Escolar Documentos
Profissional Documentos
Cultura Documentos
1 ITC, Division of Soil Sciences, PO Box 6, 7500 AA Enschede, the Netherlands (fax: +31-53-4874-379; phone: +31-53-4874-322; e-mail:
zincka@itc.nl)
2 Apartado Aereo 5057, Santa Marta, Colombia
3 Curtin University of Technology, GPO Box U1987, Perth, WA 6845, Australia
4 UNAM, lnstituto de Geografla, Ciudad Universitaria, CP 04510, D.F. Mexico
KEYWORDS: mass movements, gullies, mapping, moni- characteristics. In both cases, the soil material frequently
toring, modelling, hazard assessment, remote sensing, conditions the initiation and development of the erosion
GIS processes and, at the same time, is affected by them.
There are also dynamic relationships between gullies and
mass movements themselves, since small landslides or
ABSTRACT earth slumps often convert into gully heads, while gullies
frequently expand laterally through mass wasting. Mass
Natural as well as human-induced mass movements and gullies are
severe environmental hazards. Remote sensing data offer promising
movements and gullies are processes of multiple ori-
possibilities for identification and monitoring. But their effective use gin/causes and the initiation mechanisms might take
in mountainous areas is hampered by cloud effects and relief-con- place at the terrain surface or beneath. They are chaotic
trolled factors, which cause geometric distortions and shadow phenomena, triggered by sudden alteration of the envi-
areas, among other constraints. Nevertheless, aerial photographs ronmental equilibrium and generating catastrophic dam-
and satellite images (visible, infrared and microwave bands), or
ages Common factors, such as the complexity of the
combinations thereof, have been successfully used to discriminate
and delineate landslide and gully types. GIS modelling of mass
processes and interactions, the catastrophic character of
movements and gullies, using ancillary information in combination the events, and the difficulty in predicting their spatial
with remote sensing data, is rapidly developing. The shortcomings and temporal occurrence, contribute to making deter-
of deterministic modelling of such chaotic phenomena as mass ministic modelling cumbersome in both cases.
movements and gullies highlight the relevance of GIS-assisted
approaches to exploratory and predictive modelling. This paper
However, the development of modern earth observation
describes practical applications of remote sensing and GIS for map-
techniques, in particular the availability of multitemporal
ping, monitoring, exploring cause-effect relationships and assessing
hazards of mass movements and gullies in hilly and mountainous remote sensing data, improves the mapping and moni-
areas. toring possibilities. Similarly, GIS techniques facilitate the
integration of multiple data layers and spatial simulation
to explore cause-effect relationships. Such issues have
INTRODUCTION been addressed by a number of authors from different
Mass movements and gullies are severe environmental perspectives, focussing on mass movements and/or gul-
hazards in mountainous areas. Both erosion processes, lies specifically, or environmental hazards in general. In
especially mass movements, cause extensive material and this context, research with significant input of remote
human losses, which are often blamed in official statistics sensing and/or GIS has focused on a variety of aspects,
on primary causes such as earthquakes or hurricanes. including mapping and monitoring approaches [Kienholz
Mass movements probably constitute the single most et al, 1983; Pike, 1988; McKean et al, 1991; Palacio-
widespread hazard on the earth’s surface. For instance in Prieto & Lopez-Blanco, 1994; Verstappen, 1995;
New Zealand, one of the few countries that has a coun- Chowdhury & Flentje, 1996; lrigaray et al, 1996; Liener
trywide landslide map, 36% of the territory shows for- et a/, 1996; Rosenbaum & Popescu, 1996; Duan & Grant,
merly or presently active mass movements [Eyles, 19831. 20001, the scale at which hazard maps are prepared for
regional, local and site planning purposes [Rengers et a/,
Although the basic processes are fundamentally differ- 1992; van Westen, 1993; Leroi, 19961, the synergy
ent, mass movements and gullies share some common obtained from merging different kinds of remote sensing
43
Mapping and modelling mass movements and gullies JAG l Volume 3 - Issue 1 - 2001
data [Koopmans & Forero, 1993; Buchroithner, 1995; phenomena. For instance, a recent study conducted by
Singhroy, 1995; Metternicht & Zinck, 1998; Singhroy et Kniveton et al [ZOO01 shows that high temporal resolu-
al, 19981, the integration of remote sensing and GIS tion data can be used to provide early warning of atmos-
techniques [Leroi et a/, 1992; Garcia-Melendez et al, pheric conditions likely to initiate debris flow events.
