Developing a strategy for deriving accurate LC information in urban areas from HR imagery is a challenging task, because urban LC classification from HR-data (Ikonos, Quickbird) involves a limited number of spectral bands, a strong spectral heterogeneity of classes and the overall presence of shadow and occlusion.

A QuickBird image of Dublin acquired on August 4, 2003 was used as input for the work on HR image classification. It covers the northern part of Dublin from the CBD towards the suburban fringe. This area is characterized by strong urban dynamics

 

The strategy used for producing the Dublin LC-map can be summarized as follows:

• Fusion (used for visual interpretation during the classification and validation)  
                                   
• Rectification (based on 23 GCPs - the global RMSE is 1.21m)    
 More info ...

• Pre-classification (ISODATA and interpretation process followed by Maximum Likelihood classification @ 0.61m resolution)            
More info ...

• Classification: Maximum Likelihood @ 2.44m on 4 multispectral bands and 1 textural band (initialy computed on panchromatic band)           

• Post-classification: Two kind of rules:
   1- Rules based on contextual and morphological parameters of regionsclumps;
   2- Thematic rules operating on specific classes or sub-classes     

• Validation (by visual interpretation of the fused image)    

       

The global accuracy and the conditional Kappa are respectively of 78.1% and 0.74 for 8 classes.

These relatively weak values are drastically improved at the level of validation of MR image in 3 classes (Vegetation, Impervious, other [Bare soil, dump sites, etc.]) (WP2)