Sei sulla pagina 1di 11

Metode Kriging

Frankstein Arphan 3715100011


Shabila Gadis Halida 3715100054

Fakultas Teknik Sipil Lingkungan dan Kebumian


Institut Teknologi Sepuluh Nopember
Penerapan Kriging dalam Geofisika

Data geofisika sering digunakan dalam eksplorasi mineral untuk menggambarkan geologi suatu
daerah. Dikarena data geofisika berupa sampel, metode interpolasi digunakan untuk meng grid
kumpulan data individu. Mengingat besarnya biaya akuisisi data geofisika, teknik ini dapat membuat
pengumpulan data geofisika lebih efisien dan hemat biaya.

Contoh kasus :

Application of Kriging Technique to Seismic Intensity Data

Evaluation of seismic intensity is based on the collection of effects that an earthquake can induce on an
environment, including both temporary and permanent changes. The presence of high data variability
and lack of instrumental measures may lead to the conclusion that the results are of limited precision.
Nonetheless, there is no doubt that in some situations macroseismic information is irreplaceable, in
particular, for historical earthquakes that occurred before the instrumental era. It is also important for
very recent seismic events where large intensity datasets must be analyzed and interpreted. To render
this kind of analysis as reliable as possible, a suitable statistical approach is the methodologically correct
way to deal with the information (De Rubeis et al., 1992a; Tosi et al., 1995; Pettenati et al., 1999, and
references therein). Macroseismic data analysis aims at resolving the following major issues: 1. How
would the intensity pattern of the whole area affected by the earthquake be if it were completely
covered by urban structures in such a way that it was possible to define an intensity degree at every
location? 2. Is it possible to recognize and separate error components from true intensity values?
Although there is no definitive answer to the first point, it implicitly states that available intensity data
are a sample subset of a spatially continuous set that could potentially exist. This justifies any attempt to
interpolate an entire macroseismic field. With regard to the second point, a macroseismic field is the
result of the superposition of diverse processes; the role of the seismic source is fundamental, followed
by the geologic-tectonic setting and by the behavior of urban structures under seismic load. Different
spatial ranges of influence permit the separation of these components, from a regional to a very local
scale. Intuitively, it seems correct to attribute the overall shape of the field to the seismic-source
geometry and then, as the spatial scale decreases, the effects of local components become more
evident. It can be asserted that the main criterion to distinguish a very local pattern is the lack of spatial
correlation to field elements at larger distances. This classification of components based on their spatial
range of influence is very important, because pure error is also included; by definition, pure error is
totally uncorrelated in space. Quantification of spatial relationships is thus necessary and it influences
the choice of a proper analytical method. For these reasons we propose the application of kriging with
its main spatial-analysis tool, the semivariogram. Kriging is a statistical technique permitting to estimate
unknown values at specific points in space by using data values from known locations. The intrinsic
characteristics of this method, developed during the 1960s and 1970s, have long been acknowledged as
a good spatial interpolator (Matheron, 1963; Isaaks and Srivastava, 1989; Davis, 2002). Kriging is widely
used within the field of earth sciences, including mining, geochemistry, remote sensing, and, in general,
in any situation where spatial data must be interpolated. Among the most important features of this
method are (1) the unbiased estimate of results, (2) the minimum Application of Kriging Technique to
Seismic Intensity Data 541 Figure 1. Typical behavior of experimental semivariogram (e.s.) for data from
a regionalized variable; c(l) has low values for small distances l, it increases more or less regularly until
reaching a stable level after a specific value of lsill. This limit separates data having spatial
autocorrelation from data that are spatially independent. estimation error, and (3) uncertainty
evaluation of interpolated data points. The main assumption, when using kriging, is that data analyzed
are samples of a regionalized variable. The properties of this kind of variable lie in the range between
true spatial randomness and fully deterministic behavior. Determinism is not complete and depends on
the spatial distance between points; the closer they are, the stronger is their relation. At greater
distances determinism is lost, and spatial autocorrelation of data vanishes. The semivariogram is the
analytical tool used to evaluate and quantify the degree of spatial autocorrelation; its results constitute
the basis of the kriging interpolation.

