Assessment of Gully Erosion Susceptibility Using Multivariate Adaptive Regression Splines and Accounting for Terrain Connectivity
Abstract
In this work, we assessed gully erosion susceptibility in two adjacent cultivated catchments of Sicily (Italy) by employing multivariate adaptive regression splines and a set of geo-environmental variables. To explore the influence of hydrological connectivity on gully occurrence, we measured the changes of performance occurred when adding one by one nine predictors reflecting terrain connectivity to a base model that included contributing area and slope gradient. Receiver operating characteristic (ROC) curves and the area under the ROC curve were used to evaluate model performance. Gully predictive models were trained in both the catchments and submitted to internal (in the calibration catchment) and external (in the adjacent one) validation, using samples extracted both from all cells of the catchments and only from cells located along flow concentration axes. Model evaluation on the entire catchments shows outstanding predictive performance of models that either include or do not include the predictors selected to reflect potential hydrological connectivity. Conversely, area under the ROC curve values measured on flow concentration axes reveals that almost all the additional predictors improve the performance of the base model, but the most enhanced increase of accuracy occurs when upstream drainage density of each landscape position is included as predictor of gully occurrence. Copyright © 2017 John Wiley & Sons, Ltd.
Introduction
Gully erosion is responsible for large soil losses in a wide range of environments. The contribution of gully erosion to the total sediment budget is particularly relevant in agricultural catchments, where it causes severe damage to crops and hampers agricultural traffic (Maugnard et al., 2014; Poesen et al., 2003). Moreover, gullies connect upland areas with the stream network, modifying water and sediment connectivity, especially during intense rainstorms (Torri & Poesen, 2014).
Soil loss caused by gully erosion may be estimated by using physically based models such as Chemicals, Runoff, and Erosion from Agricultural Management Systems (Knisel, 1980); Ephemeral Gully Erosion Model (Merkel et al., 1988; Woodward, 1999) and the Water Erosion Prediction Project routine for linear erosion (Flanagan & Nearing, 1995). Reliable estimations of eroded gully volumes can also be achieved by measuring their lengths and applying empirical length–volume relationships (e.g. Caraballo-Arias et al., 2014, 2015; Nachtergaele et al., 2001). However, these models do not predict where gully erosion is more likely to occur, which is important for planning erosion control measures (Poesen et al., 2003).
Susceptibility to gully erosion can be assessed by defining functional relationships between gully locations and spatial variability of controlling factors such as topography, soil, parent material, rainfall, land use and vegetation cover. To this aim, different statistical techniques have been used, including logistic regression (e.g. Dewitte et al., 2015; Meyer & Martínez-Casasnovas, 1999; Selkimäki & González-Olabarria, 2016), classification and regression trees (e.g. Gómez-Gutiérrez et al., 2009a; Jurchescu & Grecu, 2015), multivariate adaptive regression splines (MARS; Gómez-Gutiérrez et al., 2009a, 2009b, 2015), conditional analysis (e.g. Conoscenti et al., 2013; Magliulo, 2012), information value (Conforti et al., 2011; Lucà et al., 2011) and stochastic gradient treeboost (Angileri et al., 2016). In this approach, modelling of gully occurrence is achieved by using different topographic attributes extracted from digital elevation models (DEMs), often including categorical or continuous predictor variables related to soil, bedrock, land use and vegetation. However, the employed predictors rarely account for hydrological connectivity, although this property is expected to control runoff volume and thus its erosive power.