1998; Metternicht & Fermont, 1998; Metternicht &
Zinck, 19981, and simulation modelling and hazard pre- RELIEF-INDUCED CONSTRAINTS ON THE USE OF REMOTE
diction [Bocco et a/, 1990; Wilson, 1996; Fujita et al, SENSING DATA
1996; Leroi et a/, 1996; Zinck, 1997, 1999; Fabbri & Strong relief variations cause geometric and climatic con-
Chung, 1999; Shrestha & Zinck, 1999; Kniveton et a/, straints to the efficient use of remote sensing in moun-
ZOOO]. tainous areas. Topography and elevation produce distor-
tions, which are reflected in anomalous height differ-
Our paper addresses some issues related to the use of ences, scale differences, relief displacements and shadow
remote sensing for mapping and monitoring purposes. areas, among others. They also cause climatic conditions
Further, it describes and illustrates several GIS-assisted to change from the bottom of the valleys to the summit
approaches to modelling gully and mass movement haz- of the ridges, with effects on temperature, humidity,
ards, using examples drawn from research work devel- cloud formation, snow cover and grey tones
oped at ITC by advanced students. [Buchroithner, 19951. The delineation of gullies and mass
movements is particularly affected by such geometric and
climatic restrictions, as they are usually elongated fea-
MAPPING tures, which develop parallel to the slopes over long dis-
DATA RESOLUTION REQUIREMENTS tances and may cross several bioclimatic elevation belts.
The application of remote sensing to environmental stud- A variety of technical solutions including digital elevation
ies, including the mapping and monitoring of mass and illumination models, multi-seasonal and multi-direc-
movements and gullies, is controlled by the spatial, spec- tional remote sensing data, haze and atmospheric cor-
tral and temporal resolutions of the data. Spectral data rections, radar and stereo data, interferometry and GPS,
resolution does not refer only to the number of spectral is available to correct for relief, scale, shadow, weather
bands offered by the sensor, but also to the ability of and snow effects (Table 1).
specific portions of the electromagnetic spectrum to pro-
vide enough spectral separability amongst surface fea- Changes in topographic parameters, such as slope gradi-
tures related to mass movements or gully formation ent, aspect and orientation with respect to sun elevation,
processes. This requires a good understanding of the create illumination variations which cause the reflectance
interactions between ecosystem characteristics and values from the same surface cover type to vary. This pro-
incoming solar radiation or artificially propagated elec- duces an elongation of the training samples in the fea-
tromagnetic energy, such as in the case of radar sensors ture space plots and induces a bias towards either fully
[Lunetta, 19991. illuminated or fully shaded slopes. Furthermore, the
application of multispectral classification algorithms,
The spatial resolution of the sensor determines the scale based on statistical pattern recognition, assumes that the
at which the data may be useful for mass movement or training samples be normally distributed. The latter is
gully analysis and mapping. The concept of a minimum generally not the case in areas of rugged topography,
map unit, which makes it possible to consistently delin- with frequent slope gradient and aspect variations caus-
eate the smallest ground features of interest over a ing large illumination differences, and this requires thus
selected area, is an important consideration when estab- some kind of data transformation to approach normality.
lishing spatial data requirements. Lunetta [ 19991 recog- In a case study of land use classification in the Likhu
nises that remotely sensed data should support a mini- Khola watershed, Nepal, Shrestha & Zinck [2001] used
mum map unit, characteristic of the process or feature intensity normalisation of multispectral data to remove
being mapped. His view is supported by the current trend the topography-induced constraints for land degradation
of using spatially explicit approaches based on available assessment. This approach solved the problem caused by
remote sensing, GIS and digital terrain analysis tech- biased sampling of the training pixels. The effect of data
niques, which allow researchers to consider local hetero- normalisation in the feature space plots is shown in
geneity of the landscape, such as discussed by Duan & Figure 1.
Grant [2000] and Shroder & Bishop [19981.
