Expert System House Detection in High Spatial Resolution Imagery Using Size, Shape, and Context

IKONOS 1 × 1 m panchromatic data fused with 4 × 4 m multispectral data were used for residential house detection in three 1 × 1 km study areas of Columbia, South Carolina. One study area contained houses built in the 1940s, another in the 1970s, and another in the 1990s. An expert system was developed to extract individual houses from the imagery. The system employed a data mining algorithm as the core of an automated knowledge acquisition module. Separate knowledge bases were generated for each study area using training samples and the data mining algorithm in two sequential stages. In the first stage, brightness values and NDVI yielded a knowledge base that was used to locate candidate house pixels. In the second stage, region metrics including size, shape, and a series of context variables were employed. Regions of asphalt roads mistakenly identified by the expert system as houses were removed using road buffers. Also, separate knowledge bases were generated both with and without the use of context variables. Each scenario was compared with a point map of photo‐interpreted (reference) houses. The photo‐interpreted database was verified against in situ housing counts. There was a strong increasing trend in both machine and photo‐interpreter accuracy as housing age decreased, with the highest accuracies (79‐88%) in the 1990s study area.