CROP AREA ESTIMATES USING ERS-1 SAR DATA
Gerhard Smiatek
Department of Remote Sensing and Digital Image Processing
Institute of Navigation, Stuttgart University
Keplerstr. 11 , 70174 Stuttgart
Commission VII, WG 2
KEY WORDS: Remote.Sensing, Agriculture, Land. Use, Statistics, GIS, SAR, Accuracy
ABSTRACT:
Misclassified numerical results of a multispectral classification can be corrected using the double sampling
scheme. For a small sample both the true classification, i.e. ground truth, and the fallible classification from
the classified imagery have to be provided. The comparison of both yields information on the misclassification
errors. This information is used to correct for the bias introduced by the multispectral classification. The article
describes the sampling method.
KURZFASSUNG:
Die Korrektur der numerischen Ergebnisse einer multispektralen Klassifizierung kann mit der Methode der
doppelten Stichprobe erfolgen. Hierbei wird fiir eine kleine Stichprobe die wahre Klassifizierung bereitgestellt.
Aus dem Vergleich der wahren Klassifizierung und der fehlerhaften Klassifizierung mit den Bilddaten werden
statistisch signifikante Korrekturfaktoren abgeleitet. Die Kenntnis dieser Faktoren erlaubt die Korrektur des
Klassifikationsergebnisses. Der Beitrag beschreibt das Verfahren.
1. INTRODUCTION In crop surveys the acreage of a certain crop is
usually of importance. Misclassified results of the ML
The Maximum Likelihood classifier (ML) is fallible classification can be corrected if some information on
for well-known reasons. If SAR imagery is classified, the misclassification error is known. For the study
the bias can be extremely large. Several methods have area for each ML classified pixel the true classification
been developed to assess the accuracy of classification from the site visit is also available. Thus, the results
of remotely sensed data (i.e. Congalton, 1991). There ^ can be summarized in the following contingency table:
are also methods available to correct for the bias intro-
duces by multispectral classifications (i.e. Czaplewski
and Catts, 1992). One of them is the double sampling fallible maximum
developed by Tenebein (1970) employed with remote- likelihood
ly sensed data by Card (1982), Czaplewski and Catts classifier
(1990), Smiatek (1995) and others. This article covers 1 2 k
some issues of the double sampling for misclassified l aj aı2 + Aix Ou.
ERS-1-SAR data in agricultural surveys. True 2 831 8232 ++ ag 82.
classifier 3 a3, ass --- as, as.
(GIS)
2. CORRECTION FOR THE BIAS i5
kl 8k2 '^' Bkk 8k.
41.823.550 & sri
For a study area the geometry of the fields was di-
gitized from orthophotograph and stored in a geogra- ^ where k is the number of land use category (i.e. winter
phical information system (GIS). After that the site ^ wheat, maize etc.) in the survey; aj; is the number
was visited and the actual land use was mapped and of pixels whose true category is à and whose fallible
introduced into the database. Three multitemporal category 1s j.
ERS-1 GTC SAR images were provided for classifica- Following the misclassification model developed by
tion. In addition, the Lee filtering was applied to the ~~ Bross (1954), Tenebein (1970) and Tenebein (1972)
images. The results of the multispectral classificati- ^ the result of the ML classification given as the pro-
on have shown that the area of major crops is widely portion 7 can be considered as
underestimated and that neither Lee filtering nor the
majority filter improved the results. 7 — p(1— 0) 4 qó (1)
624
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
ENON Pee
Rm AEN ant
EN: at: joues Acta 0. 0 -—-—- a
