Handbook of Ecological Indicators for Assessment of Ecosystem Health - Chapter 13 pptx

Different resilience levelsare expected to be intertwined with different scale domains of real habitats inrelation to the type and intensity of natural and human disturbances from... In

CHAPTER 13

Multi-Scale Resilience Estimates for Health Assessment of Real Habitats

in a Landscape

G Zurlini, N Zaccarelli, and I Petrosillo

Vegetation or habitat types are ecological phases that can assume multiplestates Transformations from one type of phase to another are called ecologicalphase transitions If an ecological phase maintains its condition of normality inthe linked processes and functions that constitute ecosystems then it is believed

to be healthy An adaptive cycle, such as that given in Holling’s model, hasbeen proposed as a fundamental unit for understanding complex systems Suchmodel alternates between long periods of aggregation and transformation ofresources and shorter periods that create opportunities for innovation Thelikelihood of shifts among different domains largely depends on domainresilience, measurable by the size of scale domains, but these do not provideany indication on resistance — the external pressure to displace a system by agiven amount We argue that the type, magnitude, length, and timing ofexternal pressure, its predictability, the exposure of the habitat, and thehabitat’s inherent resistance, have important interactive relationships whichdetermine resilience, and in turn, ecosystem health Different resilience levelsare expected to be intertwined with different scale domains of real habitats inrelation to the type and intensity of natural and human disturbances from

management activities and land manipulation In this paper, we provide

an operational framework to derive operational indices of short-term spective resilience of real grasslands in a northern Italy watershed, from multi-scale analysis of landscape patterns, to find scale domains for habitat edgeswhere change is most likely — that is, where resilience is lowest and fragilityhighest This is achieved through cross-scale algorithms such as fractalanalysis coupled with change detection of ecological response indices Theframework implements the integration of habitat-edge fractal geometry, thefitting of empirical power functions by piecewise regressions, and change-detection procedures as a method to find scale domains for grassland habitatedges where change is most likely and consequently resilience is lowest.Changes due to external pressure significantly related to habitat scale domains,according to their scaling properties resulting from the interaction betweenecological, physical, and social controls shaping the systems Grassland scaledomains provided evidence and support for identifying and explaining scale-invariant ecological processes at various scales, from which much insight could

retro-be gained for characterizing grassland adaptive cycles and capabilities to resistdisturbances to facilitate ecosystem health assessment

13.1 INTRODUCTION

The rapid progress made in the conceptual, technical, and organizationalrequirements for generating synoptic multi-scale views and explanations of theEarth’s surface, and for linking remote sensing at multi-resolution levels fromsatellite and airborne imageries, geographical information systems, spatialanalysis of landscape patterns, and habitat classification methods, provides anoutstanding potential support to:

1 Identify real landscape patches as habitats and land use types

2 Detect ecological processes by remotely sensed response variables

3 Relate response variables to habitats, by observing at different timesecological changes in habitat pattern as well as in the scales of habitatpattern (Simmons et al., 1992)

Multi-scale studies are increasingly conducted (Wu and Qi, 2000), whichgive emphasis to the identification of scale domains (Li, 2000; Brown et al.,2002), that are self-similarity regions of the scale spectrum over which, for aparticular phenomenon, patterns do not change or change monotonically withscale Thresholds are, in general, difficult to delineate across scales, because itremains difficult to detect multiple scales of variability in ecological data and torelate these scales to the processes generating the patterns (Levin, 1992;Ward and Salz, 1994) The likelihood of sharp shifts is linked to an ecosystem’sresilience, which is the capacity of a system to undergo disturbance andmaintain its functions and controls (Gunderson and Holling, 2002) Theimportance of a clear and measurable definition of resilience has becomeparamount (Carpenter et al., 2001) for evaluating the health of an ecosystem,

defined as being stable and sustainable, maintaining its organization andautonomy over time, and its resilience to stress (Costanza, 1992; Mageau et al.,1995) Scaling domains of habitats can be identified, for instance, by shifts inthe fractal dimension of patch edges, and can indicate a substantial change inprocesses generating and maintaining landscape patches at different scales(Krummel et al., 1987; Sugihara and May, 1990; Milne, 1991), so that differentprocesses dominate at different scales (Peterson, 2000) One way to appreciatethe interaction between pattern and processes is to look at temporal changesdetected by remote sensing, and whether they are significantly associatedwith different scale domains If such processes change in type and intensityacross scales, the ability of ecosystems to resist lasting change caused bydisturbances — their resilience (Gunderson et al., 1997; Gunderson andHolling, 2002) — will change accordingly, so that habitat resilience and scalingare expected to be intertwined (Peterson, 2000).