LANDSLIDE MAPPING
Finally, temporal resolution is determined by the revisit- A variety of approaches has been used for landslide map-
ing cycle of the sensor. Ideally, the data acquired for ping, including on-ground monitoring, remote sensing,
mass movement or gully mapping should have a tempo- factor overlay, statistical models, and geotechnical
ral resolution higher than the changes evidenced by the process models [Duan & Grant, 20001. Singhroy [19951,
44
Mapping and modelling mass movements and gullies JAG l Volume 3 - Issue 1 - 2001
TABLE 1: Problem-and-solution matrix regarding relief-induced factors influencing remote sensing data use in high-mountain regions
[Buchroithner, 19951.
REMOTE SENSING CARTOGRAPHY AND HIGH MOUNTAINS
PROBLEM SOLUTION REMARKS
Weather, Haze, Clouds Multi-seasonal data (high temporal resolution) Also multi-sensoral
Haze correction Simple approach
Atmospheric correction Problem of data availability
Radar Cloud penetration
45
Mapping and modelling mass movements and gullies JAG l Volume 3 - Issue 1 - 2001
Distribution P P E E E
Classification A E A
Factors controlling
slope stability
Geomorphology of slopes P P A E A
Lithology P P
Structure A A E E E
Neotectonics A
Landusellandcoverlinfrastructure E A A E E
features on steep valley slopes; and (3) a combination of the gullies: (1) natural vegetation (mainly shrubs), (2) fal-
airborne SAR and TM images is appropriate to monitor low land (bare during dry season), (3) slightly eroded
retrogressive slope failures. areas (mainly sheet erosion), (4) moderately eroded areas
(mainly rill erosion), (5) badlands (mainly gully erosion),
GULLY MAPPING and (6) miscellaneous land (ephemeral riverbeds and
Individual gullies are elongated, narrow features that are stone pavements).
difficult to identify at medium and small scales. More
often, gullies develop into large ramified badland areas, In the above study, gullies showed low reflectance in the
which are easier to map from remote sensing data. In a visible and near infrared. This can be attributed to a
case study developed in the intra-mountainous basin of shadow effect caused by the depth and surface irregu-
Cochabamba, eastern Bolivian Andes, the possibility of larities of the gullies, which traps the incoming light and
discriminating gullied badlands from other kinds of sur- reduces the reflectance (Figure 2). In the middle infrared,
face features was assessed using Landsat TM and JERS-1 reflectance remained lower than that of the other fea-
SAR data [Metternicht, 1996; Metternicht & Zinck, tures because of the surface roughness component.
19981. Six information classes were considered, referring After merging Landsat TM and JERS-1 SAR data, the
to terrain surface components spatially intermingled with energy backscattered by the gullied areas remained rela-
50
n Badlands
45 n Moderately eroded areas
Slightly eroded areas
40 0 Miscellaneous land
0 Fallow land
35
n Vegetation
1 2 3 4 5 6
LANDSAT TM BANDS (1 to 7)
FIGURE 2: Digital numbers of selected surface components from Landsat TM bands, Sacaba valley, Bolivia [Metternicht, 19961.The
X-axis represents the visible (1, 2, 3). near infrared (4), middle infrared (5, 7) and thermal (6) bands of the Landsat TM sensor.
46
Mapping and modelling mass movements and gullies JAG l Volume 3 - Issue 1 - 2001
tively low, but became more variable so that the separa- intricate distribution of branching gullies generates
bility of the gullies from the other erosion features and impure pixels, whose spectral reflectances are mixtures
surface components slightly improved (Figure 3). of the reflectances of the individual components. Thus
the issue is not one of spectral confusion, but one of
The effect of data fusion on spectral class separability spectral mixing of the surface components within a sin-
was assessed using transformed divergence (TDij), a mea- gle pixel. To solve this problem, a linear mixture model
sure to select the best band combination for pixel including five pure end-members (n-l), one of them
labelling. Output values are usually scaled between 0 and being pure gullies, was applied to a 6-band Landsat TM
2000. For TDij>l600, good separability between classes i data set (the thermal range being excluded). The proce-
and j can be expected [Richards, 19931. In the study case dure is described in Metternicht & Fermont [1998]. This
[Metternicht, 19961, best between-class separability was made it possible to generate an error map from spectral
obtained when combining JERS-1 SAR (L-band) and TM unmixing. The error between the original map and the
bands 1, 3, 4, 5, 6 and 7, causing all transformed diver- root-mean-square image representing the best-fit output
gence values to cross the threshold figure of 1600. This proportion map was lower than 1 percent in 36 percent
is particularly the case between badlands and miscella- of the classified area, and between 1 and 10 percent in
neous land, as well as between badlands and moderate- 63 percent of the area. Only 1 percent of the classified
ly eroded areas. The improvement of class separability is area had an error greater than 10 percent. These results
also reflected in the spatial distribution of the gullies were further used to improve the image classification.