Evaluation of seismic intensity is based on the collection of effects that an earthquake can
induce on an environment, including both temporary and permanent changes. The presence of high
data variability and lack of instrumental measures may lead to the conclusion that the results are of
limited precision. Nonetheless, there is no doubt that in some situations macroseismic information is
irreplaceable, in particular, for historical earthquakes that occurred before the instrumental era. It is
also important for very recent seismic events where large intensity datasets must be analyzed and
interpreted. To render this kind of analysis as reliable as possible, a suitable statistical approach is the
methodologically correct way to deal with the information (De Rubeis et al., 1992a; Tosi et al., 1995;
Pettenati et al., 1999, and references therein). Macroseismic data analysis aims at resolving the
following major issues: 1. How would the intensity pattern of the whole area affected by the earthquake
be if it were completely covered by urban structures in such a way that it was possible to define an
intensity degree at every location? 2. Is it possible to recognize and separate error components from
true intensity values? Although there is no definitive answer to the first point, it implicitly states that
available intensity data are a sample subset of a spatially continuous set that could potentially exist. This
justifies any attempt to interpolate an entire macroseismic field. With regard to the second point, a
macroseismic field is the result of the superposition of diverse processes; the role of the seismic source
is fundamental, followed by the geologic-tectonic setting and by the behavior of urban structures under
seismic load. Different spatial ranges of influence permit the separation of these components, from a
regional to a very local scale. Intuitively, it seems correct to attribute the overall shape of the field to the
seismic-source geometry and then, as the spatial scale decreases, the effects of local components
become more evident. It can be asserted that the main criterion to distinguish a very local pattern is the
lack of spatial correlation to field elements at larger distances. This classification of components based
on their spatial range of influence is very important, because pure error is also included; by definition,
pure error is totally uncorrelated in space. Quantification of spatial relationships is thus necessary and it
influences the choice of a proper analytical method. For these reasons we propose the application of
kriging with its main spatial-analysis tool, the semivariogram. Kriging is a statistical technique permitting
to estimate unknown values at specific points in space by using data values from known locations. The
intrinsic characteristics of this method, developed during the 1960s and 1970s, have long been
acknowledged as a good spatial interpolator (Matheron, 1963; Isaaks and Srivastava, 1989; Davis, 2002).
Kriging is widely used within the field of earth sciences, including mining, geochemistry, remote sensing,
and, in general, in any situation where spatial data must be interpolated. Among the most important
features of this method are (1) the unbiased estimate of results, (2) the minimum Application of Kriging
Technique to Seismic Intensity Data 541 Figure 1. Typical behavior of experimental semivariogram (e.s.)
for data from a regionalized variable; c(l) has low values for small distances l, it increases more or less
regularly until reaching a stable level after a specific value of lsill. This limit separates data having spatial
autocorrelation from data that are spatially independent. estimation error, and (3) uncertainty
evaluation of interpolated data points. The main assumption, when using kriging, is that data analyzed
are samples of a regionalized variable. The properties of this kind of variable lie in the range between
true spatial randomness and fully deterministic behavior. Determinism is not complete and depends on
the spatial distance between points; the closer they are, the stronger is their relation. At greater
distances determinism is lost, and spatial autocorrelation of data vanishes. The semivariogram is the
analytical tool used to evaluate and quantify the degree of spatial autocorrelation; its results constitute
the basis of the kriging interpolation.

544 V. De Rubeis, P. Tosi, C. Gasparini, and A. Solipaca Figure 3. Reconstruction of synthetic


macroseismic fields by application of the kriging methods based on the experimental semivariograms
represented in Figure 4. (a) Reconstruction by application of block kriging method from error-added
data (modeled semivariogram of Fig. 4 [filled circles]). Dashed lines delineate boundaries of original
error-free field, colored circles are point data created using equations (10,11,12), continuous colored
field is the result of filtering and smoothing of the erroradded data. Considering that error increases
from south to north, and density data increases from west to east, it is evident that the best
reconstructed zone is the southeast corner, whereas the worst results are located in the northwest part
of the field. (b) The sb error estimation (equation 9) of the interpolated field: error increases, as
expected, toward the north-northwest and a relatively high level can be seen around the epicentral zone.
(c) Reconstruction of the field based on error-free data points located as in a; original and filtered fields
have same representation as in a: the original field is almost perfectly fitted by ordinary kriging
reconstruction obtained by using the modeled semivariogram of Fig. 4 (open circles). (d) The s error
estimation (7) of the reconstructed error-free field; lower density sampling areas have the highest
relative error. 546 V. De Rubeis, P. Tosi, C. Gasparini, and A. Solipaca Figure 5. Experimental and
modeled semivariogram of macroseismic intensities of the 31 October 2002, ML 5.4, central Italy event.
Two spatial ranges are recognized: because of error data content, the shorter range reveals a higher
fractal dimension (D

3.0) until a distance of 16 km. The longer-range portion is modeled from 16 to 60 km.