Hydrological connectivity refers to the movement of water from one part of the landscape to another (Bracken & Croke, 2007). Connectivity is variable in time and space and is mainly controlled by climate, vegetation, topography, soil and bedrock properties. Two aspects of hydrological connectivity have been recognized: static/structural and dynamic/functional connectivity. The static elements of hydrological connectivity refer to pedological, topographical and geological properties, whereas dynamic connectivity is controlled by both long-term landscape development (e.g. land use changes) and short-term variations (e.g. storm intensity and duration) (Bracken & Croke, 2007). Bracken et al. (2013) propose the term terrain connectivity for the approaches focusing on topographic controls on hydrological connectivity and state that with a purely structural approach, we can only infer potential hydrological connectivity. So far, most studies have focused on the static hydrological connectivity. For example, Lane et al. (2004, 2009) proposed to model static connectivity by using the topographic wetness index (TWI). TWI is a function of contributing area and slope gradient, which allows for estimating areas of high soil moisture (Beven & Kirkby, 1979). Lane et al. (2009) concluded that a topographic parameter could be successfully used to model spatial variability of hydrological connectivity. Borselli et al. (2008) developed a topography-based index of connectivity that includes an upslope and a downslope component. These are calculated by using slope gradient, contributing area, flow distance to drainage network and the C-factor of Universal Soil Loss Equation–Revised Universal Soil Loss Equation models (Renard et al., 1997; Wischmeier & Smith, 1978) as a weighting factor. Cavalli et al. (2013) proposed to modify this index by using a weighting factor based on the roughness index (RI), which is a local measure of topographic surface roughness. Also, Gay et al. (2016) modified the Borselli et al. (2008) connectivity index by adding a parameter related to drainage density.
In this study, we evaluated susceptibility to gully erosion in two adjacent cultivated catchments of Sicily (Italy), characterized by similar geological and geomorphological settings and affected by intense erosion processes. The main objectives of this experiment were to (i) achieve accurate predictions of gully occurrence by using MARS (Friedman, 1991) as modelling technique; (ii) verify whether the accuracy of the base model is improved by using predictors related to potential hydrological connectivity and (iii) test if variations of accuracy can be more easily detected by focusing model evaluation along flow concentration axes.
Materials and Methods
Study Areas
The experiment was carried out in two small catchments located in central-western Sicily (Italy) (Figure 1). The western (hereafter called W1) and the eastern catchments (W2) extend for around 6·2 and 9 km2 respectively. Elevation ranges of W1 and W2 are 185–576 m asl (mean = 302·9 m; SD = 46·9 m) and 209–571 m asl (mean = 345·3 m; SD = 48·0 m) respectively. Both areas are characterized by an undulating topography, with mean slope angles of W1 and W2 equal to 10·1° (SD = 5·0°) and 9·7° (SD = 6·9°) respectively. The climate is Mediterranean, with mild and wet winters and hot and dry summers. Average annual rainfall recorded at the Camporeale station (350 m asl) is 564·2 mm, with minimum and maximum monthly values occurring in July (5·6 mm) and December (83·7 mm) respectively. The study areas are mainly underlain by (i) eluvial-colluvial deposits (38% of W1 and 17% of W2); (ii) sands of the late Miocene Terravecchia Formation (20% of W1 and 30% of W2) and (iii) clays of the middle-late Miocene Castellana Sicula Formation (22% of W1 and 20% of W2) and by silty-clays and sandy-silts of the Terravecchia Formation (6% of W1 and 18% of W2) (Figure 2a and Table 1). Soils are mainly regosols and vertisols with fine-medium texture (Fierotti, 1988).
