Optimization of supervised classification procedure for irrigated crop discrimination using Landsat TM images

M.

Description

Resumen
En este artículo se evalúan varios modos de clasificación supervisada con objeto de optimizar el proceso de clasificación y poder así agilizar y mejorar la estimación de superficies de cultivos. Para ello se han utilizado cuatro imágenes Landsat TM del regadío de Flumen (Huesca), dos de 1993 y dos de 1994, y se han ensayado doce clasificaciones supervisadas diferentes por año con firmas espectrales obtenidas de imágenes unitemporales y multitemporales para las ocupaciones de primavera y verano, aplicándose también tres formas diferentes de toma de áreas de entrenamiento (automática, semiautomática y manual). La bondad de las clasificaciones se ha evaluado con varias medidas de exactitud. La clasificación multitemporal automática ha resultado la más idónea. Además se ha constatado la influencia de la fecha de las imágenes en la discriminación de cultivos, indicándose cuáles son las imágenes más adecuadas para su discriminación.
Abstract
In this article different supervised classifications modes were evaluated in order to optimize the classification procedure for ease and improve the crop hectarage estimations. Four Landsat TM images from the irrigated district of Flumen ( Huesca , Spain ), two dated from 1993 and another two from 1994, were used. Twelve supervised classifications for each year were applied using spectral signatures of spring and summer land cover, obtained from unitemporal and multitemporal images, with three different kinds of training area selection (automatic, semiautomatic, and manual). After applying several accuracy indices, the automatic multitemporal classification was found to be the most sound. The influence of the image date on the crop classification was also studied, and this article shows, which images were the most suitable in crop discrimination.

Sujets

• Exactitud
• Teledetección
• Supervised learning
• Accuracy
• Remote sensing
• Crops

Informations

• Publié par: erevistas
• Publié le: 01 janvier 2004
• Nombre de lectures: 26
• Langue: Español