before and after data merging. With Landsat data alone,
gullies appear as large undifferentiated areas. The efficiency of the different portions of the spectrum
Comparatively, Landsat and JERS-1 SAR data together (ie, visible, infrared, thermal and microwave) to properly
produce a sharper spatial segmentation of the gullies. separate gullies from the other selected surface compo-
nents was assessed using a procedure based on the per-
However, in any case of band combination or data centage of spectral confusion, derived from the trans-
fusion, the class accuracy for gullies remained low (54 formed divergence analysis, in relation to the total num-
percent), when compared to the other surface compo- ber of training samples [Metternicht, 19961. Applying
nents included in the study. This is because gullies are this criterion to the study case, the different regions of
heterogeneous areas including variable mixtures of nat- the electromagnetic spectrum performed relatively well
ural vegetation, stone pavements and eroded soils. The when separating gullies from fallow land and from slight-
n Badlands
180 w Moderatelyeroded areas
H Slightlyeroded areas
1
0 Miscellaneousland
160
1 c] Fallow land
n Vegetation
140 -
120 -
3 4 5 6 7 8
LANDSAT TM SANDS (1 to 7’) and JEW-1 SAR (8) DATA
FIGURE 3: Digital numbers of selected surface components from Landsat TM and JERS-1SAR bands, Sacaba valley, Bolivia
[Metternicht, 19961.The X-axis represents merged Landsat TM and JERS-1SAR data (1 to 7), and JERS-1SAR data alone (8).
47
Mapping and modelling mass movements and gullies JAG l Volume 3 - Issue 1 - 2001
ly and moderately eroded areas. In contrast, major spec- vide the conditions for geomorphic stability [Zinck,
tral confusions occurred between gullies and surfaces 19961.
covered with rock fragments and stone pavements, espe-
cially in the visible and infrared ranges. Similarly, the vis- Mass movements, as well as gullies, are chaotic phenom-
ible and microwave ranges performed poorly in discrimi- ena. They occur when the terms of a meta-stable equi-
nating between gullies and vegetation cover (Figure 4). librium situation drastically and suddenly change. The
This study shows that regardless of the satellite sensor, activating factor might be abnormal rainfall or an earth-
spectral band combinations and/or image classification quake. The conditioning factors include properties of the
algorithm applied, accurate detection of gullies requires vegetation cover, topography, geomorphodynamics,
contextual knowledge to improve the between-class geological structure and hydrogeological behaviour. But
spectral separability. In particular, knowledge about the it is the intrinsic nature of the soil material which deter-
relationship between gullies, their characteristic surface mines its propensity to mass wasting, mainly through its
features, and geomorphic positions is needed. mechanical and hydrological properties. The rheological
contrast between consecutive soil horizons creates
planes susceptible to functioning as shear surfaces.
MODELLING When the actual moisture content of the soil mantle
BOUNDARY CONDITIONS exceeds the water holding capacity and/or the Atterberg
Attempts to model mass movements and gullies have limits, a mass movement hazard exists. A simple graphi-
been made from different perspectives [van Westen & cal comparison between water-potential or consistence
Terlien, 19961. Ideal modelling would be deterministic, profiles and the real moisture contents of the soil cover,
which aims at explaining erosion phenomena through preferably as time series, allows identification of areas
the mechanics of the processes. Mechanistic modelling where a mass movement potential builds up during criti-
would allow one to predict the geomorphic response cal periods of the year [Zinck, 1986, 19961. While deter-
from the properties of the soil or substratum material, ministic models able to explain and predict mass move-
thus from its intrinsic susceptibility. For instance, when ment hazards require further elaboration, the simple
the micro-fabric of the material is deflocculated, thus in graphical correlation between consistence and moisture
liquid state, the expected geomorphic response would be profiles can be used as a basis for designing rhexistatic
mudflow. Similarly, with a dispersed micro-fabric and a models. To improve the basic conceptual framework for
plastic consistence, the predicted mass movement would deterministic modelling of mass movements, better
be solifluction. An aggregated micro-fabric and semi- understanding of the relationships between the micro-
solid consistence would promote sliding. In contrast, a fabric of the soil material and the geomorphic response
flocculated micro-fabric and solid consistence would pro- must still be achieved. As quite some progress is needed
35-
30-
25
20
15
10
Ve: Natural vegetation Fa: Fallow land SI: Slightly eroded areas
MO: Moderately eroded areas Ba: Gullied badlands Rf: Rock fragments and stone pavements
FIGURE4: Spectral confusion among terrain surface features, Sacaba valley, Bolivia [Metternicht, 19963.
