Rates of soil erosion on agricultural land often exceed the rates at which new soil is formed (Evans et al., 2019; Montgomery, 2007). This imbalance is one which, left unchecked, can pose a critical threat to the sustainability of global soil resources and their ability to deliver vital ecosystem services across environments and society (Bot et al., 2000; Quinton et al., 2010). Moreover, as soils become thinner (i.e. truncated), the subsoil horizons and their underlying parent material become progressively closer to the land surface. This process might affect physical, chemical, and biological topsoil properties (Bouchoms et al., 2019; Papiernik et al., 2009; Vanacker et al., 2019), as well as soil water availability to plants and ultimately crop growth (Herbrich et al., 2018; Öttl et al., 2021; Schneider et al., 2021).

For instance, plot- and catena-based studies report that truncated agricultural soil profiles often display increased clay and/or sand contents in their Ap horizons (Rhoton and Tyler, 1990), compared to the soils from non-eroding positions in the landscape. Moreover, eroded Ap horizons tend to have higher bulk density, lower organic carbon content, and lower water holding capacity (Olson and Nizeyimana, 1988; Stone et al., 1985; Strauss and Klaghofer, 2001). However, such patterns are highly variable and greatly dependent on the properties of the underlying subsoil material being progressively tilled into the Ap horizon (Lowery et al., 1995).

Erosion-induced changes to soil depth and soil properties can therefore influence soil losses and runoff formation through a soil erosion feedback system (Morgan et al., 1984; Vanwalleghem et al., 2017). That is, erosion-induced changes to soil physical properties might affect soil erodibility (i.e. the susceptibility of soil to erosion), which may accelerate or slow down soil losses. Understanding how such a system might develop over time and under assorted conditions is an important step to proactively design and implement effective soil conservation strategies, as different soils are likely to be impacted by erosion in varied ways (Hoag, 1998). However, the empirical data over decadal to centennial timescales required to explore the feedbacks between soil erosion and soil thinning are currently non-existent. It follows that process-oriented soil erosion models are arguably the only available tool to simulate how erosion processes interact with truncation-induced changes in the soil system.

Process-oriented models allow for a representation of multiple mechanisms that influence soil erosion, from basic processes such as particle detachment by raindrop impact and surface runoff, to more complex interactions between soil properties, hydrological processes, climate, and plant cover (see Merritt et al., 2003 for an overview). This ability to simulate the response of specific soil erosion processes to external stimuli makes process-oriented models useful for exploring what-if scenarios. Hence, models such as the Water Erosion Prediction Project (WEPP; Nearing et al., 1989), the LImburg Soil Erosion Model (LISEM; De Roo et al., 1996), and the Morgan–Morgan–Finney model (MMF; Morgan, 2001; Morgan et al., 1984; Morgan and Duzant, 2008) have been used to explore the impacts of land use or climate change on soil erosion (e.g. Anache et al., 2018; Eekhout et al., 2021; Nearing et al., 2005).

To date, most soil erosion models and model users assume that the inherent erodibility of different soil horizons down a soil profile is constant over the period of a model simulation. As upper soil horizons are removed by erosion, thereby exposing the subsurface material, the implicit assumption in soil erosion modelling is that this erodibility is not variable, such that any changes to projected erosion rates are solely a factor of climate, land cover, and topography (e.g. Ciampalini et al., 2020; Eekhout et al., 2021; Panagos et al., 2021). However, since erodibility is a reflection of soil physical, chemical, and biological properties, and given that subsoils typically (although not exclusively) exhibit contrasting soil properties to those observed in upper horizons, it follows that erodibility is not necessarily a constant as a soil profile thins. Furthermore, soil erodibility might change over longer timescales due to the coarsening and armouring of surface soils (Sharmeen and Willgoose, 2007; Willgoose and Sharmeen, 2006) and the depletion of erodible material as a result of extreme soil truncation (Anselmetti et al., 2007).

