How Transferable Are Spatial Features for the Classification of Very High Resolution Remote Sensing Data?

Knowledge transfer for the classification of very high resolution panchromatic data over urban area is investigated. Invariant feature are extracted with some morphological processing. The well-known spectral angle mapper (SAM) is proposed as a measure of transferability. Support vector machines (SVMs) are used to fit a separating hyperplane in a vector space defined by the extracted spatial features. The hyperplane is then used to classify other data set without any new training. Several experiments are presented. Results confirm the usefulness of spatial feature when the classification of two images from two separates data set is considered.