You are on page 1of 19

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/261411369

Geostatistics and Remote Sensing: an Improvement in Image Classification

Conference Paper · September 2008

CITATIONS READS

2 160

4 authors, including:

C. Tarantino Annamaria Castrignanò


Italian National Research Council Università degli Studi G. d'Annunzio Chieti e Pescara
210 PUBLICATIONS 5,796 CITATIONS 258 PUBLICATIONS 3,144 CITATIONS

SEE PROFILE SEE PROFILE

Guido Pasquariello
Italian National Research Council
200 PUBLICATIONS 2,281 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Sentinels Synergy for Agriculture - SENSAGRI View project

Improving flood detection and monitoring through remote sensing View project

All content following this page was uploaded by C. Tarantino on 10 April 2014.

The user has requested enhancement of the downloaded file.


Geostatistics and remote sensing: an improvement in
image classification
C.Fiorentino, C.Tarantino, A.Castrignanò, G.Pasquariello.

Abstract
In the context of the use of remote sensed data for monitoring land cover it is very
important to develop methodologies to obtain reliable maps. In order to achieve this
objective a possible approach is to combine both “spectral” and “spatial” features to
characterizing each ground class. In this paper we propose the integration of a spectral
classifier for remote sensed data at medium resolution, based on a traditional statistical
supervised classifier as “Maximum Likelihood”, with the spatial information provided
by a geostatistical tool, as “Indicator Kriging” algorithm. Using this combined
approach, better results in land cover class discrimination have been obtained and the
resulting maps look more homogenous than in the case with the spectral information
only.

Introduction

The remote sensed data for monitoring land use has an important role in many
applications: prevention, planning, controlling etc.; therefore it is very important to
develop methodologies to obtain more accurate and reliable maps of land use. This
experimental work has the aim to integrate the use of a spectral classification of remote
sensed data at medium resolution in the optical spectrum, based on a traditional
statistical supervised classifier as “Maximum Likelihood”, with the spatial information
provided by a geostatistical analysis.
Traditional spectral classification of remotely sensed images, applied on a pixel-by-
pixel basis, ignores the potentially useful spatial information between the values of
proximate pixels. Infact, the main limitations of the spectral classifiers as “Maximum
Likelihood” is that the similarity of a pixel to a class is computed only in the spectral
space, without taking into account the spatial coordinates of that pixel. In this way the
features obtained might be unrealistic because there may be some isolated and not
connected pixels assigned to a particular class. These features should be corrected by
using the spatial information during the classification process.
For some years the spatial information inherent in remotely sensed images has been
employed to enhance spectral classification. This has been achieved primarily by
filtering the original imagery to derive texture ‘wavebands’ (Atkinson and Lewis, 2000)
for subsequent use in classification or smooth the imagery prior to (or after)
classification. Recently, the variogram has been used to represent formally the spatial
dependence in remotely sensed images and used in texture classification in place of
simple variance filters. However, the variogram has also been employed in soil survey
as a smoothing function for unsupervised classification.
The Indicator Kriging (IK) geostatistics tecnique is quite interesting in the optics of
incorporate spatial information in data processing. IK is an interpolation algorithm that
estimates the probability of occurrence of classes on the basis of surrounding
observation(Goovaerts,…….).
The recent developed geostatistical tecnique of Probability Kriging (PK) is an
enanchement of Indicator Kriging (IK), because IK uses only the ground truth indicator
data (spatial information), wheras PK uses both secondary information (spectral
information) provided by remote sensing data and the indicator information (primary
variable) in a multivariate approach, called co-indicator kriging.
The purpose of this paper is to integrate the IK geostatistic tecnique, which ignores
spectral information, and the traditional “Maximum Likelihood”supervised classifier,
which ignores spatial information. In this context “Indicator Kriging” algorithm has
been considered to produce the posterior probabilities for a selected data set of pixels on
the investigated scene, as an alternative to the ‘a priori’ probabilities, traditionally
provided to “Maximum Likelihood” classifier.
We have compared the performance of four classification methodologies: traditional
“Maximum Likelihood” classifier, “Indicator Kriging” tecnique, combined ML-IK
algorithm and PK tecnique.