| Catchment W1 | Catchment W2 | |||||
|---|---|---|---|---|---|---|
| Lithology | Area (km2) | Gully density (km−1) | Average slope (°) | Area (km2) | Gully density (km−1) | Average slope (°) |
| Alluvial deposits | 0·44 | 0·37 | 7·0 | 0·32 | 0·00 | 9·3 |
| Alluvial deposits (terraced) | - | - | - | 0·21 | 0·00 | 3·5 |
| Ancient landslides | 0·13 | 0·55 | 10·7 | 0·01 | 0·00 | 18·2 |
| Clays | 1·35 | 0·54 | 9·9 | 1·71 | 0·24 | 8·4 |
| Clays interbedded with sandstones | 0·07 | 0·00 | 9·2 | 0·08 | 0·90 | 8·7 |
| Conglomerates | 0·13 | 0·19 | 12·9 | 0·03 | 0·00 | 9·7 |
| Eluvial-colluvial deposits | 2·27 | 1·37 | 9·7 | 1·45 | 0·28 | 8·8 |
| Limestones | 0·01 | 0·00 | 10·7 | - | - | - |
| Sands | 1·19 | 0·36 | 10·7 | 2·56 | 0·20 | 13·0 |
| Sands interbedded with sandstones | 0·08 | 0·00 | 16·5 | 0·54 | 0·16 | 12·5 |
| Silty-clays and sandy-silts | 0·37 | 0·03 | 12·0 | 1·56 | 0·09 | 7·1 |
| Land cover | ||||||
| Arable lands | 4·89 | 0·90 | 10·0 | 4·06 | 0·37 | 8·3 |
| Complex cultivation patterns | 0·06 | 0·00 | 16·1 | 0·32 | 0·00 | 10·7 |
| Natural grassland | 0·13 | 0·00 | 16·7 | 0·71 | 0·00 | 21·9 |
| Orchards, groves and tree plantations | 0·13 | 0·00 | 13·1 | 0·45 | 0·00 | 12·8 |
| Vineyards | 0·64 | 0·06 | 8·8 | 2·54 | 0·04 | 8·7 |
| Willow brush | 0·18 | 0·48 | 7·2 | 0·22 | 0·00 | 6·3 |
Almost the entire catchments are characterized by farming activities (Figure 2b and Table 1). Land use is mainly arable lands, which are almost all devoted to cereal cropping (81% of W1 and 49% of W2), and vineyards (11% of W1 and 31% of W2). These activities are affected by intense erosion processes, which reduce agricultural production and generate a significant economic damage for the local population. In particular, gully erosion causes landscape dissection (hampering the movement of farm vehicles and animals) and soil loss from agricultural fields. Gullies mainly occur in clearly defined natural drainage lines but some also develop along man-made linear features such as parcel borders, access roads, wheel tracks or plow furrows. Most of the gullies are routinely obliterated by farming operations and can thus be classified as ephemeral gullies.
Gully Inventory
The gully database was prepared in two steps. The first one consisted of a visual interpretation of a satellite image of 3 May 2015, which is available on Google Earth (GE). Gullies were manually digitized as linear features. Discontinuous gullies were mapped as separate gullies; incisions with abrupt width changes were split into two or more separate gullies. For gullies developing from rill coalescence, an arbitrary minimum channel width of 0·5 m was used to identify gully initiation points.
In the second step, the position and shape of the gullies digitized in GE were compared with contour lines and runoff concentration axes extracted from the DEM. Gullies matching drainage lines, apart from those occurring where flow direction is affected by man-made features, were selected for the analysis of gully erosion susceptibility. Gullies occurring on planar slopes, with no evident runoff concentration axes, were not included in the analysis as their origin may be (i) not related to overland flow concentration (e.g. due to seepage or piping erosion) or (ii) related to a runoff concentration that cannot be detected with the available DEM. Then, the shape of the selected gullies was modified to exactly fit that of the corresponding drainage lines extracted from the DEM. This way, we ensured that values of A would grow downslope from the head of each gully.
The vector map of the gullies was converted into a 2 m raster with values equal to 0 (gully) and 1 (non-gully) and employed as dependent variable in the statistical analysis of gully erosion susceptibility.
Predictor Variables
In this experiment, susceptibility to gully erosion was evaluated by using bedrock lithology, land cover and a set of terrain attributes as predictor variables of gully spatial distribution.
Lithology of the study areas was derived from a 1:50,000 scale geological map (Catalano et al., 2010), whereas land cover was classified from a 1:10·000 scale map of Sicily (Gianguzzi et al., 2016) and from visual interpretation of the same GE image employed to build the gullies inventory.