Although the soil erosion feedback system has been recognised as a key challenge for modelling past and future erosion rates (Vanwalleghem et al., 2017), long-term dynamics of soil erodibility are an underexplored topic in erosion research. Exceptions come from landscape evolution models which simulate the influence of erosion and deposition on soil development over millennia (van der Meij et al., 2020; Sommer et al., 2008). However, in areas under severe erosion rates, subsoil horizons can be exposed within a matter of decades (Evans et al., 2020), which might trigger unexpected responses regarding runoff formation and soil losses. Moreover, soil truncation can introduce substantial spatial variability to soil properties, often not accounted for in static soil maps used for a variety of purposes (Świtoniak et al., 2016). Still, the lack of knowledge about the potential interactions between soil erosion and soil erodibility (i.e. in which timescale are erosion feedback mechanisms developed, how are different soil properties and soil types affected by truncation?) interpose the representation of soil truncation in the applications of soil erosion models.

Here, we hypothesise that erosion-induced changes to soil properties affect soil loss rates over long-term periods. To evaluate such a hypothesis, we performed an exploration of the soil erosion feedback system by simulating 500 years of soil losses and surface runoff on 265 agricultural soil profiles in the UK. This allowed us to investigate how soil erosion rates respond to truncation-induced changes in soil properties and unravel the processes potentially driving such responses in different soil types in the UK. To the best of our knowledge, this is the first time soil erosion models have been used to understand the interactions between soil erosion, soil thinning, and soil erodibility, and how these interactions are established in varying soil types. An enhanced understanding of such a dynamic soil erosion feedback system will be crucial for improving the calculation of soil lifespans, providing future soil loss projections, and designing long-term soil conservation strategies.

From the 265 soil profiles in the UK SOILPITS database, 262 (99 %) displayed a significant correlation (p<0.00001) between soil truncation and annual soil losses, of which 261 presented a decelerating feedback trend (Pearson's r<0). Considering the median of the simulations from all soil profiles, the temporal evolution of erosion rates was characterised by an exponential decay function (Fig. 5), which meant erosion rates initially decelerated at a higher pace until reaching a horizontal asymptote at approximately 250 years into the model runs. Based on the adjusted decay function, the initial soil loss rates (intercept = 0.81 kg m-2 yr-1) decrease in 6 % and 10 % over the first 50 and 100 years of the simulations, respectively; but there is only a decrease of 3 % in the last 100 years.

The changes in simulated soil losses between the initial and final model time steps per profile followed a normal distribution, which signified an average 20±8 % decrease in erosion rates. The steepest declines in soil losses (maximum 46 %) typically occurred in soil profiles with a negative gradient in silt contents in their subsoil horizons. Contrarily, the soil profiles with a stable erosion progression or with a gentler decay steepness were associated with the presence of silty substrata. Figure 6 illustrates the typical behaviour of these trends using data from two representative soil profiles with similar topsoil but with different subsoil properties at the beginning of the simulations.

It follows that simulated changes in erosion rates were largely explained by alterations in soil texture. This is demonstrated in Fig. 7, which displays a random forest importance ranking for predicting the differences in erosion rates between the initial and final time steps of the simulations, for each soil profile. The random forest analysis described how soil erosion responses were highly influenced by variations in silt content and stone cover. Changes in soil moisture at field capacity and bulk density had a lower impact on the modelled soil losses (Fig. 7).

The sensitivity of the simulated changes in soil erosion rates to the erosion-induced changes to soil properties can be further visualised by comparing the difference in model parameter values with the variation in soil losses over 10-year rolling means (Fig. 8). Positive and negative changes in single parameter values did not yield consistent responses regarding the simulated soil losses (e.g. a 1 % decrease in sand content can lead to both accelerating and decelerating erosion rates over the rolling means). However, the direction of the changes in model parameters explains the general decelerating trend for the profiles (Fig. 8). For instance, the profile trajectories were characterised almost exclusively by decreasing contents of silt and increasing stone cover within the rolling means which led to net decreases in soil losses over the whole simulation period. Clay and sand contents mostly increased within the 10-year rolling means, although such a pattern was less pronounced compared to the silt and stone cover progressions. Less important parameters, such as bulk density and soil moisture at field capacity, displayed a more centred pattern along the y axis (Fig. 8).