1. Materials and method

1.1.Combined Spectral and Spatial approach


The Maximum Likelihood algorithm is a conventional statistical classification
technique that allocates each pixel of an image to the class with which has the highest
likelihood or ‘a posterior’ probability of membership.
Let the spectral classes for an image be represented by the categorical variable
{ωi ; i = 1,..., M } with M mutually exclusive categories and let X=X(u α) be B-variate
random vectors (B = number of spectral bands of the image), the pattern observations
describing a point at the position u α .
In remote sensing the measurement vector X, referred to the pixel of spatial coordinates
u α ( α = 1,..., n ), is a column of brightness values for the image and the training data for
ground cover type are associated to the sample points u α .
To determine the class or category to which a generic pixel vector X(u) belongs, it is
strictly the conditional probabilities:

P(ω i X(u )) i =1,……..,M

that are of interest. This probability gives the likelihood that the class ω i prevails for the
pixel at the position u.
The Maximum Likelihood algorithm assigns each pixel to the class whose ‘a posterior‘
probability is maximised:

assign the position u at the class ω i ⇔ P( ω i | X(u ) ) = max ω P (ω X(u ))

P( ω i | X(u )) are unknown, but suppose we have sufficient training data for each class
that can be used to estimate a “spectral” probability density function p(X(u) ω i ) for a
cover type, i.e. the chance of finding a pixel from class ω i , say, at the position X(u).
P( ω i | X(u )) is then obtained by applying the Bayes rule:

p (X(u) ω i )P(ω i )
P(ω i | X(u) ) =
p (X(u) )

where P( ω i | X(u )) represents the posterior probability of pixel with data vector X(u),
belonging to class i , P(X(u) ) is the unconditional probability that the pixel occurs in
the image, P(ωi ) is the ‘a priori’ probability that class ω i prevails at the pixel u. It is
assumed that spectral probability density function is of the form of multivariate normal
model, therefore:

1
( X −m i )T ∑i−1 ( X −mi )
p (X(u) ωi ) =
1 −
2
1
e
(2π )
B
2 ∑i 2

where ∑ is the variance-covariance matrix for the class i (calculated from the training
data), B are the bands of the image, ( X − m i )Σ i−1 ( X − m i ) is the Mahalanobis distance
between the generic pixel X and the centroid of class i and mi is the mean vector for
class i.
Traditionally, in the Remote Sensing literature P(ωi ) are considered ’a priori’ class
probabilities, i.e. independent of spatial position and no neighbourhood information is
taken into account. The approach we adopt is based on the idea that ‘a priori’
probability of occurrence of each class is not actually the same everywhere but depends
on the pixel location [1], so that P(ωi ) becomes, in this combined approach, an ‘a
posterior’ probability P(ωi (u) uα ) computed by “Indicator Kriging” algorithm of
geostatistics.
When the variable is “categorical” (e.g. type of land cover), the probability of
occurrence of each category is first estimated at each interpolation grid node by using an
indicator approach, which amounts to apply Ordinary Kriging to indicator transforms of
the ground data.
At each sample location u α , a vector of M values i(u α ; ω i ) can be defined according
to the following indicator transform:

i(u α ; ω i ) = 1 if ω i occurs at u α α=1,2,…………n

i(u α ; ω i ) = 0 otherwise

If u is an unsampled location, the unknown indicator value for a given category ω i 0 may
be estimated from the neighbouring indicator data i(u α ; ω i ) by kriging. The resulting
kriging estimate i*(u ; ω i 0 ) calculates the probability of the category ω i 0 to prevail at the
location u:

i*(u ; ω i 0 ) = Prob { ω i 0 prevails at u | (n) }


where the notation (n) represents the conditional information retained in the
neighbourhood of u.
According to Ordinary Kriging the estimated probability is given by:

n
i * (u , ω i 0 ) = ∑ λα (u , ω i 0 )i (u α , ω i 0 )
α =1
uuuu

where λα ( , ω i 0 ) is the weight assigned to the indicator datum i(u α ; ω i 0 ), obtained


solving an ordinary kriging system so to minimize error variance:

