MOE-Fellowship: Dawid Olesiuk

Preparing and detailed land cover classification of hyperspectral imagery

Preparing and detailed land cover classification of hyperspectral imageryThere is no doubt, that especially vegetation cover is a perfect indicator of all other components of biosphere and should be well researched and mapped. This is available using remote sensing methods which are a highly useful tool to investigate large areas in a flexible and quick way. Application of remotely sensed techniques allows for vegetation research and mapping (state and condition, biomass production), because plant species develop specific adaptations (pigment content, plant tissue structure etc.), what have direct impact on reflectance, which can be quantified especially using hyperspectral imagery. Hyperspectral sensors provide precious information about environment contained even in hundred tight spectral channels with very good spatial and radiometric resolution. Big amount of tight spectral channels is main characteristics of hyperspectral images, because it is allow distinguish the bio-physico-chemical properties of the objects that are located on the earth area.Goal of this project is use hyperspectral images to advanced environmental analysis. The effectively use of hyperspectral images depend of pre-processing methods, which include system correction, orthographic rectification and atmospheric correction. After these process I use artificial neural networks in land cover classification, because these method shows big potential for discriminating land cover types from hyperspectral imagery. I use multilayer, one-directional network trained using a supervised back-propagation method. The training process consists of determining the neuron connection weights to make the output signal from the network as close as possible to the expected one. The training data is a pair of vectors. The first (input) vector represents the structure, which the network is to recognize. The second (output) vector represents the pattern results corresponding to output data. The training aims, by adjusting the weights, to minimize the difference between the pattern vector, and the result generated by the network.

AZ: 30009/170

Zeitraum

01.03.2009 - 30.09.2009

Land

Polen

Institut

Deutsches Zentrum für Luft- und Raumfahrt (DLR) e. V.

E-Mail

Mail

Betreuer

Dr. Andreas Müller