The employed terrain attributes were extracted as raster with a 2 m grid size from a light detection and ranging-derived DEM with the same resolution (Regione Siciliana, 2010), by using the terrain analysis tools of saga-gis software. The terrain variables include slope gradient (S), contributing area (A) and nine attributes that were selected to reflect the potential hydraulic connectivity of the upslope area of each cell.
Bedrock lithology and land cover are widely recognized as controlling factors of water erosion. S and A serve as proxies for flow velocity and volume and, thus, reflect the erosive power of runoff. They have been frequently used to define topographic thresholds for gully initiation (e.g. Montgomery & Dietrich, 1992; Torri & Poesen, 2014; Vanwalleghem et al., 2005). Moreover, as individual attributes or combined as secondary topographic attributes (e.g. stream power index, TWI), S and A have been exploited to assess susceptibility to gully erosion (e.g. Conoscenti et al., 2014; Gómez-Gutiérrez et al., 2009a, 2009b; Rahmati et al., 2016). S was computed employing the method proposed by Zevenbergen & Thorne (1987). The calculation of A was performed by using the D8 single-flow direction algorithm (O'Callaghan & Mark, 1984), after a pre-processing of the DEM, which included sink filling and removal of the cells within artificial water bodies. Artificial lakes, urban areas and roads affecting flow direction were excluded from the analysis (Figure 1b).
Hydrological connectivity in a given drainage area potentially depends on its shape and on the morphometric characteristics of the drainage network (Delmas et al., 2009). Therefore, we explored the contribution of the Horton's (1932) form factor (HFF) and that of drainage density (D, Horton, 1945). HFF was computed for each cell as the ratio of upslope contributing area to the square of maximum flow path length measured in the upslope direction. To calculate D, we extracted from the DEM four different drainage networks by using two A thresholds, alone or combined with a threshold of tangential curvature (Mitášová & Hofierka, 1993). The latter is a topographic curvature measured in the direction of tangent to contour at a given point and was used also by other authors to detect incisions or valleys (Luo & Stepinski, 2006, 2008). The two A thresholds (0·01 and 0·02 ha) were selected so that even the gully pixel with the smallest upslope contributing area (i.e. 0·032 ha) has a drainage density higher than zero. After a trial-and-error analysis, the tangential curvature threshold was set to −0·015, so that terrain is considered sufficiently concave to be detected as a drainage element if its tangential curvature is below this threshold. D was calculated as the ratio of the length of the drainage network to the contributing area of each cell. Two drainage density variables, named D1 and D2, were prepared by setting to 0·01 and 0·02 ha, respectively, the A threshold for drainage network extraction. Other two variables were produced with drainage elements starting from pixels where tangential curvature is below −0·015 and A is above 0·01 ha (D1C) and 0·02 ha (D2C).
The hydrological connectivity depends on the infiltration ratio, which, in turn, apart from rainfall characteristics, antecedent moisture and vegetation type, is strongly influenced by soil/bedrock permeability. In this regard, we used a modified version of the index of development and persistence of the drainage network (IDPR; Gay et al., 2016). This index characterizes the landscape in terms of soil/bedrock relative permeability by comparing the development of a drainage network extracted from a DEM and the development of the real river network: permeability is assumed to be lower where the real river network is more developed, whereas it is expected higher when real network is less developed than DEM-derived network. We calculated the IDPR as the ratio between flow distance to DEM-derived streams and flow distance to real streams. The DEM-derived network was extracted by using an A threshold equal to the median value measured along real streams. The average IDPR calculated upslope of each landscape position (
) was also used as gully predictor variable.
Slope gradient controls infiltration ratio, and thus, it is included in most of the models/indices of hydrological connectivity (Bracken et al., 2013). As in the upslope component of the Borselli et al. (2008) index of connectivity, we calculated the average slope gradient of the upslope contributing area (
). These attributes are expected to control volume and velocity of runoff.