λ β (u, ω )γ (u α , u β ; ω ) + µ (u;ω ) = γ (u α , u; ω )
n


β=1
i i i i i ∀α = 1,.........., n
n

∑ λ β (u, ω ) = 1 i (1)
β=1

where γ i is the theoretical variogram fitted to experimental variograms


hhhh

hhhh

2
1 N( )
∑ [i(uα ;ωi ) − i(uα + ;ωi )] ,
hhhh

γ 'i = [2]
2 N ( ) α =1
obtained from indicator data for each class ωi , and N(h) is the number of data pairs
within a given vector distance h.
The last equation of the Kriging system (1) derives from unbiasness condition of
Ordinary Kriging.
At each location u the M estimated probabilities i * (u, ω i ) must have values within the
range [0,1] but owing to negative kriging weights it is necessary resetting faulty
probabilities to the nearest bound, 0 or 1.
Moreover, because the probabilities i * (u, ω i ) represent the classes ω i to be exhaustive
and not overlapping, they must satisfy the following condition:

∑ i (u, ω ) = 1 .
i =1
*
i (2).

In this work we compare the results of land cover classification obtained using three
methods: traditional Maximum Likelihood algorithm, Indicator Kriging and the
combined approach. The last methodology is also tested against Probability Kriging
assumed as reference methodology.
1.2.Probability Kriging

Probability Kriging (PK) is an enanchement of indicator kriging (IK), because IK uses


only the indicator data, wheras PK uses both secondary information (spectral
information of the remote sensing data Z) and the primary variable (ground truth
indicator information) at the sampled points uα in a multivariate approach, called co-
indicator kriging. This tecnique is applied when there is a target variable known on a
sparse sampling and a correlated variable aviable at each node of a regular grid. Large
differences may occur between the units of measurement of secondary information and
the indicator transforms, which may cause instability problems when solving the co-
kriging system.
Since indicator data i(u α ; ωi ) are valued either 0 or 1, it is necessary preliminarly to
submit the z-values to uniform transform in the range [0.1].
Z(uα ) must be replaced by the values of their standardized ranks:
r (uα )
k (uα ) =
N

where r (uα ) ∈ (1, N ) is the rank of the datum Z(uα ) , and N the total number of the
pixels in the image.
As first step PK approach requires calculation of the indicator variograms, the rank-
order transformed variograms and the cross variograms between each pairs of variables.
PK is the co-kriging of the indicator transform i(u α ; ω i ) and the rank-order transform
k (uα ) of each secondary variable; the probability kriging estimator is then given by:

n nb n
I (u, ωi ) = ∑ λα (u, ωi )i (uα , ωi ) + ∑
*
∑ να ,j (u, ωi ) k j (uα ) .
α =1 j =1 α
=1
where nb is the number of secondary variables to take into account, the weights
λα (u, ωi ) and ν α (u, ωi ) are obtained by solving the ordinary co-kriging system
(Goovaerts, 1997).
We have preferred to perform a principal component analysis(PCA) of the remote
sensing data considering the rank-order transformation of the first 3 principal
components k (uα ) as secondary information and the ground truth indicator data as
primary variable, to be estimated by co-kriging. Ince secondary information is known at
each node of interpolation grid, we have applied a simplifield version of full co-kriging,
known as “Multi-Collocated Cokriging”. The approach is quite similar to ordinary
cokriging with the only difference in the neighbourhood search. As using all secondary
information contained within the neighbourhood may lead to an intractable solution due
to too much information, the initial solution of collocated co-kriging is to use the single
secondary value located at the target grid node location. However, in ordinary co-
kriging the weights attached to the secondary variable must add up to zero. Therefore if
only one data value is used, its single weight is zero. The original technique is then
extended so that the secondary variable is used at the target location and also at all the
locations where the primary variable is defined within the neighbourhood. This solution
has generally produced more reliable and stable results (Rivoirard, 2001). The modified
version is less precise than full co-kriging, not using all the auxiliary information
contained within the neighbourhood. However, because the co-located secondary datum
tends to screen the influence of more distant secondary data, there is actually little loss
of information. Even if the exhaustive auxiliary variables are actually three in our study,
the approach was implemented for only one collocated variable (first principal
component). Of the second and third principal components the only data collocated with
the ground truth data were retained. Finally, M probability maps were produced, from
which the global classification was obtained by assigning each pixel of the image to the
class whose probability is maximised.

