Urban area operations have become extremely important in the post 9-11 world, and tactical operations in urban areas require high fidelity geospatial information. Urban-area geospatial information is very valuable for content-based (image and text) queries, geospatial data conflation, and automated registration/fusion of other intelligence data streams. The ability to register and/or mine multiple intelligence (multi-INT) data sources improves the speed and reliability of image analysis and confidence in positive target identification.

In our research, we have developed semi-automated classification methods for the production of high-resolution urban land cover maps from commercial satellite imagery (CSI). Our basic approach uses fuzzy-logic techniques that provide a number of unique advantages for processing and exploitation of high-resolution CSI. Fuzzy-logic techniques can be developed to mimic human cognitive function by integrating and exploiting multiple visual cues (text, context, shape) in addition to spectral response. This is important for developing robust classification strategies that can accommodate feature or class spectral ambiguities present in all multi-spectral imagery.

Our initial approach used a fuzzy-logic classifier that incorporated low-level spatial information at the pixel level to resolve spectral ambiguities present between different urban land cover classes. The fuzzy-logic pixel classifier allows initial class membership in more than one class. Spatial features and information are then used to resolve spectral class ambiguities. For instance, the example below shows how an entropy texture measure can be used to resolve the spectral ambiguity between Tree and Grass classes in a high-resolution CSI sub-scene of a vegetated area.

In addition to texture measures, the fuzzy-logic pixel classifier also uses information about the directional similarity of spectrally similar pixels. The example below shows the calculation of the 2D spatial signature of a street pixel in an urban sub-scene. Length and width features for each pixel can be determined from the spatial signature's variation as a function of azimuth angle.

The length and width features are useful for resolving ambiguities between Road and Building classes that often have very similar spectral signatures. In the suburban sub-scene shown below, the length and width values for each non-vegetated pixel are shown were the color red represents the length value and the color blue represents the width value. As seen, the street pixels tend to show in red (large length values) while the single-family residential houses tend to show up in blue (length  width).

The fuzzy logic pixel classifier utilizes both the texture and length/width features to significantly improve the high-resolution land cover maps. The example shown below presents a comparison between the maximum likelihood pixel classifier (MLPC) and the fuzzy logic pixel classifier (FLPC) of a suburban residential area. Note the many MLPC misclassifications where residential houses are classified as Water, streets are classified as Building, and wooded areas are classified as Grass all because of spectral confusion. The utilization of spatial information in the FLPC eliminates that vast majority of these misclassifications.

After the FLPC, a fuzzy logic object-based classifier (FLOC) was then developed to further improve the classification results by incorporating higher-level shape and context information. First, image objects are generated by automatic segmentation using both spectral and shape homogeneity criteria. Fuzzy logic rules are applied to the objects to identify rectilinear features (buildings and parking lots) and to use contextual information like shadows to improve the classification of these features. The example below shows how shape was quantified using morphological operations applied to image objects to generate piecewise linear polygons. Fuzzy logic rules were developed to identify "approximately rectangular" features using line segment lengths and angle vertices from the object's polygon. Fuzzy rules are required because perfect right angles (e.g. 90 °) are not normally obtained from the segmented objects.

In addition to shape, the objects composite spectral signature derived from all constituent pixels and the contextual relationship with surrounding shadow objects were used in the FLOC. The example below shows a comparison between the FLPC and the FLOC for a dense urban area. The FLPC overclassifies buildings in the urban area because there is not enough spectral and spatial information at the pixel level to differentiate these from parking lots. By using object information, the FLOC is able to identify an additional Impervious Surface class and the overclassification of buildings is greatly reduced.

The FLOC produced classification accuracies in the range of 80-99% for the following urban land cover classes: Road, Building, Grass, Tree, Water, Impervious Surface, and Shadow. These are the highest accuracies achieved for high-resolution urban land cover classes reported thus far in the literature.

Both the FLPC and FLOC require training data to be input by a user that is then used by the algorithms to produce the urban land cover classification of the entire image. In ongoing research, we are seeking to combine Automatic Feature Extraction (AFE) to generate training data for the various urban land cover classes. We are developing fuzzy membership values for image features and objects for all urban land cover classes. We will then use the "high confidence" features as training data in the FLPC and FLOC. Using this approach, we will be able to automatically generate the urban land cover maps without the need for human operator input.