Relations between morphological settings and vegetation covers in a medium relief landscape of Central Italy

Morphometric units and vegetation classes were determined by applying two classification methods to the Soratte Mount area, a medium relief structure within the Italian Latium region. The study aims at defining the relationships between vegetation and landform types and highlighting the main morphological characteristics within examined land cover classes. These were the result of the application of a supervised classification method to the first 28 (VISNIR) bands of the airborne MIVIS data collected within an extensive survey campaign over Rome Province. The analysis was supported by photo-interpretation of peculiar MIVIS band combinations and by data acquired during field surveys and from a pre-existing vegetation map. The morphometric data were obtained by processing a raster DEM created from topographic maps. These data were processed by means of a new morphometric classification method based on the statistical multivariate investigation of local topographic gradients, calculated along the 8 azimuth directions of each pixel neighbourhood. Such approach quickly estimates the spatial distribution of different types of homogeneous terrain units, emphasizing the impact of erosional and tectonic processes on the overall relief. Mutual relations between morphometric units and vegetation types were assessed by performing a correspondence analysis between the results of the two classifications. Mailing address: Dr. Simone Pascucci, Laboratorio Aereo Ricerche Ambientali (LARA), IIA-CNR, Via del Fosso del Cavaliere 100, 00133 Tor Vergata (RM), Italy; email: s.pascucci@lara.rm.cnr.it


Introduction
The physical environment is often regarded as one of the most important factors controlling the spatial heterogeneity of the landscape in mountain areas (Bolstad et al., 1998;Hoersch et al., 2002).According to Hoersh et al. (2002) the ecological space of a vegetation type corresponds to its fundamental niche and, in contrast to the fundamental, the realised niche is defined through interaction with other vegetation types; finally, the actual geographic space of vegetation types or species is caused by natural factors and human impact.The realised niche and the geographic space of a vegetation type in a mountain environment is closely related to topographic relief and so landform attributes, such as elevation, slope, aspect and others, are important input parameters for spatial analysis and modelling of vegetation distribution in mountain landscapes.Thus, topography creates a patchwork-like pattern of small scale habitats and realised niches within the ecological space.Besides natural environmental factors, the his-tory of human impact in terms of agricultural land use and animal husbandry, but also former and recent natural disturbances (rockfall, mudslides, partly related to human impact, thus being only semi-natural disturbance factors), play a major role for the distribution of vegetation types.The actual vegetation distribution is therefore a result of the complex interaction of historic and recent disturbance factors.Nevertheless, in a landscape marked by human impact the overall influence of topography on vegetation types distribution is beyond doubt.
Remotely sensed data have been widely used for assisting in vegetation mapping in the last few years and have proved to be an effective tool (Rogana et al., 2002), offering the possibility of extrapolating mapping results, especially in large and hardly accessible remote areas.Ahmad et al. (1992) and Wyatt (2000) also discussed the use and restrictions of remotely sensed data for vegetation mapping.The landscape shape, which is so strictly connected to vegetation distribution, is analysed by a discipline such as geomorphometry (Nogami, 1995;Pike, 2002), which, in general, can be considered a bridge between the results of the studies based on field analysis and the physical modelling of geomorphic processes (Onorati et al., 1992).Among other applications (Pike, 2002) morphometric parameters, extracted from a DEM (Wood, 1996), were often analyzed to quantitatively compare terrain units (Giles and Franklin, 1998;Hutchinson and Gallant, 2000;Adediran et al., 2004).
Such approach was followed to investigate a medium-high relief area of the Italian Latium region i) by applying two parallel procedures to identify morphometric units and vegetation types and ii) by statistically analyzing mutual relationships.The definition of these relation- ships and controls highlights the inference of the relative occurrence of land cover types within the morphometric units.