2.Case study

Images from Landsat TM with the spatial resolution of 30 meters and 6 bands in the
visible and near/medium infrared spectrum have been used. The scene, from remote
sensed data, is the Capitanata area located in Foggia plain (in Apulia region, south-east
Italy), mostly used to agricultural crops. The land use includes 6 classes: cereal crop,
orchard or olive grove, horticultural grove, vineyard, urban areas and water sytems. The
cloud cover was removed from the scene by an automatic thresholding method.
The time of the image is July 2006 and the ground truth is based on land use
information provided by field survey, a data set of 2400 sample points: 1600 training
points and 800 to validate the resulting map. The test sample has been selected, within
each cover class, from the 2400 points randomly, but uniformly distributed on the
investigated scene using an annealing simulation approach (……..). The ground truth
points have been verified using an orthophoto dated 2005, certainly this check has been
possible only for time unvariable classes. The map of training sample points is
represented in fig.1.

Figure 1. Ground truth sampled points in the Capitanata area.


A second investigated scene is a zootechnical farm, located in Capitanata area, dated
May 1997 and the ground truth is based on land use information provided by the farmer.
This scene is characterised by land cover classes more homogeneous and regularly
shaped then the previous. The land use is characterized by 9 classes: grain-oat, oat,
durum wheat, barley, meadow, pasture, vetch-oat, burned soil, fallow.
An infrastructure area was cut from the scene.
A test set of 200 points, evenly distributed, have been extracted from the 700 ground
truth sample points for the classification validation. The remaining 500 (training points)
sample data has been used to produce the land cover maps.

Figure 2. Ground truth sampled points in the zootechnical farm.

The training data sets have been processed appling Indicator Kriging algorithm:
-as first step the variogram of the indicator transformed variable has been calculated,
so six experimental indicator semivariograms in the first investigated area and nine in
the second zone has been computed[2]. The matrix of the variograms (fig. 3 e 4) shows
how all variograms are bounded, the variation can then be assumed stationary and
occurring at two distinct spatial scales;
-as second step an isotropic Linear Model of Coregionalization (LMC) was fitted
including two basic structures: two spherical models1 (Capitanata area: range = 4000m
and range = 12000m; Zoothecnical farm: range = 300m and range = 800m;) and a
nugget effect (discontinuity at the origin) was added to avoid numerical instabilities in
the spatial interpolaton (Goovaerts, 1997).

____________________________________

Spherical model (Webster & Oliver, 2001) is the function:

  3
C  3  h  − 1  h   h≤a
γ (h ) =   2  a  2  a  
for
 

C for h > a

C is the sill; a is the range.


Figure 3. Experimental variograms of indicative variables (blu-line) and LMC (red-line) for the Capitanata area.

Figure 4. Experimental variograms of indicative variables (blu-line) and LMC (red-line) for the zootechnical
farm.
In the PK case, for each ground truth class, in addition to experimental indicator
variogram, there are evaluated also the experimental variograms of the rank tranformed
pricipal component datas associated to the the training sample locations. As an example
in fig. 5 are represented the resulting variograms for the classes cereal crop and
vineyard of the capitanata area. In fig. 6 there are two of the resulting variograms for the
classes durum wheat and grain oat of the second investigated area.

(a)

(b)
Figure 5. Matrix of the experimental variograms of indicative variables and rank tranformed principal
component variables (blu-line) and LMC (red-line) for the classes cereal crop (a) and vineyard (b) in the
Capitanata area.
All variograms are bounded and the variation can be assumed stationary, the theorical
variograms (LMC) were fitted, for each cover class, including two basic structures as
explaned in the previous cases; these plots (fig.5 & 6) are also informative on the type
of spatial correlation between the indicator primary variable and the secondary variables
(cross-variograms).

(a)

(b)

Figure 6. Matrix of the experimental variograms of indicative variables and rank tranformed principal
component variables (blu-line) and LMC (red-line) for the classes durum wheat (a) and meadow (b) in the
zootechnical farm.
3. RESULTS
A map of ‘a posteriori’ probabilities, for each cover class, has been obtained processing
the training data by Indicator Kriging algorithm. These maps have been given as input
into the Maximum Likelihood algorithm opportunely modified to introduce a pixel
variable a priori probability cause the commercial software as ENVI or ERDAS doesn’t
implement this target. This algorithm has been developed in IDL language and
integrated in ENVI software.
As first result in fig. 7 is shown the classification resulting map obtained from the
previous IK probability maps by assigning each pixel of the image to the class whose
probability is maximised.

