A neural network approach using multi-scale textural metrics from very high resolution panchromatic imagery for urban land-use classification

The successful launch of panchromatic WorldView-1 and the planned launch of WorldView-2 will make a
major contribution towards the advancement of the commercial remote sensing industry by providing
improved capabilities, more frequent revisits and greater imaging flexibility with respect to the precursor
QuickBird satellite. Remote sensing data from panchromatic systems have a potential for more detailed and
accurate mapping of the urban environment with details of sub-meter ground resolution, but at the same
time, they present additional complexities for information mining.
In this study, very high-resolution panchromatic images from QuickBird and WorldView-1 have been used to
accurately classify the land-use of four different urban environments: Las Vegas (U.S.A.), Rome (Italy),
Washington D.C. (U.S.A.) and San Francisco (U.S.A.). The proposed method is based on the analysis of firstand
second-order multi-scale textural features extracted from panchromatic data. For this purpose, textural
parameters have been systematically investigated by computing the features over five different window
sizes, three different directions and two different cell shifts for a total of 191 input features. Neural Network
Pruning and saliency measurements made it possible to determine the most important textural features for
sub-metric spatial resolution imagery of urban scenes.
The results show that with a multi-scale approach it is possible to discriminate different asphalt surfaces,
such as roads, highways and parking lots due to the different textural information content. This approach also
makes it possible to differentiate building architectures, sizes and heights, such as residential houses,
apartment blocks and towers with classification accuracies above 0.90 in terms of Kappa coefficient
computed over more than a million independent validation pixels.