therefore weakening - or in some circumstances even compromising - the application of theoretically accurate models (Saltelli et al., 2010).

[\textbackslash n] While physically based models can in principle offer scientifically sound methods for deriving soil erosion rates from a plethora of detailed input data, their practical suitability for regional- or continental-scale assessment is controversial (Bras et al., 2003). The enormous gap between the type and accuracy of the needed input parameters and the actual availability of harmonized, verifiable large-scale data sets limits the applicability of such models (Stroosnijder, 2005). In theory, when working with physically based models, possibly all the requested parameters are measurable and then could be considered as ” known”. In practice, often the parameters have to be calibrated against observed data (Beck et al., 1995; Wheater et al., 1993). This calibration adds nonnegligible uncertainty in the parameters' values. The heterogeneity, variability and uncertainty associated with input parameter values and their interpolation in spatial or temporal domains outside the observed ones should be considered as key factors (Saltelli et al., 2010; Jetten et al., 2003) which may partially explain why lumped regression-based models can perform better than more complex physically based models (Bosco et al., 2013; de Vente et al., 2013).

[\textbackslash n] If at watershed scale a trend is observed (Daniel et al., 2011) to complement or replace physically based models with machine-learning techniques (which are advanced empirical modelling techniques), at regional scale the adaptation of widely adopted empirical models and their improvement with the same techniques could play a meaningful role. Regional-scale approximations with robust empirical modelling could provide useful - even if necessarily less accurate - support for risk assessors involved in decision-making processes over wide spatial extents. The main limit of such approach is that empirical models do not necessarily model the right process and should only be used for the range of conditions they were developed for (Hessel, 2002; de Vente et al., 2013).

[\textbackslash n] Computational science is emerging as one of the central topics within environmental modelling (Casagrandi and Guariso, 2009). To overcome the above problems, reproducible computational methods based on free software and data are increasingly needed (Stallman, 2005, 2009; Peng, 2011). This may also help to reuse in a controlled way empirical equations for compensating the lack of detailed data.

[\textbackslash n] [...] [Results and discussion] [...] The well-known role of natural vegetation in mitigating soil erosion (Cerdan et al., 2010; de Rigo and Bosco, 2011; Maetens et al., 2012) may be observed by comparing the presented map with pan-European forest maps (e.g. Kempeneers et al., 2012) and vegetation maps (e.g. Martin et al., 2010, derived from CLC 2006). Brittany, northern Portugal and western Norway show high soil erosion rates that seem to be related to the pattern of the interpolated rainfalls. Especially in northern Portugal and Norway, the positive relationship between erosion rates and slope length (Cerdan et al., 2010), that increases the runoff rates, appears to be enhanced by the intense precipitation pattern. [...] 

[\textbackslash n] The map shows that approximately 14 \% of the European territory is characterized by a significant soil erosion rate (moderate/high level), which is in line with previous estimations that 15-16 \% of Europe's land area is affected by soil erosion (Cerdan et al., 2010; EEA, 2003). [\textbackslash n] It is clear from the map that soil erosion by water is a major problem in many parts of Europe. The average rate of soil erosion by water across the EU-28 is 2.76 t ha-1 yr-1, excluding Cyprus (CY), Greece (GR) and Malta (MT). Just over 7 \% of cultivated land (arable and permanent cropland) in the EU-25 (excluding GR, CY and MT) is estimated to suffer from moderate to severe erosion (i.e. OECD definition of {$>$} 11 t ha-1 yr-1 ). This corresponds approximately to the entire area of Bulgaria. In comparison, only 2 \% of permanent grasslands and pasture in the EU-25 is estimated to suffer from moderate to severe erosion. This demonstrates the importance of maintaining permanent vegetation cover as a mechanism to combat soil erosion.

[\textbackslash n] Several countries appear as not affected by notable soil erosion. Others, mainly in the southern part of Europe, are particularly susceptible to erosion, showing a soil erosion rate much higher than the European average. However, such values can be misleading: erosion rates in many areas can be considerably higher, even in those countries having a low average. The opposite is also true for countries with higher values.

[\textbackslash n] Considering the European ecozones (based on the FAO ecological zoning FAO, 2001, 2012), the mountain system shows a mean soil erosion rate 2-3 times higher than the average (from 4.06 t ha-1 yr-1 of the subtropical mountain system to 7.8 t ha-1 yr-1 of the boreal mountain system). 

[\textbackslash n] As already mentioned, there is a high probability for some of the model results to be overestimated. The R factor uncertainty, the coarse resolution of the layers used for calculating the K factor, the CLC 2006 misclassification and the presence of areas having a stoniness value much higher than the value indicated by the underlying soil database (e.g. northern Scotland) can be at the basis of many of the uncertain estimations. Further investigations are suggested on the key role of land cover changes and misclassifications (CORINE Land Cover 2006 is found currently accurate in no more than 69 \% of the sampled cells, bootstrap p ≤ 0.05) and of forests and vegetation, especially in mountainous areas with intense precipitation. [...]},
  keywords = {*imported-from-citeulike-INRMM,~INRMM-MiD:c-13508255,bias-toward-primacy-of-theory-over-reality,boreal-mountain-system,computational-science,continental-scale,data-integration,e-rusle,ecological-zones,empirical-equation,ensemble,environmental-modelling,erosivity,europe,fagus-sylvatica,geospatial-semantic-array-programming,gis,gnu-octave,integrated-modelling,knowledge-integration,larix-decidua,mastrave-modelling-library,modelling,mountainous-areas,relative-distance-similarity,reproducibility,reproducible-research,rusle,semantic-array-programming,semap,soil-erosion,soil-resources,stoniness,subtropical-mountain-system,uncertainty,visual-assessment},
  number = {2},
  options = {useprefix=true}
}
Downloads: 0