This study was designed to address some specific questions:

1 Can we objectively and accurately identify scale breaks delimiting ically equivalent scales in real habitat patches in a landscape?

ecolog-2 Are temporal changes detected significantly related to habitat scaledomains, providing evidence on the types of biophysical and socialcontrols shaping the systems?

3 If so, can we derive an operational index of short-term retrospectiveresilience, through cross-scale algorithms like fractal analysis coupledwith remotely sensed change detection, to find scale domains wherechange is most likely — that is, where resilience is lowest?

To make this approach practical, we need:

1 A really effective classification procedure for habitat recognition fromgeneral and vague categories of habitats to more specific categories

2 A statistically objective procedure for identifying shifts in scale domains

3 Suitable ecological response variables for change detection

We describe an operational framework for the accurate identification ofself-similar domains in few specific grassland habitats, and for estimating theirdisplayed short-term resilience First, we looked for scale domains in realgrassland patches of a stream watershed in northern Italy (Zurlini et al., 2001),resulting from long-term natural and man-induced interactive disturbanceregimes We then quantified short-term intensity changes of habitat scaledomains, based on remotely-sensed ecological response indices This frame-work implements the integration of edge fractal analysis, the fitting of powerlaws by piecewise regressions and hypothesis testing of scale shifts (Grossi etal., 1999; 2001), together with procedures of change detection, as a method tofind scale domains for grassland edges where change is most likely Togetherthey represent a framework for spatially defining critical landscape thresholdsand scale domains, habitat adaptive cycles (Gunderson and Holling, 2002), andhabitat resilience by which scale-dependent ecological models could bedeveloped and applied By introducing this approach, we address some basicconcepts of ecological phases and multiple states, with a general discussion on

self-similarity regions, fractal analysis, and on the statistical procedures for theobjective identification of shifts among scale domains, as well as on resilienceand its practical measure The detailed and often complex composition of reallandscape habitat mosaics in terms of habitat types and land use has beenrarely considered in the understanding of the relationships between landscapepattern and process response variables Much of the insight obtained is related

to the coupling of change-detection procedures with the availability of detailedhabitat type distribution in a stream watershed The potential of such approachfor ecosystem health assessment, planning, and management of habitatsmosaics is also discussed

13.2 RATIONALE

13.2.1 Ecological Phases, States, and Scale Domains

From several long-term observations, experimentations, and comparativestudies of many sites, it is now evident that alternate and alternative states arise

in a wide variety of ecosystems, such as lakes, marine fisheries, benthic systems,wetlands, forests, savannas, and rangelands (Gunderson and Holling, 2002) Aphase state of a system at a particular instant in time is the collection of values

of the state variables at that time (Grimm et al., 1992; Walker et al., 2002),different from other states the system can visit over and over again (alternate),

or from those typical of other systems (alternative) Vegetation or habitat typesare considered ecological phases which can assume multiple states, andtransformations from one type to another (alternative) correspond to ecolog-ical phase transitions, which change the integral structures of the systems (Li,2002) Multiple states (alternate) can be assumed by ecological phases withoutlosing their basic identity For example, a forest stand or grassland may remain

a forest patch or grassland over and over again, each with its own dynamicstates of rapid growth, conservation, collapse, and reorganization asproposed by Holling’s adaptive cycle model (Holling et al., 1995) Forgrasslands, such model proposes that as young grasses grow without grazing

or cutting, they gradually become denser and accumulate fuel, and thusbecome increasingly susceptible to fire After a fire, the system is reorganized

as vegetation resprouts from roots or seeds, producing new grassland Allthese states are deemed as multiple configuration states (attractors) of thesame ecological phase Another example is provided by a simple meta-model describing different common ecological states of coral reefs, andthe factors that may cause or maintain these states (McClanahan et al.,2002) Individual coral reefs that are exposed to a combination of humanand natural influences may be a mosaic of several states Which state theecosystem currently assumes is function of its history and of the driving forcesoperating