Fig. 7. Land cover map of the Capitanata area obtaind appling the Indicator Kriging methodology
(the white areas are affected by cloud cover).

Fig.8. Map of land cover performed by appling the “Maximum Likelihood” supervised classification
algorithm.
In fig.8 is shown the land cover map obtained appling the traditional “Maximum
Likelihood” spectral classifier implemented in ENVI software.
The simple IK approch, considering only the spatial information, produces a remarkable
smoothing effect, so the land cover classes discrimination is unrealistic; the ML
tecnique, considering only the spectral information, produces a quite noisy map because
the classes discrimination is more confused.
The combined ML-IK tecnique improves the previous results because it takes into
account not only spectral information but also the, not secondarly, pixel’s
neighbourhood information, fig.9 shows the relative map.
Fig. 10 shows the land cover map obtained appling the Probability (co)-Kriging
geostatistics tecnique.
The land cover map in fig.10 is more accurated then the simple Indicator Kriging
classification map because, as the combined approch, it takes advantage of exhuustive
radiometric information from remote sensing data (known at each pixel) in addition to
the sparse information of ground truth. The choise to compare PK tecnique with the
combined approch arises from the advantage of Probability co-Kriging to process
different tipology of data in a single step.

Fig. 9. Map of land cover performed by appling the combined ML-IK algorithm of calssification.
Fig. 10. Map of land cover performed by appling the Probability (co)-Kriging geostatistics tecnique.

From the comparison of the resulting maps in fig.8 and 9, it is quite evident the better
performance of the combined tecnique achieving a better homogeneity within classes, as
it is also confirmed by the inspection of the confusion matrices: Tab.1 and Tab.2. The
overall accuracy increases from 76.16% to 85.96%.
Tab. 1. Confusion matrix in test for the combined methodology resulting map in the capitanata area.

(TEST) GROUND TRUTH


A PRIORI Water Orch. & Cereal Urban User’s
Horticultural Vineyard Total
IK-PROBABILITY systems olive crop areas Accuracy
crop (%) (%) (%)
(%) grove (%) (%) (%) (%)
Water systems
100.00 0.00 0.00 0.00 0.00 0.00 1.06 100.00
(%)
Orch.&olive
0.00 76.92 4.97 4.23 1.67 14.52 18.54 71.43
grove (%)
Cereal crop (%) 0.00 6.92 91.44 9.86 0.00 4.84 46.75 93.77
Horticultural
CLASSIFIED 0.00 0.77 2.21 71.83 0.00 0.00 7.95 85.00
crop (%)
DATA
Urban areas (%) 0.00 0.00 0.28 0.00 98.33 0.00 7.95 98.33
Vineyard (%) 0.00 15.38 1.10 14.08 0.00 80.65 17.75 74.63
Total (%) 100.00 100.00 100.00 100.00 100.00 100.00 100.00
Producer’s
100.00 76.92 91.44 71.83 98.33 80.65
Accuracy (%)

Overall accuracy = 85.96% K coef. = 0.80

It is notable from tab. 1 and 2 that there is much more confusion between the classes
orchard or olive grove and vineyard, probably cause their quite similar spectral sign,
and between the classes horticultural crop and vineyard, probably cause the small
extention of some horicultural field. In both cases the spatial information enanches the
classification.

Tab. 2. Confusion matrix in test for the land cover map performed appling traditional Maximum Likelihood
classifier.

(TEST) GROUND TRUTH


A PRIORI Water Orch. & Urban User’s
Cereal Horticultural Vineyard Total
EQUI systems olive grove areas Accuracy
crop (%) crop (%) (%) (%)
PROBABILITY (%) (%) (%) (%)
Water systems
100.00 0.00 0.28 0.00 0.00 0.00 1.19 88.89
(%)
Orch.&olive
0.00 72.31 12.15 7.04 3.33 21.77 22.78 54.65
grove (%)

Cereal crop (%) 0.00 7.69 80.39 5.63 3.33 3.23 41.19 93.57
CLASSIFIED Horticultural crop
0.00 4.62 2.21 63.38 1.67 8.06 9.27 64.29
DATA (%)
Urban areas (%) 0.00 0.00 3.87 2.82 91.67 0.81 9.54 76.39
Vineyard (%) 0.00 15.38 1.10 21.13 0.00 66.13 16.03 67.77

Total (%) 100.00 100.00 100.00 100.00 100.00 100.00 100.00


Producer’s
100.00 72.31 80.39 63.38 91.67 66.13
Accuracy (%)

Overall accuracy = 76.16% K coef. = 0.67


Tab. 3. Confusion matrix in test for the land cover map performed appling Probability co-Kriging tecnique.