Study area
The examined area, centred about the Soratte Mt., is representative of the natural parks of Rome Province's interest, exhibiting carbonate formations, reaching elevations as high as almost 700 m, and peculiar vegetation covers heavily impacted during the 2nd world war (IGM, 1954).The Soratte Mt. is located 40 km north of Rome (fig. 1) and steeply rises over the surrounding landscape as an isolated mountain included inside a natural reserve, which covers a territory of about 410 ha.The main massif exhibits NNW-SSE orientation with a peculiar elliptic shape and reaches a summit of 691 m a.s.l.
The Soratte Mt. is mainly made up of Mesozoic limestone rocks, constituted almost exclusively by formations such as Calcare massiccio and Corniola, which were emplaced during the Apennines orogenesis (Miocene) and later underwent extensive faulting, subsequent to the Tyrrhenian Sea's opening (Upper Miocene) (Servizio Geologico d'Italia, 1975).It exhibits steep slopes close to the summit and gently sloping densely vegetated areas along the foothills.The relief has undergone strong karstic erosion activity, especially along the mountain top with the formation of small cavities, leading to surface water flow shortage.Moreover, a hummocky plain, mostly covered by olive trees and crossed by several creeks, encompasses the massif.
The peculiar geomorphological characteristics of the Soratte Mt. and the carbonate bedrock strongly bias the soil types and consequently the vegetation spatial distribution, clearly diversifying from those of the neighbouring areas (Abbate et al., 1981;Lulli et al., 1988).In the NE side of the massif mixed evergreen and deciduous woods prevail, while in its SW side, where xeric habitats markedly occur, sclerophylls-rich shrublands and thickets are widespread.Dryness conditions take place and the rocky substratum crops out, xerophytic grasslands and garigues predominate.At the foot of the mountain, the lower slant of slopes and deep soils favour the occurrence of deciduous Turkey oak-rich woods.Even in phytoclimatic terms, the study area shows complexity being in the «transition Mediterranean Region», next to the boundary with the «transition Temperate Region» (Blasi, 1994).

MIVIS data pre-processing
The vegetation classes have been obtained by applying the Maximum Likelihood (ML) supervised method to the first 28 of 102 hyperspectral bands of the airborne MIVIS (Multispectral Infrared Visible Imaging System) data (see table I) collected over Soratte Mt.Such band range (0.4-1.55 nm) was selected because it encompasses the major spectral features related to the biophysical characteristics of the plants; moreover, for the final classification ancillary information about the study area was taken into account.The hyperspectral MIVIS data were acquired during an airborne campaign carried out on 20th June 1998 at an absolute altitude of 1500 m a.s.l.: such value corresponds to elevations varying from 1350 m to 810 m with respect to the topographic surface, leading respectively to a pixel ground resolution of 2.7 m and 1.6 m.Recorded data were radiometrically calibrated to radiance (nW cm −2 sr −1 nm −1 ) and geometrically corrected at the CNR ground station seated in Pomezia (Rome, Italy) (Bianchi et al., 1994).
Owing to the morphological complexity of the study area, a raster DEM (10 m/pixel ground resolution) was employed to yield thematic maps topographically corrected.Geometric correction was performed on the basis of re-sampling tables, attained by integrating the airplane's position and attitude data recorded by MIVIS position and attitude system sensors.These tables yielded geocoded thematic maps on the desired scale of 1:10.000,safeguarding the data spectral integrity during this pre-processing phase.
Atmospheric correction procedures were then applied by using the 6S code to correct for the additive path-radiance component.
Another topographic problem affects sensor-perceived radiance: the irradiance received by the target varies with the cosine of the sunbeam incidence angle.The larger the incidence angle, the less the amount of radiation reaching the surface (Teillet et al., 1982): if no other factors change, less electro-magnetic radiance is reflected, as being less received by the target.The instrument perceived radiance is also affected by the sun elevation owing to atmospheric scattering: a lower sun elevation means a longer transportation for the sunbeam and, thus, a larger scattering effect (Ekstrand, 1993).To some extent, the target altitude also has an effect on the sensor registered radiance.This is because the optical thickness of the atmosphere decreases with altitude and also affects the scattering.The amount of radiance reflected by the target depends on the class-specific reflection in different directions, but the class-specific reflection may change with topography if the geometric structure within the class, e.g. a forest canopy, changes.Different types of vegetation respond differently to direction and illumination effects (Thomson and Jones, 1990).The relative importance of slope-aspect effects in forest reflectance is still under debate.
Among different topographic correction models, the MIVIS data gathered over Soratte Mt. were processed by applying the C-correc-tion method, which is a modified version of the cosine correction procedure, where a linear relation between radiance and cosine of the incidence angle is assumed.This procedure accounts for the local topographic conditions that influence the direct solar irradiance.To avoid overcorrection, a band specific constant, derived from regression functions between radiance and cosine of the incidence angle, is introduced: it is called C and is calculated by dividing the offset of the regression line by the slope.