Surface roughness is also expected to affect hydraulic connectivity (Bracken & Croke, 2007). Following the procedure described by Cavalli et al. (2013), we calculated the RI as the standard deviation of the residual topography at a scale of few metres. The residual topography was calculated as the difference between the original DEM and a smoothed version obtained by averaging elevation values on a 5 × 5 cell moving window. The standard deviations of residual topography values were calculated in a 5 × 5 cell moving window over the residual topography grid. Finally, we included the average RI of the upslope contributing area (
) as predictor of gully occurrence.
Statistical Modelling
The MARS analysis was carried out by using the package ‘earth’ (Milborrow et al., 2011), implemented in r software (R Core Team, 2015). To limit the complexity of the MARS models, the maximum degree of interaction was set equal to one, hence avoiding model terms given by two or more BFs. The maximum number of terms entering the models was semi-automatically determined by ‘earth’.
In addition to the base model, which contains only A and S, other MARS models were trained on each catchment by adding one by one nine predictors of terrain connectivity. Therefore, by comparing the performance of the models obtained from the ten different combinations of predictors, we evaluated the contribution to the base model of each of the additional variables.
Calibration and Validation Sampling Strategy
Calibration and validation of the MARS models require datasets made of both presences (i.e. gully cells) and absences (i.e. non-gully cells). Independent of the proportion of presences and absences in the two study catchments, all analyses presented here were based on samples with a 1:1 ratio of gully to non-gully cells.
We prepared calibration and validation samples separately for each of the catchments. Calibration samples include 25% of all presences and an equal number of absences (both randomly selected). This ratio was chosen to achieve a reasonable trade-off between minimizing the effects of auto-correlation among gully cells and selecting sufficient data to obtain stable results. Validation samples, which include all presences and a same number of absences, were extracted from two different datasets: (i) a dataset including all cells of the catchments (dataset ALL) and (ii) a dataset including only cells along flow concentration axes (dataset FLOW). These axes were identified according to an A threshold, which was calculated separately for each catchment as the minimum A value at a gully cell, after excluding potential outliers below the first percentile. The dataset FLOW allowed us to measure the predictive skill of the models by focusing on those cells where topography favours runoff concentration, and given the rainfall that triggered gully occurrence in the study areas, flow accumulation is expected to be sufficient to cause gullying.
Predictive models were trained on each of the catchments and submitted to a tenfold internal cross validation as well as to an external validation carried out in the neighbouring catchment. To evaluate the robustness of the modelling procedure, 100 MARS runs were performed in W1 and W2 for each of the ten combinations of predictors. Then, an average probability of gully occurrence was extracted on each catchment from both the model replicates trained on the same catchment and from those trained on the neighbouring area. These probabilities were compared with 100 samples of cells extracted from the dataset ALL as well as to 100 samples extracted from the dataset FLOW.
The performance of the MARS runs was estimated by calculating the area under the receiver operating characteristics curve (AUC) (Lasko et al., 2005). A receiver operating characteristics curve plots the true positive rate (sensitivity) versus the false negative rate (1 – specificity), at any given cut-off score. The closer the curve is to the upper-left corner, the larger the AUC is and the more accurate the model is. AUC values close to 0·5 demonstrate no discrimination ability, whereas AUC values equal to 1 are found when a perfect classification is achieved. In our study, intermediate AUC values were interpreted according to Hosmer & Lemeshow (2000), who classify model performance as acceptable, excellent and outstanding when AUC is higher than 0·7, 0·8 and 0·9 respectively. The differences in model performance were tested for statistical significance by using the Wilcoxon signed-rank test. Differences at p-value < 0·01 were considered significant.
Results
Gully Inventory
The mapping of gully landforms revealed that gully erosion is more severe in W1 than in W2 (Figure 3). In fact, density of gullies is 0·73 and 0·18 km km−2 in W1 (83 gullies) and W2 (32 gullies) respectively. Most of the gullies have widths of approximately 0·5–1·0 m and average length of 55 m in W1 and 51 m in W2 (Figure 4).
