Neural networks

Reusing Artificial Neural Networks to enhance satellite image land cover classification

Research associate: Lieven Verbeke

Promoter: Robert De Wulf

Duration: 01/2002 – 12/2003

 

Project objectives

  1. The development of a method for the extraction of knowledge, acquired during training, from a neural network, within the context of vegetation mapping from remotely sensed imagery. Such a methodology should lead to (1) a tool for the identification of key factors in the classification process and (2) should allow for the description and interpretation of the established relation between inputs and outputs in a trained neural network;
  2. The development of a method for the transfer of knowledge between a trained artificial neural network and an untrained neural network, within the context of vegetation mapping using remotely sensed data.

Two key publications

Verbeke, L., Van Coillie, F. and De Wulf, R., “Reusing Back-propagation Artificial Neural Networks for Land Cover Classification in Tropical Savannahs.” INTERNATIONAL JOURNAL OF REMOTE SENSING 25.14 (2004): 2747–2771.

Verbeke, L., Van Coillie, F., and De Wulf, R.. “Previously Trained Neural Networks as Building Blocks for a New Classifier: Improving Classification Performance by Knowledge Transfer.” Remote Sensing in Transition. Ed. R Goossens. Rotterdam, The Netherlands: Millpress Science, 2004. 115–119.