Data
Light detection and ranging–derived digital terrain models (DTMs) at 2-m resolution (Fig. 3) are available thanks to public authorities in Italy (Italian Ministry for Environment, Land and Sea; Treviso Province; Trentino Alto-Adige Autonomous Region). The datasets refer to the year’s range 2010–2012.
Information about land cover is available through the Corine-Land-Cover database (CLC, Coordination of Information on the Environment Land Cover) classification, as also reported by the local authorities. The considered CLC data come from an updated version of the Urban Atlas (European Environment Agency 2012) provided by the local government (Regione del Veneto 2012). The original Urban Atlas is mainly based on the combination of statistical image classification and visual interpretation of very high resolution (VHR) satellite imagery. Multispectral SPOT 5 & 6 and Formosat-2 pan-sharpened imagery with a 2-to 2.5-m spatial resolution is used as input data. The built-up classes are combined with density information on the level of sealed soil derived from the high-resolution layer imperviousness to provide more detail in the density of the urban fabric (European Environment Agency 2012). The updated version was enriched by the local government (Regione del Veneto) with functional information (road network, services, utilities…) using ancillary data sources such as regional cartography, forest inventories, road network graphs, aerial photographs and ground surveys.
For the purpose of this work, we focused on artificial surfaces, agriculture and forest (level 1 of the CLC classification). However, due to the large-scale cultivation of vineyard in the plain to hilly and hilly areas, which we expect to have a significant impact on the morphology of the surfaces, we defined vineyard as an independent classification from agriculture. As well, we considered grass as an independent land cover because it may occur naturally or as the result of human activity (pastures, park and recreational sites), and this allows us to understand better the associated anthropogenic impact on land covers. The land cover classification can be seen from Fig. 4.
Geomorphometric parameters
To make quantitative measurements of landscape properties, we considered three geomorphometric parameters: slope and mean curvature proposed by Evans (1980) and the Spc developed by Sofia et al. (2014a).
Slope and curvature
Evans (1980) describes the DTM surface is approximated to a bivariate quadratic function in the form of:
$$ Z=a{x}^2+b{y}^2+ cxy+ dx+ ey+f $$
(1)
where x, y and z are local coordinates, and a to f are quadratic coefficients.
From such a surface, it is possible to compute the first (slope, Eq. (2)) and second (curvature, Eq. (5)) derivative. Slope (Fig. 5) is calculated as:
$$ \mathrm{Slope}=\arctan \sqrt{d^2+{e}^2} $$
(2)
where d and e are coefficients from Eq. s(1).
Curvature is the second derivative of the surface, also referred to the change rate of slope gradient or direction (Wilson and Gallant 2000), and it emphasises convex and concave elements in the landscape. Evans (1980) proposes two measure of curvature, maximum and minimum, and Wood (1996) testifies that only the resolution of the DTM and the neighbouring cells relevant to these parameters and further defined as
$$ {\mathrm{curvature}}_{\mathrm{max}}=k\times g\left(-a-b+\sqrt{{\left(a-b\right)}^2+{c}^2}\right) $$
(3)
$$ {\mathrm{curvature}}_{\mathrm{min}}=k\times g\left(-a-b-\sqrt{{\left(a-b\right)}^2+{c}^2}\right) $$
(4)
where a, b and c are quadratic coefficients from Eq. (1), g is the grid resolution of the DTM (2 m) and k is the size of the moving window.
From the Eq. (3) and (4), mean curvature (Fig. 6) can be defined as:
$$ {\mathrm{curvature}}_{\mathrm{mean}}=\frac{{\mathrm{curvature}}_{\mathrm{max}}+\kern0.5em {\mathrm{curvature}}_{\mathrm{min}}\ }{2} $$
(5)
Surface peak curvature
The Spc is inversely correlated with anthropogenic pressure (Chen et al. 2015; Sofia et al. 2016; Xiang et al. 2018). Surface morphology (slope) of regions presenting anthropogenic structures tends to be well organised (low Spc) and, in general, self-similar at a long distance. The basis for the evaluation of the Spc is the Slope Local Length of Auto-Correlation (SLLAC). This index quantifies the local self-similarities of slope (Sofia et al. 2014a). It is based on the (demonstrated) assumption that natural areas present low correlations within a neighbourhood because they are inherently irregular, while artificial surfaces to satisfy human needs for mobility and machine access tend to display a higher level of self-similarity with surroundings (Sofia et al. 2014a; Xiang et al. 2019). Describing the algorithm in detail is beyond the scope of this study: the authors refer to Sofia et al. (2014a) for a complete description of the procedure and to other examples of applications (Chen et al. 2015; Sofia et al. 2016; Tarolli and Sofia 2016; Xiang et al. 2018; Xiang et al. 2019).
