Mulit-Spectral/Textural supervised classification - Land Cover Mapping with SPOT in Indonesia

Mulit-Spectral/Textural supervised classification - Land Cover Mapping with SPOT in Indonesia

Gastellu-Etchegorry
J.P.SCOT Conseil 18, Avenue Edouard Belin
31055-Toulouse Cedex France

Ducros-Gambart
D.C.E.S.R Paul Sabatier University
21029-Toulouse Cedex France

Abstract:
The capability of SPOT combined with a specifically designed classifier was investigated for computer assisted land cover/use mapping in Indonesia. Atmospheric conditions and the small size, complexity and dynamic nature of local agro-forest systems both confuse spectral analyses. In this context, conventional classifiers are inadequate. This led to the development of an original supervised classification method for discriminating between the large numbers of classes (or subclasses) that are apparent in high-resolution satellite images. Several multi-spectral/temporal classifications are initially processed with various combinations of multi-spectral/temporal channels. Then, results are fusioned with a class priority system but information about spectral confusions is preserved. These confusions are further solved by applying texture features to classes that are confused. These confusions are further solved by applying texture features to classes that are confused. The selection of these features, occasionally correlated, is difficult and subtle. A technique was developed for selecting the optimum feature for each class. But textural confusions appear, e. g in heterogeneous and interface zones; they are solved by a Specifically designed process. The final result is improved confusion matrices enable class improvement to be identified, which is considered as being from 10 to 50% depending on the particular class.

Introduction
Both the small size, complexity and dynamic nature of Indonesian agro-forest systems (Malingreau and Christiani, 1981), especially in Java, create problems for inventorying and monitoring. Studies have already been conducted by the authors for determining spectral and spatial characteristics of these systems (Gastellu-Etchegorry and Ducros-Gambart, 1989).

Two methods were used by the authors for assessing the atmospheric influence foe determining spectral characteristics; i.e. the dark ground feature calibration method (Sabins, 1978) and a method derived from piech and Wlaker (1974), based on the statistical analysis of landscape units which are partly in shadow. Several SPOT scenes of Central Java were considered. Results concerning three study areas of SPOT scene (193, 365) are listed in table 1. Two major points must be emphasized:

Atmospheric upwelling radiances have very large values. Compared to the total measured radiance, they represent between 30% and 80% for channel XS1, between 20% and 70% for channel XS2 and between 15% and 45% for channels XS3.
Atmospheric radiances are characterized by an spatial important heterogeneity; which means that identical Earth features within the same SPOT image may have quite different spectral characteristics.

Table 1:
(a)atmospheric influence

Study areas	SPOT scene	XS1	XS2	XS3
Study area 1	( 293, 362)	52	37	34
Study area 2		43	31	32
Study area 3		35	27	23

(b)Mean radiometric values of Earth feature

	XS1	XS2	XS3
Mean radiometric values Within the SPOT scen	63	58	68

Another major limitation for determining spectral characteristics of land cover units is due to the fact that there is no direct relationship between lands covers unit and spectral calluses. For example, a land cover unit corresponds to several spectral (Sub) classes, whereas a spectral (sub) class may correspond to several land cover units. This is particularly disturbing for classification processes. This aspect is of special importance with high-resolution satellites. Indeed, the number of land cover units which can be discriminated in a multispectral or multitemporal image is undoubtedly greater than with low-resolution satellites. Indeed, the number of land cover units which can be discriminated in a multispectral or multitemporal image is undoubtedly greater than with low-resolution satellites. Numerous tests have revealed the possibility of identifying twenty to thirty spectral classes for this type of

Image (Gastellu-Etchegorry and Ducros-Gambart, 1989) In reality, these classes correspond to land cover sub-units; e. g. the forest class can be broken down into subclasses of varying degrees of density.

After several tests, the best classification results are obtained by breaking down land cover units (main classes) into sub-classes. As the number of units becomes greater, so these units are spectrally resemblant. If they are broken down into subunits, spectral confusions are thus reduced. However, confusions are still present. Classification methods should therefore become progressively more accurate in order to handle a large number of classes.

This is consequently the objective of the supervised classification method presented in this paper. The method progresses through several stages. Each of these stages has been developed in order to achieve its purpose in the most efficient way possible.

Classification training

Spectral training

The first step is training which is essential and often fastidious for supervised classification: the operator should intervene and act accordingly at each level of the training program. The results of classification depend on the accuracy of the samples. The difficulty in locating the samples is proportional to the number of classes. Indentifying a sample which features two subclasses is a frequent occurrence.

In all cases samples shall be representative, well spread out over the image and in sufficient number. Statistics have revealed that the number of samples should be between 4 and 15; above this number, class characteristics are more or less constant (Ducros-Gambart 1988).