Land cover data classification
The preliminary analysis was based on the photo-interpretation of the FCC (False Colour Composite) of three MIVIS bands, as pointed out in table II.The 13, 7 and 1 channels (FCC A in table II) respectively represent for vegetation the two bands with higher spectral absorption (Red and Blue) and the one with major reflectivity (Green): this band combination better approaches the natural colour representation.The 19, 28 and 13 channels (FCC B in table II) respectively correspond, instead, to the maximum reflectance plateau in the Near-Infrared, the H2O absorption peak and chlorophyll absorption peak; hence, this FCC highlights vegetation covers, discriminating between agricultural lands and spontaneous vegetation.Moreover, single MIVIS bands such as 28 (1.50 <m<1.55 nm) and 93 (8.21<m<8.56nm) were photo-interpreted to discriminate less easily detectable cov- ers, such as bare soils, vegetation sparse or in relative dryness conditions.MIVIS corrected data were further analysed by applying a supervised classification method which needs the collection of the spectral signatures of the main vegetation types.
At first glance, the study area appears mainly exploited for agricultural use with a significant diffusion of olive-trees; there are, however, also wide territory segments characterized by relatively intact natural environment, with well structured woods associated with re-naturalization areas.In particular, the following land use classes were examined and their boundaries outlined: natural environments (woods, thickets, shrublands, grasslands, garigues and rock communities); agricultural lands (arable lands, olivetrees, orchards); urban areas (built-up centres, single dwellings); quarries; water bodies.
By applying appropriate masking procedures, only areas defined as «natural environments» were taken into account in the subsequent investigation.Ancillary information was further provided by carrying out in situ surveys to determine the interpretation keys and to ver-ify the distribution of vegetation types.The results were compared and integrated with those extracted from the pre-existing «Vegetation Map of Soratte Mount» (Abbate et al., 1981).
By considering MIVIS calibrated hyperspectral data and the ancillary information, spectrally homogeneous training areas were selected allowing the definition of an exhaustive number of Regions Of Interest (ROI, as implemented in the ENVI 4.0.software package; RSI, 2003) to be used as input data for the ML supervised classification method.This approach was followed since vegetation types, in the same phenological conditions, exhibit similar spectral trend, but shifted absorption and reflectivity peaks (see fig. 2).
The chosen ROI refer to the following land cover classes: 1. Mediterranean evergreen sclerophylls woods with predominance of Quercus ilex (nomenclature of the species follows Pignatti, 1982) and subordinate presence of deciduous trees (Ostrya carpinifolia, Fraxinus ornus, Acer monspessulanum).
16. Urban areas (build-up centres, single dwellings).17.Quarries.All identified classes were then assembled into a unique thematic map on a 1:10.000scale (see fig. 3) and compared with the vegetation types distribution shown in the pre-existing vegetation map (Abbate et al., 1981;Lattanti et al., 1981).

