Guest lecture by Dr. Xiuping Jia

University of New South Wales, Cranberra, Australia

"Mapping with Hyperspectral Data Cubes: Feature Generation, Selection and Integration"

A hyperspectral data cube is typically composed of about 100 to 200 wavebands for an imaged scene on the earth’s surface. They form the original set of the spectral features. From there, many more new features can be generated. They can be derived via linear or nonlinear transformations. Spatial texture features, such as contrast, homogeneity and energy, can be generated when neighbours’ radiometric measurements are related to each central pixel. Object-based features and attributes including area, orientation and solidity become available on a segmented or masked image. Is having more features better? Is using a large number of features a good practice? How can features be combined in machine learning?

In this talk, feature selection together with data modelling will be discussed in order to avoid the problems of over training. A few feature fusion techniques in implementing spectral-spatial based classification will be presented.

previous page
more events