Facing the archaeological looting in Peru by local spatial autocorrelation statistics of Very high resolution satellite imagery

Maria Danese (1), Rosa Lasaponara (2), Nicola Masini (1)

1 CNR-IBAM, C/da S. Loia Zona industriale, 85050, Tito Scalo (PZ), Italy2 CNR-IMAA, C/da S. Loia Zona industriale, 85050, Tito Scalo (PZ), Italy

INDEX

Satellite Remote Sensing for Archaeological research

How face the clandestine excavations by satellite remote sensing (potential and limit)

Concepts of spatial autocorrelation

Improving the detection of archaeological looting by spatial autocorrelation : application in Cahuachi, discussion and conclusions

The increasing development of Earth Observation (EO) techniques (ground, aerial and space) and the tremendous advancement of computer science has determined an increasingly importance of remote sensing

archaeological research

management and preservation of cultural resources and landscape

FOR

protection of archaeological heritage from looting

Very High Resolution (VHR) Satellite imagery

Satellite data Resolutions Panchromatic Multispectral

IKONOS (1999)

Spatial resolutions 1 mt 4 mt

Spectral range 450-900 nm

445-516 nm (blue)

506-595 nm (green)

632-698 nm (red)

757-853 nm (near IR)

QuickBird (2001)

Spatial resolutions 0,61 mt 2,44 mt

Spectral range 450-900 nm

450-520 nm (blue)

520-600 nm (green)

630-690 nm (red)

760-900 nm (near IR)

GeoEye (2008)

Spatial resolutions 0,41 mt 1,65 mt

Spectral range 450-900 nm

450-520 nm (blue)

520-600 nm (green)

625-695 nm (red)

760-900 nm (near IR)

WorldView1(2007)

Spatial resolutions 0,50 mt -

Spectral range 450-900 nm -

WorldView2 (2009)

Spatial resolutions 0,46 mt 1,84 mt

Spectral range 450-780 nm

400 - 450 nm (coastal)

450-520 nm (blue)

520-585 nm (green)

585 - 625 nm (yellow)

630-690 nm (red)

705 - 745 nm (red edge)

760-900 nm (near IR1)

860 - 1040 nm (near IR1)

Satellite data processing : methodological approach

Detection and characterization of buried remains

Panchromatic image

Multispectral imagery

Datafusion

Datafusion products

Edge detection

Edge enhancement: vegetation indices, PCA, TCT, etc.

Edg

e ex

trac

tion

Edge thinning

Reconnaissance and Interpretation

Mapping within GIS environment

Evaluation of data fusion algorithms

Evaluation of edge enhancement techniquesM

etho

dolo

gy

Assessment of Spectral capability

Paleoenvironmental studies

Archaeological landscape

R. Lasaponara, N. Masini, 2007. Detection of archaeological crop marks by using satellite QuickBird, Journal of Archaeological Science, 34: 214-221

Crop marks can appear as differences of height or color in crops which are under stress due to lack of water or deficiencies in other nutrients. Crop-marks can be formed both as negative marks above wall foundations and as positive marks above damp and nutritious soil of buried pits and ditches

Soil marks are traces of archaeological features visible in ploughed or harrowed often for very restricted periods before the crops begin to grow.

Shadow marks can be seen in the presence of micro-topographic relief variations that can be made visible by shadowing in low sunlight angle conditions.

Traces of archaeological features

buried masonry

crop-marks related to buried structures (positive presences)

stressed vegetationhealthy veget. healthy veget. healthy vegetation very healthy veget. healthy veget.

buried ditch

crop-marks related to ditches (negative presences)

sandy (or permeable) soil

clayey soil

buried masonry

soil-marks related to buried structures (positive presences)

buried ditchsandy (or permeable) soil

clayey soil

damper soilless damp soil less damp soil

soil-marks related to ditches (negative presences)

soil-marks and shadows marks related to buried structures

sunbeams

buried masonry

less damp s.damper soil damper s.shaded anddamper s.

Archaeological marks spectral response: crop marks

Panchromatic image

Multispectral imagery Datafusion

Datafusion products

Edge detection

Edge thresholding

Edge thinningEdg

e ex

trac

tion

Line extraction

NDVI

Pancromatic

Red band (R)

NIR (near infrared)NDVI=(NIR-R)/(NIR+R)

PANNIRREDNDVI

Nazca river near Cahuachi (Peru)

Basament of a buried pyramid

Direct surveillance (field survey)

Aerial surveillance

How protect the archaeological heritage from clandestine excavations?

.

time consuming,

expensive

not suitable (for remote archaeological sites,

characterized by difficult accessibility)

Are suitable?