The distribution of samples over the entire image is mandatory, due to atmospheric and climatic variations as well as to agricultural activities, which imply reflectance variations for one same land cover unit. In this case, a unit is broken down into subclasses, which can subsequently be grouped.

Sample checking is achieved through the acquisition of statistics in both graphic and tabular form. Certain corrections are automatic. Establishing class histogram threshold levels enables value extremes, originating from parasitc points within samples, to be eliminated.

A procedure compares the resemblance of a sample to its class (or subclass) through a divergence computation of the samples. This ensures that all the samples are correctly attributed to this class. If a sample is excessively uncharacteristic, it is either eliminated o affected to another subclass if it can characterize this subclass.

These processes enable the rigorous acquisition of spectral class characteristics while eliminating error as well as avoiding tedious processing and checking.

Class confusion matrices are calculated in order to identify interclass confusions in various combinations of spectral or synthetic channels (index of vegetation, of brightness etc.) Based on these matrices, subclasses may be grouped. Their number can therefore be reduced. Combinations of channels providing optimum results are also determined.
Textural training.

The second stage, and second part of the training program, consists in detecting textural characteristics, which will enable confusions subsisting between classes subsequent to the first part to be eliminated.

The originality of the method consists in only applying textural discrimination in the classification for classes, which feature confusions; in general each class is only confused with a few classes. Confusion matrices enable the required classes to be determined.

Furthermore, it appears that each class can only be characterized by a reduced number of textural parameters. Only the textural factors, which determine a given class, will be used to identify this class without a systematic computation of all textural features. This not only reduces computation times but also improves results of the textural classification.

Among textural parameters characteristic of one class, certain are essential either for the description of class or for its separation in relation to other classes. These parameters are consequently sequenced, and intervene in the classification in a pre-established order as a function or their importance, defined in this stage (Swain) and Hauska, 1977).

Conventional textural parameters, i.e. homogeneity (variance), directivity (gradient), co-occurrence matrix (Haralick et al, 1973), Walsh-Hadamard transform, are computed for each sample, either from images transformed by the abovementioned operators or from spectral channels from which the features are extracted.

In order to identify the parameters which separate all the confused classes, and the order in which they intervee, a divergence computation enabling interclass distance to be measured is implemented. For each class pari, a distance is computed by applying the following formula:

iÎ I (set of textural parameters) and j, kÎC (set classes);
tji, tki: textural parameter value i, for classes j, k;
timin, timax: minimum, textural parameter values i:
tjiÎTj (set of textural parameters of class j)
di (j,k,): distance between classes j, k for textural parameter i;
d: divergence between classes j, k;
Selected parameter i is that which is at maximum distance di

As an example, if class j at 30% from these points in class k and at 10% in class 1 (percentage given by the confusion matrix), the researched parameter will be that which provides optimum separation of theses classes. However, two parameters may be required: one same parameter does not necessarily separate the three classes. The parameter having priority is that which corresponds to the greatest divergence (gastellu-Etchegorry and Ducros-Gambart, 1987).

Multi-spectral/temporal classification
During the following stage barycentric multitemporal or multispectral classification is carried out. This method has been selected for its speed; 5 to 10 times faster than the maximum likelihood method, and is of comparable accuracy. Points are not all affected to one class, and spectral confusions are conserved; i.e., subsequent to the classification a point can be affected to several classes. Only these points will be processed during multitextural classification.

In order to improve multispectral classification, several classifications are achieved using various combinations of multispectral or multitemporal channels. In relation to these combinations, classes are relatively well separable and relatively well classified. For a classification with a given combination, certain classes may have a high percentage of well classified points, while other classes with achieve better results with a different combination. In order to minimize these confusions, a technique consisting in merging the results of several classifications has been developed.

Various results of classification (here called classification plans) with combinations of various channels are confusion, with a class priority system. These combinations are selected from class confusion matrices.

The percentages located on the confusion matrix diagonal plot correspond to well classify points. Points, viewed on the horizontal plots, reveal the confusions of the class in question. Based on these percentages, priorities are highlighted.

For example, let us consider two classifications, achieved from satisfactory and complementary combinations plan 1 and 2. If the results of plan 1 are globally better: this plan will have priority: when a point is affected to two different classes in the two plants, it is the class of the plan having priority, which is conserved. However classes of plan 2 (i, j, k for example) may have priority. If one of the classes is encountered in plan 2, it is kept. But if a class in plan 1 has priority over one of classes (i, j,k) it is therefore the class of plan 1 which is conserved.

This process reduces certain interclass overlaps and consequently improves the multispectral classification.

Multitextural Classification
This stage discriminates spectrally-confused classes by the application of textural factors (Rakaryatham, 1984) .

As per the previous stage:

One point is affected to one class only; the classification is terminated for this point.
One point is unclassified; the classification is terminated for this point.
One point belongs to several classes: let Cm be the set of class to which point P belongs.

