Possibilities and Limitations of Artificial Neural Networks for Sub-Pixel Mapping of Land Cover

Laboratory of Geo-Information Science and Remote Sensing
  Education
  Research
  Publications
  Models
  News & Calendar
  News
  Calendar
  Archive
  News
  Calendar
  2011
  2010
  2009
  2008
  2007
  2006
  2005
  2012
  Staff
  Equipment
  Contact details
  Workshops

15 Mar 2006 14:00 - 15 Mar 2006 14:20
Unit: Wageningen UR

 By: Demeke Nigussie Alemu  

Abstract 

Up-to-date and reliable land cover information is vital for many decision-making processes. Though the developments of satellite remote sensing greatly improved land cover information acquisition, there are still unaddressed problems to achieve the intended accuracy level. One of such problems is the mixed pixel. Soft-classification techniques were introduced to address the problem; but they do not show the spatial location of the classes’ proportions in a pixel. Sub-pixel mapping techniques that transform the soft-classification results into hard-class maps at sub-pixel scale were introduced to address the drawbacks of soft-classifications. In this thesis, an artificial neural network (ANN), specifically the feedforward backpropagating neural network (FFBPNN) was used for sub-pixel mapping.
To prepare the input fraction images, which are to be treated as soft-classification results, the LGN5 database was thematically aggregated into 2, 4 and 8 thematic classes. Then, these thematically aggregated data were spatially aggregated into three spatial resolution sizes, namely, 75m, 150m, and 300m. The three chosen thematic classes and three spatial resolutions end up in 9 different combinations that are considered as study cases in this thesis work. The fraction images were used to train several FFBPNN. After training and selecting the best network, each case was simulated using the fraction image of two small sites to reconstruct the 25mx25m sub-pixel hard-class map. These sites were selected from the Southwest and Southeast of the Netherlands to examine the effect of the land cover heterogeneity. The overall accuracies obtained revealed that the response of the network was highly influenced by the spatial frequency, shape and size of the different land covers. Moreover, it revealed that most of the errors are on class boundaries where highly mixed pixels are expected. The accuracies achieved had a wide range depending on the complexity of the cases. In general, the overall accuracies ranged from 38.05 % (complex cases) to 98.97 % while the Kappa coefficients ranged from 0.14 to 0.97. Although it was not possible to exhaustively explore all network architectures for the various case studies, the results achieved demonstrated the potential of the FFBPNN for sub-pixel mapping.
Keywords: Sub-pixel, land cover, neural networks, class fraction, target
Print this activity