48
Mapping and modelling mass movements and gullies JAG l Volume 3 - Issue 1 - 2001
EXPLORATORY APPROACH
Exploratory models attempt to identify non-explicit
cause-effect relationships between environmental hazard
types (eg, soil erosion processes) and affected soil types
(ie, soil map units) in order to predict, from these rela-
tionships, soilscape areas potentially exposed to degra-
dation. Because it relies on relatively simple GIS opera-
tions, cartographic modelling is an exploratory mode fre-
SLOPE IN %
quently implemented to this end [Bocco et al, 19901. The
_ KW”NrrAtlE4 ___________ ?hOFTOTAL
GUUJED
*REpI
overlay of information layers, usually represented by
series of thematic single-attribute maps, allows high- FIGURE 5: Observed gullied areas per slope gradient units at
Huasca de Ocampo, Mexico [Vazquez-Selem & Zinck, 19941.
lighting of areas of coincidence between factors presum-
ably controlling erosion processes and features resulting
from these processes.
60
Erosion features caused by gullies or mass movements
are not randomly distributed on the landscape. They
50
develop in response to a combination of controlling fac- iis
tors The simple overlay of a gully distribution map on % 40
top of maps representing environmental factors, such as 9
3
geoforms, slope gradients, soils, lithologic units and land 2 30
0
use types, highlights the degree of spatial coincidence rp
20
between gully occurrence and specific factors. Such car-
tographic coincidence helps identify the most favourable 10
combination of conditions and reveals underlying cause-
effect relationships. 0
1 2 3 4 5 6 7 6 i lb 11
was performed in a volcanic area located about one hun- m % OF UNIT AREA % OF TOTAL GULLIED AREA
Similarly, the overlay of the gully map and the soil map FIGURE 6: Observed gullied areas per dominant soil units at
Huasca de Ocampo, Mexico [Vazquez-Selem & Zinck, 19941.
shows that most gullies develop on Alfisols (Figure 6). In
fact, about 90 percent of the gullied surface within the
study area concentrates on deep Paleustalfs, although difference of permeability between the well-structured
these soils account for only 39 percent of the total area. colluvial cover and the highly clay-dispersible subsoil
Alfisols in this area are composed of two main layers: a favours horizontal water flow along percolines, from
colluvial cover material, 40 to 60 cm thick, which lies on which gullies initiate. In contrast, the areas covered with
a buried Bt horizon belonging to a truncated subsoil. The Mollisols are virtually free of gullies, although these soils
surface of discontinuity between the two materials is occur in conditions of slope and land use similar to those
marked by ancient human artefacts (2400 years BP). The of the Alfisols.
49
Mapping and modelling mass movements and gullies JAG l Volume 3 - Issue 1 - 2001
PREDICTIVE APPROACH tion to the total extent of gullies (562 ha) within the
Predictive models usually implemented in GIS are based whole study area (8009 ha). The first criterion was con-
on rules and expert knowledge. Such models lack deter- sidered more diagnostic than the second one, because it
ministic capability, because they neither simulate nor is independent of the total gullied area and thus reflects
explain the mechanisms involved in gully formation or better the intrinsic susceptibility of each thematic map
mass wasting. They are built on the results of the unit to gully formation. To establish class limits, critical
exploratory analysis, which identifies cause-effect rela- threshold values of area percentages were determined by
tionships from the spatial coincidence between observed iteration from the graphics showing the frequency of
erosion features (eg, gullies, landslides) and landscape gullies per classes of environmental factors (eg, Figures 5
factors. Such models are not able to take into considera- and 6).