A Process enables certain Cm points to be eliminated by studying points in the neighborhood of P. If certain Cm classes are not common to those of neighboring points, they are eliminated.

Let Cml be a subset of Cm after elimination of classes not belonging to the neighborhood of P, and Cvm be a set of classes to which points neighboring P belong.

Cml = Cm Cvm

If CmÇCvm = f then = Cm : all classes of Cm are conserved.

Discrimination by means of textural parameters is then applied.

Let Ti be the set of class i texture factors and Pi be the proposal "the point considered crosscheck tij Ti":

If iÎ Cml and Pi are checked - iÎ Cm2 If iÎ Cml and Pi are not checked - iÎ Cm2

Cm2 is a subset of Cml after elimination of classes which do not check Pi: if the feature value applicable to a Cml class i does not correspond to the texture feature value for point P, this class is eliminated from Cml.

Thus, after analysing the various texture parameters, one only class is affected to the point. if all classes are eliminate from Cml, they are restored: texture does not intervene in this case, which is not frequent, and signifies that the feature selected are not sufficiently accurate. when points still belong to several classes, they are discriminated by a minimum distance criterion.

This process consists of a decision tree classifier which progresses through several stage (or layers), so as to separate classes at each layer by using simple classification algorithms (Swain and Hauska, 1977).

A textural parameter is systematically applied; i.e. a homogeneity feature compute from the variance of the neighborhood of a point. After numerous tests, this parameter revealed as sure. It reduces confusions still present between minor agglomerations and certain natural entities (forest, cultivated land, bare soil). Other parameters are introduced, but in practice their selection is subtle. They are often correlated: several parameters separate the same classes.

But should the superimposition of the texture improve spectral confusions, textural confusions may result. For example, the homogeneity factor distinguishes zones of varying degrees of homogeneity, but it superimposes confusions between land cover units which are highly heterogeneous and boundaries. A technique has been developed to resolve this type of confusion: the homogeneity feature is associated with an "edge" feature. This is obtained by an edge detection procedure. A textural channel thus corresponds to an "edge" image. Consequently, if a point belongs to an extremely heterogeneous zone it is compared to the "edge" image point. If it belongs to an edge the point is affected to an "edge" class, enabling boundaries to be highlighted and improving the image definition. But should the operator do not require their plot; the point is affected to the majority class of the neighborhood of the point.

Compared to simple multi-spectral classification, a particularly interesting improvement thanks to multitextural classification is a very significative removal of the overlapping between forests and villages, and a better classification of zones of dry cropping.

Improvement of classified image
The final stage is an improvement to the previous result through a system of elimination of isolated points and enhancement of the road networks.

Conclusion
At each stage, the class confusion matrices permit classification improvement to be monitored. The mode for computing these matrices is adapted to the various stages and consists either in an evaluation of the classification or in a post- operation check of classified data, in all cases originating from samples. Tests have revealed an improvement in the region of 10 to 50%, depending on the particular class.

At present, this system requires the intervention of the operator at each stage. This will subsequently lead to establishing an expert system.

This software is developed both on a PC/AT compatible microcomputer and on a system VE CDC Cyber 990 computer.

References:

Ducros-Gambart D. et. al." Classification supervisee multispectrale, multitextural, multi-temporally- region du Nord-Est de Toulouse" Rapport CESR in 088-1244 Janvier 1988.
Gastellue-Etchegorry J.P. & Ducros-Gambart D. "Computer assisted land cover mapping with SPOT in Indonesian" INT J. Remote sensing 1989, submitted.
Gastellu-Etchegorry J.P. & Ducros-Gambart D. "Thematic Mapping of Central Java" Proceeding of Symposium SPOT 1, "image, utilization, assessment, results," Paris, November 1987, pp. 467-652.
Haralick R. M., Shanmugan K., Dinstien I. "Textural features for image classification "IEEE Transactions on Systems, Man, and Cybernetics, Vol SMC-3, N06, November 1973, pp. 610-621.
Malingreau J.P. and Christiani R., 1981, A land cover/land use classification for Indonesia, Ind. J. Geog., Vol. 4, No 2, pp 45-64.
Piech K.R. and Walker J.E., 1974, Interpretation of Soil, Photogram. Eng. 40, pp. 87-94. Rakaryatham P. "Classification multidimensionnelle (spectrale et texturale) des images de satellites" These University Paul Sabatier 1984.
Sabins F.F, 1978, Remote Sensing principles and Interpretation, W. H. Freedan and Company, San Francisco.
Swain P. H. & Hauska H. "The decision tree classifier: design and potential" IEEE Trans. On Geoscience electronics, vol. GE-15, n03, July 1977, pp. 142-147.