An Effect of High Registration Accuracy of Landsat TM Pixels on Landcover Classification Accuracy

Abstract

This research deals with the effect of pixel positioning accuracy on the accuracy of landcover classification of actual Landsat-5 TM image data by the following process.
First, IFOVs of image data over the analyzing region were projected on the ground truth map of landcover both by an usual registration method and by a new registration method which registers IFOVs more accurately than an usual method. Also, image data of the study area were classified to six categories by three kinds of supervised classification methods.
Second, the result of supervised landcover classification were compared for each pixel with the content of landcover within one IFOV of that pixel, which IFOV was projected on the ground truth map by two registration methods, and the accuracy of supervised landcover classification were calculated and compared.
Finally, the classification accuracy for pure pixels and mixed pixels were examined under usual registration accuracy by an usual method and under the higher registration accuracy by a new method.
After the procedures noted above, the results of this research are concluded as follows.
(1) The accuracy of landcover classification in remote sensing data was effected by the accuracy of the registration of IFOVs of pixels. The accuracy of supervised landcover classification for pixels of which IFOVs were registered by a new method was totally about 2% higher than that for pixels of which IFOVs were registered by an usual method in each classification method.
(2) One of the causes of misclassification of landcover in remote sensing data was concluded to be the existence of mixed pixels. The misclassification of landcovers occurred more frequently in mixed pixels than in pure pixels. And the accuracy of landcover classification of mixed pixels of which IFOVs were registered more precisely by a new method was higher than that of mixed pixels of which IFOVs were registered by an usual method.