Morphometric units discrimination
In order to quantitatively define the relief characteristics of the study area, a procedure, based on the analysis of local morphological setting, was applied to morphometric data gathered by processing a raster DEM (10 m/pixel).It is the same procedure used to geometrically correct the MIVIS images and was obtained by interpolating contour lines of a topographic map on a 1:10.000scale.The quantitative technique applied to process this DEM represents a new method of analysis, implemented also in other geomorphological studies (Parcharidis et al., 2001;Adediran et al., 2004).It is based on the application of a multivariate statistical procedure to a 8-layers stack, describing topographic gradients measured along the 8 compass orientations of each DEM pixel neighbourhood.This makes it possible to define areas characterized by similar local morphological setting, that reveal changes in shape, orientation and steepness of the relief.The method follows the approach inherent in the unsupervised classification or the spectrum shape analysis of multispectral bands imagery.For each pixel of a DEM, the 8 neighbourhood cells values are considered as defining a trend resembling that of contiguous spectral bands (left panel of fig.4).Therefore, the input data to the classification procedure are the eight elevation differences between each DEM's pixel and its neighbours, calculated along the 8 main azimuth orientations starting from the NW corner and moving clockwise (right panel of fig.4).The classification method, chosen to process the resulting gradient values, has been an unsupervised cluster analysis technique, such as ISODATA.It represents a flexible, iterative partitioning method used extensively in engineering (Hall and Khanna, 1977) and based upon estimating some reasonable assignment of the pixel vectors into candidate and, then, moving them from one cluster to another in such a way that the sum of squared error (SSE) measure of the preceding section is reduced.ISODATA stands for «Iterative Self-Organizing Data Analysis Technique», and «Self-Organizing» refers to the way in which it locates the clusters that are inherent in the data.The ISODATA utility repeats the clustering of the multi-imagery until either a maximum number of iterations have been performed, or a maximum percentage of unchanged pixels have been reached between two iterations.The output of the classification is presented as a digital thematic map showing the spatial distribution of the class membership across the study area, where each class exhibits similar morphological setting.
To apply the ISODATA procedure to the Soratte Mt. area, the following input parameters were chosen: a number of 14 classes, a change threshold percent of 1.0 and a maximum iterations number of 25, but the procedure usually converges before 15 iterations.The resulting classification map was presented assigning each class a given colour shade (fig.3) to facilitate the interpretation and to perform an accurate evaluation of the spatial distribution of different local morphological settings.The classes were then statistically analysed by computing mean and standard deviation of the eight layers, represented by the elevation differences with respect to each reference DEM's pixel.Moreover, to verify the accuracy of the corresponding map, slope and aspect values, relative to the Soratte Mt. area, were calculated from the same elevation matrix.Then, mean and st.dev. of height, slope and aspect, relatively to all 14 ISODATA classes, were computed and compared with the 8-layers statistics earlier obtained directly from the classified thematic layers.

Land cover classification
The analysis of the MIVIS remotely sensed data, supported by the contribution of ancillary information (gathered from field surveys and pre-existing vegetation map), made it possible to define the spectral signatures of the main plant typologies described for the territory of the Soratte Mt. natural reserve.The ML algorithm was chosen because it proved to be more correct than other classifiers for the recognition of spectral classes, such as vegetation species, which exhibit many correlated spectra (Congalton, 1993).The decision criterion of the ML classifier is, in fact, based on the calculation of statistical parameters (average value and covariance matrix) for each training area.The ML classification yielded valuable results in the discrimination of forest, shrub and herbaceous formations, and, in particular, among the woodland types.This discrimination among different plant typologies mainly exploits the spectral differences present in the near-infrared region (Deering et al., 1994) (see fig. 2): for example, the sclerophyllous communities have lower spectral responses than broadleaf ones.Overall accuracy of the ML classification results turned out to be of 94.4% and the Kappa coefficient was equal to 0.9259.

Morpho-units statistics and interpretation
The statistical results and interpretation for all 14 morphometric classes attained are shown in table III that illustrates the mean gradient values of the 8 ISODATA input layers, the average values of elevation, slope and aspect angles, and the relative morphostructural interpretation.Looking at these statistical results, it is possible to verify that the morpho-units, classi-   III) appear different: in fact, class 10 is steeper uphill than downhill, while class 14 is steeper downhill than uphill, determining respectively morphological concave and convex forms.