Multi-scale analysis corresponds to the detection of self-similar scaledomains of alternate states, a central point for the development of a

scalar theory in ecology (Levin, 1992; Holling, 1992; Wiens, 1995) Suchself-similarity or fractality implies a particular kind of structural composition

or dynamic behavior — that is, the fundamental features of the system exhibit

an invariant, hierarchical organization that holds over a wide range of spatialscales (Gell-Mann, 1994; Li, 2000) A spatial ecological phase transition, orecotone, is a ‘‘zone of transition between adjacent ecological systems, having

a set of characteristics uniquely defined by space and time scales and by thestrength of the interactions between adjacent ecological systems,’’ (di Castri

et al., 1988) Therefore the nature of a habitat’s edge is not just a property of

a specific habitat, but is the outcome of interactions at the landscape level.Ecological phases like vegetation or habitat types are dynamic in space andtime, each trying to expand and invade adjacent ones whenever environmentaland management conditions are beneficial to one of the adjacent ecosystems(Risser, 1995) They can have different regions of the scale spectrum overwhich there are several possible ecological states, equivalent or self-similar for

a particular phenomenon, and which do not change or change monotonicallywith changes in scale This would allow drawing the same ecologicalconclusions statistically from any scale (Sugihara and May, 1990; Milne,1991; Li, 2000)

13.2.2 Resilience and Resistance

Abrupt shifts among several very different (alternative) stable domains areplausible in local and regional ecosystems more susceptible to changes; thelikelihood of such shifts depends on resilience and resistance (see chapter 2),whereas the costs of such shifts depend on the degree of and duration forreversibility from one domain to another (Gunderson and Holling, 2002) Twosystems, or two states of the same system, may have the same resilience butdiffer in their resistance We can surmise that if the same external pressure isapplied to two systems with different intrinsic resistances, they will show adifferent ability, or resilience, to resist lasting change caused by disturbances.Resilience estimates differ from ecological indicators in that they refer to socio-ecological systems and ecosystem services (Costanza et al., 1997) they provide(Carpenter et al., 2001)

Most studies in the literature refer to theoretical approaches, usingresilience as a metaphor or a theoretical construct (Carpenter et al., 2001).Where resilience has been defined operationally, this has occurred in a fewcases within a mathematical model of a particular system (Carpenter andCottingham, 1997; Peterson et al., 1998; Janssen et al., 2000; Casagrandiand Rinaldi, 2002) In this context, bifurcation analysis of simple dynamicmodels has been often suggested or adopted, together with the size of stabilitydomains, or the magnitude of disturbance the system can tolerate and stillpersist before the system changes its structure by changing the variablesand processes that control behavior (Peterson et al., 1998; Gunderson andHolling, 2002)

However, not all those definitions, even though measurable in models, areoperationally measurable in the field In an operational sense, resilience needs

to be considered in a specific context As discussed by Carpenter et al (2001), itrequires defining the resilience of what to what One important distinction,along with those on space–time scales advanced by Carpenter et al (2001), iswhether resilience has to be measured prospectively — to predict the ability

of ecosystems to resist lasting change caused by disturbances, or tively — to evaluate such ability as observed by past exposure to externalpressures

retrospec-13.3 STUDY AREA AND METHODS

13.3.1 The Baganza Stream Watershed

The Baganza watershed was selected as pilot study area of the Map ofItalian Nature (MIN) program (Zurlini et al., 1999; 2001), since it is a goodrepresentative of the typical series of watersheds located along the same side

of the northern Apennines ridge The watershed is approximately 174.63 km2,and is located on the Emilian side of the northern Apennines (Figure 13.1),with a main stream 57 km long and a progressive elevation gradient in thesouthwest direction which varies from 57 m in the flat to the piedmont, up to

