Adaptive classification method for multispectral remote sensing data

Based on local features, a new adaptive classification approach for multispectral remote sensing data is presented. Typical classification techniques based on global features tend to degrade because all classes are projected along the same direction, e.g. the principal component direction for Principal Component Analysis (PCA) and minimum momponent direction for Minimum Component Analysis (MCA). The typical methods are under the assumption that class separability is uniform for all directions, which is not always true. The new method overcomes that disadvantage by selecting features, which give the maximum class separability, based on local information of the classes instead of global information. In the new method, a projection matrix for every class is first sought based on making its training examples well separated from the others. Every input vector is then linearly transformed into another space by every projection matrix. In the transformed spaces, it can be classified or labeled to different class by Maximum Likelihood Classification (MLC). In order to reduce computation cost, adaptive dimension reduction is also introduced. Good performance of the new method can be shown from the experimental results on the Kennedy Space Center (KSC) TM images.