The eluvial–colluvial deposits exhibit the highest gully density in W1 (1·37 km km−2) and the second density in W2 (0·28 km km−2) after clays interbedded with sandstones (0·90 km km−2), which, however, have a very small extent (Table 1). As regard land cover, gullies mostly occur on arable lands (96% of the total gully length), whereas few gullies were identified on vineyards (2·5%); in W1, some gullies also reach the valley bottom and intersect the willow brush that cover the riparian zone. Arable lands exhibit a different gully density in W1 (0·90 km km−2) and W2 (0·37 km km−2). Conversely, a similar gully density was observed in vineyards of the two catchments (0·060 and 0·045 km km−2 in W1 and W2 respectively).
Figure 5 shows a double logarithmic plot of catchment area (A) and local slope gradient (S) measured at the gully heads. S–A pairs of W1 and W2 plot approximately in the same ranges of A and S, but W1 gully heads occur on steeper slopes than those observed in W2. A power relationship between S and A was fitted separately through data observed in W1 and W2, resulting in different values for the coefficient a and the exponent b of the equation S = aA−b. Indeed, the W1 trend line is higher but less inclined than that fitted through the W2 S–A pairs. The trend line derived for gullies observed by Vandekerckhove et al. (2000) in cultivated lands of Spain and Portugal is also plotted in Figure 5.
The frequency distributions over W1 and W2 of the terrain predictor variables together with their kernel density plots calculated for the gully pixels are shown in Figure 6. The histograms and kernel density curves clearly highlight that (i) terrain attributes have very similar frequency distributions over the two catchments; (ii) gully cells tend to concentrate on values of the variables with low to very low frequency and (iii) kernel density curves in W1 and W2 gully cells are similar for A, S, D1C, D2C,
and
, whereas they are quite different for D1, D2, HFF, IDPR and
.
Validation on the Dataset ALL
Tenfold internal cross validation revealed outstanding accuracy for all the models in both the catchments with the exception of an outlier (Figure 7a). AUC ranges are small, especially in W1, indicating that the accuracy of the MARS runs is stable and our procedure is robust when changes of the calibration and validation samples are performed. The Wilcoxon test revealed that the differences of accuracy between the base model and the others are all not significant if we exclude the model incorporating
, which performs slightly worse.
The box plots of Figure 7b reveal that when an external validation is performed, all the models in both W1 and W2 achieve outstanding and quite stable accuracies. The predictors added to the base model provide rather small changes of performance. When statistically significant, these changes reveal better accuracy of the models accounting for terrain connectivity. All the changes of performance are statistical significant in W1, with the exception of that provided by IDPR, whereas in W2, significant changes are provided only by D2, IDPR,
and
.
Validation on the Dataset FLOW
The dataset FLOW includes cells located along flow concentration axes and identified according to an A threshold equal to 0·10 ha, in W1, and 0·06 ha, in W2. These cells cover only 5% and 7% of the total extent of W1 and W2 respectively.
The results of model evaluation with samples randomly extracted from dataset FLOW are summarized in Figure 8. The AUC values of internal validation reveal excellent to outstanding accuracy, which is quite stable especially in W1 (Figure 8a). The variables reflecting potential hydrological connectivity, with the exception of HFF, significantly improve the performance of the base model. The greatest increase of accuracy is provided by D1C and D2C.
The external validation on dataset FLOW reveals more enhanced changes of performance caused by the predictors of terrain connectivity. The differences of AUC are all significant apart from that produced by D2C and
in W1 and by D1 and D2 in W2. In W1, the predictive ability of the base model trained in W2 improves with D1, D2,
and
. Conversely, MARS repetitions calibrated in W1 are able to predict gullies of W2 with slightly better accuracy if they incorporate any of the additional predictors apart from HFF, IDPR or
, whereas the performance noticeably improves if the models include D1C or D2C.