The main processes driving the soil erosion feedback systems were particle detachment by raindrop impact and silt sediment supply (R2=0.85 and 0.66, respectively; Fig. 9), while changes in runoff amounts, detachment by runoff, and runoff transport capacity had a narrow effect on the simulated soil losses (Fig. 9). That is, only acute changes in discharge seem to have produced a sufficient response in detachment by runoff and in transport capacity to influence the net erosion rates (Fig. 9).

Annual runoff depths and soil losses were correlated in 187 (71 %) of the 265 soil profiles (p<0.00001); however, these correlations did not amount to causation. The positive correlations (44 profiles; 17 %) occurred for instance when the uprise of clayey subsurface horizons led to a reduction in both runoff amounts (due to an increase in soil moisture storage capacity) and soil detachment. The negative correlations (143 profiles; 54 %) occurred due to selective erosion processes, soil organic carbon depletion, and the presence of sandy soil substrata. That is, as the silty material was removed, enriching the topsoil with sand and carbon-depleted subsoil, particle detachment by raindrop impact declined, whereas runoff amounts increased due to a reduction in soil moisture at field capacity. However, this increase in runoff was not sufficient to accelerate soil losses (Fig. 10).

Our model simulations underline the strong interaction between current soil erosion dynamics and the erosion history of soil profiles. In particular, the simulations demonstrate how different soils are likely to have contrasting responses to soil thinning, depending on the properties of the surface material, as well as those of the underlying soil horizons. In our database, most of the soil profiles presented a decelerating feedback trend, which reflects both the characteristics of these soils and, importantly, our basic modelling assumptions.

For instance, as silt detachability in the MMMF model is assumed to be much higher than for other particle size fractions, silt was preferentially removed from the soil matrix. In addition, as silt contents typically remain stable or decrease in the subsurface horizons of the soil profiles in our database (Fig. 2), silt was often not replenished by the underlying substrata being mixed into the plough layer. Such behaviour is overall consistent with empirical observations of selective particle size removal by interrill erosion processes (Koiter et al., 2017) and the progressive depletion of readily erodible material from eroding surface soils (Parsons et al., 1991). However, it is worth highlighting that the soil detachability values used in MMMF have a limited empirical basis (Quansah, 1982), and the erodibility of the clay particles might be poorly described due to the neglection of other variables, such as aggregate stability (Morgan and Duzant, 2008).

The residual accumulation of rock fragments further contributed to the reduction in erosion rates in the model simulations, as stone cover is assumed to armour the land surface and to reduce soil detachment. Specifically, even a small number of rock fragments in the soil matrix can disperse the overland flow and dissipate its energy, reducing rill incision and soil losses (Rieke-Zapp et al., 2007). Moreover, decreases in water erosion rates due to a residual increment of stone cover have previously been simulated by Govers et al. (2006), who warned, however, that the accumulation of rock fragments in arable soils depends on tillage practices. Notwithstanding, increases in rock fragment contents might also affect soil hydraulic conductivity and water holding capacity (Cousin et al., 2003), and therefore influence runoff formation. None of these potential interactions were represented in our model simulations, and might therefore warrant further scrutiny.

The few soil profiles displaying a stable or slightly accelerating erosion trend were characterised by the presence of sandy or loamy surface horizons over a siltier substratum, which successively supplied the plough layer with readily erodible material as the original topsoil was removed by erosion. Moreover, these profiles were defined by the absence of a surface stone cover and by subsoil horizons with very limited amounts of rock fragments. Although accelerating erosion trends were only simulated for one profile in our modelling experiment, such behaviour might be expected in soils with highly incrementing silt contents in their C horizons, which are common, for instance, where loess is the parent material (Finke, 2012; Świtoniak et al., 2016).

