Dimensionality Reduction Techniques: An Operational Comparison On Multispectral Satellite Images Using Unsupervised Clustering

Multispectral satellite imagery provides us with useful but redundant datasets. Using Dimensionality Reduction (DR) algorithms, these datasets can be made easier to explore and to use. We present in this study an objective comparison of five DR methods, by evaluating their capacity to provide a usable input to the K-means clustering algorithm. We also suggest a method to automatically find a suitable number of classes K, using objective "cluster validity indexes" over a range of values for K. Ten Landsat images have been processed, yielding a classification rate in the 70-80% range. Our results also show that classical linear methods, though slightly outperformed by more recent nonlinear algorithms, still offer a reasonable trade-off.