Extrait

Revista de Teledetección. 2004. 22: 33-39.
Optimization of supervised classification procedure
for irrigated crop discrimination using Landsat TM
images
M. A. Casterad y T. Martín-Ordóñez
[e-mail]
Centro de Investigación y Tecnología Agroalimentaria de Aragón (CITA)
Apartado 727, 50080 Zaragoza
RESUMEN ABSTRACT
En este artículo se evalúan varios modos de clasifi- In this article different supervised classifications
cación supervisada con objeto de optimizar el proce- modes were evaluated in order to optimize the classi-
so de clasificación y poder así agilizar y mejorar la fication procedure for ease and improve the crop hec-
estimación de superficies de cultivos. Para ello se han tarage estimations. Four Landsat TM images from the
utilizado cuatro imágenes Landsat TM del regadío de irrigated district of Flumen (Huesca, Spain), two
Flumen (Huesca), dos de 1993 y dos de 1994, y se dated from 1993 and another two from 1994, were
han ensayado doce clasificaciones supervisadas dife- used. Twelve supervised classifications for each year
rentes por año con firmas espectrales obtenidas de were applied using spectral signatures of spring and
imágenes unitemporales y multitemporales para las summer land cover, obtained from unitemporal and
ocupaciones de primavera y verano, aplicándose tam- multitemporal images, with three different kinds of
bién tres formas diferentes de toma de áreas de entre- training area selection (automatic, semiautomatic,
namiento (automática, semiautomática y manual). and manual).
La bondad de las clasificaciones se ha evaluado After applying several accuracy indices, the auto-
con varias medidas de exactitud. La clasificación matic multitemporal classification was found to be
multitemporal automática ha resultado la más idónea. the most sound. The influence of the image date on
Además se ha constatado la influencia de la fecha de the crop classification was also studied, and this arti-
las imágenes en la discriminación de cultivos, indi- cle shows, which images were the most suitable in
cándose cuáles son las imágenes más adecuadas para crop discrimination.
su discriminación.
KEY WORDS: supervised classification, accuracy,
PALABRAS CLAVE: clasificación supervisada, remote sensing, crops.
exactitud, teledetección, cultivos.
give the highest classification accuracy being theINTRODUCTION
most suitable to our purposes because they tend to
Crop extent estimates in different areas of Aragón discriminate informational categories (Campbell,
(Spain), based on supervised classification of Land- 1996). The maximum likelihood classifier, by far
sat 5 TM images, were obtained for years in the the most widespread among supervised classifica-
Centro de Investigación y Tecnología Agroalimen- tion methods, was used in this work.
taria de Aragón (CITA). Classification and ground Cover type, growth stage and phenology of crops,
data were combined for crop hectarage estimation spectral bands used, training fields extraction type,
by the method of frame area sampling and regres- satellite image date acquisition, use of multitempo-
sion estimator with satellite data (Casterad et al., ral data, etc. are some factors that influence on the
1992; Barbosa et al. 1996; Casterad, 1996). classification procedure. Too many tests are requi-
The precision of the estimates obtained by the red in order to know the influence of these factors
above mentioned method is conditioned by the clas- in the classification. The majority of works cannot
sification quality, in addition to other factors, so tackle such tests using predetermined classification
that the interest is on optimizing the classification methodology. In Spain, Lobato and Moreiras
procedure. Among supervised and unsupervised (1991) analyzed different choices and variables
classification methods, the supervised ones usually (preselection of more representatives pixel, size of
N.º 22 - Diciembre 2004 33M.A. Casterad y T. Martín-Ordóñez
training pixels, number of bands used, multitmepora- cutting of sub-scenes containing the study area, and
lity, sectorization, clustering + classification, etc.) to (iii) radiometric and geometric corrections of these
be taken into account in the classification process. sub-scenes. Training fields were selected by
Following in this way, Barbosa et al. (1996) com- manual, automatic and semiautomatic techniques
pared crop hectarage estimates of the irrigated dis- from ground truth data (Figure 1), which were
trict of Flumen (Huesca, Spain) obtained by the obtained from a systematic random sampling by
method of frame area sampling and regression esti- blocks (Casterad, 1996). The sampling units were
mator with satellite data of four different supervised squares of 500 m of side.
maximum likelihood classifications: (i) manual In the manual selection, the analyst chooses the
classification of a spring image from 20 May 1991; pixels to be used as training fields based on the
(ii) manual classification of a summer image from available ground truth data. In automatic selection,
24 August 1991; (iii) manual-multitemporal classi- the mixed pixels that are in the border of the plot-
fication of both images; and (iv) automatic-multi- used (adjacent plots with the same used) were eli-
temporal classification of both images. In results, minated, and the remaining pixels were used for
for crop hectarage estimation using regression esti- training. In semiautomatic selection, the signature’s
mator with satellite data, the precision improve- frequency histograms obtained in the automatic
ments produced for each crop are different depen- selection were displayed (band by band) and
ding on the used classification procedure. However, visually inspected by the analyst. When a particular
no classification showed to be better than the others signature is present a multi-peaked histogram, dif-
in terms of global classification accuracy. Comple- ferent homogenous spectral subclasses were crea-
menting studies will be required to ratify these ted within this class, if there are sufficient number
results obtained by Barbosa et al. (1996) and to of training subclasses pixels. The correspondence
determine the best classification method to discri- of these peaks to some crop characteristics (deve-
minate the main crops in different irrigated areas. lopment stage, kind of management, etc.) was
In the present work, unitemporal and multitem- analyzed, with the help of the RGB composite
poral supervised maximum likelihood classifica- image and the ground data annotations.
tions, using three different kinds of training area Combining the three techniques of training area
selection (automatic, semiautomatic, and manual), selection and the different images, twelve classifi-
2were evaluated and compared for the 332 km irri- cations per year were tested. Six ones were unitem-
gated district of Flumen (Huesca, Spain) in 1993 poral classifications (three using the spring image,
and 1994. The influence of the image date on crop and three using the summer image); another three
discrimination was also analyzed. were multitemporal classifications (spring and
The irrigation district of Flumen is located in the summer images) using spectral signatures of spring
middle Ebro Basin, (Aragón, north-eastern Spain). land covers (land covers present in the spring ima-
Basin and border irrigation is applied to plots typi- ges); and, finally, three ones were multitemporal
cally size from 0.8 ha to 1 ha. However, groups of classifications using spectral signatures of summer
adjacent fields have often the same crop. The main land covers (land covers present in the summer
crops in the area are alfalfa, maize, rice, sunflower image).
and winter cereals (barley and wheat).
Classification accuracy
MATERIAL AND METHODS Confusion matrices were used in classification
accuracy assessment (Story and Congalton, 1986;
Congalton, 1991). They were obtained from com-Supervised classification
parison between ground truth and classified data in
One spring and one summer Landsat 5 TM scene the sampling units, once the border pixels were eli-
per year (6 March and 12 July in 1993, and 28 May minated. The overall accuracy percentage (OA%)
and 29 June in 1994), corresponding to the study and the Kappa statistic (k) were calculated in each
area, were individual and jointly classified using classification results. The OA is the simplest des-
the maximum likelihood method described in criptive statistic derived from the confusion matrix
ERDAS (1994). The pre-processing consisted of: and reports the overall proportion of correctly clas-
(i) visualization and enhancement of the image, (ii) sified pixels in the sampling units.
34 N.º 22 - Diciembre 2004Optimization of supervised classification procedure for irrigated crop discrimination using Landsat TM images
(iii) The Kappa statistic which estimate is (Bis-
hop et al., 1975; Rosenfield and Fitzpatrick-Lins,
1986)
where, n is the number of rows and columns in error
matrix, X are the diagonal entries or correctly clas-ii
sified pixels (observed number in row i column i) Fitzgerald and Lees (1994) have suggested three
and X is every element of the confusion matrixij ranges of agreement for the Kappa