Moreover, potential accelerating erosion trends might have been under-detected by the model simulations, as we did not consider how truncation can decrease the soil moisture storage capacity due to a reduction in soil depth and how this might affect runoff generation (Dunne and Black, 1970; Morgan et al., 1984). That is, as soils become shallower, saturation-excess overland flow might increase, depending on the permeability of the bedrock or the presence of an impeding horizon (Beven, 2012; Moraes et al., 2010). We did not consider these processes in our simulation due to the absence of an explicit soil thickness parameter in the MMMF equations and the lack of data regarding the permeability of the soil profiles' parent materials. Although simulated increases in runoff formation were generally not sufficient to increase particle detachment by overland flow and soil losses in the model outputs (Figs. 9 and 10), different results would be conceivable at a landscape scale. For instance, if upslope run-on is considered, increases in overland flow due to soil truncation might lead to rill initiation in flow accumulation zones, which would largely increase the simulated erosion rates.

Furthermore, very different runoff responses to soil truncation can be expected in areas where infiltration excess is the dominant overland flow mechanism. While saturation excess is common under British edaphoclimatic conditions, most subhumid and semiarid zones are prone to the formation of infiltration-excess runoff, which is a process primarily controlled at the soil surface (Smith and Goodrich, 2005). Under such circumstances, erosion-induced changes to topsoil properties might have an even greater interaction with runoff generation. In particular, soil crusting slows down infiltrability and rapidly increases the overland flow, leading to greater soil losses (Le Bissonnais, 2016; Fiener et al., 2008; Veihe et al., 2001). As the development of soil crusts is influenced by soil texture and organic carbon content (Fiener et al., 2011), the truncation-induced changes we have simulated would likely alter the susceptibility of surface soils to crusting and, consequently, to the formation of infiltration-excess runoff. Another caveat in the model structure worth highlighting is that, different to model outputs, an uprise of subsurface clayey material might in fact increase infiltration-excess overland flow, due to the lower infiltrability and saturated hydraulic conductivity of heavy-textured soils (Hao et al., 2020), which are also more susceptible to compaction, depending on their mineralogy (Bonetti et al., 2017; Hamza and Anderson, 2005).

In addition, accelerating erosion responses to soil truncation might have been more frequent if we assumed an increase in topsoil erodibility due to the lower aggregate stability and looser structure of the carbon-depleted subsurface material being incorporated into the plough layer (Le Bissonnais, 2016; Doetterl et al., 2016; Tanner et al., 2018). That is, as the soil detachability coefficients from MMMF do not take soil organic carbon into account, the sensitivity of topsoil erodibility to soil truncation was likely downplayed. For instance, Radziuk and Switoniak (2021) used an equation from the Erosion Productivity Impact Calculator (EPIC) model to estimate the erodibility of Luvisols at different truncation stages in Poland. They found that more truncated soils had higher erodibility, due to decreases in soil organic carbon and sand content in the eroded Ap horizons.

Our model simulations may have particularly underestimated erosion responses to soil thinning for the profiles which displayed an accumulation of sand and a depletion in clay and silt contents. This progressive coarsening should lead to lower soil water availability, and therefore, lower soil cover, lower crop biomass production, and less organic carbon input from plants. As sandier soils already have less carbon stabilisation mechanisms (Doetterl et al., 2016), this would lead to even greater truncation-induced depletions in soil organic carbon and therefore increases in erodibility (Auerswald et al., 2014; Fernández and Vega, 2018). In general, as organic carbon was only an indirect model input via the PTFs for estimating bulk density and soil moisture at field capacity, the interplays between soil thinning, soil organic matter, and soil erodibility were likely underrepresented here.

Although not all the complex interactions between soil erosion and soil thinning could be described in our model, the simulated trajectories of erosion-induced changes to soil properties (Fig. 10) are consistent with field-observations. That is, measured data from eroded Ap horizons typically indicate (i) incrementing clay, sand, and rock fragment contents, (ii) increasing soil bulk density, (iii) depleting soil organic carbon contents, and (iv) decreasing soil water holding capacity with increasing erosion severity (Lowery et al., 1995; Rhoton and Tyler, 1990; Stone et al., 1985; Strauss and Klaghofer, 2001). Moreover, the decelerating erosion trend in our model outputs corroborate the results from Govers et al. (2006), who simulated an exponential decay in sediment production from arable hillslopes over a period of 50 years following land use intensification due to a rapid emergence of a soil with high stone cover. Importantly, the soil losses estimated from our model outputs (median =0.67 kg m-2 yr-1; interquartile range =0.58-0.80 kg m-2 yr-1) are encompassed by the median and the upper quartile of measured erosion rates from arable land at plot scale in the UK (∼0.1–1 kg m-2 yr-1) (see Benaud et al., 2020). We compare model outputs to plot-scale measurements due to their similarity with our modelling spatial unit (i.e. 10 m eroding hillslope segment with 6∘ or 10.5 % slope gradient). It is worth highlighting that (i) the plot data described in Benaud et al. (2020) are mostly derived from slopes below 10 % and (ii) measured erosion rates at field or catchment scale in Britain are much lower than our model simulations.