1943 m at the highest peak in the Apennines mountains Mean monthlytemperature varies from
0.6 to
17.1C in the mountain range, and fromþ1.5 to þ24.7C in the lowland Mean rainfall varies from 40 to 95 mm peryear with most of the rain occurring during the fall and the spring seasons,with no deficit of evapo-transpiration during summer Snow is usually presentfor four months above 1,400 m

In the past few centuries, due to human influence on Mediterraneanecosystems and the slow abandonment of agricultural and pastoral practices,plant communities have been shaped into a mosaic-like pattern composed ofdifferent man-induced degradation and regeneration stages (Naveh andLiebermann, 1994) In the past, this watershed was almost fully covered byancient forests, still present during the ducat of Parma at the end of theeighteenth century Around the end of the nineteenth century, much of theforests in the piedmont and hills were cleared for building the many miles ofthe national railway network Many cleared areas were maintained as grass-lands with pastoral practices with sheep and cattle breeding on natural orcultivated pastures In the last century, cattle breeding on pastures prevaileddue to the increasing market success of diary products

Intensive agricultural land use is currently prevalent in the lowlands and thenearby Baganza stream, whereas abandonment of agricultural and pastoralpractices in the hills and mountains is still in progress Conservation andendangered species legislation at the national and regional level reduce thepossibility of clearing the land, whereas they are allowed to maintain pastures

in the high-hill and mountain range

13.3.2 Corine Habitats

Using synoptic multi-scale views and classifications of the Earth’s surfacenow available, researchers, land managers, and land-use planners can quantita-tively place landscape units, from general and vague categories such as

‘‘forests’’ to more specific categories such as ‘‘Illyrian Holm-oak woodland,Orno-Quercetum Hilicis dominated formations,’’ in their large-area contexts.Remote sensing technologies represent the primary data source for habitat

Figure 13.1 Location of the Val Baganza watershed and distribution of large habitat

classes (modified from Zurlini et al., 2001) F is the flat, with urban/agricultural matrix; P is the piedmont, with agricultural/grassland/woods matrix; and A is the Apennines mountain range, with grassland/forest matrix The list of main habitats corine habitat is given in Table 13.1.

identification and landscape analysis, but often suffer from the ModifiableAreal Unit Problem (MAUP, Openshaw, 1984), that is a potential source oferror that can affect spatial studies which utilise aggregate data sources Itstates that a number of different and often arbitrary ways exist by which anarea can be divided or aggregated into nonoverlapping areal units.

We used the CORINE habitat classification (EU/DG XI, 1991) to identifyecosystems as patches (Tansley, 1935) for generating digital thematic maps asGeographic Information System GIS coverages of mosaics of contiguouspatches To avoid MAUP effects, the final delineation of habitat mosaics wasperformed by an iterative process based on integrated evidence from processedsatellite imagery, aerial photos, hyperspectral imagery, existing vegetation andgeological soil maps, digital elevation models (DEM), and field reconnaissance(Zurlini et al., 1999) The detailed CORINE habitat distribution for theBaganza watershed was available at a scale of 1:25,000 (Zurlini et al., 2001), in

a revised and more detailed form with respect to the original habitatclassification used in Grossi et al (1999; 2001), with 2,327 irregular patchesbelonging to 69 different CORINE habitat and habitat mosaic types(Table 13.1)

The flat and piedmont sections are dominated by agricultural fields, withfew relatively natural habitats, represented by typical wet woodlands(Figure 13.1) Hop-horn beam (CORINE code 41.812) mixed to Quercuspubescens(41.7314) woods, are dominant in the hills, while neutrophile beechforests (41.1744) are most frequent in the mountain range above 900 to 1000 m.Three of the most frequent grassland habitats in the watershed were consideredfor subsequent analyses (EU/DG XI 1991; Sburlino et al., 1993):