Direct surveillance

Aerial surveillance

not suitable for extensive areas

non practicable in several countries due to military or political restrictions

In such conditions, Very high resolution (VHR) satellite imagery offer a suitable chance to quantify looting and damage affecting the archaeological heritage thanks to their global coverage and frequent revisitation times.

Recent applications:Iraq (Stone, 2008), other countries of Middle East (Parcak, 2007)

Umma, Iraq. 2008 QuickBird image : note extensive looting pits.

Satellite Remote Sensing for monitoring clandestine

archaeological excavations and looting

This suggest to use an approach, based on local spatial autocorrelation statistics

A time series of panchromatic and multispectral satellite images (2002-2008) allowed the mapping of looting over the years.

Looters’ holes : small and circular pits (0.7-3 m diameter) filled with sand, and by scattered remains

The reliability of the detection was evaluated by field surveys :

Rate of success high for flat areas

Unsatisfactory for other areas (mound)

Investigated area

THE GREAT PYRAMID

TEMPLO

MONTICULO

TEMPLO DEL ESCALONADO

Historical phases:(400 B.C. – 400 A.D.)I) Sanctuary II) Ceremonial Center III-IV) Theocratic

Capital V) Sacred Place

Archaeological area: 25 sqkmExcavated area: 15000 sqm (6%)

Adobe constructionsNecropolis intrusive areaContinous site evolution

Proyecto Nasca (Peru):

Ceremonial Centre of CAHUACHI

The comparative visual inspection of the available satellite dataset put in evidence that the panchromatic images are more suitable than pansharpened spectral bands to emphasize both the pitting holes and archaeological features.

2002 2005 2008

Satellite time series used to map looting in Cahuachi

Looters’ holes are usually recognizable by their small and circular pits. Some parts of the holes are illuminated, others are in shade.

Cahuachi study case

we focused only on satellite panchromatic scenes, so it was used as INTENSITY.

Consequently all these characteristics pixels with holes show very different values of reflectance, so we supposed to find a break in autocorrelated zones (soil without holes).

Spatial AutocorrelationTobler's First Law of Geography “All things are related, but nearby things are more related than distant things” (1970)

Positive Autocorrelation

(or attraction)

Negative Autocorrelation

(or repulsion)

No Autocorrelation

(or random)

Events : near and similar (clustered distribution)

between events when, even if they are near, they are not similar (uniform distribution)

no spatial effects, neither about the position of events, neither their properties

called “event” the number of spatial occurrences in the considered variable,

spatial autocorrelation measures the degree of dependency among events,

considering at the same time their similarity and their distance relationships

KDE: intensity and its measures

First order effects(Absolute location)

Second order effects

(Relative location)

ji

ji

dsdsji dsds

dsYdsYEss

ji

))()((lim),(

0,

Properties of a spatial distribution*

*Gatrell et al. (1996)

ds = the neighbourhood each point (s)E() = expected mean

Y(ds) : events number in the neighbourhood

Large scale variation in the mean value of a spatial process (global trend)

Small-scale variation around the gradient or Local dependence of a spatial process (local clustering)

Spatial autocorrelation : the nature of the problem

Quantitative nature of dataset

•understand if events are similar or dissimilar

(define the intensity of the spatial process, how strong a variable happens in the space )

Geometric nature of dataset

• the conceptualization of geometric relationships (..at which distance are events that influence each other (distance band))

Calculation method : Euclidean distance . 2)(2)(),( jyiyjxixjsisEd

Direction considered : or contiguity methods (tower c., bishop c., queen c.)

dis

tan

ce

Definition of spatial event 1

2

3

Spatial Autocorrelation (SA) in the context of image processing

the spatial event is the pixel

spatial autocorrelation statistics are calculated considering geographical coordinates of its centroid

Geometric nature : lag distance

• lag distance : the range over which autocorrelation will be calculated or the separation distance between events

Quantitative nature : spectral reflectance

• Pixel reflectance value for each band•SA measures the degree of dependency among spectral bands

3

2

1

Global indicators of autocorrelation just measure if and how much the dataset is autocorrelated.

Global indicators of Autocorrelation

Moran’s index

i j i iij

i j jiij

XXw

XXXXwNI

2)()(

))((

where, N is the total pixel number, Xi and Xj are intensity in i and j points (with i≠j), Xi is the average value, wij is an element of the weight matrix

I Є [-1; 1] if I Є[-1; 0) there’s negative autocorrelation; if I Є (0; 1] there’s positive autocorrelation; if I converges to o there’s null autocorrelation.