Moreover, our model outputs further help to identify where more empirical evidence would help constraining modelling assumptions and improve process representation. For instance, investigating how truncation affects the aggregate stability of surface soils (due to changes in soil texture, mineralogy, and organic carbon dynamics) might be an important step in order to further understand the feedbacks between erosion and soil thinning. Interactions between soil truncation, water availability to plants, and crop growth – and how these could in turn reduce soil cover and organic carbon input – might also warrant further investigations. Similarly, exploring the responses of different runoff generation mechanisms to soil thinning and erosion-induced changes to soil properties should be beneficial to increase our understanding of the erosion feedback system. Modelling-wise, an important next step would be to adapt the framework described here into spatially distributed soil erosion models, in order to evaluate the effects of soil erosion feedback systems at landscape or watershed scale. This would allow for an unravelling of different feedback processes at different positions in the landscape, which might be affected by erosion-induced changes to soil properties in different ways.

Finally, although our model simulations indicate that soil thinning has a decelerating effect on soil loss rates for eroding hillslope segments in the UK, the results also demonstrate that some of these soils' physical properties might become restrictive for agriculture before the profiles are completely truncated. That is, some simulated erosion progressions create dense Ap horizons (max. bulk density of 1.78 M m-3) with excessive rock fragment contents (max. 36 %), very high clay (max. 89 %) or sand contents (max. 99 %), and low water holding capacity (min. soil moisture at field capacity of 0.14 % w w-1). For such scenarios, rehabilitation techniques based on topsoil replacement might be necessary to sustain crop production (Schneider et al., 2021). These significant changes in soil properties, simulated in relatively short time periods (Fig. 10), further indicate how accelerated erosion in agricultural landscapes can potentially affect soil classification (Lewis and Witte, 1980; Olson and Nizeyimana, 1988; Świtoniak et al., 2016), which is important to consider when interpreting decades-old soil maps. Since soil erosion rates in Britain are much lower compared to other regions of the world (Benaud et al., 2020), our model simulations represent a somewhat conservative scenario of erosion feedback systems. Erosion-induced changes to soil properties, and their implications for modelling, should be particularly relevant in areas under severe erosion rates.

Here, we explored the soil erosion feedback system in 265 agricultural soil profiles in the UK. In particular, we simulated how selective erosion processes and the incorporation of different subsoil horizons into the plough layer affected the erodibility of surface soils. We further analysed how these processes could change erosion rates during a period of 500 years. We found that (i) soil erosion rates in 99 % of the soil profiles were sensitive to soil truncation, (ii) the truncation-sensitive profiles essentially displayed a decelerating trend, and (iii) changes in soil texture and stone cover were the main drivers of the modelled soil erosion feedbacks loops. Importantly, we found that different soils had different simulated responses to soil truncation, depending on the properties of the surface material as well as those of the underlying soil horizons.

Ultimately, our findings highlight the dynamic nature of the soil as a three-dimensional body. That is, even the so-called intrinsic properties of surface soils might change in a matter of decades in areas under accelerated erosion rates. Moreover, erosion-induced changes to soil properties can have a significant impact on the rates with which soils are eroded, which in turn affects the calculation of soil lifespans and model-based erosion projections. Therefore, understanding how erosion-induced changes to soil properties reverberate with erosion itself will be crucial for improving long-term model predictions, investigating the resilience of different soils to erosion disturbances, and for developing appropriate soil conservation strategies for a changing world.