4.0), and a longer-distance portion, up to approximately 60 km (log2(l)

5.9), with a steeper loglog linear trend. As discussed in the Method section, fractal interpretation of c(l)
is justified here because of the presence of a loglog linear fit for the two range sections. Error plays a
role within the shorter range (first portion); the second portion shows that data are spatially
autocorrelated until a range of about 60 km, after which correlation disappears. Based on these e.s.
results, block kriging was applied to the trend-free intensity data. Figure 6 shows the filtered and
interpolated macroseismic field where c(l) was modeled with block size in accordance with the first,
shorter-range portion. The epicentral area is well defined and is ascribed to M.C.S intensity degree (I) 8.
Other locations with intensity 8 are relatively scattered and are neighbored by some I 6 values. The
overall pattern of the interpolated field is roughly circular, evidencing an almost isotropic character. The
boundary between I 4 and I 5 is the most irregular, indicating possible attenuation and amplification
zones. The kriging error, Sb (here represented as one standard deviation according to equation 9), is
moderate over a great portion of the field (inside a

0.9 intensity range). Higher values are found in peripheral portions of the field, owing to the presence of
not felt values near low-intensity data. Smaller spots of high-error values are also present; these are
connected to anomalously dissimilar data values at small distances. The most useful results and
applications to be derived from macroseismic data analysis concern the study of attenuation laws of
intensity with epicentral distance. Several authors have confronted this topic, proposing simple
attenuation laws that are often reviewed (Musson, 2000; Gasperini, 2001, and references therein).
Owing to the high error content of data, empirical attenuation laws are generally of simple form.
Defining more complicated laws or trying to highlight directional anisotropies is usually avoided.
Nevertheless, a sufficiently high volume of data makes it possible to study attenuation with epicentral
distance in more detail. Figure 7a shows intensities versus epicentral distance for the original dataset.
Scattering is strong and, because error content does not afford greater detail, only a simple attenuation
law can be chosen. To check the possibility of extracting a better level of detail from the attenuation
behavior, the same representation is done on kriging filtered intensities, situated at the same locations
as the original intensity points (Fig. 7b). Data scattering is much lower and a more interesting path can
be evidenced. Attenuation of intensities is stronger until about 60 km; beyond this limit attenuation is
low until about 110 km; from this distance on it becomes stronger again until the end of the field (note
how scattering is comparatively more pronounced in this latter distance range). Changes in the
macroseismic attenuation rate between near and far field may reflect influences of crustal and
subcrustal structures (Gasperini, 2001). But for such an analysis, the attenuation pattern should be
confirmed by a multiple-event study. Here we only wish to show how proper data filtering can
significantly enhance interpretation of the attenuation pattern. Conclusions Macroseismic intensity can
be interpreted as a regionalized variable, sampled where there are structures and people to record
effects. As a result, data are point located and affected by error, and kriging is shown to be a good
method to interpolate and filter data. Moreover, as intensity is the result of different processesactive
and passiveoperating on different spatial scales, e.s. is shown to be a useful tool to differentiate
between these features and to characterize spatial autocorrelation, error level, and the fractal character
of the process. Reconstruction of a synthetic intensity field has shown that sampling density may be
important, but the effect of data errors on the reliability of the reconstructed field is stronger. The
validity of our analysis resides on the evaluation and quantification of the problem. Here again, error
quantification becomes an important evaluation tool to decide the significance of any conclusion about
the intensity field (Gasperini et al., 1999). Despite the typical uncertainties associated with macroseismic
information, we present a statistical method to analyze intensity data. Qualitative, error-contaminated,
and irregularly point-sampled data have been filtered and interpolated. The result is an estimated
continuous macroseismic field with associated error. The method was applied to a recent medium-
intensity event. Macroseismic information was principally made up of transient effects as reported by
individuals based on their experiences. Permanent damage
Metode Geostatistik dan Kriging

Istilah kriging diambil dari nama seorang ahli, yaitu D.G. Krige. Istilah kriging diperkenalkan
oleh G. Matheron untuk menonjolkan metode khusus dalam moving average terbobot (weighted
moving average) yang meminimalkan varians dari hasil estimasi.