1 Lowland hay meadows (CORINE code 38.2) present with 378 patches

2 Northern Apennine Mesobromion (CORINE code 34.3266) with 77patches

3 Brachypodium grassland (CORINE code 36.334) with 131 patches,corresponding roughly to increasing elevation gradients and to decreasinghuman influence and control (Figure 13.2)

So-called lowland hay meadows are rich mesophile grasslands in thelowland, hills and submountain ranges, regularly manured, and whennecessary irrigated, well-drained under direct human control, with speciessuch as Arrhenaterum elatius, Trisetum flavescens, and Anthriscus sylvestris.They often begin from seeding of leguminous grasses or mixed fodder, andafter are regularly cut in time for cattle breeding in farms Northern ApennineMesobromionare poor closed mesophile grasslands, sparse and rich in Bromuserectus and orchids, in local semiarid environments naturally exposed todrought and limited by the amount of organic matter in soil; they are not underdirect human disturbances, apart from infrequent cutting, and grazing andmanuring by cattle (which is an important source of organic matter) Whenlowland hay meadows are abandoned, they become Mesobromion grasslands.Brachypodium grasslands are subalpine thermophile siliceous habitats, oftenfound on skeleton soils, and are not under direct human influence apart from

sporadic grazing by cows and sheep at lower altitudes, with carpet communitieshardly browsed by cattle, and almost pure in Brachypodium genuense, typical ofhigher elevations and of the summits Fire is not currently used as a practice forcontrolling scrub formation and seldom occurs in the watershed.

13.3.3 Empirical Patterns of Self-Similarity

Domains are delimited by relatively sharp transitions or critical pointsalong the spatial scale continuum where a shift in the relative importance ofvariables influencing a process occurs (Meentemeyer, 1989; Wiens, 1989) Toidentify scales or hierarchical levels of landscape structures, some generalstatistical and spatial analysis methods, inherently multi-scaled, are availablesuch as semi-variance analysis (Burrough, 1995; Meisel and Turner, 1998;

Table 13.1 List of the main CORINE habitat type identified in the Baganza watershed (modified from Zurlini et al., 2001)

CORINE code CORINE habitat type

42.1B1 Abies alba reforestations

41.812 Supra-mediterranean hop-hornbeam woods

41.813 Montane hop-hornbeam woods

41.74 Quercus cerris woods

41.1744 Beech forests

42.67 Black pine reforestation

44.614 Italian poplar galleries

83.324 Locust tree plantations

41.731 Semi-xerophile Quercus pubescens woods

41.7312 Xerophile Quercus pubescens woods

44.122 Mediterranean purple willow scrub

31.431 Juniperion nanae scrub

31.81 Medio-European rich-soils thickets

31.811 Blackthorn-bramble scrub

31.88 Common Juniper scrub

32.A Spanish-broom fields

34.3266 Northern Apennine Mesobromion

61.3124 Submontane calcareous screes with Calamagrostis varia

61.3125 Sedo-Scleranthetea Submontane calcareous screes

61.3126 Brometalia erecti submontane calcareous screes

62.213 Hercynian serpentine cliffs

87.24 Ruderal communities with Tussilago farfara

87.23 Ruderal communities with Melilotus albus

87.29 Ruderal communities with Agropyron repens

82.11 Field crops

62.4 Bare inland cliffs

82.11 Plough field crops

86.2 Villages

86.3 Active industrial sites

86.41 Quarries

Bellehumeur and Legendre, 1998), multi-variate analysis of spatial lations (Burrough, 1983; Ver Hoef and Glen-Lewis, 1989), spectral analysis(Platt and Denman, 1975), wavelet analysis (Bradshaw and Spies, 1992),lacunarity analysis (Plotnick et al., 1993), scale variance (Wu et al., 2000),fractal analysis (Krummel et al., 1987), and fractal dimension combined withvariograms (He et al., 1994).

autocorre-Fractal analysis is a very useful tool for identifying hierarchical size scales

of patches in nature, such as how to define boundaries between hierarchicallevels and how to determine scaling rules for extrapolating within each leveldomain (Sugihara and May, 1990; Milne, 1991; Li, 2000) When natural