tion, for the purpose of prediction, the role played by the
activating factors (ie, exceptional rainfalls, earthquakes) Using these rules, six models were established with
in triggering the processes. The modelling rules are main- decreasing boundary conditions (Table 3). The first
ly based on the conditioning factors of the environment model, for instance, takes into account only the themat-
(eg, slope, vegetation cover, rock substratum) and a few ic map units with a high percentage of observed gullied
soil properties. This type of model is able to (1) repro- area (more than 10 percent or more than 15 percent
duce the spatial distribution of existing gullies for valida- according to the environmental factor considered). Such
tion purposes and (2) predict the potential occurrence of a combination of rules is highly selective, since only a
gullies in areas with favourable conditions. few units satisfy the requirements. As a consequence,
the gullied area estimated by model 1 is small. But, at the
To illustrate the above considerations, a set of nested same time, the model is efficient because a large pro-
models was developed to confirm observed gullies and portion of the calculated gullied area corresponds to
assess the hazard of potential gullies in the same Huasca existing gullies. The other models operate with less
de Ocampo area of central Mexico, where the explorato- restrictive rules in decreasing order.
ry models were developed [Vazquez-Selem & Zinck,
19941. The assessment rules mobilise only the area per- The relative efficiency of the various models in corrobo-
centages covered by observed gullies in the units of the rating observed gullies is presented in Figure 7. The total
thematic maps representing environmental factors. Six gullied surface area of 562 ha is equivalent to 7 percent
environmental factors were selected on the assumption of the study area. An ideal model would confirm 100
that they contribute, in one way or another, to gully for- percent of the gullied area with only 7 percent of the
mation: geoforms, lithologic units, slope gradients, slope study area. Thus, an efficient model is one which approx-
shapes, dominant soils and land uses. Two criteria were imates this optimum performance. To assess the efficien-
implemented to account for the area percentages: (1) the cy of the models, the gullied area calculated by each
percentage of area with observed gullies in each themat- model was compared to the area percentage of existing
ic map unit, and (2) the percentage that the observed gullies properly confirmed. Accordingly, models Zb, 3a,
gullied area in each thematic map unit represents in rela- 3b and 4a are the best predictors.
TABLE 3: Rule-basedspatial models for gully prediction at Huasca de Ocampo, Mexico [Nzquez-Selem & Zinck, 19941
Model 2 [>lO% unit area >15% unit area >5% unit area or >lO% unit area ~10% unit area >5% unit area
and >lO% total or >30% total >5% total gullied or ~20% total
gullied area] or gullied area area gullied area
>15% unit area
Model 3 >5% unit area >5% unit area >5% unit area >5% total ~5% unit area ~-3% unit area
and >5% total or >5% total gullied area
gullied area gullied area
Model 4 >2% unit area >2% unit area ~2% unit area ~2% unit area >2% unit area >2% unit area
Model 5 SO% unit area >O% unit area >O% unit area 20% unit area >O% unit area >O% unit area
(1) Only geomorphic component of the geopedologic map taken into account
Examples: “>I 5% unit area” means that only units in which observed gullres cover >15% of the area of the unit are considered rn the model’s calculations;
“>5% total gullied area” means that only unrts that contain ~-5% of the total gullied area within the study area are considered In the model’s calculatrons.
50
Mapping and modelling mass movements and gullies JAG l Volume 3 - Issue 1 - 2001
ideal model
Type “a” models: Including all thematic maps except the
geomorphic component of the geopedologic map
(see Table 3)
2a
Type “b” models: based on the geomorphic component
of the geopedologic map alone (see Table 3)
Intersection model of 3-l 6% slopes and Alfisols
Intersection model of 5-l 1% slopes and highly susceptible
geoforms
3
20
d la
10 FIGURE 7: Relative efficiency of rule-based models for gully
prediction at Huasca de Ocampo, Mexico [Vazquez-Selem &
Zinck, 19941.
I I I I I I I I I I
0 10 20 30 40 50 60 70 80 90 100
Instead of operating only on the basis of area criteria, FIGURE 9: Mass movement hazard zones in the upper Coello
predictive models can directly use concrete parameters river basin, Colombia [L6pez & Zinck, 19911.
51
Mapping and modelling mass movements and gullies JAG l Volume 3 - Issue 1 - 2001
52
Mapping and modelling mass movements and gullies JAG l Volume 3 - issue 1 - 2001
53