Gully Erosion Susceptibility Maps
Model evaluation on the dataset FLOW showed that the variables expressing drainage density upstream of each cell provide the highest additional predictive ability to the base model. In particular, the internal validation revealed that the improvement is more enhanced if variables D1C or D2C are incorporated in the predictive models. Figure 9 shows the gully erosion susceptibility maps (pixel size 2 m) obtained from the base model repetitions (Figure 9a) and from those of the model that include D2C (Figure 9b).
Discussion
The observed gully occurrence clearly highlighted more severe gully erosion in W1 than in W2. This is probably more related to a different distribution of lithology and land cover classes than to a different topographic setting. As revealed by the frequency distributions of gully predictors (Figure 6), topography of the two catchments is indeed very similar. On the other hand, the eluvial–colluvial deposits, which appear the most prone to gully occurrence, prevail in W1 but are the fourth lithology class in W2, where classes having lower gully density (i.e. sands, silty-clays and sandy-silts) cover about half catchment. Similarly, the relative extent of land cover classes showing low density of gullies, such as vineyards and natural grassland, is three to four times larger in W2 than in W1, whereas that of arable lands is significantly less (Table 1). Moreover, the different gully erosion susceptibility can be explained by the mean slope angle of the most susceptible classes of lithology (i.e. eluvial–colluvial deposits) and land cover (i.e. arable lands), which is higher in W1 (9·7 and 10·0°) than in W2 (8·8 and 8·3).
The S–A pairs measured at the gully heads of W1 and W2 (Figure 5) plot at lower S and higher A values than that provided by Vandekerckhove et al. (2000, p. 1209) (Figure 4), suggesting that in our study area, gullying process occurs on gentler slopes and thus requires larger contributing areas (and runoff discharge). Also, their trend line is more inclined than those passing through W1 and W2 data sets.
The distribution of gullies over lithological units is consistent with the findings of other authors who observed that unconsolidated and poorly sorted materials are favorable conditions for gully development (e.g. Dewitte et al., 2015; Poesen et al., 2003). As regard the high density of gullies occurring in arable lands, it can be explained by the crop cycle of cereals, which leaves the soil unprotected during the rainy season (autumn–winter). On the other hand, the low density of gullies observed in vineyards was somewhat surprising. In fact, vineyards are reported as one of the most erosive land uses in the Mediterranean area (García-Ruiz, 2010; García-Ruiz et al., 2013; Novara et al., 2011; Prosdocimi et al., 2016) because soils are almost bare for a large part of the year. In the study catchments, vineyards are generally maintained without cover plants even during summer and autumn by tillage operations (mechanical weeding) that are repeated on average two or three times per year. Moreover, vine rows are sometimes parallel to the maximum slope gradient. Despite these conditions, which might have favored erosion processes not detected in the analysed GE image (i.e. sheet erosion and rilling), few gullies were observed in vineyards of our study areas. This could be related to a higher infiltration capacity of soils as, at least in W2, grain size of parent materials is coarser in vineyards (sands: 33%; clays 17%) than in arable lands (sands: 24%; clays 25%). Moreover, as reported by Arnaez et al. (2007), roughness due to tillage (not parallel to flow direction) might cause moderate runoff coefficients on vineyards and, thus, less erosive power of concentrated flow. However, further research is needed to explore the relationships between vineyards and gully erosion as our results may reflect local variation of attributes not considered in this study. Moreover, as reported in other papers (e.g. Prosdocimi et al., 2016), an insufficient number of gully erosion studies have been carried in vineyards so far.
The internal validation of our models on the dataset ALL showed similar or better predictive skill than those achieved by other authors who evaluated gully predictions in the same catchment where calibration was performed, with samples extracted from the entire study area. For example, Gómez-Gutiérrez et al. (2009a, 2009b), who also employed MARS to predict gully occurrence, obtained AUC values in the range 0·75–0·98. The other MARS application to gully erosion susceptibility assessment was made by Gómez-Gutiérrez et al. (2015), who obtained a mean AUC of 0·859 and 0·826 in Sicily and Spain respectively. In other studies employing logistic regression as modelling algorithm (Conoscenti et al., 2014; Lucà et al., 2011), gully prediction models showed somewhat poorer performance (AUC = 0·82–0·86). Moreover, slightly lower accuracy were achieved with CART by Gómez-Gutiérrez et al. (2009b) (AUC = 0·66–0·96) and by Jurchescu & Grecu (2015) (mean AUC = 0·86) and with stochastic gradient treeboost by Angileri et al. (2016) (AUC = 0·87–0·92).