‘‘objects’’ like vegetation are not constrained by human activities and landmanipulation, or by natural obstacles, they result in highly irregular shapesdetermined by iterative and diffusive growth, which can reproduce at differentscales independently of size In theory, a perfect fractal is self-similar at allscales, and it could be scaled up and down to infinity Because of these limits toself-similarity, it is preferable to refer to these systems as fractal-like (Brown

et al., 2002) Shifts in fractal dimension of irregular patch edges have been used

to find substantial changes of spatial patterns at different scales (Krummel

et al., 1987; Grossi et al., 1999; 2001) Krummel et al (1987) were the first

to develop a method for detecting different scaling regions in a landscape for

a population of forest patches, based on perimeter-area relationships Grossi

et al (1999) conceived a general statistical procedure to detect objectivelythe change points between different scaling domains in real patch populations,based on the selection of the best piecewise regression model using a set ofstatistical tests

Given its significance within the framework of this paper, it seems worthproviding a few details Two distinct basic models were hypothesized to fitthe data: continuous piecewise linear models and discontinuous piecewiselinear models To estimate the fractal dimension D of each scale domain, we

Figure 13.2 Distributions of: (A) Mesobromiom grasslands (CORINE code 34.3266);

(B) Brachypodium grasslands (CORINE code 36.334); and (C) lowland hay meadows (CORINE code 38.2) in the Baganza watershed.

used perimeter-area relationships as suggested by Lovejoy (1982) Given areasand perimeters of n patches, we can write the relationship as follows:

We considered five alternative models If
is the breakpoint of models withone breakpoint, and
1,
2, and (
1<
2) are the first and second in models withtwo breakpoints:

where I is an indicator variable equal to one when the subscripted condition

is true and equal to 0 otherwise; 01, 001and 0001 are slopes — that is, halfthe fractal dimensions of the first, second, and third domain, respectively;

0is an intercept and " the error term

The simple continuous model given by 13.1 is called C0, whereas thediscontinuous model (13.2), called D0, is a piecewise regression (Draper andSmith, 1998) with two parallel discontinuous segments and no change of slope,thereby with one single fractal domain Models 13.3 and 13.5, called C1and C2,

have two and three continuous segments, and one and two breakpoints,respectively Models 13.4 and 13.6, called D1and D2, are models with two andthree discontinuous segments, respectively, and so on.

More generally, let Cr, with r ¼ 0, 1, , be the continuous piecewise linearmodel with r breakpoints and (r þ 1) fractal domains, in Cr the number ofparameters to be estimated is 2(r þ 1) with one intercept, r breakpoints, and(r þ 1) slopes Let Dr, with r ¼ 0, 1, 2, , be the discontinuous piecewise linearmodel with r breakpoints, in Dr the number of parameters to be estimated isthree (two intercepts and one slope) when r ¼ 0, and 3r þ 2 — one slope andone intercept for each of r þ 1 domains and r breakpoints — when r  1.Therefore we can depict a nested collection of models (Figure 13.3)

Which of the nested models is the best is a typical problem of variableselection that, in multiple linear regressions, is usually based on the F test tomeasure the statistical significance of adding variables If ! and
are twonested regression models having the same 2, with p and p þ q regressionparameters, respectively, the null hypothesis H0:! vs the alternative hypothesis

HA:
can be tested using the following LR test statistic:

l ¼ n ln SSE!^

SSE
^

where SSE! ^and SSE
^ are the residual sum of squares of ! and
, respectively

So, the rejection region can be expressed equivalently as:

Figure 13.3 Nested collection of continuous and discontinuous piecewise linear models for

hypothesis testing The number of regression parameters to be estimate is between brackets (modified from Grossi et al., 1999).