Correlation between land cover types and morphological units
In order to obtain the relation between morphometric units and land cover types, a correspondence analysis was performed (Hoersch, 2003).
i) A preliminary cross-tabulation table was computed to show the number of pixels for each class as obtained from both the morphometric and land cover maps.Then, the percent occurrence of the different land cover types within each morpho-unit was calculated by standardizing the corresponding absolute representative pixels number with respect to the total number of pixels of the considered morpho-unit (see table IV).
ii) Furthermore, the distribution of the dominance land cover types within all the morpho-metric units was assessed.Table V shows this correspondence between the morphometric units and the relative dominance land cover types.
The spatial distribution of the different ISO-DATA classes and their correlation to the land cover types were examined and shown in the tables IV and V. Looking at table IV it is evident that class 14 (arable lands) is the most present land cover type in many morpho-units; moreover, mixed sclerophylls thickets with deciduous species (class 3) and Mediterranean evergreen sclerophylls woods with predominance of Quercus ilex (class 1) are also quite abundant in the morphometric units.The analysis of table V presents a different interpretation key allowing to determine peculiar morphological characteristics related to land cover types: class 7 (Carpinus orientalis communities) is strongly correlated to morphometric class 11 (steeply sloping areas facing NW).Other remarkable links occur between: land cover class 3 (mixed sclerophylls thickets with deciduous species) and morpho-unit #2 (steeply sloping areas facing SW-W); land cover 8 (Acer monspessulanum communities) and morpho-unit #10 (averagelysteeply sloping areas facing E-NE); land cover 6 (Ostrya carpinifolia communities) and morpho-unit #3 (averagely sloping areas facing NE-E).It has been also observed that in morphometric unit#1 (very steeply sloping areas facing NE-N) the most abundant land cover types is class 3 (mixed sclerophylls thickets with deciduous species), while in morphounit#5 (almost flat areas facing S-SW) the dominant land cover is class 16 (urban areas).
The extracted mutual relations highlight the effect induced on the vegetation spatial distribution by the sun exposition and irradiation (light and heat) and by the geomorphological characteristics, responsible for a major or minor evaporation-transpiration of the soils and plants.As shown in table V, on the steeply sloping areas facing SW-W (class 2 of the morphounits) of the Soratte Mt. area, where the soils undergo erosion processes and exhibit scarce thickness (Lulli et al., 1988), the sclerophylls thickets (class 3 of the vegetation types) are abundant and constituted by thermophilus species that need lower water contribution.In- In the SW-W side at the foot of the mountain, where the thickness of the ground decreases and dryness conditions take place (class 6 of the morpho-units interpretation), the rocky substratum crops out and xerophytic grasslands and garigues (class 11 and 12 of the vegetation types) prevail.

Conclusions
The synergistic use of hyperspectral remote sensed data and DEMs was very helpful for the evaluation of the direct influence of the landscape morphology on the productivity of the vegetation covers.The unsupervised classification of a multilayer data set extracted from a DEM has allowed the automated definition of geomorphic units within the natural reserve of Soratte Mt.The application of this processing method for the evaluation of similar morphometric units has permitted to highlight the spatial distribution of morphologic features and their

Morphometric units interpretations
Prevailing land cover classes degree of intensity, providing a valuable new information source for morphological applications.The designated landform units can be easily overlaid on any digital map and imagery for further applied research.In the context of the Soratte Mt. area, the peculiar correspondences among some vegetation categories and morphounits may be also due to the overall good accuracy of the vegetation thematic map as obtained by applying the ML classification procedure to the MIVIS hyperspectral data set.An interesting development of the presented approach could be, for instance, the investigation of the mutual relation between morphology, geomorphology, soil, vegetation distribution and human activities impact, framed also into the natural processes and the natural hazard problems.A better understanding of this relation would also benefit the inventory of forest resources as well as the evaluation of the potential land productivity in areas of relief and of limited cartographic coverage.

Fig. 1 .
Fig. 1.Location of the study area.On the left the topographic map of the Soratte Mt. on a 1:10.000scale.

Fig. 2 .
Fig. 2. MIVIS at sensor radiance mean spectra of the 13 Region Of Interest (ROI) relative to the natural environment land cover classes.

Fig. 3 .
Fig. 3. Top left: map showing the prevailing vegetation classes on Soratte Mt. as derived from the ML classification.Top right: two 3D perspective views of the vegetation classes map, draped over the DEM: in the former (down) the view is from South to North, in the latter (up) from north to south.Bottom left: map showing the results of the ISODATA classification.Bottom right: two 3D views of the morphometric units map, draped over the DEM: in the former (down) the view is from north to south, in the latter (up) from south to north.Vertical exaggeration of the 3D views is 2 to better enhance the overall relief.

Fig. 4 .
Fig. 4. Sketch of the basic concept for local morphometric analysis.

Table III .
Altitude, slope and aspect mean values and morphostructural interpretation of the elevation differences between each pixel and the eight neighbours, for the 14 classes obtained by applying ISODATA clustering method to the study area.

Table IV .
Relative distribution of the 17 land cover types within each of the 14 morpho-units.

.0278 0
.0047 0.0089 0.0107 0.0047 0.0282 0.0032 0.0014 0.0124 0.0060 0.0018 stead mixed deciduous broadleaf oak-woods with Quercus cerris, Quercus frainetto, Carpinus orientalis, Fraxinus ornus, Acer campestre (class 5 of the vegetation types) are dominant in the NW-W slopes, in agreement to the soils that are deeper and with a better water retention capability, also improved by the NW slope aspect.

Table V .
The dominance land cover types in each and all morphological units.