Geary’s C

where symbols have the same meaning than the Moran’s index expression

C [0; 2]; if C [0; 1) there’s positive autocorrelation; if C(0; 2] there’s negative autocorrelation; if C converges to 1 there’s null autocorrelation

i iij

i j jiij

XXw

XXwNC

2

2

)((2

)()1( (Geary, 1954),

(Moran, 1948)

LISA allow us to understand where clustered pixels are, by measuring how much are homogeneous features inside the fixed neighbourhood

Local Indicators of Spatial Autocorrelation (LISA)

Local Moran’s index

high value of the Local Moran’s index means positive correlation both for high values both for low values of intensity (reflectance value)

N

jjij

X

ii XXw

S

XXI

12

))(()(

(Anselin, 1995),

Local Geary’s C index

Detection of areas of dissimilarity of events (pixel reflectance value)

n

i

n

jij

n

iji

n

jij

n

ii w

XXw

XX

nC

1 1

1

2

1

1

2 2

)(

)(

1

(Cliff & Ord, 1981)

Getis and Ord’s Gi index

high value of the index means positive correlation for high values of intensity, while low value of the index means positive correlation

for low values of intensity (Getis and Ord, 1992; Illian et al., 2008)

2

)()(1

)(

)()()(

2

1 1

11

N

dwdwN

iS

dwxxdwdG

n

i

n

iii

n

iiii

n

ii

i

▪ N is the events number▪ Xi ed Xj are the intensity values in the point i and j (with i≠j)▪ is the intensity mean

▪ wij is an element of the weights matrix

X

Compute the lag distance

Assume the rule of contiguyity

Calculation of Local indices

the queen’s contiguity was choosen, because the analysis should be done in all the directions also for the curve configuration of holes.

The best value is the lag that maximizes Moran’I (fig.1) and minimizes C (fig.2), allowing to captures in the best way the autocorrelation of the image.

The lag choosen for all the three years is 2.

Fig. 2. Results obtained with global Geary’s C and lag distance between 1 and 10 calculated for 2002 Quickbird image.

Fig. 1. Results obtained with global Moran’s I and lag distance between 1 and 10 calculated for 2002 Quickbird image.

Lag distance and the rule of contiguity

RGB

Panchromatic image(PAN) Zoom of PAN the Geary’s C Getis and Ord’s Gi Moran’s I

Summary of analysis procedure

1. Once lag distance is found and 2. Assumed the queen’s contiguity. 3. Local indicators of spatial association were

calculated

Results :

Geary index (d), allows to best represent the rough surface, so the pitting holes due to its capability to detect dissimilarity Getis and Ord Gi (e) needs a classification, before to be interpretated

Calculation of LISA

Getis & Ord’s Gi

Clusters that show the best results are those characterized by low reflectance intensity & corresponding low Gi values or high reflectance intensity & corresponding high Gi values show positive spatial autocorrelation These clusters were then converted to polygons with the aim to obtain the map of the looting phenomenon

n

inax ImIm n

GG minmax

in Cahuachi corresponding values were found considering equal intervals as follow

where I is the intensity, G is the index and n is the number of classes wanted in the classification

Geary’s C representation

and Getis & Ord’s Gi (classification based on) product

Ground truth (field survey in progress)

survey of hole pits Identification of hole pits

Computation of : i) rate of success (75-90% in the

considered test areas), ii) false alarms; iii) rate of unsuccess

In Cahuachi, the detection of looting pits on mounds has been significantly improved (75-90%) by applying local spatial autocorrelation statistics.

Such improvement is still more evident if we compare the panchromatic satellite time series with the correspondent time series processed by local spatial autocorrelation statistics

Cluster linked to looting pits

False alarm

Looting pits not detected by means of local spatial autocorrelation

2002 2005 2008

2002 2005 2008

RGB composition of LISA (R:Geary; G: Moran; B: Getis) applied to panchromatic images of 2002 QB (a), 2005 QB (b) and 2008 WW1 (c). RGB composition emphasize pits enhancing their edges (circled with magenta ).

The multitemporal comparison of the three RGB images clearly shows an increasing number of pits from 2002 to 2008 and, therefore, the intensification of the looting phenomenon over the years.

Panchromatic time series (2002;2005;2008)

The improvement obtained by LISA application is still more evident if we compare the panchromatic satellite time series with the correspondent time series processed by local spatial autocorrelation statistics

Clandestine excavations is one of the biggest man-made risks which affect the archaeological heritage, especially in some countries of Southern America, Asia and Middle East.

To contrast and limit this phenomenon a systematic monitoring is required. In this context, VHR satellite imagery can play a fundamental role to identify and map looted areas.

The Cahuachi study case herein presented put in evidence the limits of VHR satellite imagery in detecting features linked to looting activity.

This suggested to experience local spatial autocorrelation statistics which allowed us to improve the reliability of satellite in mapping looted area.

In Cahuachi, the detection of looting pits on mounds has been significantly improved (75-90%) by applying local spatial autocorrelation statistics.

CONCLUSIONS