Automated GIS integration in landcover classification
Satellite image-derived landcover classifications are of key importance to data acquisition and update for many operational GIS systems. Unfortunately, traditional, spectral-based methods of landcover classification do not extract all the information which is visually apparent in the image, resulting in either aggregated landcover classes or low classification accuracies. Methods which incorporate non-spectral data have been shown to improve classification accuracy and/or class distinction. Unfortunately, these methods are limited as they require significant analyst input for each application. This paper describes a system which allows the integration of traditional spectral-based classifiers with geographic information system technologies, but which greatly reduces analyst input. Spectral data, non-spectral spatial data, tabular, descriptive, and declarative data can be flexibly integrated in a landcover classification. A classification model is described via a rule-base, which may be modified incrementally. Spatial data operators are provided, such as class restriction based on thematic data and spectral likelihood classification. A test of this system in northeastern Wisconsin resulted in a significant improvement in classification accuracies when compared to a traditional maximum likelihood classification.