Kriging adalah penaksir geostatistik yang dirancang untuk penaksiran kadar blok sebagai
kombinasi linier dari contoh-contoh yang ada di dalam/sekitar blok, sedemikian rupa sehingga taksiran
ini tidak bias dan memiliki varians minimum. Secara sederhana, kriging menghasilkan seperangkat bobot
yang meminimumkan varians penaksiran (estimation variance) sesuai dengan geometri dan sifat
mineralisasi yang dinyatakan dalam fungsi variogram yang mengkuantifikasikan korelasi spatial (ruang)
antar contoh. Kriging memanfaatkan nilai spasial pada lokasi tersampel dan variogram untuk
memprediksi nilai pada lokasi lain yang belum dan/atau tidak tersampel dimana nilai prediksi tersebut
tergantung pada kedekatannya terhadap lokasi tersampel. Pada penerapannya, kriging dibawah asumsi
kestasioneran dalam ratarata () dan varians (2), sehingga jika asumsi kestasioneran tersebut
dilanggar maka kriging menghasilkan nilai prediksi yang kurang presisif. Selain itu, sebagaimana pada
semua metode analisis data non-spatial (cross-sectional, time series, panel, dll.), kriging juga dapat
menghasilkan nilai prediksi kurang presisif jika di antara data yang ada terdapat pencilan (outlier).
Outlier didefinisikan sebagai nilai yang ekstrim dari nilai amatan lainnya yang kemungkinan dapat
disebabkan oleh kesalahan pencatatan, kalibrasi alat yang tidak tepat atau kemungkinan lainnya. Kriging
sebagai interpolasi spasial optimum dapat menghasilkan nilai prediksi kurang presisif jika di antara data
yang ada terdapat pencilan (outlier). Outlier didefinisikan sebagai nilai yang ekstrim dari nilai amatan
lainnya yang kemungkinan dapat disebabkan oleh kesalahan pencatatan, kalibrasi alat yang tidak tepat
atau kemungkinan lainnya. Pengembangan ordinary kriging (kriging klasik) adalah robust kriging yang
mentransformasi bobot variogram pada variogram klasik sehingga menjadi variogram yang robust
terhadap outlier.

Metode ini menggunakan kombinasi linier atau weighted average dari data contoh lubang bor di
sekitar blok, untuk menghitung harga rata-rata blok yang ditaksir. Pembobotan tidak semata-mata
berdasarkan jarak, melainkan menggunakan korelasi statistik antar-contoh yang juga merupakan fungsi
jarak. Karena itu, cara ini lebih canggih dan perilaku anisotropik dapat dengan mudah diperhitungkan.
Cara ini memungkinkan penafsiran data kualitas secara probabilistik. Selain itu dimungkinkan pula
interpretasi statistik mengenai hal-hal seperti bias, estimation variance, dan lainnya. Dengan teknik
rata-rata tertimbang (weighted average), kriging akan memberikan bobot yang tinggi untuk contoh di
dalam/dekat blok, dan sebaliknya bobot yang rendah untuk contoh yang jauh letaknya. Selain faktor
jarak, bobot ini ditentukan pula oleh posisi contoh relatif terhadap blok dan terhadap satu sama lain.
Metode kriging yang digunakan adalah teknik linier (ordinary kriging). Ordinary kriging cenderung
menghasilkan taksiran blok yang lebih merata atau kurang bervariasi dibandingkan dengan kadar yang
sebenarnya (smoothing effect). Bobot yang diperoleh dari persamaan kriging tidak ada hubungannya
secara langsung dengan kadar contoh yang digunakan dalam penaksiran. Bobot ini hanya tergantung
pada konfigurasi contoh di sekitar blok dan satu sama lain, serta pada variogram (yang walaupun
merupakan fungsi kadar namun didefinisikan secara global).

Ada beberapa model kriging yang umum digunakan di antaranya adalah ordinary kriging dan
universal kriging yang notabenenya tidak mengakomodir adanya outlier. Lebih lanjut, pengembangan
ordinary kriging adalah robust kriging yang mentransformasi bobot variogram pada variogram klasik
sehingga menjadi variogram yang robust terhadap outlier.