(TEST) GROUND TRUTH


PK Water Orch. & Urban Total User’s
PROBABILITY Cereal Horticultural Vineyard
systems olive grove areas (%) Accuracy
crop (%) crop (%) (%)
(%) (%) (%) (%)

Water systems (%) 100.00 0.77 0.55 0.00 0.00 0.00 1.46 72.73
Orch.&olive grove
0.00 76.15 3.87 2.82 0.00 23.39 19.07 68.75
(%)
Cereal crop (%) 0.00 7.69 86.74 38.03 5.00 10.48 48.61 85.56
Horticultural crop
CLASSIFIED 0.00 0.00 1.93 19.72 0.00 0.00 2.78 66.67
(%)
DATA
Urban areas (%) 0.00 1.54 1.93 0.00 93.33 0.00 8.61 86.15
Vineyard (%) 0.00 13.85 4.97 39.44 1.67 66.13 19.47 55.78
Total (%) 100.00 100.00 100.00 100.00 100.00 100.00 100.00
Producer’s
100.00 76.15 86.74 19.72 93.33 66.13
Accuracy (%)

Overall accuracy = 75.89% K coef. = 0.65


In tab.3 is reported the confusion matrix relative to the PK tecnique, the overall
accuracy is comparable to that of traditional “ Maximum Likelihood” algorithm, but the
cause of this low value is quite different. This unexpected result arises from the spatial
complexity of the ground cover. Probably, PK tecnique and the combined ML-IK
algorithm weigh differently the spatial information, infact the land cover classes, in the
first case, look too homogenous than in the second one.
To validate the previous results a second scene has been studied: the zootechnical farm.
In this case the land cover classes are homogenous field with sharp edges, but very
similar spectral sign.
In fig.11 is shown the Indicator Kriging resulting map (elaborated as in the previous
case), while in fig.12 is shown the classification map obtained appling the ENVI
“Maximum Likelihood” classifier.

Fig. 11. Land cover map of the zootechnical farm obtaind appling the Indicator Kriging
methodology.

Fig.12. Map of land cover performed by appling the “Maximum Likelihood” supervised
classification algorithm.
In the view of homogeneous areas recognition, the enhancement of land cover map
obtained by just appling the Indicator Kriging thecnique is quite evident comparing the
results in fig.11 and 12: the number of isolated pixels in fig.12 notably decrease in
fig.11. This enanchement is also confermed by the confusion matrix since the overall
accuracy increases from 63% (case in fig.12) to 80% (case in fig.11).
In fig. 13 is shown the map obtained processing the data with the combined ML-IK
algorithm. From the comparison between fig.12 and 13 it is evident that the speckled
patterns displayed by the spectral classification might be real, but they are categorized
as misclassification because of the mapping of land cover as homogeneous classes.
In this case the overall accuracy increases from 63% to 91.5%; the relative confusion
matrix are in tab.4 and 5.

Fig. 13. Map of land cover performed by appling the combined ML-IK algorithm of calssification.
Tab. 4. Confusion matrix in test for the land cover map performed appling traditional Maximum Likelihood
classifier.

