Class-label statistics: a basis for fusing information from multispectral imagery with an application to unsupervised detection of human settlement

A new approach to fusion of information from multispectral (MS) imagery is presented. This approach is motivated by the desire to develop an unsupervised classifier which can robustly detect settled regions (i.e., regions containing man-made structures). It is well known in the remote sensing community that the ground truth used in supervised classification are not only difficult to obtain but tend to be inconsistent and difficult to validate. In this paper, we present an unsupervised approach for the problem which combines a two-step analysis of MS imagery with a spatial analysis of higher resolution panchromatic imagery. The MS analysis first combines the multispectral pixel information to create an image of pixel labels generated by the K-Means clustering algorithm. Tile-based features are then computed based on the first and second order class label statistics. These tile features are then used to classify the tiles via a second application of the K-Means clustering algorithm. The results from this MS analysis define clusters of tiles of variable texture that are highly likely to contain evidence of human settlements. Spatial information is then brought to bear by analyzing coregistered high-resolution panchromatic images. By standard detectors we find the corner and edge densities in each coregistered tile. Determination of the threshold values used to determine the presence or absence of human settlements is currently performed by a human observer. The results of this spatial analysis are then compared and combined with the MS results to finally determine the set of tiles containing signs of human settlements.