Variogram, Semivariogram, Kovariogram dan Korelogram

Pada pemodelan variogram dan kriging, data spasial diasumsikan sebagai proses stokastik {Z(S):S
E D} dengan D adalah himpunan bagian dalam ruang berdimensi Rd, d > 0. Kovarian nilai antara dua titik
sembarang si dan sj didefinisikan sebagai

dengan nilai korelasi adalah


Suatu proses dikatakan stasioner pada ratarata dan varians jika dan hanya jika (Si) = dan
2(Si)=2, akibatnya:

C(Si,Sj) = C(Si Sj) = C(h)


(Si,Sj) = (Si Sj) = (h)

di mana h adalah vektor jarak antara titik i dan j, C(h) disebut kovariogram dan (h) disebut
korelogram.

Varians nilai antara dua lokasi dengan jarak tertentu ditentukan sebagai
Var [Z(S + h) -Z(S)] = 2(h), 2(h)disebut variogram dan (h) disebut semivariogram.
Hubungan antara kovariogram, korelogram dan semivariogram berdasarkan kestasioneran
dinyatakan dengan

Semivariogram Empirik

Semivariogram empirik dihitung dari data sampel yang kemudian diplotkan sebagai fungsi
dari jarak. MisalZ(Si) adalah nilai hasil pengukuran pada lokasi i, sedangkan Si = (Xi,yi) adalah
vektor yang mengandung koordinat spasial x, y, semivariogram cloud didefinisikan sebagai
ij=0,5[Z(Si)-Z(Sj)]2

untuk semua pasangan jarak yang mungkin {(Si,Sj); i,j = 1,2,3,,n} dan diplotkan sebagai
fungsi jarak, yang dihitung dengan:
|h| = lSi -Sjl = [(Xi Xj)2 + (yi yj)2]1/2 Perhitungan ini melibatkan ribuan titik pada plot
semivariogram sehingga mengakibatkan sulitnya melihat pola tertentu.
Untuk mengatasi hal tersebut maka yij dikelompokkan (binning) berdasarkan kesamaan
jarak. Berikut rumusan semivariogram yang dikelompokkan (semivariogram empirik):

(h) =

di mana
N(h) : himpunan pasangan data pada Si dan Sj yang mempunyai selisih jarak yang sama, h E T(h),
sedangkan T(h) merupakan daerah toleransi di sekitar h. |N(h)| : banyak pasangan jarak di dalam
himpunan N(h).

Spatial Outlier

Spatial Outlier (pencilan spasial) didefinisikan sebagai nilai lokasi observasi yang tidak
konsisten (ekstrim) terhadap nilai lokasi observasi yang lainnya. Munculnya pencilan dapat
disebabkan oleh mekanisme pengambilan nilai observasi yang berbeda dengan yang lainnya, Ada
banyak metode yang digunakan untuk mendeteksi adanya pencilan salah satunya adalah dengan
spatial statistics Z test. Untuk spatial statistics Z test, didefinisikan sebagai:

Jika Zs(x) > , maka dideteksi sebagai pencilan (outlier), untuk tingkat signifikansi 5%, nilai = 2.
Daftar Pustaka

Chile`s, J. P., and P. Delfiner (1999). Geostatistics, Modeling Spatial Uncertainty,Wiley, New York.

Cressie, N. (1991). Statistics for Spatial Data, Wiley, New York.

Davis, J. C. (2002). Statistics and Data Analysis in Geology, Third Ed., Wiley, New York.

De Rubeis, V., C. Gasparini, A. Maramai, M. Murru, and A. Tertulliani (1992a). The uncertainty
and ambiguity of isoseismal maps, Int. J. Earthq. Eng. Struct. Dyn. 21, 509523.

De Rubeis, V., C. Gasparini, and P. Tosi (1992b). Determination of the macroseismic field by
means of trend and multivariate analysis of questionnaire data, Bull. Seism. Soc. Am. 82, 12061222.

Deutsch, C. V. (1996). Correcting for negative weights in ordinary kriging, Comput. Geosci. 22, no.
7, 765773.

Evans, A. K. (1998). Fourier dimension and fractal dimension, Chaos, Solitons & Fractals 9, 1977
1982.
Fox, C. G. (1989). Empirically derived relationships between fractal dimension and power law
form frequency spectra, Pure Appl. Geophys. 131, 211239.

Potrebbero piacerti anche