(in 13.8) Breakpoints in 13.2 to 13.6 were unknown parameters to be estimatedlike other regression parameters, and corresponding regression models resultswere not linear, so that in this case, the F distribution did not necessarily apply

to variable selection procedures The problem was studied using maximumlikelihood and likelihood ratio (LR) tests, and simulations were conducted inorder to check whether w2and/or F(q, n
p
q) were good approximations tothe sampling distributions of l and F For this purpose, let YiNði, 2Þ,

i ¼1, 2, , n, be the dependent variable of a linear regression model whereerrors are Gaussian with i¼ðXi, Þ, where  ¼ ð1, 2, ,pÞ0 is a vector

of unknown parameters that can vary independently of the variance 2 Themaximum likelihood estimate of  minimizes the residual sum of squares Wegenerated data from ! using habitat area as a regressor, and throughopportune transformations of the dependent variable Yinot affecting the nulldistributions (Grossi et al., 1999) Then the test statistics l and F could becomputed using SSE!^ and SSE
^ from generated data, with 6000 replicationsfor each alternative model
, when the null model is C0, and 5000 otherwise

To select the best piecewise model for each habitat type, we comparedhierarchically nested models (Figure 13.3) by computing the corresponding

LR statistic Some null models are possible:

1 With null model ! ¼ C0, the alternative model
might be any of the morecomplex models C1, C2, D0, D1and D2

2 With ! ¼ C1, the alternative model
might be any of C2, D0, D1and D2

3 With ! ¼ D1, the alternative model
could be only D2

4 With ! ¼ C2, the alternative model
could be only D2 Both F and l hadempirical distributions which could be approximated by the nominal Fand w2distributions, respectively; therefore, we limited the analysis to the

LR statistic

13.3.4 Change Intensity Detection

Change detection is the comparison of the measurements computed fromtwo co-registered remote sensing images of the same scene, by determining aquantity corresponding to the difference (or similarity) between two differenttimes at the same location A general equation for this metric may appear asfollows (Skifstad and Jain, 1989):

where D(x, y) is the difference metric, and f1is the metric computed at location(x, y) in image i, where iis a time index, and ’ denotes a linear or nonlinearoperation, which is often the absolute difference value

In this paper we refer to D(x, y) as the standardized change intensity image

if ’ denotes the standardized difference:

ð13:9Þ

where m is the mean of the differences, s1is the variance of the metric f1which

is the Normalized Difference Vegetation Index (NDVI), (Rouse et al., 1974;Kerr and Osrtovsky, 2003), calculated as (band 4
bond 3)/(band 4 þ band 3)for both images and cov12is the covariance

In order to capture mainly man-induced ecological changes, we used twofive-year different dates of Landsat Thematic Mapper (TM) images of thestudy area: August 11 1990, and July 24 1995 Reflectances were used, since arethe most correlated with ground data (Goward et al., 1991) NDVI exploitsspectral responses in the red band and in the near infrared band channels and it

is derived as the ratio of infrared minus red over infrared plus red values It isstrongly related to variables of most ecological interest such as the fraction ofphotosynthetically active radiation intercepted by vegetation (Fipar), leaf totalnitrogen content, leaf area index (LAI), and, in general, to vegetative processes

at ecosystem level (Law and Waring, 1994; Matson et al., 1994); it is alsospecies specific, and reveals health and stress conditions of vegetation cover(Guyot, 1989)

Satellite images of 1995 were almost contemporary with reconnaissanceactivities in the field Standardization was done to account for differencesbetween dates due to climatic changes and other sources of added noise, notaccounted for by prior geometric and atmospheric corrections D(x, y) inModel 13.9 was the spatially explicit response variable used for detecting eitherpositive (gains) or negative (losses) in habitat scale domains; D(x, y) wasco-registered with the raster map of CORINE habitats to assign responsevariable values to each pixel of a particular habitat patch One standarddeviation is typically used as threshold for change detection (Fung andLeDrew, 1988), however, empirical distributions of D(x, y) are not normal,but rather leptokurtic and skewed We identified absolute change intensities

 as the medians of empirical 0.20, 0.10, and 0.05 percentiles, at both tails

of the D(x, y) distribution, to be sure that real changes occurring in thewatershed were dealt with, avoiding background noise

Percentiles less than 5% were not considered in order to avoid a few localextreme values Changes detected through remote sensing techniques are realeffects of ecological significance observed in five-year intervals due to extrinsicpressure, mostly given by human activities, which could have varied spectralresponses affecting the NDVI metric