Comparison of Automatic Classification Methods for Multispectral Images

DOI Web Site Web Site 8 Citations Open Access

Bibliographic Information

Other Title
  • マルチスペクトル画像を用いた自動識別手法の比較
  • マルチ スペクトル ガゾウ オ モチイタ ジドウ シキベツ シュホウ ノ ヒカ

Search this article

Description

The use of remotely sensed multispectral data to construct a land use map was studied. The results of automatic supervised classification by the most likelihood ratio (MLH), the linear discriminant function (LDF), the minimum Euclide an distance (MED) and by the correlation coefficient methods (CCM and NCM) were compared. The data were obtained by an airborne multispectral scanner JSCAN-AT-12 M. For the comparison four channel data (blue, green, red, and near IR) were used. The training fields were selected for eight categories (pond, forest, field, marsh, bare soil, road, railway and concrete). If the distribution of the data were Gaussian, the correct classification rate (CCR) decreases as we change the methods of classification from one to another in the above mentioned order. This was the case for the training field data, but not so for those of the test fields. This means that an essential problem in practice is the generality of the training field data. The skewness and kurtosis calculated from the data of each channel for the training area show that the distribution are not Gaussian. This, however, does not have a serious effect on the classification in this case.<br>For the test data the best result (about 80% CCR) was obtained by MED, and the worst (about 60%) by MLH. In MLH, if the data had small variances slight differences were exaggerated because the data lacked generality. On the other hand, if the data had large variances, the differences between categories were likely underestimated.

Journal

Citations (8)*help

See more

Details 詳細情報について

Report a problem

Back to top