(TEST) GROUND TRUTH


A PRIORI Grain Durum User’s
Oat Barley Meadow Pasture Vetch+Oat Burned Fallow Total
EQUI Oat Wheat Accuracy
(%) (%) (%) (%) (%) Soil (%) (%) (%)
PROBABILITY (%) (%) (%)
Unclassified (%)
0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
Grain Oat (%)
86.21 18.75 26.32 16.67 0.00 0.00 0.00 20.00 0.00 23.00 54.35
Oat (%)
6.90 50.00 0.00 8.33 5.71 20.00 0.00 0.00 6.67 8.50 47.06
Durum Wheat
(%)
0.00 6.25 56.14 25.00 0.00 0.00 0.00 20.00 3.33 19.00 84.21
Barley (%)
6.90 0.00 12.28 50.00 0.00 0.00 0.00 20.00 13.33 10.00 30.00
Meadow (%)
0.00 0.00 0.00 0.00 77.14 50.00 16.67 0.00 3.33 17.00 79.41
CLASSIFIED Pasture (%)
0.00 25.00 1.75 0.00 11.43 10.00 0.00 0.00 0.00 5.00 10.00
DATA
Vetch+Oat (%)
0.00 0.00 0.00 0.00 0.00 0.00 50.00 0.00 0.00 1.50 100.00
Burned Soil (%)
0.00 0.00 0.00 0.00 0.00 0.00 0.00 40.00 0.00 1.00 100.00
Fallow (%)
0.00 0.00 3.51 0.00 5.71 20.00 33.33 0.00 73.33 15.00 73.33
Total (%)
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
Producer’s
Accuracy (%)
86.2 50.00 56.14 50.00 77.14 10.00 50.00 40.00 73.33

Overall Accuracy = 63.00% K coef.=0.56

Tab. 5. Confusion matrix in test for the combined methodology ML-IK resulting map.

(TEST) GROUND TRUTH


A PRIORI Grain Durum Vetch+ User’s
Oat Barley Meadow Pasture Burned Fallow Total
KRIGING Oat Wheat Oat Accuracy
(%) (%) (%) (%) Soil (%) (%) (%)
PROBABILITY (%) (%) (%) (%)
Unclassified (%)
0.00 6.25 1.75 0.00 0.00 0.00 0.00 0.00 0.00 1.00
Grain Oat (%)
96.55 0.00 0.00 8.33 0.00 0.00 0.00 0.00 0.00 14.50 96.55
Oat (%)
0.00 87.50 0.00 0.00 2.86 10.00 0.00 0.00 0.00 8.00 87.50
Durum
Wheat (%) 0.00 0.00 96.49 0.00 0.00 0.00 0.00 60.00 6.67 30.00 91.67
Barley (%)
3.45 0.00 0.00 91.67 0.00 0.00 0.00 0.00 0.00 6.00 91.67
Meadow (%)
0.00 0.00 1.75 0.00 94.29 0.00 0.00 0.00 3.33 17.50 94.29
CLASSIFIED Pasture (%)
0.00 6.25 0.00 0.00 0.00 90.00 0.00 0.00 0.00 5.00 90.00
DATA
Vetch+Oat (%)
0.00 0.00 0.00 0.00 0.00 0.00 66.67 0.00 0.00 2.00 100.00
Burned Soil (%)
0.00 0.00 0.00 0.00 0.00 0.00 0.00 40.00 0.00 1.00 100.00
Fallow (%)
0.00 0.00 0.00 0.00 2.86 0.00 33.33 0.00 90.00 15.00 90.00
Total (%)
100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00 100.00
Producer’s
96.55 87.50 96.49 91.67 94.29 90.00 66.67 40.00 90.00
Accuracy (%)

Overall Accuracy =91.50% K coef.=0.9


In fig.14 is shown the resulting classification map elabotrated by Probability co-Kriging
algorithm, the confusion matrix is in tab.6.

Fig. 14. Map of land cover performed by appling the Probability (co)-Kriging geostatistics tecnique.

In this case, the overall accuracy of PK classification map (92%) is comparable to that
of combined ML-IK land cover map.

4. Conclusions

This paper presents a procedure to combine Indicator Kriging based probabilities with a
supervised Maximum Likelihood spectral classifier.
In both cases studied in this work, accounting for pixel locations improves the accuracy
of the traditional spectral classification algorithm.
In the Capitanata area, where the land cover pattern is more complex and the ground
truth are not so densely sampled, the combined approch gives the better results, while in
the second scena caracterised by an high spatial continuty, but very similar spectral sign
of classification field the high performnce of combined ML-IK algoritm, in front of the
very noisy traditional Maximum Likelihood result, is comparable to that of Probability
co-Kriging tecnique. This confirms the important role of spatial information in image
classification.
Other tecniques, as multitemporal approch, should be used to improve spectral
supervised classification, the advantage of this combined methodology is that it exploits
to the utmost the same basical information.

View publication stats

You might also like