The models trained in one catchment and validated in the other exhibited outstanding and very stable predictive performances when evaluation was performed on the dataset ALL. The differences of performance between the base model and those including the other variables are rather small. Therefore, we can infer that the base model is sufficient to accurately predict gully occurrence. On the other hand, by focusing the model evaluation on the dataset FLOW, we were able to detect important changes of predictive ability that did not arise on the dataset ALL. This can be explained on the basis of the following considerations. As commonly observed, also in our catchments, the ratio of cells along flow concentration lines to total number of cells is quite low. Therefore, if absences are randomly selected from the entire catchments, most of them are very likely to come from areas with very low A values. In these landscape positions, topography is more favorable to diffuse overland flow and a low probability of gully occurrence can be expected a priori. Thus, in this case, discriminating between gully and non-gully cells can be quite simplified and changes of performance between models are more difficult to detect.
Model evaluation focused along flow concentration axes revealed an added predictive skill of most predictors selected to reflect potential hydrological connectivity. The most enhanced increase of accuracy was detected when drainage density variables were included in the models. In particular, apart from external validation in W1, the best performance was achieved by including the variables D1C or D2C.
Conclusions
In this study, we assessed susceptibility to gully erosion in two small adjacent agricultural catchments of Sicily (Italy). To this aim, we employed MARS as modelling technique and a set of geo-environmental attributes. To explore the influence on gully erosion of terrain connectivity, we measured the changes of performance occurred when adding one by one set of predictors reflecting potential hydrological connectivity to a base model that included contributing area, slope gradient, lithology and land cover. Model evaluation was carried out inside and outside the calibration catchments, using samples extracted both from all cells of the catchments and only from cells located along flow concentration axes.
The results of our experiment showed that gully susceptibility is quite different in the two study catchments. As topography is similar, this is possibly due to a different distribution of lithology and land cover. Density of gullies is higher in poorly consolidated materials such as eluvial–colluvial deposits and clays. As regard land cover, gullies occur mostly in arable lands, whereas, unexpectedly, low gully density was observed in vineyards although no soil conservation measures are adopted in our study areas. This requires further investigation as it may reflect variation in soil properties, topography or other variables that were not considered in this study. Moreover, as reported by other authors, there is still little information about the relationships between vineyards and gully erosion in the Mediterranean area.
Validation of the MARS models performed on the entire catchments revealed outstanding and quite stable accuracy, either within or outside the calibration catchment. Very small changes of performance were observed when adding to the base model the variables selected to reflect potential hydrological connectivity. When validation was focused on flow concentration axes, the accuracy of the gully predictions decreased, revealing important differences among the models. In the study areas, we found that almost all the additional predictors related to terrain connectivity improved significantly the performance of the base model, but the most enhanced increase of accuracy occurred when drainage density variables (especially D1C or D2C) were added to the base model. Therefore, when high-resolution topographic data are available, we recommend including terrain connectivity variables as predictors of gully erosion susceptibility.
Acknowledgements
This research was developed in the framework of the project FLUMEN (project number: 318969), funded by the EC Seventh Framework Programme (call identifier: FP7-PEOPLE-2012-IRSES), scientific responsible for the University of Palermo: Prof. Christian Conoscenti. All the other authors have commonly shared the whole research phases. This manuscript benefited greatly from the kind suggestions and comments of Prof. Chris Barrow, Prof. Tammo Steenhuis and two anonymous reviewers.