Briefly, the steps to obtain the Spc (Fig. 7) are as follows:
- 1)
Evaluate correlation between a moving window (W) and a patch (T) centred at the centre of the moving window. The implemented algorithm computes a normalised cross-correlation between a template and the patch, in the spatial frequency domain, and reports a standardised value that ranges between 0 (no correlation) and 1 (perfect correlation). The larger the absolute values, the stronger of the correlation.
$$ {\mathrm{Corr}}_{\left(i,j\right)}=\frac{\sum_{u,v}\left({W}_{\left(i+u,j+v\right)}-{\overline{W}}_{i,j}\right)\left({T}_{u,v}-\overline{T}\right)}{{\left({\sum}_{u.v}{\left({W}_{\left(i+u,j+v\right)}-{\overline{W}}_{i,j}\right)}^2{\sum}_{u,v}{\left({T}_{u,v}-\overline{T}\right)}^2\right)}^{0.5}} $$
(6)
-
2)
Evaluate the correlation length (L) thresholding at 37% (ISO 2013; Whitehouse 2011), the maximum correlation value (Eq. (6)). The length of correlation is the length of the longest line passing through the central pixel and connecting two boundary pixels on the extracted area connected to the central pixel (SLLAC map in Additional file 1).
-
3)
Evaluate the Spc (surface peak curvature) of the SLLAC map defined as:
$$ \mathrm{Spc}=-\frac{1}{2n}{\sum}_{i=1}^n\left[\left(\frac{\partial^2z\left(x,y\right)}{\partial^2x}\right)+\left(\frac{\partial^2z\left(x,y\right)}{\partial^2y}\right)\right], $$
(7)
for every peak (pixel higher than its eight nearest neighbours). Where z stands for SLLAC value, x and y represent the cell spacing, n is the number of considered peaks.
Please refer to the supplement to infer statistic values (mean, median, STD, MAD, skewness…) of each geomorphometric parameters within each land cover.
Statistical analysis
We expect that the topographic signature of anthropogenic activities may be more subtle than the presence of a specific landform and that it would likely be a signature on the frequency of occurrence of the various degrees of the investigated landscape properties (slope, curvature, Spc). That is, the frequency distributions of these measurements would be very different, even though all observed landform types would be found in both natural or anthropogenically modified landscapes. Therefore, we observed the probability density function (PDF) of the considered landscape parameters to (1) investigate statistical differences in geomorphological surfaces between land covers under different landforms contexts and (2) explore the specific topographic signatures of land uses. For this work, the PDFs are a probability density estimate for the sampled data. The estimate is based on a normal kernel function and is evaluated at equally spaced points that cover the range of the sampled data. The distance between points is chosen automatically, based on the range of values. This means that it can be very narrow (< 0.001) for landscape parameter with small magnitude. In these cases, the PDFs can reach values much greater than 1, but their integral over any interval is always less or equal to 1.
After statistically ensuring that the datasets did not present a normal distribution and they exhibit heteroscedasticity, we decided to consider a Kruskal-Wallis test (McKight and Najab 2010) to evaluate whether there were significant differences between landscape properties underneath a specific land cover, across multiple landscapes, and we set a p value threshold of 0.05 for significance. The null hypothesis for this test is that the data for each group are statistically equal.
To investigate the similarities in PDFs between land covers, we applied the two-sample Kolmogorov-Smirnov test, which specifies the equality of probability distribution between two samples (Wilcox 2005; Razali and Wah 2011). One thousand points within each land cover were randomly selected and tested ten times to ensure the robustness of the results.