Classification of land use and spatial indicators using satellite images - Innovation, development, and comparison

Partner: Metria Miljöanalys

National and international policies today require environmental monitoring and follow-up systems that detect, in a quality assured way, changes over time in land use and landscape indicators. Spatial elements that indicate high environmental quality are usually quite rare and important changes are sometimes very subtle. Existing monitoring programs linked to landscape surveillance, e.g. the Swedish NFI (National Forest Inventory), do not permit evaluation of data at a local level or of smaller regions. Remote sensing of satellite images offers great potential to assess wall-to-wall changes in the health of ecosystems, and identify risks. So far, however, little of this potential has been realized.

The objectives with this project are
  • to develop innovative classification methods by using advanced stochastic modelling and appropriate statistical methodologies (such as the wavelet transform, the Markov random field approaches, data editing, information theory, etc.)
  • to extend traditional and conventional parametric classification methods by taking care of the multimodality of class densities using advanced parametric estimation methods.
  • to evaluate and improve traditional classification methods, such as LDA, QDA, k-NN, etc.
  • to produce unbiased estimates of confusion matrices with high precision by using Monte Carlo techniques.
  • to supply our end-users with recommendations of highly reliable classification algorithms in different real applications.

    Comparative studies will be conducted to investigate how classification performance is affected by
  • frequency of occurrence of classes (from rare to dominating)
  • number of field plots in the training set
  • size of feature vector
  • non-normality of the feature vector

    The comparisons will be based on real data and not on simulated data.