Testing the “physical model concept” by soil loss data measured in Sicily

Abstract

The best possible model to predict the erosion from an area of land has been suggested to be a physical model of the area that has similar soil type, land use, size, shape, slope and erosive inputs. Therefore, a replicated plot has to be considered the best possible, unbiased, real world model. In this paper the physical model concept was tested by using soil loss data collected on plots of different length at the experimental station of Sparacia, in Sicily (South Italy). This investigation supported the conclusions that i) a coefficient of determination between measured and predicted soil loss values of 0.77 has to be considered as the best-case prediction scenario and ii) an uncalibrated deterministic erosion model would not give more accurate results than those obtained by a replicated plot measurement. An effectiveness coefficient of 0.47–0.49 was obtained by applying the original USLE to predict event soil losses at Sparacia. The difference between the value of 0.6, corresponding to what we can expect from an uncalibrated erosion model, and the effectiveness coefficient of the selected model represents the maximum gap that has to be covered to obtain the realistically best estimate of plot soil loss at the event temporal scale.

Introduction

A plot soil erosion model is a tool for identifying sources of soil erosion variances as a function of measurable quantities of the system of interest (Nearing, 1998). Predicting soil loss due to water erosion by a model has a great practical importance because it allows to establish the severity of the phenomenon in an area of interest and also assess the effect of alternative soil erosion control practices (Bagarello et al., 2011b, Bagarello et al., in press). Testing the performances of a soil erosion model is necessary to determine the expected reliability of the predictions (Foster, 1987, Nearing, 1998, Quinton, 1994). Determining the quality of the predictions needs establishing a criterion to discriminate between acceptable and unacceptable soil loss estimates.

The most common approach to establish the reliability of soil erosion predictions is to compare estimated against measured values of plot soil loss. In general, this comparison is carried out by using a single or a few replicated data for a given treatment. However, plots that would be modeled with identical model input parameters yield variable data (Bagarello and Ferro, 2004, Bagarello and Ferro, 2010, Wendt et al., 1986). Generally, this phenomenon is not quantitatively taken into account during evaluation of a model, because knowledge of natural variability between plots having the same treatment is limited (Nearing, 2000). Therefore, it has to be expected that a portion of any difference between measured and predicted erosion rates will be due to model error, but another portion will be due to unexplained variance of the measured sample value from the representative mean value for a particular treatment (Nearing et al., 1999). Using many replicated plots within a given treatment reduces or eliminates this type of uncertainty, because it allows calculation of a more representative mean soil loss value (Wendt et al., 1986). However, equipping and operating many plots with a given treatment at a single experimental station has to be considered occasional, being onerous from an economic point of view and also impractical. Therefore, evaluation of soil erosion models that are deterministic in nature, such as the USLE/RUSLE or the WEPP (Flanagan and Nearing, 1995, Renard et al., 1997, Wischmeier and Smith, 1965, Wischmeier et al., 1978), commonly ignores variability in the measured data (Nearing, 1998). A limit to the accuracy of deterministic models (empirical or process oriented) should be expected because of the variation in soil erosion rates, which may be considered random from a practical point of view (Nearing, 2000).

Nearing (1998) suggested that the best possible model to predict the erosion from an area of land is a physical model of the area that has similar soil type, land use, size, shape, slope and erosive inputs. Therefore, the physical model represented by a replicated plot has to be considered the best possible, unbiased, real world model. Using event soil loss data for approximately 3000 pairs of replicated plots, Nearing (1998) established a comparison between the measured soil losses and the predicted ones by the physical model represented by the replicated plot. The coefficient of determination, R², was equal to 0.77 and the conclusion was that it should not be expected that an uncalibrated erosion model would give a better overall result.

More recently, Nearing (2000) proposed an evaluation method of soil erosion models assuming that the prediction has to be considered acceptable if the difference between the model prediction and the measured plot data value lies within the population of differences between pairs of measured values (Fig. 1). Taking into account that the relationship between variance and magnitude of the measured soil loss was independent of the temporal scale, the analysis was carried out using event values of soil loss from seven sites in the United States (U.S.), with 2061 replicated storm events in the data set, and also annual values of soil loss from 13 sites, with a total of 797 replicated pairs of plots. The plots ranged from 2 to 8 m in width and 3 to 16% in steepness, and most of the plots were 22 m long. Intervals for probability of occurrence were considered to be equal to the frequency of occurrence of the data points, since the considered sample size was large. The wide range of geographic conditions, rainfall regimes, erosion rates, and soil types represented in the data set suggested that the results should be generally applicable for erosion plot model validation. However, testing the conclusions by Nearing, 1998, Nearing, 2000 with other data sets is desirable for several reasons, including the scientific and practical importance of an established and objective criterion to evaluate performances of a soil erosion model, the empirical nature of the developed criterion, the prevalence of a particular plot length (22 m) in the investigation by Nearing (2000), and the fact that all data were collected in a specific area of the world, although large and varied.

The objective of this investigation was to test the physical model concept by using soil loss data collected on plots of different length at the experimental station of Sparacia, in Sicily (South Italy).